Joscha Bach: Nature of Reality, Dreams, and Consciousness
AI 与机器学习生物与进化技术与编程哲学与宗教心理与人性
📋 章节目录
暂无章节信息
🔑 关键词
donmodelablecannotconsciousnesshumancontrolgoingmeanscertainsystemsdoesndoingsocietyagentphysicalstatehumansstoryreality
💬 精彩语录
暂无语录
🎙️ 完整对话（4310 条）
Lex Fridman (00:00.000)
The following is a conversation with Yosha Bach,
以下是与约沙·巴赫的对话，
Lex Fridman (00:02.720)
his second time on the podcast.
他第二次参加播客。
Lex Fridman (00:04.940)
Yosha is one of the most fascinating minds in the world,
Yosha 是世界上最迷人的思想家之一，
Lex Fridman (00:08.540)
exploring the nature of intelligence,
探索智能的本质，
Lex Fridman (00:10.620)
cognition, computation, and consciousness.
认知、计算和意识。
Joscha Bach (00:14.500)
To support this podcast, please check out our sponsors,
为了支持这个播客，请查看我们的赞助商，
Lex Fridman (00:17.700)
Coinbase, Codecademy, Linode, NetSuite, and ExpressVPN.
Coinbase、Codecademy、Linode、NetSuite 和 ExpressVPN。
Joscha Bach (00:23.940)
Their links are in the description.
他们的链接在描述中。
Lex Fridman (00:26.740)
This is the Lex Friedman podcast,
这是莱克斯·弗里德曼的播客，
Lex Fridman (00:28.980)
and here is my conversation with Yosha Bach.
这是我和尤沙·巴赫的对话。
Lex Fridman (00:33.340)
Thank you for once again coming on
感谢您再次光临
Joscha Bach (00:35.180)
to this particular Russian program
对于这个特定的俄罗斯计划
Lex Fridman (00:38.220)
and sticking to the theme of a Russian program.
并坚持俄罗斯节目的主题。
Joscha Bach (00:40.740)
Let's start with the darkest of topics.
让我们从最黑暗的话题开始。
Lex Fridman (00:43.100)
Kriviyat.
克里维亚特。
Lex Fridman (00:45.220)
So this is inspired by one of your tweets.
所以这是受到你的一条推文的启发。
Lex Fridman (00:48.380)
You wrote that, quote,
你写的，引用，
Joscha Bach (00:50.900)
when life feels unbearable,
当生活感到难以忍受时
Lex Fridman (00:53.740)
I remind myself that I'm not a person.
我提醒自己，我不是一个人。
Joscha Bach (00:56.620)
I am a piece of software running on the brain
我是一个运行在大脑上的软件
Lex Fridman (00:58.940)
of a random ape for a few decades.
Joscha Bach (01:01.500)
It's not the worst brain to run on.
Lex Fridman (01:04.540)
Have you experienced low points in your life?
Lex Fridman (01:07.740)
Have you experienced depression?
Lex Fridman (01:09.780)
Of course, we all experience low points in our life,
Lex Fridman (01:12.140)
and we get appalled by the things,
Lex Fridman (01:15.340)
by the ugliness of stuff around us.
Joscha Bach (01:17.060)
We might get desperate about our lack of self regulation,
Lex Fridman (01:21.300)
and sometimes life is hard,
Lex Fridman (01:24.580)
and I suspect you don't get to your life,
Lex Fridman (01:27.900)
nobody does, to get through their life without low points
Lex Fridman (01:30.700)
and without moments where they're despairing.
Lex Fridman (01:33.740)
And I thought that let's capture this state
Lex Fridman (01:37.800)
and how to deal with that state.
Lex Fridman (01:40.140)
And I found that very often you realize
Joscha Bach (01:43.120)
that when you stop taking things personally,
Lex Fridman (01:44.860)
when you realize that this notion of a person is a fiction,
Joscha Bach (01:48.980)
similar as it is in Westworld,
Lex Fridman (01:50.700)
where the robots realize that their memories and desires
Joscha Bach (01:53.300)
are the stuff that keeps them in the loop,
Lex Fridman (01:55.840)
and they don't have to act on those memories and desires,
Joscha Bach (01:59.100)
that our memories and expectations is what make us unhappy.
Lex Fridman (02:02.540)
And the present really does.
Lex Fridman (02:04.200)
The day in which we are, for the most part, it's okay, right?
Lex Fridman (02:08.300)
When we are sitting here, right here, right now,
Joscha Bach (02:11.260)
we can choose how we feel.
Lex Fridman (02:13.100)
And the thing that affects us is the expectation
Joscha Bach (02:16.740)
that something is going to be different
Lex Fridman (02:18.760)
from what we want it to be,
Joscha Bach (02:19.920)
or the memory that something was different
Lex Fridman (02:21.860)
from what you wanted it to be.
Lex Fridman (02:24.140)
And once we basically zoom out from all this,
Lex Fridman (02:27.340)
what's left is not a person.
Joscha Bach (02:28.980)
What's left is this state of being conscious,
Lex Fridman (02:32.300)
which is a software state.
Lex Fridman (02:33.620)
And software doesn't have an identity.
Lex Fridman (02:35.680)
It's a physical law.
Lex Fridman (02:37.820)
And it's a law that acts in all of us,
Lex Fridman (02:39.820)
and it's embedded in a suitable substrate.
Lex Fridman (02:42.300)
And we didn't pick that substrate, right?
Lex Fridman (02:43.780)
We are mostly randomly instantiated on it.
Lex Fridman (02:46.940)
And they're all these individuals,
Lex Fridman (02:48.900)
and everybody has to be one of them.
Lex Fridman (02:51.740)
And eventually you're stuck on one of them,
Lex Fridman (02:54.220)
and have to deal with that.
Lex Fridman (02:56.340)
So you're like a leaf floating down the river.
Lex Fridman (02:59.080)
You just have to accept that there's a river,
Lex Fridman (03:01.340)
and you just float wherever it takes you.
Lex Fridman (03:03.820)
You don't have to do this.
Joscha Bach (03:04.660)
The thing is that the illusion that you are an agent
Lex Fridman (03:08.140)
is a construct.
Lex Fridman (03:09.500)
What part of that is actually under your control?
Lex Fridman (03:13.100)
And I think that our consciousness
Joscha Bach (03:15.260)
is largely a control model for our own attention.
Lex Fridman (03:18.460)
So we notice where we are looking,
Lex Fridman (03:21.160)
and we can influence what we're looking,
Lex Fridman (03:22.740)
how we are disambiguating things,
Lex Fridman (03:24.180)
how we put things together in our mind.
Lex Fridman (03:26.580)
And the whole system that runs us
Joscha Bach (03:28.780)
is this big cybernetic motivational system.
Lex Fridman (03:30.940)
So we're basically like a little monkey
Joscha Bach (03:32.940)
sitting on top of an elephant,
Lex Fridman (03:34.940)
and we can put this elephant here and there
Joscha Bach (03:37.540)
to go this way or that way.
Lex Fridman (03:39.360)
And we might have the illusion that we are the elephant,
Joscha Bach (03:42.020)
or that we are telling it what to do.
Lex Fridman (03:43.460)
And sometimes we notice that it walks
Joscha Bach (03:45.620)
into a completely different direction.
Lex Fridman (03:47.460)
And we didn't set this thing up.
Joscha Bach (03:49.000)
It just is the situation that we find ourselves in.
Lex Fridman (03:52.620)
How much prodding can we actually do of the elephant?
Joscha Bach (03:56.420)
A lot.
Lex Fridman (03:57.380)
But I think that our consciousness
Joscha Bach (04:00.660)
cannot create the motive force.
Lex Fridman (04:03.000)
Is the elephant consciousness in this metaphor?
Joscha Bach (04:05.340)
No, the monkey is the consciousness.
Lex Fridman (04:07.940)
The monkey is the attentional system
Joscha Bach (04:09.340)
that is observing things.
Lex Fridman (04:10.460)
There is a large perceptual system
Joscha Bach (04:12.380)
combined with a motivational system
Lex Fridman (04:14.300)
that is actually providing the interface to everything
Lex Fridman (04:17.260)
and our own consciousness.
Lex Fridman (04:18.660)
I think is the tool that directs the attention
Joscha Bach (04:21.820)
of that system, which means it singles out features
Lex Fridman (04:24.740)
and performs conditional operations
Joscha Bach (04:26.980)
for which it needs an index memory.
Lex Fridman (04:28.900)
But this index memory is what we perceive
Joscha Bach (04:31.420)
as our stream of consciousness.
Lex Fridman (04:32.740)
But the consciousness is not in charge.
Joscha Bach (04:34.900)
That's an illusion.
Lex Fridman (04:35.860)
So everything outside of that consciousness
Joscha Bach (04:40.300)
is the elephant.
Lex Fridman (04:41.380)
So it's the physics of the universe,
Lex Fridman (04:43.060)
but it's also society that's outside of your...
Lex Fridman (04:46.140)
I would say the elephant is the agent.
Lex Fridman (04:48.300)
So there is an environment to which the agent is stomping
Lex Fridman (04:51.300)
and you are influencing a little part of that agent.
Lex Fridman (04:55.100)
So is the agent a single human being?
Lex Fridman (04:58.980)
Which object has agency?
Joscha Bach (05:02.340)
That's an interesting question.
Lex Fridman (05:03.820)
I think a way to think about an agent
Joscha Bach (05:06.140)
is that it's a controller with a set point generator.
Lex Fridman (05:10.500)
The notion of a controller comes from cybernetics
Lex Fridman (05:13.020)
and control theory.
Lex Fridman (05:14.380)
Control system consists out of a system
Joscha Bach (05:17.780)
that is regulating some value
Lex Fridman (05:20.860)
and the deviation of that value from a set point.
Lex Fridman (05:23.980)
And it has a sensor that measures the system's deviation
Lex Fridman (05:27.500)
from that set point and an effector
Joscha Bach (05:30.060)
that can be parametrized by the controller.
Lex Fridman (05:32.580)
So the controller tells the effector to do a certain thing.
Lex Fridman (05:35.500)
And the goal is to reduce the distance
Lex Fridman (05:38.460)
between the set point and the current value of the system.
Lex Fridman (05:40.860)
And there's an environment
Lex Fridman (05:41.700)
which disturbs the regulated system,
Joscha Bach (05:43.580)
which brings it away from that set point.
Lex Fridman (05:45.580)
So simplest case is a thermostat.
Joscha Bach (05:47.860)
The thermostat is really simple
Lex Fridman (05:49.100)
because it doesn't have a model.
Joscha Bach (05:50.260)
The thermostat is only trying to minimize
Lex Fridman (05:52.380)
the set point deviation in the next moment.
Lex Fridman (05:55.780)
And if you want to minimize the set point deviation
Lex Fridman (05:58.740)
over a longer time span, you need to integrate it.
Joscha Bach (06:00.860)
You need to model what is going to happen.
Lex Fridman (06:03.660)
So for instance, when you think about
Joscha Bach (06:05.700)
that your set point is to be comfortable in life,
Lex Fridman (06:08.220)
maybe you need to make yourself uncomfortable first, right?
Lex Fridman (06:11.420)
So you need to make a model of what's going to happen when.
Lex Fridman (06:14.060)
And this is task of the controller is to use its sensors
Joscha Bach (06:18.020)
to measure the state of the environment
Lex Fridman (06:20.540)
and the system that is being regulated
Lex Fridman (06:22.900)
and figure out what to do.
Lex Fridman (06:24.900)
And if the task is complex enough,
Joscha Bach (06:27.620)
the set points are complicated enough.
Lex Fridman (06:30.100)
And if the controller has enough capacity
Lex Fridman (06:32.540)
and enough sensor feedback,
Lex Fridman (06:34.940)
then the task of the controller is to make a model
Joscha Bach (06:37.340)
of the entire universe that it's in,
Lex Fridman (06:39.180)
the conditions under which it exists and of itself.
Lex Fridman (06:42.300)
And this is a very complex agent.
Lex Fridman (06:43.940)
And we are in that category.
Lex Fridman (06:45.820)
And an agent is not necessarily a thing in the universe.
Lex Fridman (06:49.460)
It's a class of models that we use
Joscha Bach (06:51.700)
to interpret aspects of the universe.
Lex Fridman (06:54.580)
And when we notice the environment around us,
Joscha Bach (06:57.820)
a lot of things only make sense
Lex Fridman (06:59.460)
at the level that should be entangled with them
Joscha Bach (07:01.060)
if we interpret them as control systems
Lex Fridman (07:03.340)
that make models of the world
Lex Fridman (07:04.700)
and try to minimize their own set points.
Lex Fridman (07:07.380)
So the models are the agents.
Joscha Bach (07:10.580)
The agent is a class of model.
Lex Fridman (07:12.580)
And we notice that we are an agent ourselves.
Joscha Bach (07:14.660)
We are the agent that is using our own control model
Lex Fridman (07:17.780)
to perform actions.
Joscha Bach (07:18.820)
We notice we produce a change in the model
Lex Fridman (07:22.100)
and things in the world change.
Lex Fridman (07:23.460)
And this is how we discover the idea that we have a body,
Lex Fridman (07:26.740)
that we are situated environment,
Lex Fridman (07:28.260)
and that we have a first person perspective.
Lex Fridman (07:31.140)
Still don't understand what's the best way to think
Joscha Bach (07:34.980)
of which object has agency with respect to human beings.
Lex Fridman (07:39.820)
Is it the body?
Lex Fridman (07:41.540)
Is it the brain?
Lex Fridman (07:43.420)
Is it the contents of the brain as agency?
Lex Fridman (07:46.020)
Like what's the actuators that you're referring to?
Lex Fridman (07:49.020)
What is the controller and where does it reside?
Lex Fridman (07:52.100)
Or is it these impossible things?
Lex Fridman (07:54.100)
Because I keep trying to ground it to space time,
Joscha Bach (07:57.740)
the three dimension of space and the one dimension of time.
Lex Fridman (08:01.580)
What's the agent in that for humans?
Joscha Bach (08:04.580)
There is not just one.
Lex Fridman (08:06.020)
It depends on the way in which you're looking
Joscha Bach (08:08.260)
at this thing in which you're framing it.
Lex Fridman (08:10.060)
Imagine that you are, say Angela Merkel,
Lex Fridman (08:13.540)
and you are acting on behalf of Germany.
Lex Fridman (08:16.660)
Then you could say that Germany is the agent.
Lex Fridman (08:19.820)
And in the mind of Angela Merkel,
Lex Fridman (08:21.580)
she is Germany to some extent,
Joscha Bach (08:23.540)
because in the way in which she acts,
Lex Fridman (08:25.660)
the destiny of Germany changes.
Joscha Bach (08:28.060)
There are things that she can change
Lex Fridman (08:29.700)
that basically affect the behavior of that nation state.
Joscha Bach (08:33.580)
Okay, so it's hierarchies of,
Lex Fridman (08:35.060)
to go to another one of your tweets
Joscha Bach (08:37.460)
with I think you were playfully mocking Jeff Hawkins
Lex Fridman (08:42.380)
with saying his brain's all the way down.
Lex Fridman (08:45.820)
So it's like, it's agents all the way down.
Lex Fridman (08:49.020)
It's agents made up of agents, made up of agents.
Joscha Bach (08:51.780)
Like if Angela Merkel's Germany
Lex Fridman (08:54.700)
and Germany's made up a bunch of people
Lex Fridman (08:56.540)
and the people are themselves agents
Lex Fridman (08:58.860)
in some kind of context.
Lex Fridman (09:01.060)
And then people are made up of cells, each individual.
Lex Fridman (09:04.900)
So is it agents all the way down?
Joscha Bach (09:07.220)
I suspect that has to be like this
Lex Fridman (09:08.900)
in a world where things are self organizing.
Joscha Bach (09:12.900)
Most of the complexity that we are looking at,
Lex Fridman (09:15.660)
everything in life is about self organization.
Lex Fridman (09:18.540)
So I think up from the level of life, you have agents.
Lex Fridman (09:24.220)
And below life, you rarely have agents
Joscha Bach (09:27.460)
because sometimes you have control systems
Lex Fridman (09:30.060)
that emerge randomly in nature
Lex Fridman (09:31.580)
and try to achieve a set point,
Lex Fridman (09:33.740)
but they're not that interesting agents that make models.
Lex Fridman (09:36.660)
And because to make an interesting model of the world,
Lex Fridman (09:39.580)
you typically need a system that is true and complete.
Lex Fridman (09:42.380)
Can I ask you a personal question?
Lex Fridman (09:46.140)
What's the line between life and non life?
Joscha Bach (09:48.740)
It's personal because you're a life form.
Lex Fridman (09:52.300)
So what do you think in this emerging complexity,
Joscha Bach (09:55.740)
at which point does the things that are being living
Lex Fridman (09:57.980)
and have agency?
Joscha Bach (10:00.220)
Personally, I think that the simplest answer
Lex Fridman (10:01.900)
that is that life is cells because...
Lex Fridman (10:04.540)
Life is what?
Lex Fridman (10:05.380)
Cells.
Joscha Bach (10:06.220)
Biological cells.
Lex Fridman (10:07.060)
Biological cells.
Lex Fridman (10:07.900)
So it's a particular kind of principle
Lex Fridman (10:09.420)
that we have discovered to exist in nature.
Joscha Bach (10:11.860)
It's modular stuff that consists
Lex Fridman (10:14.540)
out of basically this DNA tape
Joscha Bach (10:17.820)
with a read write head on top of it,
Lex Fridman (10:20.180)
that is able to perform arbitrary computations
Lex Fridman (10:23.220)
and state transitions within the cell.
Lex Fridman (10:25.300)
And it's combined with a membrane
Joscha Bach (10:27.380)
that insulates the cell from its environment.
Lex Fridman (10:30.540)
And there are chemical reactions inside of the cell
Joscha Bach (10:34.780)
that are in disequilibrium.
Lex Fridman (10:36.300)
And the cell is running in such a way
Joscha Bach (10:38.740)
that this disequilibrium doesn't disappear.
Lex Fridman (10:41.460)
And the cell goes into an equilibrium state, it dies.
Lex Fridman (10:46.260)
And it requires something like an neck entropy extractor
Lex Fridman (10:50.100)
to maintain this disequilibrium.
Lex Fridman (10:51.900)
So it's able to harvest like entropy from its environment
Lex Fridman (10:55.700)
and keep itself running.
Joscha Bach (10:57.820)
Yeah, so there's information and there's a wall
Lex Fridman (11:00.540)
to maintain this disequilibrium.
Lex Fridman (11:04.100)
But isn't this very earth centric?
Lex Fridman (11:06.660)
Like what you're referring to as a...
Joscha Bach (11:08.740)
I'm not making a normative notion.
Lex Fridman (11:10.660)
You could say that there are probably other things
Joscha Bach (11:13.100)
in the universe that are cell like and life like,
Lex Fridman (11:16.380)
and you could also call them life,
Lex Fridman (11:17.860)
but eventually it's just a willingness
Lex Fridman (11:21.220)
to find an agreement of how to use the terms.
Joscha Bach (11:23.820)
I like cells because it's completely coextential
Lex Fridman (11:26.580)
with the way that we use the word
Joscha Bach (11:28.500)
even before we knew about cells.
Lex Fridman (11:30.340)
So people were pointing at some stuff
Lex Fridman (11:32.380)
and saying, this is somehow animate.
Lex Fridman (11:34.260)
And this is very different from the non animate stuff.
Lex Fridman (11:36.620)
And what's the difference between the living
Lex Fridman (11:38.900)
and the dead stuff.
Lex Fridman (11:40.140)
And it's mostly whether the cells are working or not.
Lex Fridman (11:42.860)
And also this boundary of life,
Joscha Bach (11:45.340)
where we say that for instance, the virus
Lex Fridman (11:46.820)
is basically an information packet
Joscha Bach (11:48.820)
that is subverting the cell and not life by itself.
Lex Fridman (11:52.500)
That makes sense to me.
Lex Fridman (11:54.100)
And it's somewhat arbitrary.
Lex Fridman (11:55.860)
You could of course say that systems
Joscha Bach (11:57.940)
that permanently maintain a disequilibrium
Lex Fridman (12:00.140)
and can self replicate are always life.
Lex Fridman (12:03.340)
And maybe that's a useful definition too,
Lex Fridman (12:06.460)
but this is eventually just how you want to use the word.
Joscha Bach (12:10.420)
Is it so useful for conversation,
Lex Fridman (12:12.620)
but is it somehow fundamental to the universe?
Lex Fridman (12:17.340)
Do you think there's a actual line
Lex Fridman (12:19.300)
to eventually be drawn between life and non life?
Lex Fridman (12:21.860)
Or is it all a kind of continuum?
Lex Fridman (12:24.300)
I don't think it's a continuum,
Lex Fridman (12:25.460)
but there's nothing magical that is happening.
Lex Fridman (12:28.140)
Living systems are a certain type of machine.
Lex Fridman (12:31.140)
What about non living systems?
Lex Fridman (12:32.980)
Is it also a machine?
Joscha Bach (12:34.300)
There are non living machines,
Lex Fridman (12:35.980)
but the question is at which point is a system
Joscha Bach (12:38.540)
able to perform arbitrary state transitions
Lex Fridman (12:43.100)
to make representations.
Lex Fridman (12:44.940)
And living things can do this.
Lex Fridman (12:46.820)
And of course we can also build non living things
Joscha Bach (12:48.980)
that can do this, but we don't know anything in nature
Lex Fridman (12:52.220)
that is not a cell and is not created by still alive
Joscha Bach (12:56.180)
that is able to do that.
Lex Fridman (12:58.580)
Not only do we not know,
Joscha Bach (13:02.860)
I don't think we have the tools to see otherwise.
Lex Fridman (13:05.980)
I always worry that we look at the world too narrowly.
Joscha Bach (13:11.140)
Like there could be life of a very different kind
Lex Fridman (13:14.900)
right under our noses that we're just not seeing
Joscha Bach (13:18.860)
because we're not either limitations
Lex Fridman (13:21.700)
of our cognitive capacity,
Joscha Bach (13:23.260)
or we're just not open minded enough
Lex Fridman (13:26.980)
either with the tools of science
Joscha Bach (13:28.380)
or just the tools of our mind.
Lex Fridman (13:32.020)
Yeah, that's possible.
Joscha Bach (13:33.020)
I find this thought very fascinating.
Lex Fridman (13:35.060)
And I suspect that many of us ask ourselves since childhood,
Lex Fridman (13:39.020)
what are the things that we are missing?
Lex Fridman (13:40.580)
What kind of systems and interconnections exist
Lex Fridman (13:43.660)
that are outside of our gaze?
Lex Fridman (13:47.900)
But we are looking for it
Lex Fridman (13:51.140)
and physics doesn't have much room at the moment
Lex Fridman (13:55.140)
for opening up something that would not violate
Joscha Bach (13:59.580)
the conservation of information as we know it.
Lex Fridman (14:03.300)
Yeah, but I wonder about time scale and scale,
Joscha Bach (14:06.860)
spatial scale, whether we just need to open up our idea
Lex Fridman (14:11.860)
of what, like how life presents itself.
Joscha Bach (14:15.300)
It could be operating in a much slower time scale,
Lex Fridman (14:17.860)
a much faster time scale.
Lex Fridman (14:20.060)
And it's almost sad to think that there's all this life
Lex Fridman (14:23.940)
around us that we're not seeing
Joscha Bach (14:25.340)
because we're just not like thinking
Lex Fridman (14:29.220)
in terms of the right scale, both time and space.
Lex Fridman (14:34.380)
What is your definition of life?
Lex Fridman (14:36.060)
What do you understand as life?
Joscha Bach (14:40.780)
Entities of sufficiently high complexity
Lex Fridman (14:44.500)
that are full of surprises.
Joscha Bach (14:46.140)
I don't know, I don't have a free will.
Lex Fridman (14:53.980)
So that just came out of my mouth.
Joscha Bach (14:55.620)
I'm not sure that even makes sense.
Lex Fridman (14:57.260)
There's certain characteristics.
Lex Fridman (14:59.180)
So complexity seems to be a necessary property of life.
Lex Fridman (15:04.140)
And I almost want to say it has ability
Joscha Bach (15:09.980)
to do something unexpected.
Lex Fridman (15:13.340)
It seems to me that life is the main source
Joscha Bach (15:15.460)
of complexity on earth.
Lex Fridman (15:18.660)
Yes.
Lex Fridman (15:19.500)
And complexity is basically a bridgehead
Lex Fridman (15:22.060)
that order builds into chaos by modeling,
Joscha Bach (15:27.220)
by processing information in such a way
Lex Fridman (15:29.020)
that you can perform reactions
Joscha Bach (15:31.220)
that would not be possible for dump systems.
Lex Fridman (15:33.780)
And this means that you can harvest neck entropy
Joscha Bach (15:36.020)
that dump systems cannot harvest.
Lex Fridman (15:37.780)
And this is what complexity is mostly about.
Joscha Bach (15:40.140)
In some sense, the purpose of life is to create complexity.
Lex Fridman (15:45.060)
Yeah.
Joscha Bach (15:46.020)
Increasing.
Lex Fridman (15:46.860)
I mean, there seems to be some kind of universal drive
Joscha Bach (15:52.340)
towards increasing pockets of complexity.
Lex Fridman (15:56.180)
I don't know what that is.
Joscha Bach (15:57.580)
That seems to be like a fundamental,
Lex Fridman (16:00.020)
I don't know if it's a property of the universe
Joscha Bach (16:02.340)
or it's just a consequence of the way the universe works,
Lex Fridman (16:05.860)
but there seems to be this small pockets
Joscha Bach (16:08.700)
of emergent complexity that builds on top of each other
Lex Fridman (16:11.380)
and starts having like greater and greater complexity
Joscha Bach (16:15.380)
by having like a hierarchy of complexity.
Lex Fridman (16:17.980)
Little organisms building up a little society
Joscha Bach (16:20.700)
that then operates almost as an individual organism itself.
Lex Fridman (16:24.060)
And all of a sudden you have Germany and Merkel.
Joscha Bach (16:27.660)
Well, that's not obvious to me.
Lex Fridman (16:28.860)
Everything that goes up has to come down at some point.
Lex Fridman (16:32.500)
So if you see this big exponential curve somewhere,
Lex Fridman (16:36.500)
it's usually the beginning of an S curve
Joscha Bach (16:39.380)
where something eventually reaches saturation.
Lex Fridman (16:41.420)
And the S curve is the beginning of some kind of bump
Joscha Bach (16:43.820)
that goes down again.
Lex Fridman (16:45.500)
And there is just this thing that when you are
Joscha Bach (16:49.180)
in sight of an evolution of life,
Lex Fridman (16:53.220)
you are on top of a puddle of negentropy
Joscha Bach (16:55.820)
that is being sucked dry by life.
Lex Fridman (16:58.900)
And during that happening,
Joscha Bach (17:00.660)
you see an increase in complexity
Lex Fridman (17:02.940)
because life forms are competing with each other
Joscha Bach (17:04.780)
to get more and more finer and finer corner
Lex Fridman (17:09.340)
of that negentropy extraction.
Joscha Bach (17:11.620)
I feel like that's a gradual beautiful process
Lex Fridman (17:13.900)
like that's almost follows a process akin to evolution.
Lex Fridman (17:18.100)
And the way it comes down is not the same way it came up.
Lex Fridman (17:22.900)
The way it comes down is usually harshly and quickly.
Lex Fridman (17:27.380)
So usually there's some kind of catastrophic event.
Lex Fridman (17:30.620)
The Roman Empire took a long time.
Lex Fridman (17:32.380)
But would that be,
Lex Fridman (17:36.340)
would you classify this as a decrease in complexity though?
Joscha Bach (17:39.420)
Yes.
Lex Fridman (17:40.260)
I think that this size of the cities that could be fed
Joscha Bach (17:42.900)
has decreased dramatically.
Lex Fridman (17:44.740)
And you could see that the quality of the art decreased
Lex Fridman (17:47.820)
and it did so gradually.
Lex Fridman (17:49.940)
And maybe future generations,
Joscha Bach (17:53.260)
when they look at the history of the United States
Lex Fridman (17:55.660)
in the 21st century,
Joscha Bach (17:57.380)
will also talk about the gradual decline,
Lex Fridman (17:59.140)
not something that suddenly happens.
Lex Fridman (18:05.620)
Do you have a sense of where we are?
Lex Fridman (18:07.700)
Are we on the exponential rise?
Lex Fridman (18:09.740)
Are we at the peak?
Lex Fridman (18:11.260)
Or are we at the downslope of the United States empire?
Joscha Bach (18:15.740)
It's very hard to say from a single human perspective,
Lex Fridman (18:18.460)
but it seems to me that we are probably at the peak.
Joscha Bach (18:25.380)
I think that's probably the definition of like optimism
Lex Fridman (18:28.100)
and cynicism.
Lex Fridman (18:29.620)
So my nature of optimism is,
Lex Fridman (18:31.540)
I think we're on the rise.
Joscha Bach (18:36.940)
I think this is just all a matter of perspective.
Lex Fridman (18:39.300)
Nobody knows,
Lex Fridman (18:40.140)
but I do think that erring on the side of optimism,
Lex Fridman (18:43.260)
like you need a sufficient number,
Joscha Bach (18:45.420)
you need a minimum number of optimists
Lex Fridman (18:47.460)
in order to make that up thing actually work.
Lex Fridman (18:50.980)
And so I tend to be on the side of the optimists.
Lex Fridman (18:53.620)
I think that we are basically a species of grasshoppers
Joscha Bach (18:56.540)
that have turned into locusts.
Lex Fridman (18:58.620)
And when you are in that locust mode,
Joscha Bach (19:00.740)
you see an amazing rise of population numbers
Lex Fridman (19:04.100)
and of the complexity of the interactions
Joscha Bach (19:07.020)
between the individuals.
Lex Fridman (19:08.780)
But it's ultimately the question is, is it sustainable?
Joscha Bach (19:12.860)
See, I think we're a bunch of lions and tigers
Lex Fridman (19:16.140)
that have become domesticated cats,
Joscha Bach (19:20.220)
to use a different metaphor.
Lex Fridman (19:21.420)
As I'm not exactly sure we're so destructive,
Joscha Bach (19:24.300)
we're just softer and nicer and lazier.
Lex Fridman (19:27.820)
But I think we have monkeys and not the cats.
Lex Fridman (19:29.900)
And if you look at the monkeys, they are very busy.
Lex Fridman (19:33.620)
The ones that have a lot of sex, those monkeys?
Joscha Bach (19:35.820)
Not just the bonobos.
Lex Fridman (19:37.180)
I think that all the monkeys are basically
Joscha Bach (19:38.940)
a discontent species that always needs to meddle.
Lex Fridman (19:42.700)
Well, the gorillas seem to have
Joscha Bach (19:44.180)
a little bit more of a structure,
Lex Fridman (19:45.900)
but it's a different part of the tree.
Joscha Bach (19:50.620)
Okay, you mentioned the elephant
Lex Fridman (19:52.900)
and the monkey riding the elephant.
Lex Fridman (19:55.660)
And consciousness is the monkey.
Lex Fridman (1:00:02.020)
I don't see where the awareness that we're aware,
Joscha Bach (1:00:07.740)
the hard problem doesn't feel like it's solved.
Lex Fridman (1:00:10.620)
I mean, it's called a hard problem for a reason,
Joscha Bach (1:00:14.820)
because it seems like there needs to be a major leap.
Lex Fridman (1:00:19.380)
Yeah, I think the major leap is to understand
Lex Fridman (1:00:21.660)
how it is possible that a machine can dream,
Lex Fridman (1:00:25.300)
that a physical system is able to create a representation
Joscha Bach (1:00:29.540)
that the physical system is acting on,
Lex Fridman (1:00:31.260)
and that is spun force and so on.
Lex Fridman (1:00:33.980)
But once you accept the fact that you are not in physics,
Lex Fridman (1:00:36.700)
but that you exist inside of the story,
Joscha Bach (1:00:39.220)
I think the mystery disappears.
Lex Fridman (1:00:40.620)
Everything is possible in the story.
Joscha Bach (1:00:41.940)
You exist inside the story.
Lex Fridman (1:00:43.340)
Okay, so the machine.
Joscha Bach (1:00:44.180)
Your consciousness is being written into the story.
Lex Fridman (1:00:45.780)
The fact that you experience things
Joscha Bach (1:00:47.340)
is written to the side of the story.
Lex Fridman (1:00:48.860)
You ask yourself, is this real what I'm seeing?
Lex Fridman (1:00:51.300)
And your brain writes into the story, yes, it's real.
Lex Fridman (1:00:53.860)
So what about the perception of consciousness?
Lex Fridman (1:00:56.340)
So to me, you look conscious.
Lex Fridman (1:00:59.540)
So the illusion of consciousness,
Joscha Bach (1:01:02.500)
the demonstration of consciousness.
Lex Fridman (1:01:04.340)
I ask for the legged robot.
Lex Fridman (1:01:07.700)
How do we make this legged robot conscious?
Lex Fridman (1:01:10.580)
So there's two things,
Lex Fridman (1:01:12.820)
and maybe you can tell me if they're neighboring ideas.
Lex Fridman (1:01:16.340)
One is actually make it conscious,
Lex Fridman (1:01:18.860)
and the other is make it appear conscious to others.
Lex Fridman (1:01:22.620)
Are those related?
Joscha Bach (1:01:25.660)
Let's ask it from the other direction.
Lex Fridman (1:01:27.380)
What would it take to make you not conscious?
Lex Fridman (1:01:31.140)
So when you are thinking about how you perceive the world,
Lex Fridman (1:01:35.180)
can you decide to switch from looking at qualia
Lex Fridman (1:01:39.820)
to looking at representational states?
Lex Fridman (1:01:43.060)
And it turns out you can.
Joscha Bach (1:01:44.900)
There is a particular way in which you can look at the world
Lex Fridman (1:01:48.340)
and recognize its machine nature, including your own.
Lex Fridman (1:01:51.420)
And in that state,
Lex Fridman (1:01:52.420)
you don't have that conscious experience
Joscha Bach (1:01:54.260)
in this way anymore.
Lex Fridman (1:01:55.740)
It becomes apparent as a representation.
Joscha Bach (1:01:59.660)
Everything becomes opaque.
Lex Fridman (1:02:01.580)
And I think this thing that you recognize,
Joscha Bach (1:02:04.020)
everything is a representation.
Lex Fridman (1:02:05.380)
This is typically what we mean with enlightenment states.
Lex Fridman (1:02:09.100)
And it can happen on the motivational level,
Lex Fridman (1:02:11.700)
but you can also do this on the experiential level,
Joscha Bach (1:02:14.820)
on the perceptual level.
Lex Fridman (1:02:16.220)
See, but then I can come back to a conscious state.
Joscha Bach (1:02:20.420)
Okay, I particularly,
Lex Fridman (1:02:23.780)
I'm referring to the social aspect
Joscha Bach (1:02:26.940)
that the demonstration of consciousness
Lex Fridman (1:02:30.100)
is a really nice thing at a party
Joscha Bach (1:02:32.140)
when you're trying to meet a new person.
Lex Fridman (1:02:34.460)
It's a nice thing to know that they're conscious
Lex Fridman (1:02:38.300)
and they can,
Lex Fridman (1:02:41.020)
I don't know how fundamental consciousness
Joscha Bach (1:02:42.700)
is in human interaction,
Lex Fridman (1:02:43.900)
but it seems like to be at least an important part.
Lex Fridman (1:02:48.020)
And I ask that in the same kind of way for robots.
Lex Fridman (1:02:53.620)
In order to create a rich, compelling
Joscha Bach (1:02:56.340)
human robot interaction,
Lex Fridman (1:02:58.380)
it feels like there needs to be elements of consciousness
Joscha Bach (1:03:00.740)
within that interaction.
Lex Fridman (1:03:02.660)
My cat is obviously conscious.
Lex Fridman (1:03:04.900)
And so my cat can do this party trick.
Lex Fridman (1:03:07.380)
She also knows that I am conscious,
Joscha Bach (1:03:09.220)
be able to have feedback about the fact
Lex Fridman (1:03:11.380)
that we are both acting on models of our own awareness.
Joscha Bach (1:03:14.860)
The question is how hard is it for the robot,
Lex Fridman (1:03:19.660)
artificially created robot to achieve cat level
Lex Fridman (1:03:22.060)
and party tricks?
Lex Fridman (1:03:24.380)
Yes, so the issue for me is currently not so much
Joscha Bach (1:03:27.300)
on how to build a system that creates a story
Lex Fridman (1:03:30.300)
about a robot that lives in the world,
Lex Fridman (1:03:32.860)
but to make an adequate representation of the world.
Lex Fridman (1:03:36.540)
And the model that you and me have is a unified one.
Joscha Bach (1:03:40.260)
It's one where you basically make sense of everything
Lex Fridman (1:03:44.060)
that you can perceive.
Joscha Bach (1:03:44.980)
Every feature in the world that enters your perception
Lex Fridman (1:03:47.940)
can be relationally mapped to a unified model of everything.
Lex Fridman (1:03:51.780)
And we don't have an AI that is able to construct
Lex Fridman (1:03:54.060)
such a unified model yet.
Lex Fridman (1:03:56.220)
So you need that unified model to do the party trick?
Lex Fridman (1:03:58.820)
Yes, I think that it doesn't make sense
Joscha Bach (1:04:01.780)
if this thing is conscious,
Lex Fridman (1:04:03.060)
but not in the same universe as you,
Joscha Bach (1:04:04.660)
because you could not relate to each other.
Lex Fridman (1:04:06.780)
So what's the process, would you say,
Lex Fridman (1:04:08.980)
of engineering consciousness in the machine?
Lex Fridman (1:04:12.060)
Like what are the ideas here?
Lex Fridman (1:04:14.580)
So you probably want to have some kind of perceptual system.
Lex Fridman (1:04:19.060)
This perceptual system is a processing agent
Joscha Bach (1:04:21.300)
that is able to track sensory data
Lex Fridman (1:04:23.860)
and predict the next frame in the sensory data
Joscha Bach (1:04:26.740)
from the previous frames of the sensory data
Lex Fridman (1:04:29.780)
and the current state of the system.
Lex Fridman (1:04:31.740)
So the current state of the system is, in perception,
Lex Fridman (1:04:34.580)
instrumental to predicting what happens next.
Lex Fridman (1:04:37.580)
And this means you build lots and lots of functions
Lex Fridman (1:04:39.780)
that take all the blips that you feel on your skin
Lex Fridman (1:04:42.100)
and that you see on your retina, or that you hear,
Lex Fridman (1:04:45.540)
and puts them into a set of relationships
Joscha Bach (1:04:48.100)
that allows you to predict what kind of sensory data,
Lex Fridman (1:04:51.180)
what kind of sensor of blips, vector of blips,
Joscha Bach (1:04:53.780)
you're going to perceive in the next frame.
Lex Fridman (1:04:56.060)
This is tuned and it's constantly tuned
Joscha Bach (1:04:59.180)
until it gets as accurate as it can.
Lex Fridman (1:05:01.940)
You build a very accurate prediction mechanism
Joscha Bach (1:05:05.100)
that is step one of the perception.
Lex Fridman (1:05:08.060)
So first you predict, then you perceive
Lex Fridman (1:05:09.900)
and see the error in your prediction.
Lex Fridman (1:05:11.740)
And you have to do two things to make that happen.
Joscha Bach (1:05:13.820)
One is you have to build a network of relationships
Lex Fridman (1:05:16.900)
that are constraints,
Joscha Bach (1:05:18.460)
that take all the variants in the world
Lex Fridman (1:05:21.020)
and put each of the variances into a variable
Joscha Bach (1:05:24.420)
that is connected with relationships to other variables.
Lex Fridman (1:05:27.940)
And these relationships are computable functions
Joscha Bach (1:05:30.060)
that constrain each other.
Lex Fridman (1:05:31.140)
So when you see a nose
Joscha Bach (1:05:32.260)
that points in a certain direction in space,
Lex Fridman (1:05:34.900)
you have a constraint that says
Joscha Bach (1:05:36.100)
there should be a face nearby that has the same direction.
Lex Fridman (1:05:39.260)
And if that is not the case,
Joscha Bach (1:05:40.380)
you have some kind of contradiction
Lex Fridman (1:05:41.700)
that you need to resolve
Joscha Bach (1:05:42.540)
because it's probably not a nose what you're looking at.
Lex Fridman (1:05:44.620)
It just looks like one.
Lex Fridman (1:05:45.940)
So you have to reinterpret the data
Lex Fridman (1:05:48.580)
until you get to a point where your model converges.
Lex Fridman (1:05:52.460)
And this process of making the sensory data
Lex Fridman (1:05:54.940)
fit into your model structure
Joscha Bach (1:05:56.700)
is what Piaget calls the assimilation.
Lex Fridman (1:06:01.140)
And accommodation is the change of the models
Joscha Bach (1:06:04.060)
where you change your model in such a way
Lex Fridman (1:06:05.700)
that you can assimilate everything.
Lex Fridman (1:06:08.140)
So you're talking about building
Lex Fridman (1:06:09.860)
a hell of an awesome perception system
Joscha Bach (1:06:12.380)
that's able to do prediction and perception
Lex Fridman (1:06:14.700)
and correct and keep improving.
Joscha Bach (1:06:15.980)
No, wait, that's...
Lex Fridman (1:06:17.740)
Wait, there's more.
Joscha Bach (1:06:18.660)
Yes, there's more.
Lex Fridman (1:06:19.580)
So the first thing that we wanted to do
Joscha Bach (1:06:21.500)
is we want to minimize the contradictions in the model.
Lex Fridman (1:06:24.660)
And of course, it's very easy to make a model
Joscha Bach (1:06:26.700)
in which you minimize the contradictions
Lex Fridman (1:06:28.220)
just by allowing that it can be
Lex Fridman (1:06:29.700)
in many, many possible states, right?
Lex Fridman (1:06:31.500)
So if you increase degrees of freedom,
Joscha Bach (1:06:33.980)
you will have fewer contradictions.
Lex Fridman (1:06:35.860)
But you also want to reduce the degrees of freedom
Joscha Bach (1:06:37.820)
because degrees of freedom mean uncertainty.
Lex Fridman (1:06:40.260)
You want your model to reduce uncertainty
Joscha Bach (1:06:42.420)
as much as possible,
Lex Fridman (1:06:44.380)
but reducing uncertainty is expensive.
Lex Fridman (1:06:46.540)
So you have to have a trade off
Lex Fridman (1:06:47.780)
between minimizing contradictions
Lex Fridman (1:06:50.340)
and reducing uncertainty.
Lex Fridman (1:06:52.380)
And you have only a finite amount of compute
Lex Fridman (1:06:54.620)
and experimental time and effort
Lex Fridman (1:06:57.020)
available to reduce uncertainty in the world.
Lex Fridman (1:06:59.260)
So you need to assign value to what you observe.
Lex Fridman (1:07:02.740)
So you need some kind of motivational system
Joscha Bach (1:07:05.060)
that is estimating what you should be looking at
Lex Fridman (1:07:07.660)
and what you should be thinking about it,
Lex Fridman (1:07:09.180)
how you should be applying your resources
Lex Fridman (1:07:10.900)
to model what that is, right?
Lex Fridman (1:07:12.900)
So you need to have something like convergence links
Lex Fridman (1:07:15.940)
that tell you how to get from the present state
Joscha Bach (1:07:17.540)
of the model to the next one.
Lex Fridman (1:07:19.020)
You need to have these compatibility links
Joscha Bach (1:07:20.620)
that tell you which constraints exist
Lex Fridman (1:07:23.500)
and which constraint violations exist.
Lex Fridman (1:07:25.500)
And you need to have some kind of motivational system
Lex Fridman (1:07:28.900)
that tells you what to pay attention to.
Lex Fridman (1:07:30.700)
So now we have a second agent next to the perceptual agent.
Lex Fridman (1:07:32.980)
We have a motivational agent.
Joscha Bach (1:07:34.860)
This is a cybernetic system
Lex Fridman (1:07:36.260)
that is modeling what the system needs,
Joscha Bach (1:07:38.740)
what's important for the system,
Lex Fridman (1:07:40.460)
and that interacts with the perceptual system
Joscha Bach (1:07:42.100)
to maximize the expected reward.
Lex Fridman (1:07:44.540)
And you're saying the motivational system
Lex Fridman (1:07:46.020)
is some kind of like, what is it?
Lex Fridman (1:07:49.580)
A high level narrative over some lower level.
Joscha Bach (1:07:52.500)
No, it's just your brainstem stuff,
Lex Fridman (1:07:53.980)
the limbic system stuff that tells you,
Joscha Bach (1:07:55.660)
okay, now you should get something to eat
Lex Fridman (1:07:57.660)
because I've just measured your blood sugar.
Lex Fridman (1:07:59.380)
So you mean like motivational system,
Lex Fridman (1:08:00.940)
like the lower level stuff, like hungry.
Joscha Bach (1:08:03.060)
Yes, there's basically physiological needs
Lex Fridman (1:08:05.700)
and some cognitive needs and some social needs
Lex Fridman (1:08:07.500)
and they all interact.
Lex Fridman (1:08:08.420)
And they're all implemented at different parts
Joscha Bach (1:08:10.220)
in your nervous system as the motivational system.
Lex Fridman (1:08:12.660)
But they're basically cybernetic feedback loops.
Joscha Bach (1:08:14.700)
It's not that complicated.
Lex Fridman (1:08:16.420)
It's just a lot of code.
Lex Fridman (1:08:18.260)
And so you now have a motivational agent
Lex Fridman (1:08:21.420)
that makes your robot go for the ball
Joscha Bach (1:08:23.100)
or that makes your worm go to eat food and so on.
Lex Fridman (1:08:27.580)
And you have the perceptual system
Joscha Bach (1:08:29.140)
that lets it predict the environment
Lex Fridman (1:08:30.580)
so it's able to solve that control problem to some degree.
Lex Fridman (1:08:33.620)
And now what we learned is that it's very hard
Lex Fridman (1:08:35.780)
to build a machine learning system
Joscha Bach (1:08:37.220)
that looks at all the data simultaneously
Lex Fridman (1:08:39.340)
to see what kind of relationships
Joscha Bach (1:08:41.300)
could exist between them.
Lex Fridman (1:08:43.260)
So you need to selectively model the world.
Joscha Bach (1:08:45.580)
You need to figure out where can I make the biggest difference
Lex Fridman (1:08:48.300)
if I would put the following things together.
Joscha Bach (1:08:50.980)
Sometimes you find a gradient for that.
Lex Fridman (1:08:53.020)
When you have a gradient,
Joscha Bach (1:08:54.180)
you don't need to remember where you came from.
Lex Fridman (1:08:56.460)
You just follow the gradient
Joscha Bach (1:08:57.540)
until it doesn't get any better.
Lex Fridman (1:08:59.340)
But if you have a world where the problems are discontinuous
Lex Fridman (1:09:02.140)
and the search spaces are discontinuous,
Lex Fridman (1:09:04.260)
you need to retain memory of what you explored.
Joscha Bach (1:09:07.300)
You need to construct a plan of what to explore next.
Lex Fridman (1:09:10.540)
And this thing means that you have next
Joscha Bach (1:09:13.300)
to this perceptual construction system
Lex Fridman (1:09:15.340)
and the motivational cybernetics,
Joscha Bach (1:09:17.620)
an agent that is paying attention
Lex Fridman (1:09:20.220)
to what it should select at any given moment
Joscha Bach (1:09:22.700)
to maximize reward.
Lex Fridman (1:09:24.300)
And this scanning system, this attention agent,
Joscha Bach (1:09:27.460)
is required for consciousness
Lex Fridman (1:09:28.900)
and consciousness is its control model.
Lex Fridman (1:09:32.580)
So it's the index memories that this thing retains
Lex Fridman (1:09:36.140)
when it manipulates the perceptual representations
Joscha Bach (1:09:39.180)
to maximize the value and minimize the conflicts
Lex Fridman (1:09:43.020)
and to increase coherence.
Lex Fridman (1:09:44.820)
So the purpose of consciousness is to create coherence
Lex Fridman (1:09:47.740)
in your perceptual representations,
Joscha Bach (1:09:49.500)
remove conflicts, predict the future,
Lex Fridman (1:09:52.140)
construct counterfactual representations
Lex Fridman (1:09:54.100)
so you can coordinate your actions and so on.
Lex Fridman (1:09:57.460)
And in order to do this, it needs to form memories.
Joscha Bach (1:10:00.220)
These memories are partial binding states
Lex Fridman (1:10:02.340)
of the working memory contents
Joscha Bach (1:10:04.100)
that are being revisited later on to backtrack,
Lex Fridman (1:10:07.100)
to undo certain states, to look for alternatives.
Lex Fridman (1:10:10.140)
And these index memories that you can recall,
Lex Fridman (1:10:13.020)
that is what you perceive as your stream of consciousness.
Lex Fridman (1:10:15.940)
And being able to recall these memories,
Lex Fridman (1:10:17.860)
this is what makes you conscious.
Joscha Bach (1:10:19.420)
If you could not remember what you paid attention to,
Lex Fridman (1:10:21.660)
you wouldn't be conscious.
Lex Fridman (1:10:26.180)
So consciousness is the index in the memory database.
Lex Fridman (1:10:29.140)
Okay.
Lex Fridman (1:10:31.380)
But let me sneak up to the questions of consciousness
Lex Fridman (1:10:35.500)
a little further.
Lex Fridman (1:10:37.180)
So we usually relate suffering to consciousness.
Lex Fridman (1:10:42.660)
So the capacity to suffer.
Joscha Bach (1:10:46.260)
I think to me, that's a really strong sign of consciousness
Lex Fridman (1:10:49.700)
is a thing that can suffer.
Lex Fridman (1:10:52.460)
How is that useful?
Lex Fridman (1:10:55.100)
Suffering.
Lex Fridman (1:10:57.140)
And like in your model where you just described,
Lex Fridman (1:10:59.540)
which is indexing of memories and what is the coherence
Joscha Bach (1:11:03.740)
with the perception, with this predictive thing
Lex Fridman (1:11:07.220)
that's going on in the perception,
Lex Fridman (1:11:09.260)
how does suffering relate to any of that?
Lex Fridman (1:11:13.060)
The higher level suffering that humans do.
Joscha Bach (1:11:16.580)
Basically pain is a reinforcement signal.
Lex Fridman (1:11:20.020)
Pain is a signal that one part of your brain
Joscha Bach (1:11:23.380)
sends to another part of your brain,
Lex Fridman (1:11:25.140)
or in an abstract sense, part of your mind
Joscha Bach (1:11:27.940)
sends to another part of the mind to regulate its behavior,
Lex Fridman (1:11:30.860)
to tell it the behavior that you're currently exhibiting
Joscha Bach (1:11:33.540)
should be improved.
Lex Fridman (1:11:34.940)
And this is the signal that I tell you to move away
Joscha Bach (1:11:39.020)
from what you're currently doing
Lex Fridman (1:11:40.180)
and push into a different direction.
Lex Fridman (1:11:42.300)
So pain gives you a part of you an impulse
Lex Fridman (1:11:46.060)
to do something differently.
Lex Fridman (1:11:47.940)
But sometimes this doesn't work
Lex Fridman (1:11:49.940)
because the training part of your brain
Joscha Bach (1:11:52.140)
is talking to the wrong region,
Lex Fridman (1:11:54.220)
or because it has the wrong model
Joscha Bach (1:11:55.900)
of the relationships in the world.
Lex Fridman (1:11:57.180)
Maybe you're mismodeling yourself
Joscha Bach (1:11:58.580)
or you're mismodeling the relationship of yourself
Lex Fridman (1:12:00.860)
to the world,
Joscha Bach (1:12:01.700)
or you're mismodeling the dynamics of the world.
Lex Fridman (1:12:03.500)
So you're trying to improve something
Joscha Bach (1:12:04.940)
that cannot be improved by generating more pain.
Lex Fridman (1:12:07.900)
But the system doesn't have any alternative.
Lex Fridman (1:12:10.380)
So it doesn't get better.
Lex Fridman (1:12:12.340)
What do you do if something doesn't get better
Lex Fridman (1:12:14.220)
and you want it to get better?
Lex Fridman (1:12:15.540)
You increase the strengths of the signal.
Lex Fridman (1:12:17.940)
And then the signal becomes chronic
Lex Fridman (1:12:19.580)
when it becomes permanent without a change inside.
Joscha Bach (1:12:22.220)
This is what we call suffering.
Lex Fridman (1:12:24.300)
And the purpose of consciousness
Joscha Bach (1:12:26.420)
is to deal with contradictions,
Lex Fridman (1:12:28.180)
with things that cannot be resolved.
Joscha Bach (1:12:30.300)
The purpose of consciousness,
Lex Fridman (1:12:31.740)
I think is similar to a conductor in an orchestra.
Joscha Bach (1:12:35.060)
When everything works well,
Lex Fridman (1:12:36.420)
the orchestra doesn't need much of a conductor
Joscha Bach (1:12:38.580)
as long as it's coherent.
Lex Fridman (1:12:40.260)
But when there is a lack of coherence
Joscha Bach (1:12:42.020)
or something is consistently producing
Lex Fridman (1:12:44.340)
disharmony and mismatches,
Joscha Bach (1:12:46.220)
then the conductor becomes alert and interacts with it.
Lex Fridman (1:12:48.980)
So suffering attracts the activity of our consciousness.
Lex Fridman (1:12:52.660)
And the purpose of that is ideally
Lex Fridman (1:12:54.740)
that we bring new layers online,
Joscha Bach (1:12:56.660)
new layers of modeling that are able to create
Lex Fridman (1:13:00.460)
a model of the dysregulation so we can deal with it.
Lex Fridman (1:13:04.500)
And this means that we typically get
Lex Fridman (1:13:06.860)
higher level consciousness, so to speak, right?
Joscha Bach (1:13:08.820)
We get some consciousness above our pay grade maybe
Lex Fridman (1:13:11.420)
if we have some suffering early in our life.
Joscha Bach (1:13:13.260)
Most of the interesting people
Lex Fridman (1:13:14.820)
had trauma early on in their childhood.
Lex Fridman (1:13:17.060)
And trauma means that you are suffering an injury
Lex Fridman (1:13:20.940)
for which the system is not prepared,
Joscha Bach (1:13:23.060)
which it cannot deal with,
Lex Fridman (1:13:24.380)
which it cannot insulate itself from.
Lex Fridman (1:13:26.260)
So something breaks.
Lex Fridman (1:13:27.940)
And this means that the behavior of the system
Joscha Bach (1:13:29.860)
is permanently disturbed in a way
Lex Fridman (1:13:34.020)
that some mismatch exists now in the regulation
Joscha Bach (1:13:37.500)
that just by following your impulses,
Lex Fridman (1:13:39.100)
by following the pain in the direction where it hurts,
Joscha Bach (1:13:41.860)
the situation doesn't improve but get worse.
Lex Fridman (1:13:44.380)
And so what needs to happen is that you grow up.
Lex Fridman (1:13:47.940)
And that's part that has grown up
Lex Fridman (1:13:49.540)
is able to deal with the part
Joscha Bach (1:13:51.180)
that is stuck in this earlier phase.
Lex Fridman (1:13:53.340)
Yeah, so at least to grow,
Lex Fridman (1:13:54.580)
so you're adding extra layers to your cognition.
Lex Fridman (1:13:58.060)
And let me ask you then,
Joscha Bach (1:14:00.380)
because I gotta stick on suffering,
Lex Fridman (1:14:02.380)
the ethics of the whole thing.
Lex Fridman (1:14:05.420)
So not our consciousness, but the consciousness of others.
Lex Fridman (1:14:08.900)
You've tweeted, one of my biggest fears
Joscha Bach (1:14:13.380)
is that insects could be conscious.
Lex Fridman (1:14:16.260)
The amount of suffering on earth would be unthinkable.
Lex Fridman (1:14:20.300)
So when we think of other conscious beings,
Lex Fridman (1:14:24.380)
is suffering a property of consciousness
Lex Fridman (1:14:30.300)
that we're most concerned about?
Lex Fridman (1:14:32.660)
So I'm still thinking about robots,
Lex Fridman (1:14:40.140)
how to make sense of other nonhuman things
Lex Fridman (1:14:44.540)
that appear to have the depth of experience
Joscha Bach (1:14:48.380)
that humans have.
Lex Fridman (1:14:50.700)
And to me, that means consciousness
Lex Fridman (1:14:54.020)
and the darkest side of that, which is suffering,
Lex Fridman (1:14:57.460)
the capacity to suffer.
Lex Fridman (1:15:00.380)
And so I started thinking,
Lex Fridman (1:15:02.420)
how much responsibility do we have
Lex Fridman (1:15:04.100)
for those other conscious beings?
Lex Fridman (1:15:06.620)
That's where the definition of consciousness
Joscha Bach (1:15:10.980)
becomes most urgent.
Lex Fridman (1:15:13.100)
Like having to come up with a definition of consciousness
Joscha Bach (1:15:15.140)
becomes most urgent,
Lex Fridman (1:15:16.900)
is who should we and should we not be torturing?
Joscha Bach (1:15:24.740)
There's no general answer to this.
Lex Fridman (1:15:26.300)
Was Genghis Khan doing anything wrong?
Joscha Bach (1:15:29.100)
It depends right on how you look at it.
Lex Fridman (1:15:31.900)
Well, he drew a line somewhere
Joscha Bach (1:15:36.300)
where this is us and that's them.
Lex Fridman (1:15:38.820)
It's the circle of empathy.
Joscha Bach (1:15:40.820)
It's like these,
Lex Fridman (1:15:42.860)
you don't have to use the word consciousness,
Lex Fridman (1:15:44.820)
but these are the things that matter to me
Lex Fridman (1:15:48.980)
if they suffer or not.
Lex Fridman (1:15:50.100)
And these are the things that don't matter to him.
Lex Fridman (1:15:52.340)
Yeah, but when one of his commanders failed him,
Joscha Bach (1:15:54.580)
he broke his spine and let him die in a horrible way.
Lex Fridman (1:15:59.140)
And so in some sense,
Joscha Bach (1:16:01.420)
I think he was indifferent to suffering
Lex Fridman (1:16:03.860)
or he was not different in the sense
Joscha Bach (1:16:05.820)
that he didn't see it as useful if he inflicted suffering,
Lex Fridman (1:16:10.380)
but he did not see it as something that had to be avoided.
Joscha Bach (1:16:14.100)
That was not the goal.
Lex Fridman (1:16:15.460)
The question was, how can I use suffering
Lex Fridman (1:16:18.860)
and the infliction of suffering to reach my goals
Lex Fridman (1:16:21.260)
from his perspective?
Joscha Bach (1:16:23.900)
I see.
Lex Fridman (1:16:24.740)
So like different societies throughout history
Joscha Bach (1:16:26.700)
put different value on the...
Lex Fridman (1:16:29.940)
Different individuals, different psyches.
Lex Fridman (1:16:31.580)
But also even the objective of avoiding suffering,
Lex Fridman (1:16:35.100)
like some societies probably,
Joscha Bach (1:16:37.540)
I mean, this is where like religious belief really helps
Lex Fridman (1:16:40.740)
that afterlife, that it doesn't matter
Joscha Bach (1:16:44.700)
that you suffer or die,
Lex Fridman (1:16:45.980)
what matters is you suffer honorably, right?
Lex Fridman (1:16:49.300)
So that you enter the afterlife as a hero.
Lex Fridman (1:16:52.260)
It seems to be superstitious to me,
Joscha Bach (1:16:53.860)
basically beliefs that assert things
Lex Fridman (1:16:57.580)
for which no evidence exists
Joscha Bach (1:17:00.020)
are incompatible with sound epistemology.
Lex Fridman (1:17:02.180)
And I don't think that religion has to be superstitious,
Joscha Bach (1:17:04.620)
otherwise it should be condemned in all cases.
Lex Fridman (1:17:06.860)
You're somebody who's saying we live in a dream world,
Joscha Bach (1:17:09.140)
we have zero evidence for anything.
Lex Fridman (1:17:11.340)
So...
Joscha Bach (1:17:12.180)
That's not the case.
Lex Fridman (1:17:13.500)
There are limits to what languages can be constructed.
Joscha Bach (1:17:16.060)
Mathematics brings solid evidence for its own structure.
Lex Fridman (1:17:19.500)
And once we have some idea of what languages exist
Lex Fridman (1:17:23.260)
and how a system can learn
Lex Fridman (1:17:24.460)
and what learning itself is in the first place.
Lex Fridman (1:17:26.580)
And so we can begin to realize that our intuitions
Lex Fridman (1:17:31.900)
that we are able to learn about the regularities
Joscha Bach (1:17:34.620)
of the world and minimize surprise
Lex Fridman (1:17:36.300)
and understand the nature of our own agency
Joscha Bach (1:17:38.900)
to some degree of abstraction.
Lex Fridman (1:17:40.660)
That's not an illusion.
Lex Fridman (1:17:42.140)
So it's a useful approximation.
Lex Fridman (1:17:44.140)
Just because we live in a dream world
Joscha Bach (1:17:46.860)
doesn't mean mathematics can't give us a consistent glimpse
Lex Fridman (1:17:51.780)
of physical, of objective reality.
Joscha Bach (1:17:54.940)
We can basically distinguish useful encodings
Lex Fridman (1:17:57.340)
from useless encodings.
Lex Fridman (1:17:58.980)
And when we apply our truth seeking to the world,
Lex Fridman (1:18:03.460)
we know we usually cannot find out
Joscha Bach (1:18:05.460)
whether a certain thing is true.
Lex Fridman (1:18:07.460)
What we typically do is we take the state vector
Joscha Bach (1:18:10.060)
of the universe separated into separate objects
Lex Fridman (1:18:12.100)
that interact with each other through interfaces.
Lex Fridman (1:18:14.420)
And this distinction that we are making
Lex Fridman (1:18:16.180)
is not completely arbitrary.
Joscha Bach (1:18:17.420)
It's done to optimize the compression
Lex Fridman (1:18:21.140)
that we can apply to our models of the universe.
Lex Fridman (1:18:23.380)
So we can predict what's happening
Lex Fridman (1:18:25.660)
with our limited resources.
Joscha Bach (1:18:27.300)
In this sense, it's not arbitrary.
Lex Fridman (1:18:29.220)
But the separation of the world into objects
Joscha Bach (1:18:32.020)
that are somehow discrete and interacting with each other
Lex Fridman (1:18:34.940)
is not the true reality, right?
Joscha Bach (1:18:36.900)
The boundaries between the objects
Lex Fridman (1:18:38.420)
are projected into the world, not arbitrarily projected.
Lex Fridman (1:18:41.660)
But still, it's only an approximation
Lex Fridman (1:18:44.020)
of what's actually the case.
Lex Fridman (1:18:46.460)
And we sometimes notice that we run into contradictions
Lex Fridman (1:18:48.980)
when we try to understand high level things
Joscha Bach (1:18:50.980)
like economic aspects of the world
Lex Fridman (1:18:53.100)
and so on, or political aspects, or psychological aspects
Joscha Bach (1:18:56.980)
where we make simplifications.
Lex Fridman (1:18:58.300)
And the objects that we are using to separate the world
Joscha Bach (1:19:00.780)
are just one of many possible projections
Lex Fridman (1:19:03.100)
of what's going on.
Lex Fridman (1:19:04.820)
So it's not, in this postmodernist sense,
Lex Fridman (1:19:07.180)
completely arbitrary, and you're free to pick
Lex Fridman (1:19:09.260)
what you want or dismiss what you don't like
Lex Fridman (1:19:11.100)
because it's all stories.
Joscha Bach (1:19:12.260)
No, that's not true.
Lex Fridman (1:19:13.660)
You have to show for every model
Joscha Bach (1:19:15.380)
of how well it predicts the world.
Lex Fridman (1:19:17.340)
So the confidence that you should have
Joscha Bach (1:19:19.220)
in the entities of your models
Lex Fridman (1:19:21.020)
should correspond to the evidence that you have.
Joscha Bach (1:19:24.460)
Can I ask you on a small tangent
Lex Fridman (1:19:27.660)
to talk about your favorite set of ideas and people,
Joscha Bach (1:19:32.660)
which is postmodernism.
Lex Fridman (1:19:35.180)
What?
Lex Fridman (1:19:37.900)
What is postmodernism?
Lex Fridman (1:19:39.980)
How would you define it?
Lex Fridman (1:19:40.980)
And why to you is it not a useful framework of thought?
Lex Fridman (1:19:48.860)
Postmodernism is something that I'm really not an expert on.
Lex Fridman (1:19:52.340)
And postmodernism is a set of philosophical ideas
Lex Fridman (1:19:57.340)
that is difficult to lump together,
Joscha Bach (1:19:58.980)
that is characterized by some useful thinkers,
Lex Fridman (1:20:01.980)
some of them poststructuralists and so on.
Lex Fridman (1:20:04.180)
And I'm mostly not interested in it
Lex Fridman (1:20:05.740)
because I think that it's not leading me anywhere
Joscha Bach (1:20:08.660)
that I find particularly useful.
Lex Fridman (1:20:11.340)
It's mostly, I think, born out of the insight
Joscha Bach (1:20:13.820)
that the ontologies that we impose on the world
Lex Fridman (1:20:17.380)
are not literally true.
Lex Fridman (1:20:18.780)
And that we can often get to a different interpretation
Lex Fridman (1:20:20.980)
by the world by using a different ontology
Joscha Bach (1:20:22.780)
that is different separation of the world
Lex Fridman (1:20:25.060)
into interacting objects.
Lex Fridman (1:20:26.540)
But the idea that this makes the world a set of stories
Lex Fridman (1:20:30.860)
that are arbitrary, I think, is wrong.
Lex Fridman (1:20:33.180)
And the people that are engaging in this type of philosophy
Lex Fridman (1:20:37.540)
are working in an area that I largely don't find productive.
Joscha Bach (1:20:40.900)
There's nothing useful coming out of this.
Lex Fridman (1:20:43.060)
So this idea that truth is relative
Joscha Bach (1:20:45.060)
is not something that has, in some sense,
Lex Fridman (1:20:46.980)
informed physics or theory of relativity.
Lex Fridman (1:20:49.620)
And there is no feedback between those.
Lex Fridman (1:20:51.540)
There is no meaningful information
Joscha Bach (1:20:54.020)
of this type of philosophy on the sciences
Lex Fridman (1:20:56.860)
or on engineering or in politics.
Lex Fridman (1:20:59.340)
But there is a very strong information on ideology
Lex Fridman (1:21:04.820)
because it basically has become an ideology
Joscha Bach (1:21:07.620)
that is justifying itself by the notion
Lex Fridman (1:21:11.100)
that truth is a relative concept.
Lex Fridman (1:21:13.420)
And it's not being used in such a way
Lex Fridman (1:21:15.620)
that the philosophers or sociologists
Joscha Bach (1:21:18.540)
that take up these ideas say,
Lex Fridman (1:21:20.340)
oh, I should doubt my own ideas because maybe my separation of the world
Joscha Bach (1:21:24.140)
into objects is not completely valid.
Lex Fridman (1:21:25.740)
And I should maybe use a different one
Lex Fridman (1:21:27.580)
and be open to a pluralism of ideas.
Lex Fridman (1:21:30.460)
But it mostly exists to dismiss the ideas of other people.
Joscha Bach (1:21:34.300)
It becomes, yeah, it becomes a political weapon of sorts
Lex Fridman (1:21:37.540)
to achieve power.
Joscha Bach (1:21:39.220)
Basically, there's nothing wrong, I think,
Lex Fridman (1:21:42.580)
with developing a philosophy around this.
Lex Fridman (1:21:46.060)
But to develop a philosophy around this,
Lex Fridman (1:21:49.180)
to develop norms around the idea
Joscha Bach (1:21:51.820)
that truth is something that is completely negotiable,
Lex Fridman (1:21:54.940)
is incompatible with the scientific project.
Lex Fridman (1:21:57.860)
And I think if the academia has no defense
Lex Fridman (1:22:02.140)
against the ideological parts of the postmodernist movement,
Joscha Bach (1:22:06.740)
it's doomed.
Lex Fridman (1:22:07.860)
Right, you have to acknowledge the ideological part
Joscha Bach (1:22:11.740)
of any movement, actually, including postmodernism.
Lex Fridman (1:22:15.260)
Well, the question is what an ideology is.
Lex Fridman (1:22:17.500)
And to me, an ideology is basically a viral memeplex
Lex Fridman (1:22:21.060)
that is changing your mind in such a way that reality gets warped.
Joscha Bach (1:22:25.980)
It gets warped in such a way that you're being cut off
Lex Fridman (1:22:28.180)
from the rest of human thought space.
Lex Fridman (1:22:29.540)
And you cannot consider things outside of the range of ideas
Lex Fridman (1:22:33.420)
of your own ideology as possibly true.
Joscha Bach (1:22:35.780)
Right, so, I mean, there's certain properties to an ideology
Lex Fridman (1:22:38.380)
that make it harmful.
Joscha Bach (1:22:39.340)
One of them is that dogmatism of just certainty,
Lex Fridman (1:22:44.060)
dogged certainty in that you're right,
Joscha Bach (1:22:46.660)
you have the truth, and nobody else does.
Lex Fridman (1:22:48.780)
Yeah, but what is creating the certainty?
Joscha Bach (1:22:50.220)
It's very interesting to look at the type of model
Lex Fridman (1:22:53.060)
that is being produced.
Joscha Bach (1:22:54.100)
Is it basically just a strong prior, and you tell people,
Lex Fridman (1:22:57.500)
oh, this idea that you consider to be very true,
Joscha Bach (1:22:59.980)
the evidence for this is actually just much weaker
Lex Fridman (1:23:02.220)
than you thought, and look, here are some studies.
Joscha Bach (1:23:04.380)
No, this is not how it works.
Lex Fridman (1:23:06.100)
It's usually normative, which means some thoughts
Joscha Bach (1:23:09.260)
are unthinkable because they would change your identity
Lex Fridman (1:23:13.780)
into something that is no longer acceptable.
Lex Fridman (1:23:17.100)
And this cuts you off from considering an alternative.
Lex Fridman (1:23:20.100)
And many de facto religions use this trick
Joscha Bach (1:23:23.220)
to lock people into a certain mode of thought,
Lex Fridman (1:23:25.700)
and this removes agency over your own thoughts.
Lex Fridman (1:23:27.700)
And it's very ugly to me.
Lex Fridman (1:23:28.660)
It's basically not just a process of domestication,
Lex Fridman (1:23:32.580)
but it's actually an intellectual castration
Lex Fridman (1:23:35.180)
that happens.
Joscha Bach (1:23:36.220)
It's an inability to think creatively
Lex Fridman (1:23:39.140)
and to bring forth new thoughts.
Joscha Bach (1:23:40.820)
I can ask you about substances, chemical substances
Lex Fridman (1:23:48.300)
that affect the video game, the dream world.
Lex Fridman (1:23:53.140)
So psychedelics that increasingly have been getting
Lex Fridman (1:23:57.140)
a lot of research done on them.
Lex Fridman (1:23:58.820)
So in general, psychedelics, psilocybin, MDMA,
Lex Fridman (1:24:02.620)
but also a really interesting one, the big one, which is DMT.
Lex Fridman (1:24:06.300)
What and where are the places that these substances
Lex Fridman (1:24:10.820)
take the mind that is operating in the dream world?
Lex Fridman (1:24:16.620)
Do you have an interesting sense how this throws a wrinkle
Lex Fridman (1:24:20.340)
into the prediction model?
Joscha Bach (1:24:22.260)
Is it just some weird little quirk
Lex Fridman (1:24:24.500)
or is there some fundamental expansion
Lex Fridman (1:24:27.820)
of the mind going on?
Lex Fridman (1:24:31.700)
I suspect that a way to look at psychedelics
Joscha Bach (1:24:34.140)
is that they induce particular types
Lex Fridman (1:24:36.420)
of lucid dreaming states.
Lex Fridman (1:24:38.540)
So it's a state in which certain connections
Lex Fridman (1:24:41.620)
are being severed in your mind.
Joscha Bach (1:24:43.820)
They're no longer active.
Lex Fridman (1:24:45.340)
Your mind basically gets free to move in a certain direction
Joscha Bach (1:24:48.860)
because some inhibition, some particular inhibition
Lex Fridman (1:24:51.060)
doesn't work anymore.
Lex Fridman (1:24:52.740)
And as a result, you might stop having a self
Lex Fridman (1:24:55.340)
or you might stop perceiving the world as three dimensional.
Lex Fridman (1:25:00.340)
And you can explore that state.
Lex Fridman (1:25:04.500)
And I suppose that for every state
Joscha Bach (1:25:06.780)
that can be induced with psychedelics,
Lex Fridman (1:25:08.300)
there are people that are naturally in that state.
Lex Fridman (1:25:10.980)
So sometimes psychedelics to shift you
Lex Fridman (1:25:13.340)
through a range of possible mental states.
Lex Fridman (1:25:15.860)
And they can also shift you out of the range
Lex Fridman (1:25:17.820)
of permissible mental states
Joscha Bach (1:25:19.100)
that is where you can make predictive models of reality.
Lex Fridman (1:25:22.620)
And what I observe in people that use psychedelics a lot
Joscha Bach (1:25:26.980)
is that they tend to be overfitting.
Lex Fridman (1:25:29.580)
Overfitting means that you are using more bits
Joscha Bach (1:25:34.540)
for modeling the dynamics of a function than you should.
Lex Fridman (1:25:38.060)
And so you can fit your curve
Joscha Bach (1:25:40.220)
to extremely detailed things in the past,
Lex Fridman (1:25:42.780)
but this model is no longer predictive for the future.
Lex Fridman (1:25:45.860)
What is it about psychedelics that forces that?
Lex Fridman (1:25:49.660)
I thought it would be the opposite.
Joscha Bach (1:25:51.060)
I thought that it's a good mechanism
Lex Fridman (1:25:54.860)
for generalization, for regularization.
Lex Fridman (1:25:59.300)
So it feels like psychedelics expansion of the mind,
Lex Fridman (1:26:03.220)
like taking you outside of,
Joscha Bach (1:26:04.820)
like forcing your model to be non predictive
Lex Fridman (1:26:08.820)
is a good thing.
Joscha Bach (1:26:11.180)
Meaning like, it's almost like, okay,
Lex Fridman (1:26:14.340)
what I would say psychedelics are akin to
Joscha Bach (1:26:16.700)
is traveling to a totally different environment.
Lex Fridman (1:26:19.820)
Like going, if you've never been to like India
Joscha Bach (1:26:21.980)
or something like that from the United States,
Lex Fridman (1:26:24.220)
very different set of people, different culture,
Joscha Bach (1:26:26.180)
different food, different roads and values
Lex Fridman (1:26:30.340)
and all those kinds of things.
Joscha Bach (1:26:31.420)
Yeah, so psychedelics can, for instance,
Lex Fridman (1:26:33.540)
teleport people into a universe that is hyperbolic,
Joscha Bach (1:26:37.820)
which means that if you imagine a room that you're in,
Lex Fridman (1:26:41.300)
you can turn around 360 degrees
Lex Fridman (1:26:43.580)
and you didn't go full circle.
Lex Fridman (1:26:44.660)
You need to go 720 degrees to go full circle.
Joscha Bach (1:26:47.180)
Exactly.
Lex Fridman (1:26:48.020)
So the things that people learn in that state
Joscha Bach (1:26:50.820)
cannot be easily transferred
Lex Fridman (1:26:52.180)
in this universe that we are in.
Joscha Bach (1:26:54.260)
It could be that if they're able to abstract
Lex Fridman (1:26:56.420)
and understand what happened to them,
Joscha Bach (1:26:58.260)
that they understand that some part
Lex Fridman (1:27:00.300)
of their spatial cognition has been desynchronized
Lex Fridman (1:27:03.500)
and has found a different synchronization.
Lex Fridman (1:27:05.660)
And this different synchronization
Lex Fridman (1:27:06.900)
happens to be a hyperbolic one, right?
Lex Fridman (1:27:08.620)
So you learn something interesting about your brain.
Joscha Bach (1:27:10.980)
It's difficult to understand what exactly happened,
Lex Fridman (1:27:13.140)
but we get a pretty good idea
Joscha Bach (1:27:14.580)
once we understand how the brain is representing geometry.
Lex Fridman (1:27:17.700)
Yeah, but doesn't it give you a fresh perspective
Lex Fridman (1:27:20.180)
on the physical reality?
Lex Fridman (1:27:26.060)
Who's making that sound?
Lex Fridman (1:27:27.780)
Is it inside my head or is it external?
Lex Fridman (1:27:30.980)
Well, there is no sound outside of your mind,
Lex Fridman (1:27:33.180)
but it's making sense of phenomenon physics.
Lex Fridman (1:27:39.660)
Yeah, in the physical reality, there's sound waves
Joscha Bach (1:27:44.780)
traveling through air.
Lex Fridman (1:27:45.860)
Okay.
Joscha Bach (1:27:47.020)
That's our model of what happened.
Lex Fridman (1:27:48.580)
That's our model of what happened, right.
Joscha Bach (1:27:53.380)
Doesn't Psychedelics give you a fresh perspective
Lex Fridman (1:27:57.220)
on this physical reality?
Joscha Bach (1:27:59.100)
Like, not this physical reality, but this more...
Lex Fridman (1:28:05.980)
What do you call the dream world that's mapped directly to...
Joscha Bach (1:28:09.780)
The purpose of dreaming at night, I think,
Lex Fridman (1:28:11.580)
is data augmentation.
Joscha Bach (1:28:13.660)
Exactly.
Lex Fridman (1:28:14.900)
So that's very different.
Joscha Bach (1:28:16.300)
That's very similar to Psychedelics.
Lex Fridman (1:28:18.780)
It's changed parameters about the things that you have learned.
Joscha Bach (1:28:21.660)
And, for instance, when you are young,
Lex Fridman (1:28:24.140)
you have seen things from certain perspectives,
Lex Fridman (1:28:26.060)
but not from others.
Lex Fridman (1:28:27.300)
So your brain is generating new perspectives of objects
Joscha Bach (1:28:30.180)
that you already know,
Lex Fridman (1:28:31.540)
which means you can learn to recognize them later
Joscha Bach (1:28:34.100)
from different perspectives.
Lex Fridman (1:28:35.180)
And I suspect that's the reason that many of us
Joscha Bach (1:28:37.660)
remember to have flying dreams as children,
Lex Fridman (1:28:39.700)
because it's just different perspectives of the world
Joscha Bach (1:28:41.700)
that you already know,
Lex Fridman (1:28:43.020)
and that it starts to generate these different
Joscha Bach (1:28:46.540)
perspective changes,
Lex Fridman (1:28:47.860)
and then it fluidly turns this into a flying dream
Lex Fridman (1:28:50.540)
to make sense of what's happening, right?
Lex Fridman (1:28:52.260)
So you fill in the gaps,
Lex Fridman (1:28:53.620)
and suddenly you see yourself flying.
Lex Fridman (1:28:55.860)
And similar things can happen with semantic relationships.
Lex Fridman (1:28:58.820)
So it's not just spatial relationships,
Lex Fridman (1:29:00.580)
but it can also be the relationships between ideas
Joscha Bach (1:29:03.420)
that are being changed.
Lex Fridman (1:29:05.180)
And it seems that the mechanisms that make that happen
Joscha Bach (1:29:08.140)
during dreaming are interacting
Lex Fridman (1:29:12.060)
with these same receptors
Joscha Bach (1:29:14.300)
that are being stimulated by psychedelics.
Lex Fridman (1:29:17.220)
So I suspect that there is a thing
Joscha Bach (1:29:19.780)
that I haven't read really about.
Lex Fridman (1:29:22.020)
The way in which dreams are induced in the brain
Joscha Bach (1:29:24.380)
is not just that the activity of the brain gets tuned down
Lex Fridman (1:29:28.500)
because your eyes are closed
Lex Fridman (1:29:30.620)
and you no longer get enough data from your eyes,
Lex Fridman (1:29:33.980)
but there is a particular type of neurotransmitter
Joscha Bach (1:29:37.180)
that is saturating your brain during these phases,
Lex Fridman (1:29:40.140)
during the REM phases, and you produce
Joscha Bach (1:29:42.980)
controlled hallucinations.
Lex Fridman (1:29:44.740)
And psychedelics are linking into these mechanisms,
Joscha Bach (1:29:48.700)
I suspect.
Lex Fridman (1:29:49.860)
So isn't that another trickier form of data augmentation?
Joscha Bach (1:29:54.060)
Yes, but it's also data augmentation
Lex Fridman (1:29:57.740)
that can happen outside of the specification
Joscha Bach (1:29:59.860)
that your brain is tuned to.
Lex Fridman (1:30:00.940)
So basically people are overclocking their brains
Lex Fridman (1:30:03.420)
and that produces states
Lex Fridman (1:30:05.780)
that are subjectively extremely interesting.
Joscha Bach (1:30:09.260)
Yeah, I just.
Lex Fridman (1:30:10.540)
But from the outside, very suspicious.
Lex Fridman (1:30:12.860)
So I think I'm over applying the metaphor
Lex Fridman (1:30:15.660)
of a neural network in my own mind,
Lex Fridman (1:30:17.860)
which I just think that doesn't lead to overfitting, right?
Lex Fridman (1:30:22.460)
But you were just sort of anecdotally saying
Joscha Bach (1:30:26.380)
my experiences with people that have done psychedelics
Lex Fridman (1:30:28.660)
are that kind of quality.
Joscha Bach (1:30:30.460)
I think it typically happens.
Lex Fridman (1:30:31.580)
So if you look at people like Timothy Leary,
Lex Fridman (1:30:34.420)
and he has written beautiful manifestos
Lex Fridman (1:30:36.700)
about the effect of LSD on people.
Joscha Bach (1:30:40.220)
He genuinely believed, he writes in these manifestos,
Lex Fridman (1:30:42.820)
that in the future, science and art
Joscha Bach (1:30:44.860)
will only be done on psychedelics
Lex Fridman (1:30:46.300)
because it's so much more efficient and so much better.
Lex Fridman (1:30:49.020)
And he gave LSD to children in this community
Lex Fridman (1:30:52.660)
of a few thousand people that he had near San Francisco.
Lex Fridman (1:30:55.780)
And basically he was losing touch with reality.
Lex Fridman (1:31:00.540)
He did not understand the effects
Joscha Bach (1:31:02.220)
that the things that he was doing
Lex Fridman (1:31:04.260)
would have on the reception of psychedelics
Joscha Bach (1:31:06.620)
by society because he was unable to think critically
Lex Fridman (1:31:09.900)
about what happened.
Lex Fridman (1:31:10.740)
What happened was that he got in a euphoric state,
Lex Fridman (1:31:13.500)
that euphoric state happened because he was overfitting.
Joscha Bach (1:31:16.620)
He was taking this sense of euphoria
Lex Fridman (1:31:19.460)
and translating it into a model
Lex Fridman (1:31:21.500)
of actual success in the world, right?
Lex Fridman (1:31:23.660)
He was feeling better.
Joscha Bach (1:31:25.500)
Limitations had disappeared,
Lex Fridman (1:31:26.940)
that he experienced to be existing,
Lex Fridman (1:31:29.580)
but he didn't get superpowers.
Lex Fridman (1:31:30.740)
I understand what you mean by overfitting now.
Joscha Bach (1:31:33.860)
There's a lot of interpretation to the term
Lex Fridman (1:31:36.020)
overfitting in this case, but I got you.
Lex Fridman (1:31:38.660)
So he was getting positive rewards
Lex Fridman (1:31:42.740)
from a lot of actions that he shouldn't have been doing.
Joscha Bach (1:31:44.220)
Yeah, but not just this.
Lex Fridman (1:31:45.060)
So if you take, for instance, John Lilly,
Joscha Bach (1:31:46.620)
who was studying dolphin languages and aliens and so on,
Lex Fridman (1:31:52.140)
a lot of people that use psychedelics became very loopy.
Lex Fridman (1:31:55.900)
And the typical thing that you notice
Lex Fridman (1:31:58.700)
when people are on psychedelics is that they are in a state
Joscha Bach (1:32:00.940)
where they feel that everything can be explained now.
Lex Fridman (1:32:03.660)
Everything is clear, everything is obvious.
Lex Fridman (1:32:06.620)
And sometimes they have indeed discovered
Lex Fridman (1:32:09.660)
a useful connection, but not always.
Joscha Bach (1:32:12.060)
Very often these connections are overinterpretations.
Lex Fridman (1:32:15.380)
I wonder, you know, there's a question
Joscha Bach (1:32:17.740)
of correlation versus causation.
Lex Fridman (1:32:21.060)
And also I wonder if it's the psychedelics
Joscha Bach (1:32:23.340)
or if it's more the social, like being the outsider
Lex Fridman (1:32:28.580)
and having a strong community of outside
Lex Fridman (1:32:31.140)
and having a leadership position in an outsider cult
Lex Fridman (1:32:34.300)
like community that could have a much stronger effect
Joscha Bach (1:32:37.420)
of overfitting than do psychedelics themselves,
Lex Fridman (1:32:39.940)
the actual substances, because it's a counterculture thing.
Lex Fridman (1:32:43.340)
So it could be that as opposed to the actual substance.
Lex Fridman (1:32:46.540)
If you're a boring person who wears a suit and tie
Lex Fridman (1:32:49.700)
and works at a bank and takes psychedelics,
Lex Fridman (1:32:53.220)
that could be a very different effect
Joscha Bach (1:32:55.140)
of psychedelics on your mind.
Lex Fridman (1:32:57.820)
I'm just sort of raising the point
Joscha Bach (1:32:59.660)
that the people you referenced are already weirdos.
Lex Fridman (1:33:02.860)
I'm not sure exactly.
Joscha Bach (1:33:04.180)
No, not necessarily.
Lex Fridman (1:33:05.220)
A lot of the people that tell me
Joscha Bach (1:33:07.540)
that they use psychedelics in a useful way
Lex Fridman (1:33:10.980)
started out as squares and were liberating themselves
Joscha Bach (1:33:14.580)
because they were stuck.
Lex Fridman (1:33:16.060)
They were basically stuck in local optimum
Joscha Bach (1:33:17.980)
of their own self model, of their relationship to the world.
Lex Fridman (1:33:20.980)
And suddenly they had data augmentation.
Joscha Bach (1:33:23.180)
They basically saw and experienced a space of possibilities.
Lex Fridman (1:33:26.740)
They experienced what it would be like to be another person.
Lex Fridman (1:33:29.540)
And they took important lessons
Lex Fridman (1:33:32.260)
from that experience back home.
Joscha Bach (1:33:36.660)
Yeah, I mean, I love the metaphor of data augmentation
Lex Fridman (1:33:40.660)
because that's been the primary driver
Joscha Bach (1:33:44.900)
of self supervised learning in the computer vision domain
Lex Fridman (1:33:48.900)
is data augmentation.
Lex Fridman (1:33:50.100)
So it's funny to think of data augmentation,
Lex Fridman (1:33:53.100)
like chemically induced data augmentation in the human mind.
Joscha Bach (1:33:58.100)
There's also a very interesting effect that I noticed.
Lex Fridman (1:34:02.100)
I know several people who are sphere to me
Joscha Bach (1:34:06.100)
that LSD has cured their migraines.
Lex Fridman (1:34:09.780)
So severe cluster headaches or migraines
Joscha Bach (1:34:13.060)
that didn't respond to standard medication
Lex Fridman (1:34:15.900)
that disappeared after a single dose.
Lex Fridman (1:34:18.220)
And I don't recommend anybody doing this,
Lex Fridman (1:34:20.820)
especially not in the US where it's illegal.
Lex Fridman (1:34:23.260)
And there are no studies on this for that reason.
Lex Fridman (1:34:26.340)
But it seems that anecdotally
Joscha Bach (1:34:28.900)
that it basically can reset the serotonergic system.
Lex Fridman (1:34:33.420)
So it's basically pushing them
Joscha Bach (1:34:36.380)
outside of their normal boundaries.
Lex Fridman (1:34:38.220)
And as a result, it needs to find a new equilibrium.
Lex Fridman (1:34:41.020)
And in some people that equilibrium is better,
Lex Fridman (1:34:43.300)
but it also follows that in other people it might be worse.
Lex Fridman (1:34:46.260)
So if you have a brain that is already teetering
Lex Fridman (1:34:49.500)
on the boundary to psychosis,
Joscha Bach (1:34:51.980)
it can be permanently pushed over that boundary.
Lex Fridman (1:34:54.740)
Well, that's why you have to do good science,
Joscha Bach (1:34:56.540)
which they're starting to do on all these different
Lex Fridman (1:34:58.340)
substances of how well it actually works
Joscha Bach (1:34:59.940)
for the different conditions like MDMA seems to help
Lex Fridman (1:35:02.660)
with PTSD, same with psilocybin.
Joscha Bach (1:35:05.580)
You need to do good science,
Lex Fridman (1:35:08.340)
meaning large studies of large N.
Joscha Bach (1:35:10.860)
Yeah, so based on the existing studies of MDMA,
Lex Fridman (1:35:14.060)
it seems that if you look at Rick Doblin's work
Lex Fridman (1:35:17.460)
and what he has published about this and talks about,
Lex Fridman (1:35:20.780)
MDMA seems to be a psychologically relatively safe drug.
Lex Fridman (1:35:24.180)
But it's physiologically not very safe.
Lex Fridman (1:35:26.740)
That is, there is neurotoxicity
Joscha Bach (1:35:30.060)
if you would use a too large dose.
Lex Fridman (1:35:31.820)
And if you combine this with alcohol,
Joscha Bach (1:35:34.420)
which a lot of kids do in party settings during raves
Lex Fridman (1:35:37.540)
and so on, it's very hepatotoxic.
Lex Fridman (1:35:40.260)
So basically you can kill your liver.
Lex Fridman (1:35:42.220)
And this means that it's probably something that is best
Lex Fridman (1:35:45.340)
and most productively used in a clinical setting
Lex Fridman (1:35:48.340)
by people who really know what they're doing.
Lex Fridman (1:35:50.020)
And I suspect that's also true for the other psychedelics
Lex Fridman (1:35:53.580)
that is while the other psychedelics are probably not
Joscha Bach (1:35:56.620)
as toxic as say alcohol,
Lex Fridman (1:35:59.460)
the effects on the psyche can be much more profound
Lex Fridman (1:36:02.220)
and lasting.
Lex Fridman (1:36:03.460)
Yeah, well, as far as I know psilocybin,
Lex Fridman (1:36:05.940)
so mushrooms, magic mushrooms,
Lex Fridman (1:36:09.020)
as far as I know in terms of the studies they're running,
Joscha Bach (1:36:11.780)
I think have no, like they're allowed to do
Lex Fridman (1:36:15.060)
what they're calling heroic doses.
Lex Fridman (1:36:17.060)
So that one does not have a toxicity.
Lex Fridman (1:36:18.980)
So they could do like huge doses in a clinical setting
Joscha Bach (1:36:21.700)
when they're doing study on psilocybin,
Lex Fridman (1:36:23.620)
which is kind of fun.
Joscha Bach (1:36:25.140)
Yeah, it seems that most of the psychedelics
Lex Fridman (1:36:27.100)
work in extremely small doses,
Joscha Bach (1:36:29.300)
which means that the effect on the rest of the body
Lex Fridman (1:36:32.180)
is relatively low.
Lex Fridman (1:36:33.660)
And MDMA is probably the exception.
Lex Fridman (1:36:36.180)
Maybe ketamine can be dangerous in larger doses
Joscha Bach (1:36:38.300)
because it can depress breathing and so on.
Lex Fridman (1:36:41.260)
But the LSD and psilocybin work in very, very small doses,
Joscha Bach (1:36:45.940)
at least the active part of them,
Lex Fridman (1:36:47.820)
of psilocybin LSD is only the active part.
Lex Fridman (1:36:50.580)
And the, but the effect that it can have
Lex Fridman (1:36:54.100)
on your mental wiring can be very dangerous, I think.
Joscha Bach (1:36:57.940)
Let's talk about AI a little bit.
Lex Fridman (1:37:00.540)
What are your thoughts about GPT3 and language models
Lex Fridman (1:37:05.300)
trained with self supervised learning?
Lex Fridman (1:37:09.980)
It came out quite a bit ago,
Lex Fridman (1:37:11.420)
but I wanted to get your thoughts on it.
Lex Fridman (1:37:13.180)
Yeah.
Joscha Bach (1:37:14.580)
In the nineties, I was in New Zealand
Lex Fridman (1:37:16.900)
and I had an amazing professor, Ian Witten,
Joscha Bach (1:37:21.140)
who realized I was bored in class and put me in his lab.
Lex Fridman (1:37:25.220)
And he gave me the task to discover grammatical structure
Joscha Bach (1:37:28.780)
in an unknown language.
Lex Fridman (1:37:31.420)
And the unknown language that I picked was English
Joscha Bach (1:37:33.740)
because it was the easiest one
Lex Fridman (1:37:35.100)
to find a corpus for construct one.
Lex Fridman (1:37:37.980)
And he gave me the largest computer at the whole university.
Lex Fridman (1:37:41.980)
It had two gigabytes of RAM, which was amazing.
Lex Fridman (1:37:44.140)
And I wrote everything in C
Lex Fridman (1:37:45.380)
with some in memory compression to do statistics
Joscha Bach (1:37:47.740)
over the language.
Lex Fridman (1:37:49.340)
And I first would create a dictionary of all the words,
Joscha Bach (1:37:53.900)
which basically tokenizes everything and compresses things
Lex Fridman (1:37:57.300)
so that I don't need to store the whole word,
Lex Fridman (1:37:58.820)
but just a code for every word.
Lex Fridman (1:38:02.300)
And then I was taking this all apart in sentences
Lex Fridman (1:38:05.900)
and I was trying to find all the relationships
Lex Fridman (1:38:09.140)
between all the words in the sentences
Lex Fridman (1:38:10.860)
and do statistics over them.
Lex Fridman (1:38:12.940)
And that proved to be impossible
Joscha Bach (1:38:15.180)
because the complexity is just too large.
Lex Fridman (1:38:18.020)
So if you want to discover the relationship
Joscha Bach (1:38:20.460)
between an article and a noun,
Lex Fridman (1:38:21.860)
and there are three adjectives in between,
Joscha Bach (1:38:23.860)
you cannot do ngram statistics
Lex Fridman (1:38:25.420)
and look at all the possibilities that can exist,
Joscha Bach (1:38:28.060)
at least not with the resources that we had back then.
Lex Fridman (1:38:30.740)
So I realized I need to make some statistics
Joscha Bach (1:38:33.300)
over what I need to make statistics over.
Lex Fridman (1:38:35.220)
So I wrote something that was pretty much a hack
Joscha Bach (1:38:38.620)
that did this for at least first order relationships.
Lex Fridman (1:38:42.380)
And I came up with some kind of mutual information graph
Joscha Bach (1:38:45.100)
that was indeed discovering something that looks exactly
Lex Fridman (1:38:48.380)
like the grammatical structure of the sentence,
Joscha Bach (1:38:50.500)
just by trying to encode the sentence
Lex Fridman (1:38:52.660)
in such a way that the words would be written
Joscha Bach (1:38:54.740)
in the optimal order inside of the model.
Lex Fridman (1:38:58.100)
And what I also found is that if we would be able
Joscha Bach (1:39:02.140)
to increase the resolution of that
Lex Fridman (1:39:03.820)
and not just use this model
Joscha Bach (1:39:06.620)
to reproduce grammatically correct sentences,
Lex Fridman (1:39:09.060)
we would also be able
Joscha Bach (1:39:09.980)
to correct stylistically correct sentences
Lex Fridman (1:39:12.020)
by just having more bits in these relationships.
Lex Fridman (1:39:14.580)
And if we wanted to have meaning,
Lex Fridman (1:39:16.300)
we would have to go much higher order.
Lex Fridman (1:39:18.740)
And I didn't know how to make higher order models back then
Lex Fridman (1:39:21.460)
without spending way more years in research
Joscha Bach (1:39:23.860)
on how to make the statistics
Lex Fridman (1:39:25.580)
over what we need to make statistics over.
Lex Fridman (1:39:28.660)
And this thing that we cannot look at the relationships
Lex Fridman (1:39:31.540)
between all the bits in your input is being solved
Joscha Bach (1:39:34.020)
in different domains in different ways.
Lex Fridman (1:39:35.780)
So in computer graphics, computer vision,
Joscha Bach (1:39:39.380)
standard methods for many years now
Lex Fridman (1:39:41.380)
is convolutional neural networks.
Joscha Bach (1:39:43.620)
Convolutional neural networks are hierarchies of filters
Lex Fridman (1:39:46.620)
that exploit the fact that neighboring pixels
Joscha Bach (1:39:48.980)
in images are usually semantically related
Lex Fridman (1:39:51.100)
and distance pixels in images
Joscha Bach (1:39:53.020)
are usually not semantically related.
Lex Fridman (1:39:55.500)
So you can just by grouping the pixels
Joscha Bach (1:39:57.700)
that are next to each other,
Lex Fridman (1:39:59.100)
hierarchically together reconstruct the shape of objects.
Lex Fridman (1:40:02.780)
And this is an important prior
Lex Fridman (1:40:04.620)
that we built into these models
Lex Fridman (1:40:06.140)
so they can converge quickly.
Lex Fridman (1:40:08.460)
But this doesn't work in language
Joscha Bach (1:40:09.820)
for the reason that adjacent words are often
Lex Fridman (1:40:12.940)
but not always related and distant words
Joscha Bach (1:40:14.980)
are sometimes related while the words in between are not.
Lex Fridman (1:40:19.380)
So how can you learn the topology of language?
Lex Fridman (1:40:22.660)
And I think for this reason that this difficulty existed,
Lex Fridman (1:40:26.460)
the transformer was invented
Joscha Bach (1:40:28.700)
in natural language processing, not in vision.
Lex Fridman (1:40:32.780)
And what the transformer is doing,
Joscha Bach (1:40:34.900)
it's a hierarchy of layers where every layer learns
Lex Fridman (1:40:38.420)
what to pay attention to in the given context
Joscha Bach (1:40:40.900)
in the previous layer.
Lex Fridman (1:40:43.220)
So what to make the statistics over.
Lex Fridman (1:40:46.020)
And the context is significantly larger
Lex Fridman (1:40:49.980)
than the adjacent word.
Joscha Bach (1:40:51.540)
Yes, so the context that GPT3 has been using,
Lex Fridman (1:40:55.980)
the transformer itself is from 2017
Lex Fridman (1:40:58.500)
and it wasn't using that large of a context.
Lex Fridman (1:41:02.180)
OpenAI has basically scaled up this idea
Joscha Bach (1:41:05.060)
as far as they could at the time.
Lex Fridman (1:41:06.940)
And the context is about 2048 symbols,
Joscha Bach (1:41:11.300)
tokens in the language.
Lex Fridman (1:41:12.860)
These symbols are not characters,
Lex Fridman (1:41:15.060)
but they take the words and project them
Lex Fridman (1:41:17.020)
into a vector space where words
Joscha Bach (1:41:20.100)
that are statistically co occurring a lot
Lex Fridman (1:41:22.020)
are neighbors already.
Lex Fridman (1:41:23.220)
So it's already a simplification
Lex Fridman (1:41:24.780)
of the problem a little bit.
Lex Fridman (1:41:26.580)
And so every word is basically a set of coordinates
Lex Fridman (1:41:29.260)
in a high dimensional space.
Lex Fridman (1:41:31.060)
And then they use some kind of trick
Lex Fridman (1:41:33.100)
to also encode the order of the words in a sentence
Joscha Bach (1:41:36.340)
or in the not just sentence,
Lex Fridman (1:41:37.820)
but 2048 tokens is about a couple of pages of text
Joscha Bach (1:41:41.780)
or two and a half pages of text.
Lex Fridman (1:41:43.620)
And so they managed to do pretty exhaustive statistics
Joscha Bach (1:41:46.860)
over the potential relationships
Lex Fridman (1:41:49.140)
between two pages of text, which is tremendous.
Joscha Bach (1:41:51.740)
I was just using a single sentence back then.
Lex Fridman (1:41:55.020)
And I was only looking for first order relationships.
Lex Fridman (1:41:58.780)
And they were really looking
Lex Fridman (1:42:00.380)
for much, much higher level relationships.
Lex Fridman (1:42:02.740)
And what they discover after they fed this
Lex Fridman (1:42:05.300)
with an enormous amount of training,
Joscha Bach (1:42:06.580)
they are pretty much the written internet
Lex Fridman (1:42:08.980)
or a subset of it that had some quality,
Lex Fridman (1:42:12.140)
but substantial portion of the common core
Lex Fridman (1:42:15.180)
that they're not only able to reproduce style,
Lex Fridman (1:42:18.180)
but they're also able to reproduce
Lex Fridman (1:42:19.860)
some pretty detailed semantics,
Joscha Bach (1:42:21.660)
like being able to add three digit numbers
Lex Fridman (1:42:24.700)
and multiply two digit numbers
Joscha Bach (1:42:26.220)
or to translate between programming languages
Lex Fridman (1:42:28.820)
and things like that.
Lex Fridman (1:42:30.220)
So the results that GPT3 got, I think were amazing.
Lex Fridman (1:42:34.060)
By the way, I actually didn't check carefully.
Joscha Bach (1:42:38.620)
It's funny you just mentioned
Lex Fridman (1:42:40.540)
how you coupled semantics to the multiplication.
Lex Fridman (1:42:42.940)
Is it able to do some basic math on two digit numbers?
Lex Fridman (1:42:46.700)
Yes.
Joscha Bach (1:42:47.940)
Okay, interesting.
Lex Fridman (1:42:48.820)
I thought there's a lot of failure cases.
Joscha Bach (1:42:53.100)
Yeah, it basically fails if you take larger digit numbers.
Lex Fridman (1:42:56.140)
So four digit numbers and so on makes carrying mistakes
Lex Fridman (1:42:59.780)
and so on.
Lex Fridman (1:43:00.620)
And if you take larger numbers,
Joscha Bach (1:43:02.580)
you don't get useful results at all.
Lex Fridman (1:43:04.980)
And this could be an issue of the training set
Joscha Bach (1:43:09.260)
where there are not many examples
Lex Fridman (1:43:10.940)
of successful long form addition
Lex Fridman (1:43:13.300)
and standard human written text.
Lex Fridman (1:43:15.300)
And humans aren't very good
Joscha Bach (1:43:16.780)
at doing three digit numbers either.
Lex Fridman (1:43:19.460)
Yeah, you're not writing a lot about it.
Lex Fridman (1:43:22.340)
And the other thing is that the loss function
Lex Fridman (1:43:24.740)
that is being used is only minimizing surprise.
Lex Fridman (1:43:27.020)
So it's predicting what comes next in the typical text.
Lex Fridman (1:43:29.580)
It's not trying to go for causal closure first
Joscha Bach (1:43:32.260)
as we do.
Lex Fridman (1:43:33.100)
Yeah.
Lex Fridman (1:43:35.100)
But the fact that that kind of prediction works
Lex Fridman (1:43:39.620)
to generate text that's semantically rich
Lex Fridman (1:43:42.740)
and consistent is interesting.
Lex Fridman (1:43:45.020)
Yeah.
Lex Fridman (1:43:45.860)
So yeah, so it's amazing that it's able
Lex Fridman (1:43:47.220)
to generate semantically consistent text.
Joscha Bach (1:43:50.780)
It's not consistent.
Lex Fridman (1:43:51.940)
So the problem is that it loses coherence at some point,
Lex Fridman (1:43:54.660)
but it's also, I think, not correct to say
Lex Fridman (1:43:57.140)
that GPT3 is unable to deal with semantics at all
Joscha Bach (1:44:01.340)
because you ask it to perform certain transformations
Lex Fridman (1:44:04.100)
in text and it performs these transformation in text.
Lex Fridman (1:44:07.220)
And the kind of additions that it's able
Lex Fridman (1:44:09.220)
to perform are transformations in text, right?
Lex Fridman (1:44:12.540)
And there are proper semantics involved.
Lex Fridman (1:44:15.340)
You can also do more.
Joscha Bach (1:44:16.420)
There was a paper that was generating lots
Lex Fridman (1:44:19.900)
and lots of mathematically correct text
Lex Fridman (1:44:24.180)
and was feeding this into a transformer.
Lex Fridman (1:44:26.340)
And as a result, it was able to learn
Lex Fridman (1:44:29.340)
how to do differentiation integration in race
Lex Fridman (1:44:32.460)
that according to the authors, Mathematica could not.
Joscha Bach (1:44:37.340)
To which some of the people in Mathematica responded
Lex Fridman (1:44:39.860)
that they were not using Mathematica in the right way
Lex Fridman (1:44:42.700)
and so on.
Lex Fridman (1:44:43.540)
I have not really followed the resolution of this conflict.
Joscha Bach (1:44:46.380)
This part, as a small tangent,
Lex Fridman (1:44:48.700)
I really don't like in machine learning papers,
Joscha Bach (1:44:51.500)
which they often do anecdotal evidence.
Lex Fridman (1:44:56.620)
They'll find like one example
Joscha Bach (1:44:58.300)
in some kind of specific use of Mathematica
Lex Fridman (1:45:00.460)
and demonstrate, look, here's,
Joscha Bach (1:45:01.940)
they'll show successes and failures,
Lex Fridman (1:45:04.140)
but they won't have a very clear representation
Joscha Bach (1:45:07.660)
of how many cases this actually represents.
Lex Fridman (1:45:09.380)
Yes, but I think as a first paper,
Joscha Bach (1:45:11.260)
this is a pretty good start.
Lex Fridman (1:45:12.660)
And so the take home message, I think,
Joscha Bach (1:45:15.460)
is that the authors could get better results
Lex Fridman (1:45:19.420)
from this in their experiments
Joscha Bach (1:45:21.540)
than they could get from the vein,
Lex Fridman (1:45:23.460)
which they were using computer algebra systems,
Joscha Bach (1:45:25.940)
which means that was not nothing.
Lex Fridman (1:45:29.100)
And it's able to perform substantially better
Joscha Bach (1:45:32.340)
than GPT's V can based on a much larger amount
Lex Fridman (1:45:35.660)
of training data using the same underlying algorithm.
Joscha Bach (1:45:38.980)
Well, let me ask, again,
Lex Fridman (1:45:41.300)
so I'm using your tweets as if this is like Plato, right?
Joscha Bach (1:45:47.060)
As if this is well thought out novels that you've written.
Lex Fridman (1:45:51.780)
You tweeted, GPT4 is listening to us now.
Joscha Bach (1:45:58.660)
This is one way of asking,
Lex Fridman (1:46:00.300)
what are the limitations of GPT3 when it scales?
Lex Fridman (1:46:04.220)
So what do you think will be the capabilities
Lex Fridman (1:46:06.500)
of GPT4, GPT5, and so on?
Lex Fridman (1:46:10.260)
What are the limits of this approach?
Lex Fridman (1:46:11.780)
So obviously when we are writing things right now,
Joscha Bach (1:46:15.100)
everything that we are writing now
Lex Fridman (1:46:16.460)
is going to be training data
Joscha Bach (1:46:18.020)
for the next generation of machine learning models.
Lex Fridman (1:46:20.100)
So yes, of course, GPT4 is listening to us.
Lex Fridman (1:46:23.100)
And I think the tweet is already a little bit older
Lex Fridman (1:46:25.620)
and we now have Voodao
Lex Fridman (1:46:27.500)
and we have a number of other systems
Lex Fridman (1:46:30.140)
that basically are placeholders for GPT4.
Joscha Bach (1:46:33.620)
Don't know what open AIS plans are in this regard.
Lex Fridman (1:46:35.980)
I read that tweet in several ways.
Lex Fridman (1:46:39.060)
So one is obviously everything you put on the internet
Lex Fridman (1:46:42.700)
is used as training data.
Lex Fridman (1:46:44.660)
But in a second way I read it is in a,
Lex Fridman (1:46:49.580)
we talked about agency.
Joscha Bach (1:46:51.620)
I read it as almost like GPT4 is intelligent enough
Lex Fridman (1:46:55.460)
to be choosing to listen.
Lex Fridman (1:46:58.260)
So not only like did a programmer tell it
Lex Fridman (1:47:00.460)
to collect this data and use it for training,
Joscha Bach (1:47:03.700)
I almost saw the humorous angle,
Lex Fridman (1:47:06.220)
which is like it has achieved AGI kind of thing.
Lex Fridman (1:47:09.100)
Well, the thing is, could we be already be living in GPT5?
Lex Fridman (1:47:13.300)
So GPT4 is listening and GPT5 actually constructing
Joscha Bach (1:47:18.300)
the entirety of the reality where we...
Lex Fridman (1:47:20.420)
Of course, in some sense,
Lex Fridman (1:47:22.340)
what everybody is trying to do right now in AI
Lex Fridman (1:47:24.500)
is to extend the transformer to be able to deal with video.
Lex Fridman (1:47:28.980)
And there are very promising extensions, right?
Lex Fridman (1:47:31.860)
There's a work by Google that is called Perceiver
Lex Fridman (1:47:36.060)
and that is overcoming some of the limitations
Lex Fridman (1:47:39.300)
of the transformer by letting it learn the topology
Joscha Bach (1:47:41.980)
of the different modalities separately.
Lex Fridman (1:47:44.900)
And by training it to find better input features.
Lex Fridman (1:47:50.060)
So basically feature abstractions that are being used
Lex Fridman (1:47:52.540)
by this successor to GPT3 are chosen such a way
Joscha Bach (1:47:58.140)
that it's able to deal with video input.
Lex Fridman (1:48:00.780)
And there is more to be done.
Lex Fridman (1:48:02.220)
So one of the limitations of GPT3 is that it's amnesiac.
Lex Fridman (1:48:07.820)
So it forgets everything beyond the two pages
Joscha Bach (1:48:09.980)
that it currently reads also during generation,
Lex Fridman (1:48:12.340)
not just during learning.
Lex Fridman (1:48:14.420)
Do you think that's fixable
Lex Fridman (1:48:16.580)
within the space of deep learning?
Lex Fridman (1:48:18.700)
Can you just make a bigger, bigger, bigger input?
Lex Fridman (1:48:21.340)
No, I don't think that our own working memory
Joscha Bach (1:48:24.500)
is infinitely large.
Lex Fridman (1:48:25.620)
It's probably also just a few thousand bits.
Lex Fridman (1:48:28.020)
But what you can do is you can structure
Lex Fridman (1:48:31.060)
this working memory.
Lex Fridman (1:48:31.900)
So instead of just force feeding this thing,
Lex Fridman (1:48:34.980)
a certain thing that it has to focus on,
Lex Fridman (1:48:37.060)
and it's not allowed to focus on anything else
Lex Fridman (1:48:39.460)
as its network,
Joscha Bach (1:48:41.100)
you allow it to construct its own working memory as we do.
Lex Fridman (1:48:44.860)
When we are reading a book,
Joscha Bach (1:48:46.780)
it's not that we are focusing our attention
Lex Fridman (1:48:48.700)
in such a way that we can only remember the current page.
Joscha Bach (1:48:52.380)
We will also try to remember other pages
Lex Fridman (1:48:54.660)
and try to undo what we learned from them
Joscha Bach (1:48:56.860)
or modify what we learned from them.
Lex Fridman (1:48:58.660)
We might get up and take another book from the shelf.
Joscha Bach (1:49:01.020)
We might go out and ask somebody,
Lex Fridman (1:49:03.060)
we can edit our working memory in any way that is useful
Joscha Bach (1:49:06.860)
to put a context together that allows us
Lex Fridman (1:49:09.260)
to draw the right inferences and to learn the right things.
Lex Fridman (1:49:13.100)
So this ability to perform experiments on the world
Lex Fridman (1:49:16.380)
based on an attempt to become fully coherent
Lex Fridman (1:49:20.420)
and to achieve causal closure,
Lex Fridman (1:49:22.260)
to achieve a certain aesthetic of your modeling,
Joscha Bach (1:49:24.860)
that is something that eventually needs to be done.
Lex Fridman (1:49:28.300)
And at the moment we are skirting this in some sense
Joscha Bach (1:49:31.140)
by building systems that are larger and faster
Lex Fridman (1:49:33.420)
so they can use dramatically larger resources
Lex Fridman (1:49:36.100)
and human beings can do and much more training data
Lex Fridman (1:49:38.700)
to get to models that in some sense
Joscha Bach (1:49:40.420)
are already way superhuman
Lex Fridman (1:49:42.380)
and in other ways are laughingly incoherent.
Lex Fridman (1:49:45.500)
So do you think sort of making the systems like,
Lex Fridman (1:49:50.060)
what would you say, multi resolutional?
Lex Fridman (1:49:51.900)
So like some of the language models
Lex Fridman (1:49:56.980)
are focused on two pages,
Joscha Bach (1:49:59.820)
some are focused on two books,
Lex Fridman (1:50:03.380)
some are focused on two years of reading,
Joscha Bach (1:50:06.580)
some are focused on a lifetime,
Lex Fridman (1:50:08.100)
so it's like stacks of GPT3s all the way down.
Joscha Bach (1:50:11.900)
You want to have gaps in between them.
Lex Fridman (1:50:13.700)
So it's not necessarily two years, there's no gaps.
Joscha Bach (1:50:17.060)
It's things out of two years or out of 20 years
Lex Fridman (1:50:19.980)
or 2,000 years or 2 billion years
Joscha Bach (1:50:22.220)
where you are just selecting those bits
Lex Fridman (1:50:24.580)
that are predicted to be the most useful ones
Joscha Bach (1:50:27.540)
to understand what you're currently doing.
Lex Fridman (1:50:29.700)
And this prediction itself requires a very complicated model
Lex Fridman (1:50:32.780)
and that's the actual model that you need to be making.
Lex Fridman (1:50:34.780)
It's not just that you are trying to understand
Joscha Bach (1:50:36.940)
the relationships between things,
Lex Fridman (1:50:38.340)
but what you need to make relationships,
Joscha Bach (1:50:40.740)
discover relationships over.
Lex Fridman (1:50:42.540)
I wonder what that thing looks like,
Lex Fridman (1:50:45.500)
what the architecture for the thing
Lex Fridman (1:50:47.820)
that's able to have that kind of model.
Joscha Bach (1:50:50.100)
I think it needs more degrees of freedom
Lex Fridman (1:50:52.460)
than the current models have.
Lex Fridman (1:50:54.300)
So it starts out with the fact that you possibly
Lex Fridman (1:50:57.420)
don't just want to have a feed forward model,
Lex Fridman (1:50:59.700)
but you want it to be fully recurrent.
Lex Fridman (1:51:02.700)
And to make it fully recurrent,
Joscha Bach (1:51:04.540)
you probably need to loop it back into itself
Lex Fridman (1:51:06.660)
and allow it to skip connections.
Joscha Bach (1:51:08.220)
Once you do this,
Lex Fridman (1:51:09.980)
when you're predicting the next frame
Lex Fridman (1:51:12.020)
and your internal next frame in every moment,
Lex Fridman (1:51:15.180)
and you are able to skip connection,
Joscha Bach (1:51:17.500)
it means that signals can travel from the output
Lex Fridman (1:51:21.300)
of the network into the middle of the network
Joscha Bach (1:51:24.300)
faster than the inputs do.
Lex Fridman (1:51:25.900)
Do you think it can still be differentiable?
Lex Fridman (1:51:28.780)
Do you think it still can be a neural network?
Lex Fridman (1:51:30.660)
Sometimes it can and sometimes it cannot.
Lex Fridman (1:51:32.980)
So it can still be a neural network,
Lex Fridman (1:51:35.500)
but not a fully differentiable one.
Lex Fridman (1:51:37.180)
And when you want to deal with non differentiable ones,
Lex Fridman (1:51:40.900)
you need to have an attention system
Joscha Bach (1:51:42.780)
that is discreet and two dimensional
Lex Fridman (1:51:44.780)
and can perform grammatical operations.
Joscha Bach (1:51:46.660)
You need to be able to perform program synthesis.
Lex Fridman (1:51:49.300)
You need to be able to backtrack
Joscha Bach (1:51:51.220)
in this operations that you perform on this thing.
Lex Fridman (1:51:54.020)
And this thing needs a model of what it's currently doing.
Lex Fridman (1:51:56.460)
And I think this is exactly the purpose
Lex Fridman (1:51:58.420)
of our own consciousness.
Joscha Bach (1:52:01.220)
Yeah, the program things are tricky on neural networks.
Lex Fridman (1:52:05.380)
So let me ask you, it's not quite program synthesis,
Lex Fridman (1:52:09.020)
but the application of these language models
Lex Fridman (1:52:12.060)
to generation, to program synthesis,
Lex Fridman (1:52:15.060)
but generation of programs.
Lex Fridman (1:52:16.580)
So if you look at GitHub OpenPilot,
Joscha Bach (1:52:19.180)
which is based on OpenAI's codecs,
Lex Fridman (1:52:21.220)
I don't know if you got a chance to look at it,
Lex Fridman (1:52:22.780)
but it's the system that's able to generate code
Lex Fridman (1:52:26.180)
once you prompt it with, what is it?
Joscha Bach (1:52:30.100)
Like the header of a function with some comments.
Lex Fridman (1:52:32.700)
And it seems to do an incredibly good job
Joscha Bach (1:52:36.060)
or not a perfect job, which is very important,
Lex Fridman (1:52:39.220)
but an incredibly good job of generating functions.
Lex Fridman (1:52:42.780)
What do you make of that?
Lex Fridman (1:52:44.220)
Are you, is this exciting
Lex Fridman (1:52:45.500)
or is this just a party trick, a demo?
Lex Fridman (1:52:48.980)
Or is this revolutionary?
Joscha Bach (1:52:51.860)
I haven't worked with it yet.
Lex Fridman (1:52:52.980)
So it's difficult for me to judge it,
Lex Fridman (1:52:55.140)
but I would not be surprised
Lex Fridman (1:52:57.100)
if it turns out to be a revolutionary.
Lex Fridman (1:52:59.540)
And that's because the majority of programming tasks
Lex Fridman (1:53:01.740)
that are being done in the industry right now
Joscha Bach (1:53:04.260)
are not creative.
Lex Fridman (1:53:05.940)
People are writing code that other people have written,
Joscha Bach (1:53:08.500)
or they're putting things together from code fragments
Lex Fridman (1:53:10.620)
that others have had.
Lex Fridman (1:53:11.540)
And a lot of the work that programmers do in practice
Lex Fridman (1:53:14.300)
is to figure out how to overcome the gaps
Joscha Bach (1:53:17.780)
in their current knowledge
Lex Fridman (1:53:18.900)
and the things that people have already done.
Lex Fridman (1:53:20.940)
How to copy and paste from Stack Overflow, that's right.
Lex Fridman (1:53:24.060)
And so of course we can automate that.
Joscha Bach (1:53:26.380)
Yeah, to make it much faster to copy and paste
Lex Fridman (1:53:29.900)
from Stack Overflow.
Joscha Bach (1:53:30.860)
Yes, but it's not just copying and pasting.
Lex Fridman (1:53:32.820)
It's also basically learning which parts you need to modify
Joscha Bach (1:53:36.580)
to make them fit together.
Lex Fridman (1:53:38.140)
Yeah, like literally sometimes as simple
Joscha Bach (1:53:41.180)
as just changing the variable names.
Lex Fridman (1:53:43.380)
So it fits into the rest of your code.
Joscha Bach (1:53:45.060)
Yes, but this requires that you understand the semantics
Lex Fridman (1:53:47.460)
of what you're doing to some degree.
Lex Fridman (1:53:49.660)
And you can automate some of those things.
Lex Fridman (1:53:51.580)
The thing that makes people nervous of course
Joscha Bach (1:53:53.500)
is that a little bit wrong in a program
Lex Fridman (1:53:57.780)
can have a dramatic effect on the actual final operation
Joscha Bach (1:54:02.700)
of that program.
Lex Fridman (1:54:03.540)
So that's one little error,
Joscha Bach (1:54:05.380)
which in the space of language doesn't really matter,
Lex Fridman (1:54:08.780)
but in the space of programs can matter a lot.
Joscha Bach (1:54:11.980)
Yes, but this is already what is happening
Lex Fridman (1:54:14.100)
when humans program code.
Joscha Bach (1:54:15.820)
Yeah, this is.
Lex Fridman (1:54:16.820)
So we have a technology to deal with this.
Joscha Bach (1:54:20.220)
Somehow it becomes scarier when you know
Lex Fridman (1:54:23.500)
that a program generated code
Joscha Bach (1:54:25.100)
that's running a nuclear power plant.
Lex Fridman (1:54:27.940)
It becomes scarier.
Joscha Bach (1:54:29.100)
You know, humans have errors too.
Lex Fridman (1:54:31.380)
Exactly.
Lex Fridman (1:54:32.220)
But it's scarier when a program is doing it
Lex Fridman (1:54:35.100)
because why, why?
Joscha Bach (1:54:38.700)
I mean, there's a fear that a program,
Lex Fridman (1:54:43.660)
like a program may not be as good as humans
Joscha Bach (1:54:48.020)
to know when stuff is important to not mess up.
Lex Fridman (1:54:51.340)
Like there's a misalignment of priorities of values
Joscha Bach (1:55:00.260)
that's potential.
Lex Fridman (1:55:01.300)
Maybe that's the source of the worry.
Joscha Bach (1:55:03.500)
I mean, okay, if I give you code generated
Lex Fridman (1:55:06.380)
by GitHub open pilot and code generated by a human
Lex Fridman (1:55:12.500)
and say here, use one of these,
Lex Fridman (1:55:16.020)
which how do you select today and in the next 10 years
Lex Fridman (1:55:20.380)
which code do you use?
Lex Fridman (1:55:21.860)
Wouldn't you still be comfortable with the human?
Joscha Bach (1:55:25.380)
At the moment when you go to Stanford to get an MRI,
Lex Fridman (1:55:29.580)
they will write a bill to the insurance over $20,000.
Lex Fridman (1:55:34.580)
And of this, maybe half of that gets paid by the insurance
Lex Fridman (1:55:38.300)
and a quarter gets paid by you.
Lex Fridman (1:55:40.580)
And the MRI cost them $600 to make maybe probably less.
Lex Fridman (1:55:44.940)
And what are the values of the person
Lex Fridman (1:55:47.700)
that writes the software and deploys this process?
Lex Fridman (1:55:51.820)
It's very difficult for me to say whether I trust people.
Joscha Bach (1:55:56.260)
I think that what happens there is a mixture
Lex Fridman (1:55:58.700)
of proper Anglo Saxon Protestant values
Joscha Bach (1:56:01.980)
where somebody is trying to serve an abstract radar hole
Lex Fridman (1:56:04.860)
and organize crime.
Joscha Bach (1:56:06.300)
Well, that's a very harsh,
Lex Fridman (1:56:10.980)
I think that's a harsh view of humanity.
Joscha Bach (1:56:15.500)
There's a lot of bad people, whether incompetent
Lex Fridman (1:56:18.780)
or just malevolent in this world, yes.
Lex Fridman (1:56:21.700)
But it feels like the more malevolent,
Lex Fridman (1:56:25.820)
so the more damage you do to the world,
Joscha Bach (1:56:29.580)
the more resistance you have in your own human heart.
Lex Fridman (1:56:34.420)
Yeah, but don't explain with malevolence or stupidity
Lex Fridman (1:56:37.140)
what can be explained by just people
Lex Fridman (1:56:38.780)
acting on their incentives.
Joscha Bach (1:56:41.540)
Right, so what happens in Stanford
Lex Fridman (1:56:42.940)
is not that somebody is evil.
Joscha Bach (1:56:45.100)
It's just that they do what they're being paid for.
Lex Fridman (1:56:48.740)
No, it's not evil.
Joscha Bach (1:56:50.500)
That's, I tend to, no, I see that as malevolence.
Lex Fridman (1:56:53.740)
I see as I, even like being a good German,
Joscha Bach (1:56:58.780)
as I told you offline, is some,
Lex Fridman (1:57:01.540)
it's not absolute malevolence,
Lex Fridman (1:57:05.060)
but it's a small amount, it's cowardice.
Lex Fridman (1:57:07.460)
I mean, when you see there's something wrong with the world,
Joscha Bach (1:57:10.580)
it's either incompetence and you're not able to see it,
Lex Fridman (1:57:15.100)
or it's cowardice that you're not able to stand up,
Joscha Bach (1:57:17.780)
not necessarily in a big way, but in a small way.
Lex Fridman (1:57:21.620)
So I do think that is a bit of malevolence.
Joscha Bach (1:57:25.940)
I'm not sure the example you're describing
Lex Fridman (1:57:27.660)
is a good example of that.
Lex Fridman (1:57:28.500)
So the question is, what is it that you are aiming for?
Lex Fridman (1:57:31.220)
And if you don't believe in the future,
Joscha Bach (1:57:34.900)
if you, for instance, think that the dollar is going to crash,
Lex Fridman (1:57:37.460)
why would you try to save dollars?
Joscha Bach (1:57:39.540)
If you don't think that humanity will be around
Lex Fridman (1:57:42.580)
in a hundred years from now,
Joscha Bach (1:57:43.740)
because global warming will wipe out civilization,
Lex Fridman (1:57:47.500)
why would you need to act as if it were?
Joscha Bach (1:57:50.220)
Right, so the question is,
Lex Fridman (1:57:51.660)
is there an overarching aesthetics
Joscha Bach (1:57:53.980)
that is projecting you and the world into the future,
Lex Fridman (1:57:56.900)
which I think is the basic idea of religion,
Joscha Bach (1:57:59.020)
that you understand the interactions
Lex Fridman (1:58:01.220)
that we have with each other
Joscha Bach (1:58:02.340)
as some kind of civilization level agent
Lex Fridman (1:58:04.780)
that is projecting itself into the future.
Joscha Bach (1:58:07.180)
If you don't have that shared purpose,
Lex Fridman (1:58:10.420)
what is there to be ethical for?
Lex Fridman (1:58:12.940)
So I think when we talk about ethics and AI,
Lex Fridman (1:58:16.380)
we need to go beyond the insane bias discussions and so on,
Joscha Bach (1:58:20.020)
where people are just measuring the distance
Lex Fridman (1:58:22.060)
between a statistic to their preferred current world model.
Joscha Bach (1:58:27.940)
The optimism, wait, wait, wait,
Lex Fridman (1:58:29.340)
I was a little confused by the previous thing,
Joscha Bach (1:58:31.180)
just to clarify.
Lex Fridman (1:58:32.260)
There is a kind of underlying morality
Joscha Bach (1:58:39.820)
to having an optimism that human civilization
Lex Fridman (1:58:43.620)
will persist for longer than a hundred years.
Joscha Bach (1:58:45.580)
Like I think a lot of people believe
Lex Fridman (1:58:50.060)
that it's a good thing for us to keep living.
Joscha Bach (1:58:53.220)
Yeah, of course.
Lex Fridman (1:58:54.060)
And thriving.
Joscha Bach (1:58:54.900)
This morality itself is not an end to itself.
Lex Fridman (1:58:56.900)
It's instrumental to people living in a hundred years
Joscha Bach (1:58:59.660)
from now or 500 years from now.
Lex Fridman (1:59:03.100)
So it's only justifiable if you actually think
Joscha Bach (1:59:06.540)
that it will lead to people or increase the probability
Lex Fridman (1:59:09.780)
of people being around in that timeframe.
Lex Fridman (1:59:12.500)
And a lot of people don't actually believe that,
Lex Fridman (1:59:14.980)
at least not actively.
Lex Fridman (1:59:16.860)
But believe what exactly?
Lex Fridman (1:59:17.980)
So I was...
Joscha Bach (1:59:19.180)
Most people don't believe
Lex Fridman (1:59:20.620)
that they can afford to act on such a model.
Joscha Bach (1:59:23.500)
Basically what happens in the US
Lex Fridman (1:59:25.340)
is I think that the healthcare system
Joscha Bach (1:59:26.940)
is for a lot of people no longer sustainable,
Lex Fridman (1:59:28.940)
which means that if they need the help
Joscha Bach (1:59:30.580)
of the healthcare system,
Lex Fridman (1:59:31.540)
they're often not able to afford it.
Lex Fridman (1:59:33.460)
And when they cannot help it,
Lex Fridman (1:59:35.060)
they are often going bankrupt.
Joscha Bach (1:59:37.300)
I think the leading cause of personal bankruptcy
Lex Fridman (1:59:40.220)
in the US is the healthcare system.
Lex Fridman (1:59:42.740)
And that would not be necessary.
Lex Fridman (1:59:44.820)
It's not because people are consuming
Joscha Bach (1:59:46.660)
more and more medical services
Lex Fridman (1:59:48.780)
and are achieving a much, much longer life as a result.
Joscha Bach (1:59:51.460)
That's not actually the story that is happening
Lex Fridman (1:59:53.620)
because you can compare it to other countries.
Lex Fridman (1:59:55.300)
And life expectancy in the US is currently not increasing
Lex Fridman (1:59:58.420)
and it's not as high as in all the other
Lex Fridman (20:00.300)
And there's some prodding that the monkey gets to do.
Lex Fridman (20:03.180)
And sometimes the elephant listens.
Joscha Bach (20:06.180)
I heard you got into some contentious,
Lex Fridman (20:08.940)
maybe you can correct me,
Lex Fridman (20:09.820)
but I heard you got into some contentious
Lex Fridman (20:11.540)
free will discussions.
Lex Fridman (20:13.900)
Is this with Sam Harris or something like that?
Lex Fridman (20:16.100)
Not that I know of.
Joscha Bach (20:18.700)
Some people on Clubhouse told me
Lex Fridman (20:20.460)
you made a bunch of big debate points about free will.
Joscha Bach (20:25.940)
Well, let me just then ask you where,
Lex Fridman (20:28.860)
in terms of the monkey and the elephant,
Lex Fridman (20:31.700)
do you think we land in terms of the illusion of free will?
Lex Fridman (20:35.300)
How much control does the monkey have?
Joscha Bach (20:38.580)
We have to think about what the free will is
Lex Fridman (20:41.460)
in the first place.
Joscha Bach (20:43.260)
We are not the machine.
Lex Fridman (20:44.420)
We are not the thing that is making the decisions.
Joscha Bach (20:46.820)
We are a model of that decision making process.
Lex Fridman (20:49.900)
And there is a difference between making your own decisions
Lex Fridman (20:54.220)
and predicting your own decisions.
Lex Fridman (20:56.180)
And that difference is the first person perspective.
Lex Fridman (20:59.860)
And what basically makes decision making
Lex Fridman (21:04.820)
and the conditions of free will distinct
Joscha Bach (21:06.620)
from just automatically doing the best thing is
Lex Fridman (21:10.340)
that we often don't know what the best thing is.
Joscha Bach (21:13.260)
We make decisions under uncertainty.
Lex Fridman (21:15.540)
We make informed bets using a betting algorithm
Joscha Bach (21:17.900)
that we don't yet understand
Lex Fridman (21:19.140)
because we haven't reverse engineered
Joscha Bach (21:20.900)
our own minds sufficiently.
Lex Fridman (21:22.340)
We don't know the expected rewards.
Joscha Bach (21:23.900)
We don't know the mechanism
Lex Fridman (21:24.940)
by which we estimate the rewards and so on.
Lex Fridman (21:27.180)
But there is an algorithm.
Lex Fridman (21:28.300)
We observe ourselves performing
Joscha Bach (21:30.580)
where we see that we weight facts and factors
Lex Fridman (21:34.820)
and the future, and then some kind of possibility,
Joscha Bach (21:39.300)
some motive gets raised to an intention.
Lex Fridman (21:41.620)
And that's informed bet that the system is making.
Lex Fridman (21:44.500)
And that making of the informed bet,
Lex Fridman (21:46.420)
the representation of that is what we call free will.
Lex Fridman (21:49.460)
And it seems to be paradoxical
Lex Fridman (21:51.580)
because we think that the crucial thing is
Joscha Bach (21:53.700)
about it that it's somehow indeterministic.
Lex Fridman (21:56.500)
And yet if it was indeterministic, it would be random.
Lex Fridman (22:00.340)
And it cannot be random because if it was random,
Lex Fridman (22:03.380)
if just dice were being thrown in the universe,
Joscha Bach (22:05.300)
randomly forces you to do things, it would be meaningless.
Lex Fridman (22:08.180)
So the important part of the decisions
Joscha Bach (22:10.420)
is always the deterministic stuff.
Lex Fridman (22:12.700)
But it appears to be indeterministic to you
Joscha Bach (22:15.220)
because it's unpredictable.
Lex Fridman (22:16.820)
Because if it was predictable,
Joscha Bach (22:18.500)
you wouldn't experience it as a free will decision.
Lex Fridman (22:21.460)
You would experience it as just doing
Joscha Bach (22:23.260)
the necessary right thing.
Lex Fridman (22:25.580)
And you see this continuum between the free will
Lex Fridman (22:28.740)
and the execution of automatic behavior
Lex Fridman (22:31.740)
when you're observing other people.
Lex Fridman (22:33.220)
So for instance, when you are observing your own children,
Lex Fridman (22:36.220)
if you don't understand them,
Joscha Bach (22:37.580)
you will abuse this agent model
Lex Fridman (22:40.060)
where you have an agent with a set point generator.
Lex Fridman (22:43.260)
And the agent is doing the best it can
Lex Fridman (22:45.420)
to minimize the difference to the set point.
Lex Fridman (22:47.420)
And it might be confused and sometimes impulsive or whatever,
Lex Fridman (22:51.220)
but it's acting on its own free will.
Lex Fridman (22:53.340)
And when you understand what's happens
Lex Fridman (22:55.420)
in the mind of the child, you see that it's automatic.
Lex Fridman (22:58.540)
And you can outmodel the child,
Lex Fridman (23:00.300)
you can build things around the child
Joscha Bach (23:02.300)
that will lead the child to making exactly the decision
Lex Fridman (23:05.260)
that you are predicting.
Lex Fridman (23:06.740)
And under these circumstances,
Lex Fridman (23:08.740)
like when you are a stage musician
Joscha Bach (23:10.500)
or somebody who is dealing with people
Lex Fridman (23:13.420)
that you sell a car to,
Lex Fridman (23:15.260)
and you completely understand the psychology
Lex Fridman (23:17.300)
and the impulses and the space of thoughts
Joscha Bach (23:19.660)
that this individual can have at that moment.
Lex Fridman (23:21.580)
Under these circumstances,
Joscha Bach (23:22.620)
it makes no sense to attribute free will.
Lex Fridman (23:26.060)
Because it's no longer decision making under uncertainty.
Joscha Bach (23:28.220)
You are already certain.
Lex Fridman (23:29.220)
For them, there's uncertainty,
Lex Fridman (23:30.500)
but you already know what they're doing.
Lex Fridman (23:33.780)
But what about for you?
Lex Fridman (23:34.980)
So is this akin to like systems like cellular automata
Lex Fridman (23:40.500)
where it's deterministic,
Lex Fridman (23:43.300)
but when you squint your eyes a little bit,
Lex Fridman (23:46.940)
it starts to look like there's agents making decisions
Joscha Bach (23:50.780)
at the higher sort of when you zoom out
Lex Fridman (23:53.780)
and look at the entities
Joscha Bach (23:55.020)
that are composed by the individual cells.
Lex Fridman (23:58.460)
Even though there's underlying simple rules
Joscha Bach (24:02.060)
that make the system evolve in deterministic ways,
Lex Fridman (24:07.540)
it looks like there's organisms making decisions.
Joscha Bach (24:10.780)
Is that where the illusion of free will emerges,
Lex Fridman (24:14.500)
that jump in scale?
Joscha Bach (24:16.740)
It's a particular type of model,
Lex Fridman (24:18.500)
but this jump in scale is crucial.
Joscha Bach (24:20.700)
The jump in scale happens whenever
Lex Fridman (24:22.380)
you have too many parts to count
Lex Fridman (24:23.820)
and you cannot make a model at that level
Lex Fridman (24:25.780)
and you try to find some higher level regularity.
Lex Fridman (24:28.780)
And the higher level regularity is a pattern
Lex Fridman (24:30.900)
that you project into the world to make sense of it.
Lex Fridman (24:34.660)
And agency is one of these patterns, right?
Lex Fridman (24:36.460)
You have all these cells that interact with each other
Lex Fridman (24:39.700)
and the cells in our body are set up in such a way
Lex Fridman (24:42.220)
that they benefit if their behavior is coherent,
Joscha Bach (24:45.060)
which means that they act
Lex Fridman (24:46.580)
as if they were serving a common goal.
Lex Fridman (24:49.180)
And that means that they will evolve regulation mechanisms
Lex Fridman (24:52.340)
that act as if they were serving a common goal.
Lex Fridman (24:55.300)
And now you can make sense of all these cells
Lex Fridman (24:57.620)
by projecting the common goal into them.
Joscha Bach (24:59.900)
Right, so for you then, free will is an illusion.
Lex Fridman (25:03.340)
No, it's a model and it's a construct.
Joscha Bach (25:06.460)
It's basically a model that the system is making
Lex Fridman (25:08.580)
of its own behavior.
Lex Fridman (25:09.420)
And it's the best model that it can come up with
Lex Fridman (25:11.500)
under the circumstances.
Lex Fridman (25:12.740)
And it can get replaced by a different model,
Lex Fridman (25:14.740)
which is automatic behavior,
Joscha Bach (25:16.420)
when you fully understand the mechanism
Lex Fridman (25:17.980)
under which you are acting.
Joscha Bach (25:19.180)
Yeah, but another word for model is what, story.
Lex Fridman (25:23.860)
So it's the story you're telling.
Lex Fridman (25:25.300)
I mean, do you actually have control?
Lex Fridman (25:27.340)
Is there such a thing as a you
Lex Fridman (25:29.420)
and is there such a thing as you have in control?
Lex Fridman (25:33.980)
So like, are you manifesting your evolution as an entity?
Joscha Bach (25:42.020)
In some sense, the you is the model of the system
Lex Fridman (25:44.380)
that is in control.
Joscha Bach (25:45.660)
It's a story that the system tells itself
Lex Fridman (25:47.860)
about somebody who is in control.
Joscha Bach (25:50.340)
Yeah.
Lex Fridman (25:51.180)
And the contents of that model are being used
Joscha Bach (25:53.060)
to inform the behavior of the system.
Lex Fridman (25:56.940)
Okay.
Lex Fridman (25:57.780)
So the system is completely mechanical
Lex Fridman (26:00.500)
and the system creates that story like a loom.
Lex Fridman (26:03.300)
And then it uses the contents of that story
Lex Fridman (26:06.020)
to inform its actions
Lex Fridman (26:07.460)
and writes the results of that actions into the story.
Lex Fridman (26:11.220)
So how's that not an illusion?
Joscha Bach (26:13.380)
The story is written then,
Lex Fridman (26:16.220)
or rather we're not the writers of the story.
Joscha Bach (26:21.260)
Yes, but we always knew that.
Lex Fridman (26:24.060)
No, we don't know that.
Lex Fridman (26:25.300)
When did we know that?
Lex Fridman (26:26.740)
I think that's mostly a confusion about concepts.
Joscha Bach (26:29.260)
The conceptual illusion in our culture
Lex Fridman (26:31.980)
comes from the idea that we live in physical reality
Lex Fridman (26:35.700)
and that we experience physical reality
Lex Fridman (26:37.460)
and that you have ideas about it.
Lex Fridman (26:39.500)
And then you have this dualist interpretation
Lex Fridman (26:41.660)
where you have two substances, res extensa,
Joscha Bach (26:45.060)
the world that you can touch
Lex Fridman (26:46.940)
and that is made of extended things
Lex Fridman (26:48.980)
and res cogitans, which is the world of ideas.
Lex Fridman (26:51.620)
And in fact, both of them are mental representations.
Joscha Bach (26:54.580)
One is the representations of the world as a game engine
Lex Fridman (26:57.900)
that your mind generates to make sense of the perceptual data.
Lex Fridman (27:01.100)
And the other one,
Lex Fridman (27:02.260)
yes, that's what we perceive as the physical world.
Lex Fridman (27:04.460)
But we already know that the physical world
Lex Fridman (27:05.940)
is nothing like that, right?
Joscha Bach (27:07.020)
Quantum mechanics is very different
Lex Fridman (27:08.860)
from what you and me perceive as the world.
Joscha Bach (27:11.340)
The world that you and me perceive as a game engine.
Lex Fridman (27:14.820)
And there are no colors and sounds in the physical world.
Joscha Bach (27:17.180)
They only exist in the game engine generated by your brain.
Lex Fridman (27:20.100)
And then you have ideas
Lex Fridman (27:21.500)
that cannot be mapped onto extended regions, right?
Lex Fridman (27:24.740)
So the objects that have a spatial extension
Joscha Bach (27:26.940)
in the game engine, res extensa,
Lex Fridman (27:29.500)
and the objects that don't have a physical extension
Joscha Bach (27:31.460)
in the game engine are ideas.
Lex Fridman (27:34.540)
And they both interact in our mind
Joscha Bach (27:36.140)
to produce models of the world.
Lex Fridman (27:38.220)
Yep, but, you know, when you play video games,
Joscha Bach (27:42.780)
I understand that what's actually happening
Lex Fridman (27:45.020)
is zeros and ones inside of a computer,
Joscha Bach (27:50.020)
inside of a CPU and a GPU,
Lex Fridman (27:52.820)
but you're still seeing like the rendering of that.
Lex Fridman (27:58.140)
And you're still making decisions,
Lex Fridman (28:00.700)
whether to shoot, to turn left or to turn right,
Joscha Bach (28:03.740)
if you're playing a shooter,
Lex Fridman (28:04.660)
or every time I started thinking about Skyrim
Lex Fridman (28:07.100)
and Elder Scrolls and walking around in beautiful nature
Lex Fridman (28:09.860)
and swinging a sword.
Lex Fridman (28:10.900)
But it feels like you're making decisions
Lex Fridman (28:13.100)
inside that video game.
Lex Fridman (28:15.060)
So even though you don't have direct access
Lex Fridman (28:17.220)
in terms of perception to the bits,
Joscha Bach (28:21.220)
to the zeros and ones,
Lex Fridman (28:22.660)
it still feels like you're making decisions
Lex Fridman (28:24.860)
and your decisions actually feels
Lex Fridman (28:27.900)
like they're being applied all the way down
Joscha Bach (28:30.740)
to the zeros and ones.
Lex Fridman (28:32.300)
So it feels like you have control,
Joscha Bach (28:33.460)
even though you don't have direct access to reality.
Lex Fridman (28:36.540)
So there is basically a special character
Joscha Bach (28:38.780)
in the video game that is being created
Lex Fridman (28:40.420)
by the video game engine.
Lex Fridman (28:42.100)
And this character is serving the aesthetics
Lex Fridman (28:43.820)
of the video game, and that is you.
Joscha Bach (28:47.060)
Yes, but I feel like I have control inside the video game.
Lex Fridman (28:50.900)
Like all those like 12 year olds
Joscha Bach (28:53.060)
that kick my ass on the internet.
Lex Fridman (28:55.420)
So when you play the video game,
Lex Fridman (28:57.700)
it doesn't really matter that there's zeros and ones, right?
Lex Fridman (28:59.900)
You don't care about the bits of the past.
Joscha Bach (29:01.700)
You don't care about the nature of the CPU
Lex Fridman (29:03.540)
that it runs on.
Lex Fridman (29:04.460)
What you care about are the properties of the game
Lex Fridman (29:06.700)
that you're playing.
Lex Fridman (29:07.780)
And you hope that the CPU is good enough.
Lex Fridman (29:10.060)
Yes.
Lex Fridman (29:10.900)
And a similar thing happens when we interact with physics.
Lex Fridman (29:13.300)
The world that you and me are in is not the physical world.
Joscha Bach (29:15.980)
The world that you and me are in is a dream world.
Lex Fridman (29:19.580)
How close is it to the real world though?
Joscha Bach (29:23.420)
We know that it's not very close,
Lex Fridman (29:25.020)
but we know that the dynamics of the dream world
Joscha Bach (29:27.500)
match the dynamics of the physical world
Lex Fridman (29:29.300)
to a certain degree of resolution.
Lex Fridman (29:31.060)
But the causal structure of the dream world is different.
Lex Fridman (29:35.220)
So you see for instance waves crashing on your feet, right?
Lex Fridman (29:38.180)
But there are no waves in the ocean.
Lex Fridman (29:39.420)
There's only water molecules that have tangents
Joscha Bach (29:42.420)
between the molecules that are the result of electrons
Lex Fridman (29:47.340)
in the molecules interacting with each other.
Lex Fridman (29:50.060)
Aren't they like very consistent?
Lex Fridman (29:52.140)
We're just seeing a very crude approximation.
Joscha Bach (29:55.700)
Isn't our dream world very consistent,
Lex Fridman (29:59.340)
like to the point of being mapped directly one to one
Joscha Bach (2:00:00.420)
industrialized countries.
Lex Fridman (2:00:01.700)
So some industrialized countries are doing better
Joscha Bach (2:00:03.820)
with a much cheaper healthcare system.
Lex Fridman (2:00:06.260)
And what you can see is for instance,
Joscha Bach (2:00:08.580)
administrative bloat.
Lex Fridman (2:00:09.980)
The healthcare system has maybe to some degree
Joscha Bach (2:00:13.660)
deliberately set up as a job placement program
Lex Fridman (2:00:17.220)
to allow people to continue living
Joscha Bach (2:00:19.740)
in middle class existence,
Lex Fridman (2:00:20.940)
despite not having useful use case in productivity.
Lex Fridman (2:00:25.940)
So they are being paid to push paper around.
Lex Fridman (2:00:28.700)
And the number of administrator in the healthcare system
Joscha Bach (2:00:31.540)
has been increasing much faster
Lex Fridman (2:00:33.180)
than the number of practitioners.
Lex Fridman (2:00:35.020)
And this is something that you have to pay for.
Lex Fridman (2:00:37.060)
And also the revenues that are being generated
Joscha Bach (2:00:40.460)
in the healthcare system are relatively large
Lex Fridman (2:00:42.220)
and somebody has to pay for them.
Lex Fridman (2:00:43.700)
And the result why they are so large
Lex Fridman (2:00:45.900)
is because market mechanisms are not working.
Joscha Bach (2:00:48.460)
The FDA is largely not protecting people
Lex Fridman (2:00:51.860)
from malpractice of healthcare providers.
Joscha Bach (2:00:55.380)
The FDA is protecting healthcare providers
Lex Fridman (2:00:58.660)
from competition.
Joscha Bach (2:00:59.940)
Right, okay.
Lex Fridman (2:01:00.780)
So this is a thing that has to do with values.
Lex Fridman (2:01:03.380)
And this is not because people are malicious on all levels.
Lex Fridman (2:01:06.500)
It's because they are not incentivized
Joscha Bach (2:01:08.460)
to act on a greater whole on this idea
Lex Fridman (2:01:11.380)
that you treat somebody who comes to you as a patient,
Joscha Bach (2:01:14.340)
like you would treat a family member.
Lex Fridman (2:01:15.660)
Yeah, but we're trying, I mean,
Joscha Bach (2:01:18.020)
you're highlighting a lot of the flaws
Lex Fridman (2:01:20.020)
of the different institutions,
Joscha Bach (2:01:21.220)
the systems we're operating under,
Lex Fridman (2:01:23.100)
but I think there's a continued throughout history
Joscha Bach (2:01:25.940)
mechanism design of trying to design incentives
Lex Fridman (2:01:29.380)
in such a way that these systems behave
Joscha Bach (2:01:31.420)
better and better and better.
Lex Fridman (2:01:32.780)
I mean, it's a very difficult thing
Joscha Bach (2:01:34.220)
to operate a society of hundreds of millions of people
Lex Fridman (2:01:38.140)
effectively with.
Lex Fridman (2:01:39.220)
Yes, so do we live in a society that is ever correcting?
Lex Fridman (2:01:42.820)
Is this, do we observe that our models
Joscha Bach (2:01:46.740)
of what we are doing are predictive of the future
Lex Fridman (2:01:49.420)
and when they are not, we improve them.
Joscha Bach (2:01:51.540)
Are our laws adjudicated with clauses
Lex Fridman (2:01:54.780)
that you put into every law,
Lex Fridman (2:01:56.020)
what is meant to be achieved by that law
Lex Fridman (2:01:57.900)
and the law will be automatically repealed
Lex Fridman (2:02:00.060)
if it's not achieving that, right?
Lex Fridman (2:02:01.340)
If you are optimizing your own laws,
Joscha Bach (2:02:03.220)
if you're writing your own source code,
Lex Fridman (2:02:05.140)
you probably make an estimate of what is this thing
Lex Fridman (2:02:08.180)
that's currently wrong in my life?
Lex Fridman (2:02:09.420)
What is it that I should change about my own policies?
Lex Fridman (2:02:12.180)
What is the expected outcome?
Lex Fridman (2:02:14.100)
And if that outcome doesn't manifest,
Lex Fridman (2:02:16.580)
I will change the policy back, right?
Lex Fridman (2:02:18.460)
Or I would change it to something different.
Lex Fridman (2:02:20.260)
Are we doing this on a societal level?
Lex Fridman (2:02:22.220)
I think so.
Joscha Bach (2:02:23.060)
I think it's easy to sort of highlight the,
Lex Fridman (2:02:25.580)
I think we're doing it in the way that,
Joscha Bach (2:02:29.020)
like I operate my current life.
Lex Fridman (2:02:30.380)
I didn't sleep much last night.
Joscha Bach (2:02:32.580)
You would say that Lex,
Lex Fridman (2:02:34.740)
the way you need to operate your life
Joscha Bach (2:02:35.980)
is you need to always get sleep.
Lex Fridman (2:02:37.340)
The fact that you didn't sleep last night
Joscha Bach (2:02:39.060)
is totally the wrong way to operate in your life.
Lex Fridman (2:02:43.060)
Like you should have gotten all your shit done in time
Lex Fridman (2:02:46.460)
and gotten to sleep because sleep is very important
Lex Fridman (2:02:48.940)
for health and you're highlighting,
Joscha Bach (2:02:50.540)
look, this person is not sleeping.
Lex Fridman (2:02:52.500)
Look, the medical, the healthcare system is operating poor.
Lex Fridman (2:02:56.380)
But the point is we just,
Lex Fridman (2:02:59.140)
it seems like this is the way,
Joscha Bach (2:03:00.460)
especially in the capitalist society, we operate.
Lex Fridman (2:03:02.700)
We keep running into trouble and last minute,
Joscha Bach (2:03:05.740)
we try to get our way out through innovation
Lex Fridman (2:03:09.260)
and it seems to work.
Joscha Bach (2:03:10.740)
You have a lot of people that ultimately are trying
Lex Fridman (2:03:13.380)
to build a better world and get urgency about them
Joscha Bach (2:03:18.380)
when the problem becomes more and more imminent.
Lex Fridman (2:03:22.900)
And that's the way this operates.
Lex Fridman (2:03:24.380)
But if you look at the long arc of history,
Lex Fridman (2:03:29.820)
it seems like that operating on deadlines
Joscha Bach (2:03:34.180)
produces progress and builds better and better systems.
Lex Fridman (2:03:36.980)
You probably agree with me that the US
Joscha Bach (2:03:39.060)
should have engaged in mask production in January 2020
Lex Fridman (2:03:44.580)
and that we should have shut down the airports early on
Lex Fridman (2:03:47.900)
and that we should have made it mandatory
Lex Fridman (2:03:50.940)
that the people that work in nursing homes
Joscha Bach (2:03:53.860)
are living on campus rather than living at home
Lex Fridman (2:03:57.940)
and then coming in and infecting people in the nursing homes
Joscha Bach (2:04:01.460)
that had no immune response to COVID.
Lex Fridman (2:04:03.900)
And that is something that was, I think, visible back then.
Joscha Bach (2:04:08.180)
The correct decisions haven't been made.
Lex Fridman (2:04:10.540)
We would have the same situation again.
Lex Fridman (2:04:12.580)
How do we know that these wrong decisions
Lex Fridman (2:04:14.340)
are not being made again?
Joscha Bach (2:04:15.780)
Have the people that made the decisions
Lex Fridman (2:04:17.620)
to not protect the nursing homes been punished?
Joscha Bach (2:04:20.580)
Have the people that made the wrong decisions
Lex Fridman (2:04:23.180)
with respect to testing that prevented the development
Joscha Bach (2:04:26.780)
of testing by startup companies and the importing
Lex Fridman (2:04:29.500)
of tests from countries that already had them,
Lex Fridman (2:04:32.140)
have these people been held responsible?
Lex Fridman (2:04:34.460)
First of all, so what do you wanna put
Lex Fridman (2:04:37.380)
before the firing squad?
Lex Fridman (2:04:38.780)
I think they are being held responsible.
Joscha Bach (2:04:39.860)
No, just make sure that this doesn't happen again.
Lex Fridman (2:04:41.820)
No, but it's not that, yes, they're being held responsible
Joscha Bach (2:04:46.220)
by many voices, by people being frustrated.
Lex Fridman (2:04:48.820)
There's new leaders being born now
Joscha Bach (2:04:50.740)
that we're going to see rise to the top in 10 years.
Lex Fridman (2:04:54.220)
This moves slower than, there's obviously
Joscha Bach (2:04:57.380)
a lot of older incompetence and bureaucracy
Lex Fridman (2:05:01.220)
and these systems move slowly.
Joscha Bach (2:05:03.660)
They move like science, one death at a time.
Lex Fridman (2:05:06.860)
So yes, I think the pain that's been felt
Joscha Bach (2:05:11.340)
in the previous year is reverberating throughout the world.
Lex Fridman (2:05:15.540)
Maybe I'm getting old, I suspect that every generation
Joscha Bach (2:05:18.340)
in the US after the war has lost the plot even more.
Lex Fridman (2:05:21.500)
I don't see this development.
Lex Fridman (2:05:23.140)
The war, World War II?
Lex Fridman (2:05:24.700)
Yes, so basically there was a time when we were modernist
Lex Fridman (2:05:29.140)
and in this modernist time, the US felt actively threatened
Lex Fridman (2:05:33.620)
by the things that happened in the world.
Joscha Bach (2:05:35.740)
The US was worried about possibility of failure
Lex Fridman (2:05:39.660)
and this imminence of possible failure led to decisions.
Joscha Bach (2:05:44.580)
There was a time when the government would listen
Lex Fridman (2:05:47.340)
to physicists about how to do things
Lex Fridman (2:05:50.540)
and the physicists were actually concerned
Lex Fridman (2:05:52.100)
about what the government should be doing.
Lex Fridman (2:05:53.580)
So they would be writing letters to the government
Lex Fridman (2:05:56.100)
and so for instance, the decision for the Manhattan Project
Joscha Bach (2:05:58.860)
was something that was driven in a conversation
Lex Fridman (2:06:01.740)
between physicists and the government.
Joscha Bach (2:06:04.020)
I don't think such a discussion would take place today.
Lex Fridman (2:06:06.900)
I disagree, I think if the virus was much deadlier,
Joscha Bach (2:06:10.500)
we would see a very different response.
Lex Fridman (2:06:12.620)
I think the virus was not sufficiently deadly
Lex Fridman (2:06:14.900)
and instead because it wasn't very deadly,
Lex Fridman (2:06:17.420)
what happened is the current system
Joscha Bach (2:06:20.420)
started to politicize it.
Lex Fridman (2:06:21.980)
The mask, this is what I realized with masks early on,
Joscha Bach (2:06:25.300)
they were not, very quickly became not as a solution
Lex Fridman (2:06:29.540)
but they became a thing that politicians used
Joscha Bach (2:06:32.620)
to divide the country.
Lex Fridman (2:06:33.900)
So the same things happened with vaccines, same thing.
Lex Fridman (2:06:36.940)
So like nobody's really,
Lex Fridman (2:06:38.740)
people weren't talking about solutions to this problem
Joscha Bach (2:06:41.100)
because I don't think the problem was bad enough.
Lex Fridman (2:06:43.060)
When you talk about the war,
Joscha Bach (2:06:45.020)
I think our lives are too comfortable.
Lex Fridman (2:06:48.740)
I think in the developed world, things are too good
Lex Fridman (2:06:52.220)
and we have not faced severe dangers.
Lex Fridman (2:06:54.900)
When the danger, the severe dangers,
Joscha Bach (2:06:57.500)
existential threats are faced, that's when we step up
Lex Fridman (2:07:00.740)
on a small scale and a large scale.
Joscha Bach (2:07:02.980)
Now, I don't, that's sort of my argument here
Lex Fridman (2:07:07.980)
but I did think the virus is, I was hoping
Joscha Bach (2:07:11.700)
that it was actually sufficiently dangerous
Lex Fridman (2:07:16.060)
for us to step up because especially in the early days,
Joscha Bach (2:07:18.660)
it was unclear, it still is unclear because of mutations,
Lex Fridman (2:07:23.260)
how bad it might be, right?
Lex Fridman (2:07:25.820)
And so I thought we would step up and even,
Lex Fridman (2:07:30.700)
so the masks point is a tricky one because to me,
Joscha Bach (2:07:35.540)
the manufacture of masks isn't even the problem.
Lex Fridman (2:07:38.780)
I'm still to this day and I was involved
Joscha Bach (2:07:41.020)
with a bunch of this work, have not seen good science done
Lex Fridman (2:07:44.300)
on whether masks work or not.
Joscha Bach (2:07:46.860)
Like there still has not been a large scale study.
Lex Fridman (2:07:49.420)
To me, that should be, there should be large scale studies
Lex Fridman (2:07:51.820)
and every possible solution, like aggressive
Lex Fridman (2:07:55.140)
in the same way that the vaccine development
Joscha Bach (2:07:56.780)
was aggressive.
Lex Fridman (2:07:57.740)
There should be masks, which tests,
Lex Fridman (2:07:59.860)
what kind of tests work really well, what kind of,
Lex Fridman (2:08:03.740)
like even the question of how the virus spreads.
Joscha Bach (2:08:06.860)
There should be aggressive studies on that to understand.
Lex Fridman (2:08:09.820)
I'm still, as far as I know, there's still a lot
Joscha Bach (2:08:12.860)
of uncertainty about that.
Lex Fridman (2:08:14.180)
Nobody wants to see this as an engineering problem
Joscha Bach (2:08:17.100)
that needs to be solved.
Lex Fridman (2:08:18.540)
It's that I was surprised about, but I wouldn't.
Lex Fridman (2:08:21.940)
So I find that our views are largely convergent
Lex Fridman (2:08:24.580)
but not completely.
Lex Fridman (2:08:25.460)
So I agree with the thing that because our society
Lex Fridman (2:08:29.340)
in some sense perceives itself as too big to fail.
Joscha Bach (2:08:32.580)
Right.
Lex Fridman (2:08:33.420)
The virus did not alert people to the fact
Joscha Bach (2:08:35.940)
that we are facing possible failure
Lex Fridman (2:08:38.820)
that basically put us into the postmodernist mode.
Lex Fridman (2:08:41.540)
And I don't mean in a philosophical sense
Lex Fridman (2:08:43.260)
but in a societal sense.
Joscha Bach (2:08:45.220)
The difference between the postmodern society
Lex Fridman (2:08:47.940)
and the modern society is that the modernist society
Joscha Bach (2:08:50.540)
has to deal with the ground truth
Lex Fridman (2:08:52.340)
and the postmodernist society has to deal with appearances.
Joscha Bach (2:08:55.540)
Politics becomes a performance
Lex Fridman (2:08:57.820)
and the performance is done for an audience
Lex Fridman (2:08:59.780)
and the organized audience is the media.
Lex Fridman (2:09:02.260)
And the media evaluates itself via other media, right?
Lex Fridman (2:09:05.380)
So you have an audience of critics that evaluate themselves.
Lex Fridman (2:09:09.100)
And I don't think it's so much the failure
Joscha Bach (2:09:10.660)
of the politicians because to get in power
Lex Fridman (2:09:12.820)
and to stay in power, you need to be able
Joscha Bach (2:09:15.700)
to deal with the published opinion.
Lex Fridman (2:09:17.580)
Well, I think it goes in cycles
Joscha Bach (2:09:19.220)
because what's going to happen is all
Lex Fridman (2:09:22.300)
of the small business owners, all the people
Joscha Bach (2:09:24.820)
who truly are suffering and will suffer more
Lex Fridman (2:09:27.900)
because the effects of the closure of the economy
Lex Fridman (2:09:31.940)
and the lack of solutions to the virus,
Lex Fridman (2:09:34.220)
they're going to apprise.
Lex Fridman (2:09:36.300)
And hopefully, I mean, this is where charismatic leaders
Lex Fridman (2:09:40.180)
can get the world in trouble
Lex Fridman (2:09:42.620)
but hopefully will elect great leaders
Lex Fridman (2:09:47.900)
that will break through this postmodernist idea
Joscha Bach (2:09:51.100)
of the media and the perception
Lex Fridman (2:09:55.420)
and the drama on Twitter and all that kind of stuff.
Lex Fridman (2:09:57.660)
But you know, this can go either way.
Lex Fridman (2:09:59.340)
Yeah.
Joscha Bach (2:10:00.300)
When the Weimar Republic was unable to deal
Lex Fridman (2:10:03.620)
with the economic crisis that Germany was facing,
Joscha Bach (2:10:07.540)
there was an option to go back.
Lex Fridman (2:10:10.340)
But there were people which thought,
Joscha Bach (2:10:11.620)
let's get back to a constitutional monarchy
Lex Fridman (2:10:14.420)
and let's get this to work because democracy doesn't work.
Lex Fridman (2:10:18.780)
And eventually, there was no way back.
Lex Fridman (2:10:21.740)
People decided there was no way back.
Joscha Bach (2:10:23.300)
They needed to go forward.
Lex Fridman (2:10:24.420)
And the only options for going forward
Joscha Bach (2:10:26.700)
was to become Stalinist communist,
Lex Fridman (2:10:29.460)
basically an option to completely expropriate
Joscha Bach (2:10:34.260)
the factories and so on and nationalize them
Lex Fridman (2:10:36.820)
and to reorganize Germany in communist terms
Lex Fridman (2:10:40.540)
and ally itself with Stalin and fascism.
Lex Fridman (2:10:44.660)
And both options were obviously very bad.
Lex Fridman (2:10:47.980)
And the one that the Germans picked
Lex Fridman (2:10:49.980)
led to a catastrophe that devastated Europe.
Lex Fridman (2:10:54.300)
And I'm not sure if the US has an immune response
Lex Fridman (2:10:57.180)
against that.
Joscha Bach (2:10:58.020)
I think that the far right is currently very weak in the US,
Lex Fridman (2:11:01.340)
but this can easily change.
Lex Fridman (2:11:05.740)
Do you think from a historical perspective,
Lex Fridman (2:11:08.780)
Hitler could have been stopped
Lex Fridman (2:11:10.820)
from within Germany or from outside?
Lex Fridman (2:11:14.020)
Or this, well, depends on who you wanna focus,
Joscha Bach (2:11:17.780)
whether you wanna focus on Stalin or Hitler,
Lex Fridman (2:11:20.220)
but it feels like Hitler was the one
Joscha Bach (2:11:22.380)
as a political movement that could have been stopped.
Lex Fridman (2:11:25.180)
I think that the point was that a lot of people
Joscha Bach (2:11:28.700)
wanted Hitler, so he got support from a lot of quarters.
Lex Fridman (2:11:32.340)
There was a number of industrialists who supported him
Joscha Bach (2:11:35.100)
because they thought that the democracy
Lex Fridman (2:11:36.780)
is obviously not working and unstable
Lex Fridman (2:11:38.460)
and you need a strong man.
Lex Fridman (2:11:40.620)
And he was willing to play that part.
Joscha Bach (2:11:43.220)
There were also people in the US who thought
Lex Fridman (2:11:45.460)
that Hitler would stop Stalin
Lex Fridman (2:11:47.860)
and would act as a bulwark against Bolshevism,
Lex Fridman (2:11:52.300)
which he probably would have done, right?
Lex Fridman (2:11:54.140)
But at which cost?
Lex Fridman (2:11:56.220)
And then many of the things that he was going to do,
Joscha Bach (2:11:59.740)
like the Holocaust, was something where people thought
Lex Fridman (2:12:03.780)
this is rhetoric, he's not actually going to do this.
Joscha Bach (2:12:07.140)
Especially many of the Jews themselves, which were humanists.
Lex Fridman (2:12:10.020)
And for them, this was outside of the scope
Joscha Bach (2:12:12.300)
that was thinkable.
Lex Fridman (2:12:13.260)
Right.
Joscha Bach (2:12:14.260)
I mean, I wonder if Hitler is uniquely,
Lex Fridman (2:12:20.140)
I wanna carefully use this term, but uniquely evil.
Lex Fridman (2:12:23.500)
So if Hitler was never born,
Lex Fridman (2:12:26.420)
if somebody else would come in this place.
Lex Fridman (2:12:29.100)
So like, just thinking about the progress of history,
Lex Fridman (2:12:33.780)
how important are those singular figures
Lex Fridman (2:12:36.700)
that lead to mass destruction and cruelty?
Lex Fridman (2:12:40.980)
Because my sense is Hitler was unique.
Joscha Bach (2:12:47.540)
It wasn't just about the environment
Lex Fridman (2:12:49.420)
and the context that gave him,
Joscha Bach (2:12:51.620)
like another person would not come in his place
Lex Fridman (2:12:54.740)
to do as destructive of the things that he did.
Joscha Bach (2:12:58.220)
There was a combination of charisma, of madness,
Lex Fridman (2:13:02.860)
of psychopathy, of just ego, all those things,
Joscha Bach (2:13:07.180)
which are very unlikely to come together
Lex Fridman (2:13:09.540)
in one person in the right time.
Joscha Bach (2:13:12.540)
It also depends on the context of the country
Lex Fridman (2:13:14.660)
that you're operating in.
Joscha Bach (2:13:16.500)
If you tell the Germans that they have a historical destiny
Lex Fridman (2:13:22.220)
in this romantic country,
Joscha Bach (2:13:23.820)
the effect is probably different
Lex Fridman (2:13:25.500)
than it is in other countries.
Lex Fridman (2:13:27.180)
But Stalin has killed a few more people than Hitler did.
Lex Fridman (2:13:33.620)
And if you look at the probability
Joscha Bach (2:13:35.820)
that you survived under Stalin,
Lex Fridman (2:13:39.220)
Hitler killed people if he thought
Joscha Bach (2:13:43.140)
they were not worth living,
Lex Fridman (2:13:45.140)
or if they were harmful to his racist project.
Joscha Bach (2:13:49.260)
He basically felt that the Jews would be too cosmopolitan
Lex Fridman (2:13:52.580)
and would not be willing to participate
Joscha Bach (2:13:55.140)
in the racist redefinition of society
Lex Fridman (2:13:57.500)
and the value of society,
Lex Fridman (2:13:58.780)
and there is no state in this way
Lex Fridman (2:14:01.420)
that he wanted to have it.
Lex Fridman (2:14:03.260)
So he saw them as harmful danger,
Lex Fridman (2:14:06.980)
especially since they played such an important role
Joscha Bach (2:14:09.460)
in the economy and culture of Germany.
Lex Fridman (2:14:13.300)
And so basically he had some radical
Lex Fridman (2:14:18.020)
but rational reason to murder them.
Lex Fridman (2:14:20.780)
And Stalin just killed everyone.
Joscha Bach (2:14:23.420)
Basically the Stalinist purges were such a random thing
Lex Fridman (2:14:26.140)
where he said that there's a certain possibility
Joscha Bach (2:14:31.580)
that this particular part of the population
Lex Fridman (2:14:34.660)
has a number of German collaborators or something,
Lex Fridman (2:14:36.740)
and we just kill them all, right?
Lex Fridman (2:14:38.820)
Or if you look at what Mao did,
Joscha Bach (2:14:40.660)
the number of people that were killed
Lex Fridman (2:14:42.980)
in absolute numbers were much higher under Mao
Joscha Bach (2:14:45.620)
than they were under Stalin.
Lex Fridman (2:14:47.660)
So it's super hard to say.
Joscha Bach (2:14:49.820)
The other thing is that you look at Genghis Khan and so on,
Lex Fridman (2:14:53.540)
how many people he killed.
Joscha Bach (2:14:56.100)
When you see there are a number of things
Lex Fridman (2:14:58.900)
that happen in human history
Joscha Bach (2:14:59.900)
that actually really put a substantial dent
Lex Fridman (2:15:02.500)
in the existing population, or Napoleon.
Lex Fridman (2:15:05.900)
And it's very difficult to eventually measure it
Lex Fridman (2:15:09.500)
because what's happening is basically evolution
Joscha Bach (2:15:12.020)
on a human scale where one monkey figures out
Lex Fridman (2:15:17.220)
a way to become viral and is using this viral technology
Joscha Bach (2:15:22.380)
to change the patterns of society
Lex Fridman (2:15:24.500)
at the very, very large scale.
Lex Fridman (2:15:26.500)
And what we find so abhorrent about these changes
Lex Fridman (2:15:29.860)
is the complexity that is being destroyed by this.
Joscha Bach (2:15:32.340)
That's basically like a big fire that burns out
Lex Fridman (2:15:34.940)
a lot of the existing culture and structure
Joscha Bach (2:15:36.780)
that existed before.
Lex Fridman (2:15:38.060)
Yeah, and it all just starts with one monkey.
Joscha Bach (2:15:42.580)
One charismatic ape.
Lex Fridman (2:15:44.460)
And there's a bunch of them throughout history.
Joscha Bach (2:15:46.060)
Yeah, but it's in a given environment.
Lex Fridman (2:15:47.940)
It's basically similar to wildfires in California, right?
Joscha Bach (2:15:51.100)
The temperature is rising.
Lex Fridman (2:15:53.260)
There is less rain falling.
Lex Fridman (2:15:55.540)
And then suddenly a single spark can have an effect
Lex Fridman (2:15:57.900)
that in other times would be contained.
Joscha Bach (2:16:00.780)
Okay, speaking of which, I love how we went
Lex Fridman (2:16:04.620)
to Hitler and Stalin from 20, 30 minutes ago,
Joscha Bach (2:16:09.020)
GPT3 generating, doing programs that this is.
Lex Fridman (2:16:13.620)
The argument was about morality of AI versus human.
Lex Fridman (2:16:23.580)
And specifically in the context of writing programs,
Lex Fridman (2:16:26.220)
specifically in the context of programs
Joscha Bach (2:16:28.540)
that can be destructive.
Lex Fridman (2:16:29.940)
So running nuclear power plants
Joscha Bach (2:16:31.860)
or autonomous weapons systems, for example.
Lex Fridman (2:16:35.100)
And I think your inclination was to say that
Joscha Bach (2:16:39.580)
it's not so obvious that AI would be less moral than humans
Lex Fridman (2:16:43.460)
or less effective at making a world
Joscha Bach (2:16:46.740)
that would make humans happy.
Lex Fridman (2:16:48.660)
So I'm not talking about self directed systems
Joscha Bach (2:16:52.660)
that are making their own goals at a global scale.
Lex Fridman (2:16:57.100)
If you just talk about the deployment
Joscha Bach (2:16:59.140)
of technological systems that are able to see order
Lex Fridman (2:17:03.380)
and patterns and use this as control models
Joscha Bach (2:17:05.580)
to act on the goals that we give them,
Lex Fridman (2:17:08.420)
then if we have the correct incentives
Joscha Bach (2:17:11.140)
to set the correct incentives for these systems,
Lex Fridman (2:17:13.180)
I'm quite optimistic.
Lex Fridman (2:17:16.060)
So humans versus AI, let me give you an example.
Lex Fridman (2:17:20.660)
Autonomous weapon system.
Joscha Bach (2:17:23.900)
Let's say there's a city somewhere in the Middle East
Lex Fridman (2:17:26.900)
that has a number of terrorists.
Lex Fridman (2:17:30.380)
And the question is,
Lex Fridman (2:17:32.180)
what's currently done with drone technologies,
Joscha Bach (2:17:35.020)
you have information about the location
Lex Fridman (2:17:37.100)
of a particular terrorist and you have a targeted attack,
Joscha Bach (2:17:40.620)
you have a bombing of that particular building.
Lex Fridman (2:17:43.980)
And that's all directed by humans
Joscha Bach (2:17:45.900)
at the high level strategy
Lex Fridman (2:17:47.980)
and also at the deployment of individual bombs and missiles
Joscha Bach (2:17:50.580)
like the actual, everything is done by human
Lex Fridman (2:17:53.940)
except the final targeting.
Lex Fridman (2:17:56.700)
And it's like spot, similar thing, like control the flight.
Lex Fridman (2:18:01.820)
Okay, what if you give AI control and saying,
Joscha Bach (2:18:07.860)
write a program that says,
Lex Fridman (2:18:10.300)
here's the best information I have available
Joscha Bach (2:18:12.180)
about the location of these five terrorists,
Lex Fridman (2:18:14.780)
here's the city, make sure all the bombing you do
Joscha Bach (2:18:17.820)
is constrained to the city, make sure it's precision based,
Lex Fridman (2:18:21.820)
but you take care of it.
Lex Fridman (2:18:22.860)
So you do one level of abstraction out
Lex Fridman (2:18:25.660)
and saying, take care of the terrorists in the city.
Joscha Bach (2:18:29.580)
Which are you more comfortable with,
Lex Fridman (2:18:31.420)
the humans or the JavaScript GPT3 generated code
Lex Fridman (2:18:35.700)
that's doing the deployment?
Lex Fridman (2:18:38.220)
I mean, this is the kind of question I'm asking,
Joscha Bach (2:18:42.340)
is the kind of bugs that we see in human nature,
Lex Fridman (2:18:47.100)
are they better or worse than the kind of bugs we see in AI?
Joscha Bach (2:18:51.220)
There are different bugs.
Lex Fridman (2:18:52.460)
There is an issue that if people are creating
Joscha Bach (2:18:55.980)
an imperfect automation of a process
Lex Fridman (2:18:59.900)
that normally requires a moral judgment,
Lex Fridman (2:19:02.860)
and this moral judgment is the reason
Lex Fridman (2:19:05.940)
why it cannot be automated often,
Joscha Bach (2:19:07.460)
it's not because the computation is too expensive,
Lex Fridman (2:19:12.180)
but because the model that you give the AI
Joscha Bach (2:19:14.300)
is not an adequate model of the dynamics of the world,
Lex Fridman (2:19:17.500)
because the AI does not understand the context
Joscha Bach (2:19:19.780)
that it's operating in the right way.
Lex Fridman (2:19:21.940)
And this is something that already happens with Excel.
Joscha Bach (2:19:24.940)
You don't need to have an AI system to do this.
Lex Fridman (2:19:27.860)
You have an automated process in place
Joscha Bach (2:19:30.340)
where humans decide using automated criteria
Lex Fridman (2:19:33.180)
whom to kill when and whom to target when,
Joscha Bach (2:19:36.020)
which already happens.
Lex Fridman (2:19:38.220)
And you have no way to get off the kill list
Joscha Bach (2:19:40.300)
once that happens, once you have been targeted
Lex Fridman (2:19:42.860)
according to some automatic criterion
Joscha Bach (2:19:44.860)
by people in a bureaucracy, that is the issue.
Lex Fridman (2:19:48.980)
The issue is not the AI, it's the automation.
Lex Fridman (2:19:52.260)
So there's something about, right, it's automation,
Lex Fridman (2:19:56.380)
but there's something about the,
Joscha Bach (2:19:58.820)
there's a certain level of abstraction
Lex Fridman (2:20:00.660)
where you give control to AI to do the automation.
Joscha Bach (2:20:04.340)
There's a scale that can be achieved
Lex Fridman (2:20:07.100)
that it feels like the scale of bug and scale mistake
Lex Fridman (2:20:10.780)
and scale of destruction that can be achieved
Lex Fridman (2:20:14.580)
of the kind that humans cannot achieve.
Lex Fridman (2:20:16.860)
So AI is much more able to destroy
Lex Fridman (2:20:19.620)
an entire country accidentally versus humans.
Joscha Bach (2:20:22.580)
It feels like the more civilians die as they react
Lex Fridman (2:20:27.300)
or suffer as the consequences of your decisions,
Joscha Bach (2:20:30.900)
the more weight there is on the human mind
Lex Fridman (2:20:34.260)
to make that decision.
Lex Fridman (2:20:36.380)
And so like, it becomes more and more unlikely
Lex Fridman (2:20:39.060)
to make that decision for humans.
Joscha Bach (2:20:41.460)
For AI, it feels like it's harder to encode
Lex Fridman (2:20:44.900)
that kind of weight.
Joscha Bach (2:20:47.020)
In a way, the AI that we're currently building
Lex Fridman (2:20:49.620)
is automating statistics, right?
Joscha Bach (2:20:51.900)
Intelligence is the ability to make models
Lex Fridman (2:20:53.860)
so you can act on them,
Lex Fridman (2:20:55.220)
and AI is the tool to make better models.
Lex Fridman (2:20:58.340)
So in principle, if you're using AI wisely,
Joscha Bach (2:21:01.540)
you're able to prevent more harm.
Lex Fridman (2:21:04.220)
And I think that the main issue is not on the side of the AI,
Joscha Bach (2:21:07.860)
it's on the side of the human command hierarchy
Lex Fridman (2:21:09.940)
that is using technology irresponsibly.
Lex Fridman (2:21:12.300)
So the question is how hard is it to encode,
Lex Fridman (2:21:15.740)
to properly encode the right incentives into the AI?
Lex Fridman (2:21:19.060)
So for instance, there's this idea
Lex Fridman (2:21:21.420)
of what happens if we let our airplanes being flown
Joscha Bach (2:21:24.460)
with AI systems and the neural network is a black box
Lex Fridman (2:21:27.620)
and so on.
Lex Fridman (2:21:28.460)
And it turns out our neural networks
Lex Fridman (2:21:30.220)
are actually not black boxes anymore.
Joscha Bach (2:21:32.300)
There are function approximators using linear algebra,
Lex Fridman (2:21:36.620)
and there are performing things that we can understand.
Lex Fridman (2:21:40.020)
But we can also, instead of letting the neural network
Lex Fridman (2:21:42.820)
fly the airplane, use the neural network
Joscha Bach (2:21:44.940)
to generate a provably correct program.
Lex Fridman (2:21:47.420)
There's a degree of accuracy of the proof
Joscha Bach (2:21:49.900)
that a human could not achieve.
Lex Fridman (2:21:51.860)
And so we can use our AI by combining
Joscha Bach (2:21:54.100)
different technologies to build systems
Lex Fridman (2:21:56.460)
that are much more reliable than the systems
Joscha Bach (2:21:58.340)
that a human being could create.
Lex Fridman (2:22:00.420)
And so in this sense, I would say that
Joscha Bach (2:22:03.900)
if you use an early stage of technology to save labor
Lex Fridman (2:22:08.340)
and don't employ competent people,
Lex Fridman (2:22:11.340)
but just to hack something together because you can,
Lex Fridman (2:22:14.180)
that is very dangerous.
Lex Fridman (2:22:15.260)
And if people are acting under these incentives
Lex Fridman (2:22:17.220)
that they get away with delivering shoddy work
Joscha Bach (2:22:20.380)
more cheaply using AI with less human oversight than before,
Lex Fridman (2:22:23.980)
that's very dangerous.
Joscha Bach (2:22:25.140)
The thing is though, AI is still going to be unreliable,
Lex Fridman (2:22:28.980)
perhaps less so than humans,
Lex Fridman (2:22:30.420)
but it'll be unreliable in novel ways.
Lex Fridman (2:22:33.820)
And...
Joscha Bach (2:22:35.340)
Yeah, but this is an empirical question.
Lex Fridman (2:22:37.180)
And it's something that we can figure out and work with.
Lex Fridman (2:22:39.860)
So the issue is, do we trust the systems,
Lex Fridman (2:22:43.100)
the social systems that we have in place
Lex Fridman (2:22:45.340)
and the social systems that we can build and maintain
Lex Fridman (2:22:48.020)
that they're able to use AI responsibly?
Joscha Bach (2:22:50.420)
If they can, then AI is good news.
Lex Fridman (2:22:52.980)
If they cannot,
Joscha Bach (2:22:54.100)
then it's going to make the existing problems worse.
Lex Fridman (2:22:57.220)
Well, and also who creates the AI, who controls it,
Joscha Bach (2:23:00.100)
who makes money from it because it's ultimately humans.
Lex Fridman (2:23:03.140)
And then you start talking about
Lex Fridman (2:23:05.060)
how much you trust the humans.
Lex Fridman (2:23:06.940)
So the question is, what does who mean?
Joscha Bach (2:23:08.740)
I don't think that we have identity per se.
Lex Fridman (2:23:11.140)
I think that the story of a human being is somewhat random.
Lex Fridman (2:23:15.500)
What happens is more or less that everybody is acting
Lex Fridman (2:23:18.420)
on their local incentives,
Lex Fridman (2:23:19.780)
what they perceive to be their incentives.
Lex Fridman (2:23:21.980)
And the question is, what are the incentives
Lex Fridman (2:23:24.620)
that the one that is pressing the button is operating under?
Lex Fridman (2:23:28.500)
Yeah.
Joscha Bach (2:23:30.140)
It's nice for those incentives to be transparent.
Lex Fridman (2:23:32.620)
So, for example, I'll give you an example.
Joscha Bach (2:23:36.060)
There seems to be a significant distrust
Lex Fridman (2:23:38.700)
of a tech, like entrepreneurs in the tech space
Joscha Bach (2:23:44.380)
or people that run, for example, social media companies
Lex Fridman (2:23:47.380)
like Mark Zuckerberg.
Joscha Bach (2:23:49.980)
There's not a complete transparency of incentives
Lex Fridman (2:23:53.060)
under which that particular human being operates.
Joscha Bach (2:23:58.940)
We can listen to the words he says
Lex Fridman (2:24:00.700)
or what the marketing team says for a company,
Lex Fridman (2:24:02.980)
but we don't know.
Lex Fridman (2:24:04.220)
And that becomes a problem when the algorithms
Lex Fridman (2:24:08.260)
and the systems created by him and other people
Lex Fridman (2:24:12.780)
in that company start having more and more impact
Joscha Bach (2:24:15.780)
on society.
Lex Fridman (2:24:17.180)
And that it starts, if the incentives were somehow
Joscha Bach (2:24:21.940)
the definition and the explainability of the incentives
Lex Fridman (2:24:26.020)
was decentralized such that nobody can manipulate it,
Joscha Bach (2:24:30.860)
no propaganda type manipulation of like
Lex Fridman (2:24:35.580)
how these systems actually operate could be done,
Joscha Bach (2:24:38.020)
then yes, I think AI could achieve much fairer,
Lex Fridman (2:24:45.340)
much more effective sort of like solutions
Joscha Bach (2:24:50.580)
to difficult ethical problems.
Lex Fridman (2:24:53.260)
But when there's like humans in the loop,
Joscha Bach (2:24:55.780)
manipulating the dissemination, the communication
Lex Fridman (2:25:00.580)
of how the system actually works,
Joscha Bach (2:25:02.420)
that feels like you can run into a lot of trouble.
Lex Fridman (2:25:05.300)
And that's why there's currently a lot of distrust
Joscha Bach (2:25:07.740)
for people at the heads of companies
Lex Fridman (2:25:10.180)
that have increasingly powerful AI systems.
Joscha Bach (2:25:13.900)
I suspect what happened traditionally in the US
Lex Fridman (2:25:16.860)
was that since our decision making
Joscha Bach (2:25:18.700)
is much more decentralized than in an authoritarian state,
Lex Fridman (2:25:22.980)
people are making decisions autonomously
Joscha Bach (2:25:24.780)
at many, many levels in a society.
Lex Fridman (2:25:26.980)
What happened that was we created coherence
Lex Fridman (2:25:30.260)
and cohesion in society by controlling what people thought
Lex Fridman (2:25:33.940)
and what information they had.
Joscha Bach (2:25:35.740)
The media synchronized public opinion
Lex Fridman (2:25:38.740)
and social media have disrupted this.
Joscha Bach (2:25:40.340)
It's not, I think so much Russian influence or something,
Lex Fridman (2:25:43.780)
it's everybody's influence.
Joscha Bach (2:25:45.460)
It's that a random person can come up
Lex Fridman (2:25:47.860)
with a conspiracy theory and disrupt what people think.
Lex Fridman (2:25:52.460)
And if that conspiracy theory is more compelling
Lex Fridman (2:25:55.460)
or more attractive than the standardized
Joscha Bach (2:25:58.180)
public conspiracy theory that we give people as a default,
Lex Fridman (2:26:01.860)
then it might get more traction, right?
Joscha Bach (2:26:03.460)
You suddenly have the situation that a single individual
Lex Fridman (2:26:05.940)
somewhere on a farm in Texas has more listeners than CNN.
Lex Fridman (2:26:11.140)
Which particular farmer are you referring to in Texas?
Lex Fridman (2:26:17.380)
Probably no.
Joscha Bach (2:26:19.180)
Yes, I had dinner with him a couple of times, okay.
Lex Fridman (2:26:21.700)
Right, it's an interesting situation
Joscha Bach (2:26:23.420)
because you cannot get to be an anchor in CNN
Lex Fridman (2:26:25.940)
if you don't go through a complicated gatekeeping process.
Lex Fridman (2:26:30.420)
And suddenly you have random people
Lex Fridman (2:26:32.460)
without that gatekeeping process,
Joscha Bach (2:26:34.900)
just optimizing for attention.
Lex Fridman (2:26:36.980)
Not necessarily with a lot of responsibility
Joscha Bach (2:26:39.540)
for the longterm effects of projecting these theories
Lex Fridman (2:26:42.700)
into the public.
Lex Fridman (2:26:43.900)
And now there is a push of making social media
Lex Fridman (2:26:46.980)
more like traditional media,
Joscha Bach (2:26:48.380)
which means that the opinion that is being projected
Lex Fridman (2:26:51.380)
in social media is more limited to an acceptable range.
Joscha Bach (2:26:54.660)
With the goal of getting society into safe waters
Lex Fridman (2:26:58.380)
and increase the stability and cohesion of society again,
Joscha Bach (2:27:00.820)
which I think is a laudable goal.
Lex Fridman (2:27:03.140)
But of course it also is an opportunity
Joscha Bach (2:27:05.100)
to seize the means of indoctrination.
Lex Fridman (2:27:08.340)
And the incentives that people are under when they do this
Joscha Bach (2:27:11.420)
are in such a way that the AI ethics that we would need
Lex Fridman (2:27:17.140)
becomes very often something like AI politics,
Joscha Bach (2:27:20.620)
which is basically partisan and ideological.
Lex Fridman (2:27:23.380)
And this means that whatever one side says,
Lex Fridman (2:27:26.180)
another side is going to be disagreeing with, right?
Lex Fridman (2:27:28.380)
In the same way as when you turn masks or the vaccine
Joscha Bach (2:27:31.740)
into a political issue,
Lex Fridman (2:27:33.140)
if you say that it is politically virtuous
Joscha Bach (2:27:35.700)
to get vaccinated,
Lex Fridman (2:27:36.660)
it will mean that the people that don't like you
Lex Fridman (2:27:39.260)
will not want to get vaccinated, right?
Lex Fridman (2:27:41.020)
And as soon as you have this partisan discourse,
Joscha Bach (2:27:43.620)
it's going to be very hard to make the right decisions
Lex Fridman (2:27:47.140)
because the incentives get to be the wrong ones.
Joscha Bach (2:27:48.860)
AI ethics needs to be super boring.
Lex Fridman (2:27:51.180)
It needs to be done by people who do statistics
Joscha Bach (2:27:53.300)
all the time and have extremely boring,
Lex Fridman (2:27:56.540)
long winded discussions that most people cannot follow
Joscha Bach (2:27:59.620)
because they are too complicated,
Lex Fridman (2:28:00.900)
but that are dead serious.
Joscha Bach (2:28:02.540)
These people need to be able to be better at statistics
Lex Fridman (2:28:05.820)
than the leading machine learning researchers.
Lex Fridman (2:28:07.940)
And at the moment, the AI ethics debate is the one
Lex Fridman (2:28:12.060)
where you don't have any barrier to entry, right?
Joscha Bach (2:28:14.460)
Everybody who has a strong opinion
Lex Fridman (2:28:16.820)
and is able to signal that opinion in the right way
Joscha Bach (2:28:18.860)
can enter it.
Lex Fridman (2:28:19.700)
And to me, that is a very frustrating thing
Joscha Bach (2:28:24.340)
because the field is so crucially important
Lex Fridman (2:28:26.260)
to our future.
Joscha Bach (2:28:27.100)
It's so crucially important,
Lex Fridman (2:28:28.260)
but the only qualification you currently need
Joscha Bach (2:28:31.860)
is to be outraged by the injustice in the world.
Lex Fridman (2:28:34.740)
It's more complicated, right?
Joscha Bach (2:28:36.220)
Everybody seems to be outraged.
Lex Fridman (2:28:37.860)
But let's just say that the incentives
Joscha Bach (2:28:40.740)
are not always the right ones.
Lex Fridman (2:28:42.020)
So basically, I suspect that a lot of people
Joscha Bach (2:28:45.500)
that enter this debate don't have a vision
Lex Fridman (2:28:48.140)
for what society should be looking like
Joscha Bach (2:28:50.020)
in a way that is nonviolent,
Lex Fridman (2:28:51.380)
where we preserve liberal democracy,
Joscha Bach (2:28:53.580)
where we make sure that we all get along
Lex Fridman (2:28:56.300)
and we are around in a few hundred years from now,
Joscha Bach (2:29:00.420)
preferably with a comfortable
Lex Fridman (2:29:02.180)
technological civilization around us.
Joscha Bach (2:29:04.820)
I generally have a very foggy view of that world,
Lex Fridman (2:29:10.060)
but I tend to try to follow,
Lex Fridman (2:29:12.060)
and I think society should in some degree
Lex Fridman (2:29:13.900)
follow the gradient of love,
Joscha Bach (2:29:16.340)
increasing the amount of love in the world.
Lex Fridman (2:29:18.940)
And whenever I see different policies
Joscha Bach (2:29:21.100)
or algorithms or ideas that are not doing so,
Lex Fridman (2:29:24.460)
obviously, that's the ones that kind of resist.
Lex Fridman (2:29:27.900)
So the thing that terrifies me about this notion
Lex Fridman (2:29:30.740)
is I think that German fascism was driven by love.
Joscha Bach (2:29:35.660)
It was just a very selective love.
Lex Fridman (2:29:37.820)
It was a love that basically...
Joscha Bach (2:29:39.140)
Now you're just manipulating.
Lex Fridman (2:29:40.460)
I mean, that's, you have to be very careful.
Joscha Bach (2:29:45.460)
You're talking to the wrong person in this way about love.
Lex Fridman (2:29:50.580)
So let's talk about what love is.
Lex Fridman (2:29:52.540)
And I think that love is the discovery of shared purpose.
Lex Fridman (2:29:55.980)
It's the recognition of the sacred in the other.
Lex Fridman (2:29:59.700)
And this enables non transactional interactions.
Lex Fridman (2:30:02.780)
But the size of the other that you include
Joscha Bach (2:30:07.740)
needs to be maximized.
Lex Fridman (2:30:09.740)
So it's basically appreciation,
Joscha Bach (2:30:14.700)
like deep appreciation of the world around you fully,
Lex Fridman (2:30:23.540)
including the people that are very different than you,
Joscha Bach (2:30:25.940)
people that disagree with you completely,
Lex Fridman (2:30:27.700)
including people, including living creatures
Joscha Bach (2:30:30.180)
outside of just people, including ideas.
Lex Fridman (2:30:33.460)
And it's like appreciation of the full mess of it.
Lex Fridman (2:30:36.580)
And also it has to do with like empathy,
Lex Fridman (2:30:40.380)
which is coupled with a lack of confidence
Lex Fridman (2:30:44.020)
and certainty of your own rightness.
Lex Fridman (2:30:47.140)
It's like a radical open mindedness to the way forward.
Joscha Bach (2:30:51.140)
I agree with every part of what you said.
Lex Fridman (2:30:53.460)
And now if you scale it up,
Lex Fridman (2:30:54.980)
what you recognize is that Lafist is in some sense,
Lex Fridman (2:30:58.540)
the service to next level agency,
Joscha Bach (2:31:01.380)
to the highest level agency that you can recognize.
Lex Fridman (2:31:04.220)
It could be for instance, life on earth or beyond that,
Joscha Bach (2:31:07.860)
where you could say intelligent complexity in the universe
Lex Fridman (2:31:11.620)
that you try to maximize in a certain way.
Lex Fridman (2:31:14.100)
But when you think it's true,
Lex Fridman (2:31:15.860)
it basically means a certain aesthetic.
Lex Fridman (2:31:18.980)
And there is not one possible aesthetic,
Lex Fridman (2:31:20.820)
there are many possible aesthetics.
Lex Fridman (2:31:22.660)
And once you project an aesthetic into the future,
Lex Fridman (2:31:25.420)
you can see that there are some which defect from it,
Joscha Bach (2:31:29.260)
which are in conflict with it,
Lex Fridman (2:31:30.900)
that are corrupt, that are evil.
Joscha Bach (2:31:33.860)
You and me would probably agree that Hitler was evil
Lex Fridman (2:31:37.100)
because the aesthetic of the world that he wanted
Joscha Bach (2:31:39.980)
is in conflict with the aesthetic of the world
Lex Fridman (2:31:41.940)
that you and me have in mind.
Lex Fridman (2:31:44.540)
And so they think that he destroyed,
Lex Fridman (2:31:48.500)
we want to keep them in the world.
Joscha Bach (2:31:50.660)
There's a kind of, there's kind of ways to deal,
Lex Fridman (2:31:55.220)
I mean, Hitler is an easier case,
Lex Fridman (2:31:56.660)
but perhaps he wasn't so easy in the 30s, right?
Lex Fridman (2:31:59.180)
To understand who is Hitler and who is not.
Joscha Bach (2:32:02.380)
No, it was just there was no consensus
Lex Fridman (2:32:04.580)
that the aesthetics that he had in mind were unacceptable.
Joscha Bach (2:32:07.500)
Yeah, I mean, it's difficult, love is complicated
Lex Fridman (2:32:12.900)
because you can't just be so open minded
Joscha Bach (2:32:17.300)
that you let evil walk into the door,
Lex Fridman (2:32:20.660)
but you can't be so self assured
Joscha Bach (2:32:24.420)
that you can always identify evil perfectly
Lex Fridman (2:32:29.580)
because that's what leads to Nazi Germany.
Joscha Bach (2:32:32.620)
Having a certainty of what is and wasn't evil,
Lex Fridman (2:32:34.860)
like always drawing lines of good versus evil.
Joscha Bach (2:32:38.660)
There seems to be, there has to be a dance
Lex Fridman (2:32:42.940)
between like hard stances extending up
Joscha Bach (2:32:49.900)
against what is wrong.
Lex Fridman (2:32:51.340)
And at the same time, empathy and open mindedness
Joscha Bach (2:32:55.420)
of towards not knowing what is right and wrong
Lex Fridman (2:32:59.580)
and like a dance between those.
Joscha Bach (2:33:01.420)
I found that when I watched the Miyazaki movies
Lex Fridman (2:33:03.620)
that there is nobody who captures my spirituality
Joscha Bach (2:33:06.060)
as well as he does.
Lex Fridman (2:33:07.940)
It's very interesting and just vicious, right?
Joscha Bach (2:33:10.620)
There is something going on in his movies
Lex Fridman (2:33:13.100)
that is very interesting.
Lex Fridman (2:33:14.140)
So for instance, Mononoke is discussing
Lex Fridman (2:33:17.140)
not only an answer to Disney's simplistic notion of Mowgli,
Joscha Bach (2:33:22.380)
the jungle boy was raised by wolves.
Lex Fridman (2:33:24.980)
And as soon as he sees people realizes that he's one of them
Lex Fridman (2:33:27.780)
and the way in which the moral life and nature
Lex Fridman (2:33:32.780)
is simplified and romanticized and turned into kitsch.
Joscha Bach (2:33:36.020)
It's disgusting in the Disney movie.
Lex Fridman (2:33:37.700)
And he answers to this, you see,
Joscha Bach (2:33:39.820)
he's replaced by Mononoke, this wolf girl
Lex Fridman (2:33:42.260)
who was raised by wolves and was fierce and dangerous
Lex Fridman (2:33:44.860)
and who cannot be socialized because she cannot be tamed.
Lex Fridman (2:33:48.780)
You cannot be part of human society.
Lex Fridman (2:33:50.460)
And you see human society,
Lex Fridman (2:33:51.900)
it's something that is very, very complicated.
Joscha Bach (2:33:53.700)
You see people extracting resources and destroying nature.
Lex Fridman (2:33:57.780)
But the purpose is not to be evil,
Lex Fridman (2:34:00.740)
but to be able to have a life that is free from,
Lex Fridman (2:34:04.740)
for instance, oppression and violence
Lex Fridman (2:34:07.140)
and to curb death and disease.
Lex Fridman (2:34:10.860)
And you basically see this conflict
Joscha Bach (2:34:13.260)
which cannot be resolved in a certain way.
Lex Fridman (2:34:15.180)
You see this moment when nature is turned into a garden
Lex Fridman (2:34:18.340)
and it loses most of what it actually is
Lex Fridman (2:34:20.980)
and humans no longer submitting to life and death
Lex Fridman (2:34:23.420)
and nature and to these questions, there is no easy answer.
Lex Fridman (2:34:26.820)
So it just turns it into something that is being observed
Joscha Bach (2:34:29.980)
as a journey that happens.
Lex Fridman (2:34:31.180)
And that happens with a certain degree of inevitability.
Lex Fridman (2:34:34.940)
And the nice thing about all his movies
Lex Fridman (2:34:37.100)
is there's a certain main character
Lex Fridman (2:34:38.740)
and it's the same in all movies.
Lex Fridman (2:34:41.260)
It's this little girl that is basically Heidi.
Lex Fridman (2:34:45.740)
And I suspect that happened because when he did field work
Lex Fridman (2:34:50.540)
for working on the Heidi movies back then,
Joscha Bach (2:34:53.020)
the Heidi animations, before he did his own movies,
Lex Fridman (2:34:55.700)
he traveled to Switzerland and South Eastern Europe
Lex Fridman (2:35:00.220)
and the Adriatic and so on and got an idea
Lex Fridman (2:35:03.220)
about a certain aesthetic and a certain way of life
Joscha Bach (2:35:05.340)
that informed his future thinking.
Lex Fridman (2:35:08.140)
And Heidi has a very interesting relationship
Joscha Bach (2:35:11.020)
to herself and to the world.
Lex Fridman (2:35:13.300)
There's nothing that she takes for herself.
Joscha Bach (2:35:15.940)
She's in a way fearless because she is committed
Lex Fridman (2:35:18.780)
to a service, to a greater whole.
Joscha Bach (2:35:20.860)
Basically, she is completely committed to serving God.
Lex Fridman (2:35:24.100)
And it's not an institutionalized God.
Joscha Bach (2:35:26.300)
It has nothing to do with the Roman Catholic Church
Lex Fridman (2:35:28.500)
or something like this.
Lex Fridman (2:35:30.420)
But in some sense, Heidi is an embodiment
Lex Fridman (2:35:32.660)
of the spirit of European Protestantism.
Joscha Bach (2:35:35.780)
It's this idea of a being that is completely perfect
Lex Fridman (2:35:38.780)
and pure.
Lex Fridman (2:35:40.180)
And it's not a feminist vision
Lex Fridman (2:35:42.060)
because she is not a girl boss or something like this.
Joscha Bach (2:35:48.620)
She is the justification for the men in the audience
Lex Fridman (2:35:52.460)
to protect her, to build a civilization around her
Joscha Bach (2:35:54.780)
that makes her possible.
Lex Fridman (2:35:56.580)
So she is not just the sacrifice of Jesus
Joscha Bach (2:35:59.260)
who is innocent and therefore nailed to the cross.
Lex Fridman (2:36:02.740)
She is not being sacrificed.
Joscha Bach (2:36:04.060)
She is being protected by everybody around her
Lex Fridman (2:36:07.020)
who recognizes that she is sacred.
Lex Fridman (2:36:08.620)
And there are enough around her to see that.
Lex Fridman (2:36:12.060)
So this is a very interesting perspective.
Joscha Bach (2:36:14.020)
There's a certain notion of innocence.
Lex Fridman (2:36:16.340)
And this notion of innocence is not universal.
Joscha Bach (2:36:18.500)
It's not in all cultures.
Lex Fridman (2:36:20.140)
Hitler wasn't innocent.
Joscha Bach (2:36:21.500)
His idea of Germany was not that there is an innocence
Lex Fridman (2:36:25.620)
that is being protected.
Joscha Bach (2:36:26.900)
There was a predator that was going to triumph.
Lex Fridman (2:36:29.700)
And it's also something that is not at the core
Joscha Bach (2:36:31.420)
of every religion.
Lex Fridman (2:36:32.260)
There are many religions which don't care about innocence.
Joscha Bach (2:36:34.860)
They might care about increasing the status of something.
Lex Fridman (2:36:41.020)
And that's a very interesting notion that is quite unique
Lex Fridman (2:36:44.980)
and not claiming it's the optimal one.
Lex Fridman (2:36:47.620)
It's just a particular kind of aesthetic
Joscha Bach (2:36:49.940)
which I think makes Miyazaki
Lex Fridman (2:36:51.780)
into the most relevant Protestant philosopher today.
Lex Fridman (2:36:55.500)
And you're saying in terms of all the ways
Lex Fridman (2:36:59.780)
that a society can operate perhaps the preservation
Joscha Bach (2:37:02.020)
of innocence might be one of the best.
Lex Fridman (2:37:07.140)
No, it's just my aesthetic.
Lex Fridman (2:37:09.780)
So it's a particular way in which I feel
Lex Fridman (2:37:13.620)
that I relate to the world that is natural
Joscha Bach (2:37:15.420)
to my own socialization.
Lex Fridman (2:37:16.700)
And maybe it's not an accident
Joscha Bach (2:37:18.300)
that I have cultural roots in Europe
Lex Fridman (2:37:22.380)
in a particular world.
Lex Fridman (2:37:23.380)
And so maybe it's a natural convergence point
Lex Fridman (2:37:26.620)
and it's not something that you will find
Joscha Bach (2:37:28.500)
in all other times in history.
Lex Fridman (2:37:30.980)
So I'd like to ask you about Solzhenitsyn
Lex Fridman (2:37:33.980)
and our individual role as ants in this very large society.
Lex Fridman (2:37:39.460)
So he says that some version of the line
Joscha Bach (2:37:42.060)
between good and evil runs to the heart of every man.
Lex Fridman (2:37:44.700)
Do you think all of us are capable of good and evil?
Joscha Bach (2:37:47.340)
Like what's our role in this play
Lex Fridman (2:37:53.500)
in this game we're all playing?
Lex Fridman (2:37:55.580)
Is all of us capable to play any role?
Lex Fridman (2:37:59.020)
Like, is there an ultimate responsibility
Joscha Bach (2:38:00.980)
to you mentioned maintaining innocence
Lex Fridman (2:38:04.300)
or whatever the highest ideal for a society you want
Lex Fridman (2:38:09.140)
are all of us capable of living up to that?
Lex Fridman (2:38:11.540)
And that's our responsibility
Joscha Bach (2:38:13.340)
or is there significant limitations
Lex Fridman (2:38:15.900)
to what we're able to do in terms of good and evil?
Lex Fridman (2:38:21.340)
So there is a certain way if you are not terrible,
Lex Fridman (2:38:24.060)
if you are committed to some kind of civilizational agency,
Joscha Bach (2:38:29.460)
a next level agent that you are serving,
Lex Fridman (2:38:31.140)
some kind of transcendent principle.
Joscha Bach (2:38:34.260)
In the eyes of that transcendental principle,
Lex Fridman (2:38:36.300)
you are able to discern good from evil.
Joscha Bach (2:38:38.060)
Otherwise you cannot,
Lex Fridman (2:38:39.020)
otherwise you have just individual aesthetics.
Joscha Bach (2:38:41.660)
The cat that is torturing a mouse is not evil
Lex Fridman (2:38:44.060)
because the cat does not envision
Joscha Bach (2:38:46.340)
or no part of the world of the cat is envisioning a world
Lex Fridman (2:38:50.660)
where there is no violence and nobody is suffering.
Joscha Bach (2:38:53.740)
If you have an aesthetic where you want
Lex Fridman (2:38:55.500)
to protect innocence,
Joscha Bach (2:38:56.940)
then torturing somebody needlessly is evil,
Lex Fridman (2:39:00.900)
but only then.
Joscha Bach (2:39:02.740)
No, but within, I guess the question is within the aesthetic,
Lex Fridman (2:39:05.660)
like within your sense of what is good and evil,
Joscha Bach (2:39:10.260)
are we still, it seems like we're still able
Lex Fridman (2:39:14.460)
to commit evil.
Joscha Bach (2:39:17.140)
Yes, so basically if you are committing
Lex Fridman (2:39:19.340)
to this next level agent,
Lex Fridman (2:39:20.820)
you are not necessarily are this next level agent, right?
Lex Fridman (2:39:23.580)
You are a part of it.
Joscha Bach (2:39:24.420)
You have a relationship to it,
Lex Fridman (2:39:26.020)
like the cell does to its organism, its hyperorganism.
Lex Fridman (2:39:29.700)
And it only exists to the degree
Lex Fridman (2:39:31.340)
that it's being implemented by you and others.
Lex Fridman (2:39:34.580)
And that means that you're not completely fully serving it.
Lex Fridman (2:39:38.540)
You have freedom in what you decide,
Joscha Bach (2:39:40.340)
whether you are acting on your impulses
Lex Fridman (2:39:42.100)
and local incentives and your farewell impulses,
Lex Fridman (2:39:44.500)
so to speak, or whether you're committing to it.
Lex Fridman (2:39:47.140)
And what you perceive then is a tension
Joscha Bach (2:39:49.980)
between what you would be doing with respect
Lex Fridman (2:39:53.100)
to the thing that you recognize as the sacred, if you do,
Lex Fridman (2:39:57.300)
and what you're actually doing.
Lex Fridman (2:39:58.820)
And this is the line between good and evil,
Joscha Bach (2:40:01.460)
right where you see, oh, I'm here acting
Lex Fridman (2:40:03.100)
on my local incentives or impulses,
Lex Fridman (2:40:05.700)
and here I'm acting on what I consider to be sacred.
Lex Fridman (2:40:08.100)
And there's a tension between those.
Lex Fridman (2:40:09.780)
And this is the line between good and evil
Lex Fridman (2:40:11.940)
that might run through your heart.
Lex Fridman (2:40:14.380)
And if you don't have that,
Lex Fridman (2:40:15.700)
if you don't have this relationship
Joscha Bach (2:40:17.180)
to a transcendental agent,
Lex Fridman (2:40:18.660)
you could call this relationship
Lex Fridman (2:40:19.980)
to the next level agent soul, right?
Lex Fridman (2:40:21.700)
It's not a thing.
Joscha Bach (2:40:22.540)
It's not an immortal thing that is intrinsically valuable.
Lex Fridman (2:40:25.780)
It's a certain kind of relationship
Joscha Bach (2:40:27.460)
that you project to understand what's happening.
Lex Fridman (2:40:29.580)
Somebody is serving this transcendental sacredness
Joscha Bach (2:40:31.900)
or they're not.
Lex Fridman (2:40:33.220)
If you don't have a soul, you cannot be evil.
Joscha Bach (2:40:35.860)
You're just a complex natural phenomenon.
Lex Fridman (2:40:39.620)
So if you look at life, like starting today
Joscha Bach (2:40:42.140)
or starting tomorrow, when we leave here today,
Lex Fridman (2:40:46.020)
there's a bunch of trajectories
Joscha Bach (2:40:48.180)
that you can take through life, maybe countless.
Lex Fridman (2:40:53.780)
Do you think some of these trajectories,
Joscha Bach (2:40:57.300)
in your own conception of yourself,
Lex Fridman (2:40:59.700)
some of those trajectories are the ideal life,
Joscha Bach (2:41:04.220)
a life that if you were to be the hero of your life story,
Lex Fridman (2:41:09.620)
you would want to be?
Lex Fridman (2:41:10.860)
Like, is there some Josh or Bhakti you're striving to be?
Lex Fridman (2:41:14.500)
Like, this is the question I ask myself
Joscha Bach (2:41:15.980)
as an individual trying to make a better world
Lex Fridman (2:41:20.260)
in the best way that I could conceive of.
Lex Fridman (2:41:22.540)
What is my responsibility there?
Lex Fridman (2:41:24.660)
And how much am I responsible for the failure to do so?
Joscha Bach (2:41:28.260)
Because I'm lazy and incompetent too often.
Lex Fridman (2:41:33.260)
In my own perception.
Joscha Bach (2:41:35.740)
In my own worldview, I'm not very important.
Lex Fridman (2:41:38.260)
So it's, I don't have place for me as a hero
Joscha Bach (2:41:41.540)
in my own world.
Lex Fridman (2:41:43.460)
I'm trying to do the best that I can,
Joscha Bach (2:41:45.980)
which is often not very good.
Lex Fridman (2:41:48.060)
And so it's not important for me to have status
Joscha Bach (2:41:52.820)
or to be seen in a particular way.
Lex Fridman (2:41:55.500)
It's helpful if others can see me
Joscha Bach (2:41:57.380)
or a few people can see me that can be my friends.
Lex Fridman (2:41:59.780)
No, sorry, I want to clarify,
Joscha Bach (2:42:01.460)
the hero I didn't mean status or perception
Lex Fridman (2:42:05.220)
or like some kind of marketing thing,
Lex Fridman (2:42:09.660)
but more in private, in the quiet of your own mind.
Lex Fridman (2:42:14.060)
Is there the kind of man you want to be
Lex Fridman (2:42:16.940)
and would consider it a failure if you don't become that?
Lex Fridman (2:42:20.460)
That's what I meant by hero.
Joscha Bach (2:42:21.940)
Yeah, not really.
Lex Fridman (2:42:23.300)
I don't perceive myself as having such an identity.
Lex Fridman (2:42:26.140)
And it's also sometimes frustrating,
Lex Fridman (2:42:32.340)
but it's basically a lack of having this notion
Joscha Bach (2:42:37.940)
of father that I need to be emulating.
Lex Fridman (2:42:44.020)
It's interesting.
Joscha Bach (2:42:44.980)
I mean, it's the leaf floating down the river.
Lex Fridman (2:42:48.660)
I worry that...
Joscha Bach (2:42:50.220)
Sometimes it's more like being the river.
Lex Fridman (2:42:59.020)
I'm just a fat frog sitting on a leaf
Joscha Bach (2:43:02.740)
on a dirty, muddy lake.
Lex Fridman (2:43:06.620)
I wish I was waiting for a princess to kiss me.
Joscha Bach (2:43:13.540)
Or the other way, I forgot which way it goes.
Lex Fridman (2:43:15.780)
Somebody kisses somebody.
Joscha Bach (2:43:17.180)
I can ask you, I don't know if you know
Lex Fridman (2:43:20.420)
who Michael Malice is,
Lex Fridman (2:43:21.700)
but in terms of constructing since systems of incentives,
Lex Fridman (2:43:27.060)
it's interesting to ask.
Joscha Bach (2:43:29.540)
I don't think I've talked to you about this before.
Lex Fridman (2:43:33.060)
Malice espouses anarchism.
Lex Fridman (2:43:35.700)
So he sees all government as fundamentally
Lex Fridman (2:43:40.660)
getting in the way or even being destructive
Joscha Bach (2:43:42.940)
to collaborations between human beings thriving.
Lex Fridman (2:43:49.660)
What do you think?
Lex Fridman (2:43:50.500)
What's the role of government in a society that thrives?
Lex Fridman (2:43:56.900)
Is anarchism at all compelling to you as a system?
Lex Fridman (2:44:00.580)
So like not just small government,
Lex Fridman (2:44:02.980)
but no government at all.
Joscha Bach (2:44:05.940)
Yeah, I don't see how this would work.
Lex Fridman (2:44:09.860)
The government is an agent that imposes an offset
Joscha Bach (2:44:12.700)
on your reward function, on your payout metrics.
Lex Fridman (2:44:15.580)
So your behavior becomes compatible with the common good.
Lex Fridman (2:44:20.860)
So the argument there is that you can have collectives
Lex Fridman (2:44:25.620)
like governing organizations, but not government,
Joscha Bach (2:44:28.620)
like where you're born in a particular set of land
Lex Fridman (2:44:32.540)
and therefore you must follow this rule or else.
Joscha Bach (2:44:38.420)
You're forced by what they call violence
Lex Fridman (2:44:41.820)
because there's an implied violence here.
Lex Fridman (2:44:44.900)
So the key aspect of government is it protects you
Lex Fridman (2:44:52.020)
from the rest of the world with an army and with police.
Lex Fridman (2:44:56.700)
So it has a monopoly on violence.
Lex Fridman (2:45:00.020)
It's the only one that's able to do violence.
Lex Fridman (2:45:02.060)
So there are many forms of government,
Lex Fridman (2:45:03.540)
not all governments do that.
Lex Fridman (2:45:05.020)
But we find that in successful countries,
Lex Fridman (2:45:09.660)
the government has a monopoly on violence.
Lex Fridman (2:45:12.740)
And that means that you cannot get ahead
Lex Fridman (2:45:15.700)
by starting your own army because the government
Joscha Bach (2:45:17.740)
will come down on you and destroy you
Lex Fridman (2:45:19.340)
if you try to do that.
Lex Fridman (2:45:20.940)
And in countries where you can build your own army
Lex Fridman (2:45:23.260)
and get away with it, some people will do it.
Lex Fridman (2:45:25.700)
And these countries is what we call failed countries
Lex Fridman (2:45:28.580)
in a way.
Lex Fridman (2:45:30.060)
And if you don't want to have violence,
Lex Fridman (2:45:33.500)
the point is not to appeal to the moral intentions of people
Joscha Bach (2:45:36.860)
because some people will use strategies
Lex Fridman (2:45:39.180)
if they get ahead with them that feel a particular kind
Joscha Bach (2:45:41.820)
of ecological niche.
Lex Fridman (2:45:42.740)
So you need to destroy that ecological niche.
Lex Fridman (2:45:45.260)
And if effective government has a monopoly on violence,
Lex Fridman (2:45:50.060)
it can create a world where nobody is able to use violence
Lex Fridman (2:45:53.460)
and get ahead.
Lex Fridman (2:45:54.820)
So you want to use that monopoly on violence,
Joscha Bach (2:45:57.060)
not to exert violence, but to make violence impossible,
Lex Fridman (2:46:00.100)
to raise the cost of violence.
Lex Fridman (2:46:02.140)
So people need to get ahead with nonviolent means.
Lex Fridman (2:46:06.100)
So the idea is that you might be able to achieve that
Joscha Bach (2:46:09.300)
in an anarchist state with companies.
Lex Fridman (2:46:12.220)
So with the forces of capitalism is create security companies
Joscha Bach (2:46:18.260)
where the one that's most ethically sound rises to the top.
Lex Fridman (2:46:21.980)
Basically, it would be a much better representative
Joscha Bach (2:46:24.220)
of the people because there is a less sort of stickiness
Lex Fridman (2:46:29.220)
to the big military force sticking around
Joscha Bach (2:46:33.220)
even though it's long overlived, outlived.
Lex Fridman (2:46:36.420)
So you have groups of militants that are hopefully
Joscha Bach (2:46:40.060)
efficiently organized because otherwise they're going
Lex Fridman (2:46:41.940)
to lose against the other groups of militants
Lex Fridman (2:46:44.580)
and they are coordinating themselves with the rest
Lex Fridman (2:46:47.060)
of society until they are having a monopoly on violence.
Lex Fridman (2:46:51.220)
How is that different from a government?
Lex Fridman (2:46:53.940)
So it's basically converging to the same thing.
Lex Fridman (2:46:56.220)
So I was trying to argue with Malice,
Lex Fridman (2:47:00.020)
I feel like it always converges towards government at scale,
Lex Fridman (2:47:03.060)
but I think the idea is you can have a lot of collectives
Lex Fridman (2:47:06.100)
that are, you basically never let anything scale too big.
Lex Fridman (2:47:11.820)
So one of the problems with governments is it gets too big
Lex Fridman (2:47:15.460)
in terms of like the size of the group
Joscha Bach (2:47:19.820)
over which it has control.
Lex Fridman (2:47:23.980)
My sense is that would happen anyway.
Lex Fridman (2:47:27.060)
So a successful company like Amazon or Facebook,
Lex Fridman (2:47:30.660)
I mean, it starts forming a monopoly
Joscha Bach (2:47:33.060)
over the entire populations,
Lex Fridman (2:47:36.060)
not over just the hundreds of millions,
Lex Fridman (2:47:37.900)
but billions of people.
Lex Fridman (2:47:39.340)
So I don't know, but there is something
Joscha Bach (2:47:43.540)
about the abuses of power the government can have
Lex Fridman (2:47:46.060)
when it has a monopoly on violence, right?
Lex Fridman (2:47:49.020)
And so that's a tension there, but...
Lex Fridman (2:47:53.020)
So the question is how can you set the incentives
Lex Fridman (2:47:55.180)
for government correctly?
Lex Fridman (2:47:56.420)
And this mostly applies at the highest levels of government
Lex Fridman (2:47:59.940)
and because we haven't found a way to set them correctly,
Lex Fridman (2:48:02.940)
we made the highest levels of government relatively weak.
Lex Fridman (2:48:06.300)
And this is, I think, part of the reason
Lex Fridman (2:48:08.580)
why we had difficulty to coordinate the pandemic response
Lex Fridman (2:48:12.260)
and China didn't have that much difficulty.
Lex Fridman (2:48:14.940)
And there is, of course, a much higher risk
Joscha Bach (2:48:17.500)
of the abuse of power that exists in China
Lex Fridman (2:48:19.980)
because the power is largely unchecked.
Lex Fridman (2:48:22.740)
And that's basically what happens
Lex Fridman (2:48:25.260)
in the next generation, for instance.
Joscha Bach (2:48:26.540)
Imagine that we would agree
Lex Fridman (2:48:28.380)
that the current government of China is largely correct
Lex Fridman (2:48:30.460)
and benevolent, and maybe we don't agree on this,
Lex Fridman (2:48:33.180)
but if we did, how can we make sure
Lex Fridman (2:48:36.100)
that this stays like this?
Lex Fridman (2:48:37.540)
And if you don't have checks and balances,
Joscha Bach (2:48:40.300)
division of power, it's hard to achieve.
Lex Fridman (2:48:42.980)
You don't have a solution for that problem.
Lex Fridman (2:48:45.300)
But the abolishment of government
Lex Fridman (2:48:47.420)
basically would remove the control structure.
Joscha Bach (2:48:49.540)
From a cybernetic perspective,
Lex Fridman (2:48:51.540)
there is an optimal point in the system
Lex Fridman (2:48:54.740)
that the regulation should be happening, right?
Lex Fridman (2:48:56.460)
That you can measure the current incentives
Lex Fridman (2:48:59.780)
and the regulator would be properly incentivized
Lex Fridman (2:49:01.940)
to make the right decisions
Lex Fridman (2:49:03.740)
and change the payout metrics of everything below it
Lex Fridman (2:49:06.340)
in such a way that the local prisoners dilemmas
Lex Fridman (2:49:08.620)
get resolved, right?
Lex Fridman (2:49:09.900)
You cannot resolve the prisoners dilemma
Joscha Bach (2:49:12.060)
without some kind of eternal control
Lex Fridman (2:49:14.900)
that emulates an infinite game in a way.
Joscha Bach (2:49:19.060)
Yeah, I mean, there's a sense in which
Lex Fridman (2:49:22.380)
it seems like the reason government,
Joscha Bach (2:49:24.940)
the parts of government that don't work well currently
Lex Fridman (2:49:27.780)
is because there's not good mechanisms
Joscha Bach (2:49:34.020)
through which to interact,
Lex Fridman (2:49:35.380)
for the citizenry to interact with government
Joscha Bach (2:49:37.300)
is basically it hasn't caught up in terms of technology.
Lex Fridman (2:49:41.500)
And I think once you integrate
Joscha Bach (2:49:43.860)
some of the digital revolution
Lex Fridman (2:49:46.100)
of being able to have a lot of access to data,
Joscha Bach (2:49:48.420)
be able to vote on different ideas at a local level,
Lex Fridman (2:49:52.060)
at all levels, at the optimal level
Joscha Bach (2:49:54.780)
like you're saying that can resolve the prisoner dilemmas
Lex Fridman (2:49:58.580)
and to integrate AI to help you automate things
Joscha Bach (2:50:01.380)
that don't require the human ingenuity.
Lex Fridman (2:50:07.460)
I feel like that's where government could operate that well
Lex Fridman (2:50:10.340)
and can also break apart the inefficient bureaucracies
Lex Fridman (2:50:14.620)
if needed.
Joscha Bach (2:50:15.460)
There'll be a strong incentive to be efficient and successful.
Lex Fridman (2:50:20.620)
So out human history, we see an evolution
Lex Fridman (2:50:23.020)
and evolutionary competition of modes of government
Lex Fridman (2:50:25.660)
and of individual governments is in these modes.
Lex Fridman (2:50:28.180)
And every nation state in some sense
Lex Fridman (2:50:29.900)
is some kind of organism that has found different solutions
Joscha Bach (2:50:33.180)
for the problem of government.
Lex Fridman (2:50:34.980)
And you could look at all these different models
Lex Fridman (2:50:37.500)
and the different scales at which it exists
Lex Fridman (2:50:39.420)
as empirical attempts to validate the idea
Joscha Bach (2:50:43.020)
of how to build a better government.
Lex Fridman (2:50:45.780)
And I suspect that the idea of anarchism
Joscha Bach (2:50:49.180)
similar to the idea of communism
Lex Fridman (2:50:51.900)
is the result of being disenchanted
Joscha Bach (2:50:54.860)
with the ugliness of the real existing solutions
Lex Fridman (2:50:57.340)
and the attempt to get to an utopia.
Lex Fridman (2:51:00.980)
And I suspect that communism originally was not a utopia.
Lex Fridman (2:51:04.540)
I think that in the same way as original Christianity,
Joscha Bach (2:51:07.580)
it had a particular kind of vision.
Lex Fridman (2:51:10.020)
And this vision is a society,
Joscha Bach (2:51:12.540)
a mode of organization within the society
Lex Fridman (2:51:15.300)
in which humans can coexist at scale without coercion.
Lex Fridman (2:51:20.300)
In the same way as we do in a healthy family, right?
Lex Fridman (2:51:23.660)
In a good family,
Joscha Bach (2:51:24.500)
you don't terrorize each other into compliance,
Lex Fridman (2:51:28.060)
but you understand what everybody needs
Lex Fridman (2:51:30.300)
and what everybody is able to contribute
Lex Fridman (2:51:32.220)
and what the intended future of the whole thing is.
Lex Fridman (2:51:35.260)
And everybody coordinates their behavior in the right way
Lex Fridman (2:51:38.380)
and informs each other about how to do this.
Lex Fridman (2:51:40.820)
And all the interactions that happen
Lex Fridman (2:51:42.540)
are instrumental to making that happen, right?
Lex Fridman (2:51:45.820)
Could this happen at scale?
Lex Fridman (2:51:47.260)
And I think this is the idea of communism.
Joscha Bach (2:51:49.180)
Communism is opposed to the idea
Lex Fridman (2:51:51.420)
that we need economic terror
Joscha Bach (2:51:53.340)
or other forms of terror to make that happen.
Lex Fridman (2:51:55.700)
But in practice, what happened
Joscha Bach (2:51:56.860)
is that the proto communist countries,
Lex Fridman (2:51:59.300)
the real existing socialism,
Joscha Bach (2:52:01.140)
replaced a part of the economic terror with moral terror,
Lex Fridman (2:52:04.620)
right?
Lex Fridman (2:52:05.460)
So we were told to do the right thing for moral reasons.
Lex Fridman (2:52:07.540)
And of course it didn't really work
Lex Fridman (2:52:09.180)
and the economy eventually collapsed.
Lex Fridman (2:52:11.620)
And the moral terror had actual real cost, right?
Joscha Bach (2:52:14.540)
People were in prison
Lex Fridman (2:52:15.820)
because they were morally noncompliant.
Lex Fridman (2:52:17.820)
And the other thing is that the idea of communism
Lex Fridman (2:52:22.900)
became a utopia.
Lex Fridman (2:52:24.060)
So it basically was projected into the afterlife.
Lex Fridman (2:52:26.140)
We were told in my childhood
Joscha Bach (2:52:28.660)
that communism was a hypothetical society
Lex Fridman (2:52:31.220)
to which we were in a permanent revolution
Joscha Bach (2:52:33.380)
that justified everything
Lex Fridman (2:52:34.660)
that was presently wrong with society morally.
Lex Fridman (2:52:37.500)
But it was something that our grandchildren
Lex Fridman (2:52:39.540)
probably would not ever see
Joscha Bach (2:52:41.140)
because it was too ideal and too far in the future
Lex Fridman (2:52:43.860)
to make it happen right now.
Lex Fridman (2:52:44.980)
And people were just not there yet morally.
Lex Fridman (2:52:47.300)
And the same thing happened with Christianity, right?
Joscha Bach (2:52:50.300)
This notion of heaven was mythologized
Lex Fridman (2:52:52.820)
and projected into an afterlife.
Lex Fridman (2:52:54.380)
And I think this was just the idea of God's kingdom
Lex Fridman (2:52:56.900)
of this world in which we instantiate
Joscha Bach (2:52:59.180)
the next level transcendental agent in the perfect form.
Lex Fridman (2:53:01.980)
So everything goes smoothly and without violence
Lex Fridman (2:53:04.660)
and without conflict and without this human messiness
Lex Fridman (2:53:07.980)
on this economic messiness and the terror and coercion
Joscha Bach (2:53:11.340)
that existed in the present societies.
Lex Fridman (2:53:13.820)
And the idea of that the humans can exist at some point
Joscha Bach (2:53:16.980)
exist at scale in a harmonious way and noncoercively
Lex Fridman (2:53:20.180)
is untested, right?
Joscha Bach (2:53:21.700)
A lot of people tested it
Lex Fridman (2:53:23.140)
but didn't get it to work so far.
Lex Fridman (2:53:25.340)
And the utopia is a world in where you get
Lex Fridman (2:53:27.740)
all the good things without any of the bad things.
Lex Fridman (2:53:30.900)
And you are, I think very susceptible to believe in utopias
Lex Fridman (2:53:34.500)
when you are very young and don't understand
Joscha Bach (2:53:36.860)
that everything has to happen in causal patterns,
Lex Fridman (2:53:39.580)
that there's always feedback loops
Joscha Bach (2:53:40.940)
that ultimately are closed.
Lex Fridman (2:53:42.620)
There's nothing that just happens
Joscha Bach (2:53:44.020)
because it's good or bad.
Lex Fridman (2:53:45.460)
Good or bad don't exist in isolation.
Joscha Bach (2:53:47.220)
They only exist with respect to larger systems.
Lex Fridman (2:53:50.660)
So can you intuit why utopias fail as systems?
Lex Fridman (2:53:57.780)
So like having a utopia that's out there beyond the horizon
Lex Fridman (2:54:01.620)
is it because then,
Joscha Bach (2:54:04.980)
it's not only because it's impossible to achieve utopias
Lex Fridman (2:54:08.220)
but it's because what certain humans,
Joscha Bach (2:54:11.940)
certain small number of humans start to sort of greedily
Lex Fridman (2:54:20.220)
attain power and money and control and influence
Joscha Bach (2:54:25.540)
as they become,
Lex Fridman (2:54:28.980)
as they see the power in using this idea of a utopia
Joscha Bach (2:54:34.420)
for propaganda.
Lex Fridman (2:54:35.260)
It's a bit like saying, why is my garden not perfect?
Joscha Bach (2:54:37.260)
It's because some evil weeds are overgrowing it
Lex Fridman (2:54:39.780)
and they always do, right?
Lex Fridman (2:54:41.540)
But this is not how it works.
Lex Fridman (2:54:43.220)
A good garden is a system that is in balance
Lex Fridman (2:54:45.420)
and requires minimal interactions by the gardener.
Lex Fridman (2:54:48.620)
And so you need to create a system
Joscha Bach (2:54:51.860)
that is designed to self stabilize.
Lex Fridman (2:54:54.220)
And the design of social systems
Joscha Bach (2:54:55.860)
requires not just the implementation
Lex Fridman (2:54:57.500)
of the desired functionality,
Lex Fridman (2:54:58.780)
but the next level design, also in biological systems.
Lex Fridman (2:55:01.820)
You need to create a system that wants to converge
Joscha Bach (2:55:04.140)
to the intended function.
Lex Fridman (2:55:06.100)
And so instead of just creating an institution like the FDA
Joscha Bach (2:55:09.380)
that is performing a particular kind of role in society,
Lex Fridman (2:55:13.180)
you need to make sure that the FDA is actually driven
Joscha Bach (2:55:15.780)
by a system that wants to do this optimally,
Lex Fridman (2:55:18.180)
that is incentivized to do it optimally
Lex Fridman (2:55:19.860)
and then makes the performance that is actually enacted
Lex Fridman (2:55:23.340)
in every generation instrumental to that thing,
Lex Fridman (2:55:26.220)
that actual goal, right?
Lex Fridman (2:55:27.660)
And that is much harder to design and to achieve.
Joscha Bach (2:55:30.100)
See if the design a system where,
Lex Fridman (2:55:32.460)
and listen communism also was quote unquote incentivized
Joscha Bach (2:55:36.940)
to be a feedback loop system that achieves that utopia.
Lex Fridman (2:55:43.500)
It's just, it wasn't working given human nature.
Joscha Bach (2:55:45.820)
The incentives were not correct given human nature.
Lex Fridman (2:55:47.980)
How do you incentivize people
Joscha Bach (2:55:50.460)
when they are getting coal off the ground
Lex Fridman (2:55:52.340)
to work as hard as possible?
Joscha Bach (2:55:53.900)
Because it's a terrible job
Lex Fridman (2:55:55.540)
and it's very bad for your health.
Lex Fridman (2:55:57.060)
And right, how do you do this?
Lex Fridman (2:55:59.540)
And you can give them prices and medals and status
Lex Fridman (2:56:03.620)
to some degree, right?
Lex Fridman (2:56:04.580)
There's only so much status to give for that.
Lex Fridman (2:56:06.900)
And most people will not fall for this, right?
Lex Fridman (2:56:09.340)
Or you can pay them and you probably have to pay them
Joscha Bach (2:56:12.940)
in an asymmetric way because if you pay everybody the same
Lex Fridman (2:56:15.700)
and you nationalize the coal mines,
Joscha Bach (2:56:19.100)
eventually people will figure out
Lex Fridman (2:56:20.620)
that they can game the system.
Joscha Bach (2:56:21.940)
Yes, so you're describing capitalism.
Lex Fridman (2:56:25.820)
So capitalism is the present solution to the system.
Lex Fridman (2:56:28.620)
And what we also noticed that I think that Marx was correct
Lex Fridman (2:56:32.140)
in saying that capitalism is prone to crisis,
Joscha Bach (2:56:35.140)
that capitalism is a system that in its dynamics
Lex Fridman (2:56:38.460)
is not convergent, but divergent.
Joscha Bach (2:56:40.780)
It's not a stable system.
Lex Fridman (2:56:42.860)
And that eventually it produces an enormous potential
Joscha Bach (2:56:47.380)
for productivity, but it also is systematically
Lex Fridman (2:56:50.820)
misallocating resources.
Lex Fridman (2:56:52.140)
So a lot of people cannot participate
Lex Fridman (2:56:54.700)
in the production and consumption anymore, right?
Lex Fridman (2:56:57.300)
And this is what we observed.
Lex Fridman (2:56:58.420)
We observed that the middle class in the US is tiny.
Joscha Bach (2:57:01.460)
It's a lot of people think that they're middle class,
Lex Fridman (2:57:05.500)
but if you are still flying economy,
Lex Fridman (2:57:07.460)
you're not middle class, right?
Lex Fridman (2:57:11.660)
Every class is a magnitude smaller than the previous class.
Lex Fridman (2:57:14.700)
And I think about classes is really like airline class.
Lex Fridman (2:57:23.060)
I like class.
Joscha Bach (2:57:25.700)
A lot of people are economy class, business class,
Lex Fridman (2:57:28.580)
and very few are first class and some are budget.
Joscha Bach (2:57:30.900)
I mean, some, I understand.
Lex Fridman (2:57:32.860)
I think there's, yeah, maybe some people,
Joscha Bach (2:57:36.940)
probably I would push back
Lex Fridman (2:57:38.140)
against that definition of the middle class.
Joscha Bach (2:57:39.660)
It does feel like the middle class is pretty large,
Lex Fridman (2:57:41.460)
but yes, there's a discrepancy in terms of wealth.
Lex Fridman (2:57:45.740)
So if you think about in terms of the productivity
Lex Fridman (2:57:48.620)
that our society could have,
Lex Fridman (2:57:50.900)
there is no reason for anybody to fly economy, right?
Lex Fridman (2:57:53.980)
We would be able to let everybody travel in style.
Joscha Bach (2:57:57.940)
Well, but also some people like to be frugal
Lex Fridman (2:58:00.220)
even when they're billionaires, okay?
Lex Fridman (2:58:01.620)
So like that, let's take that into account.
Lex Fridman (2:58:04.580)
I mean, we probably don't need to be a traveling lavish,
Lex Fridman (2:58:07.260)
but you also don't need to be tortured, right?
Lex Fridman (2:58:09.780)
There is a difference between frugal
Lex Fridman (2:58:11.820)
and subjecting yourself to torture.
Lex Fridman (2:58:14.140)
Listen, I love economy.
Joscha Bach (2:58:15.220)
I don't understand why you're comparing
Lex Fridman (2:58:16.780)
a fly economy to torture.
Joscha Bach (2:58:19.420)
I don't, although the fight here,
Lex Fridman (2:58:22.500)
there's two crying babies next to me.
Lex Fridman (2:58:24.380)
So that, but that has nothing to do with economy.
Lex Fridman (2:58:26.460)
It has to do with crying babies.
Joscha Bach (2:58:28.540)
They're very cute though.
Lex Fridman (2:58:29.380)
So they kind of.
Joscha Bach (2:58:30.220)
Yeah, I have two kids
Lex Fridman (2:58:31.260)
and sometimes I have to go back to visit the grandparents.
Lex Fridman (2:58:35.020)
And that means going from the west coast to Germany
Lex Fridman (2:58:41.300)
and that's a long flight.
Joscha Bach (2:58:42.700)
Is it true that, so when you're a father,
Lex Fridman (2:58:45.300)
you grow immune to the crying and all that kind of stuff,
Joscha Bach (2:58:48.540)
like the, because like me just not having kids,
Lex Fridman (2:58:52.260)
it can be other people's kids can be quite annoying
Joscha Bach (2:58:54.620)
when they're crying and screaming
Lex Fridman (2:58:55.820)
and all that kind of stuff.
Joscha Bach (2:58:57.220)
When you have children and you are wired up
Lex Fridman (2:58:59.540)
in the default natural way,
Joscha Bach (2:59:01.460)
you're lucky in this regard, you fall in love with them.
Lex Fridman (2:59:04.340)
And this falling in love with them means
Joscha Bach (2:59:06.980)
that you basically start to see the world through their eyes
Lex Fridman (2:59:10.180)
and you understand that in a given situation,
Joscha Bach (2:59:12.500)
they cannot do anything but being expressing despair.
Lex Fridman (2:59:17.740)
And so it becomes more differentiated.
Joscha Bach (2:59:19.700)
I noticed that for instance,
Lex Fridman (2:59:21.020)
my son is typically acting on a pure experience
Joscha Bach (2:59:25.940)
of what things are like right now
Lex Fridman (2:59:28.540)
and he has to do this right now.
Lex Fridman (2:59:30.380)
And you have this small child that is,
Lex Fridman (2:59:33.740)
when he was a baby and so on,
Joscha Bach (2:59:35.020)
where he was just immediately expressing what he felt.
Lex Fridman (2:59:37.580)
And if you cannot regulate this from the outside,
Lex Fridman (2:59:39.940)
there's no point to be upset about it, right?
Lex Fridman (2:59:42.260)
It's like dealing with weather or something like this.
Joscha Bach (2:59:45.060)
You all have to get through it
Lex Fridman (2:59:46.620)
and it's not easy for him either.
Lex Fridman (2:59:48.620)
But if you also have a daughter,
Lex Fridman (2:59:51.820)
maybe she is planning for that.
Joscha Bach (2:59:53.300)
Maybe she understands that she's sitting in the car
Lex Fridman (2:59:57.420)
behind you and she's screaming at the top of her lungs
Lex Fridman (2:59:59.860)
and you're almost doing an accident
Lex Fridman (30:02.980)
to the actual physical world
Lex Fridman (30:04.260)
as opposed to us being completely tricked?
Lex Fridman (30:07.660)
Is this is like where you have like Donald?
Joscha Bach (30:09.220)
It's not a trick.
Lex Fridman (30:10.060)
That's my point.
Joscha Bach (30:10.900)
It's not an illusion.
Lex Fridman (30:11.860)
It's a form of data compression.
Joscha Bach (30:13.700)
It's an attempt to deal with the dynamics
Lex Fridman (30:15.420)
of too many parts to count
Joscha Bach (30:16.940)
at the level at which we are entangled
Lex Fridman (30:18.700)
with the best model that you can find.
Joscha Bach (30:20.740)
Yeah, so we can act in that dream world
Lex Fridman (30:22.700)
and our actions have impact in the real world,
Joscha Bach (30:26.140)
in the physical world to which we don't have access.
Lex Fridman (30:28.620)
Yes, but it's basically like accepting the fact
Joscha Bach (30:31.860)
that the software that we live in,
Lex Fridman (30:33.180)
the dream that we live in is generated
Joscha Bach (30:35.380)
by something outside of this world that you and me are in.
Lex Fridman (30:38.060)
So is the software deterministic
Lex Fridman (30:40.060)
and do we not have any control?
Lex Fridman (30:42.260)
Do we have, so free will is having a conscious being.
Joscha Bach (30:49.620)
Free will is the monkey being able to steer the elephant.
Lex Fridman (30:55.300)
No, it's slightly different.
Joscha Bach (30:58.060)
Basically in the same way as you are modeling
Lex Fridman (31:00.540)
the water molecules in the ocean that engulf your feet
Joscha Bach (31:03.460)
when you are walking on the beach as waves
Lex Fridman (31:05.980)
and there are no waves,
Lex Fridman (31:07.380)
but only the atoms on more complicated stuff
Lex Fridman (31:09.780)
underneath the atoms and so on.
Lex Fridman (31:11.820)
And you know that, right?
Lex Fridman (31:14.020)
You would accept, yes,
Joscha Bach (31:15.300)
there is a certain abstraction that happens here.
Lex Fridman (31:17.660)
It's a simplification of what happens
Lex Fridman (31:19.420)
and the simplification that is designed
Lex Fridman (31:22.100)
in such a way that your brain can deal with it,
Joscha Bach (31:24.260)
temporarily and spatially in terms of resources
Lex Fridman (31:27.020)
and tuned for the predictive value.
Lex Fridman (31:28.740)
So you can predict with some accuracy
Lex Fridman (31:31.220)
whether your feet are going to get wet or not.
Lex Fridman (31:33.380)
But it's a really good interface and approximation.
Lex Fridman (31:37.620)
It says E equals MC squared is a good,
Joscha Bach (31:40.340)
equations are good approximation for,
Lex Fridman (31:43.100)
they're much better approximation.
Lex Fridman (31:45.780)
So to me, waves is a really nice approximation
Lex Fridman (31:49.380)
of what's all the complexity that's happening underneath.
Joscha Bach (31:51.940)
Basically it's a machine learning model
Lex Fridman (31:53.140)
that is constantly tuned to minimize surprises.
Lex Fridman (31:55.580)
So it basically tries to predict as well as it can
Lex Fridman (31:58.540)
what you're going to perceive next.
Lex Fridman (31:59.780)
Are we talking about, which is the machine learning?
Lex Fridman (32:02.620)
Our perception system or the dream world?
Joscha Bach (32:05.700)
The machine world, dream world is the result
Lex Fridman (32:08.260)
of the machine learning process of the perceptual system.
Joscha Bach (32:11.180)
That's doing the compression.
Lex Fridman (32:12.220)
Yes.
Lex Fridman (32:13.060)
And the model of you as an agent
Lex Fridman (32:15.860)
is not a different type of model or it's a different type,
Lex Fridman (32:19.460)
but not different as in its model like nature
Lex Fridman (32:23.180)
from the model of the ocean, right?
Joscha Bach (32:25.540)
Some things are oceans, some things are agents.
Lex Fridman (32:28.260)
And one of these agents is using your own control model,
Joscha Bach (32:31.620)
the output of your model,
Lex Fridman (32:32.780)
the things that you perceive yourself as doing.
Lex Fridman (32:36.260)
And that is you.
Lex Fridman (32:38.180)
What about the fact that when you're standing
Joscha Bach (32:44.100)
with the water on your feet and you're looking out
Lex Fridman (32:47.340)
into the vast open water of the ocean
Lex Fridman (32:51.980)
and then there's a beautiful sunset
Lex Fridman (32:54.460)
and the fact that it's beautiful
Lex Fridman (32:56.540)
and then maybe you have friends or a loved one with you
Lex Fridman (32:59.180)
and you feel love, what is that?
Lex Fridman (33:00.900)
As the dream world or what is that?
Lex Fridman (33:02.700)
Yes, it's all happening inside of the dream.
Joscha Bach (33:05.620)
Okay.
Lex Fridman (33:06.860)
But see, the word dream makes it seem like it's not real.
Joscha Bach (33:11.380)
No, of course it's not real.
Lex Fridman (33:14.940)
The physical universe is real,
Lex Fridman (33:16.540)
but the physical universe is incomprehensible
Lex Fridman (33:18.620)
and it doesn't have any feeling of realness.
Joscha Bach (33:21.060)
The feeling of realness that you experience
Lex Fridman (33:22.900)
gets attached to certain representations
Joscha Bach (33:25.420)
where your brain assesses,
Lex Fridman (33:26.620)
this is the best model of reality that I have.
Lex Fridman (33:28.500)
So the only thing that's real to you
Lex Fridman (33:30.820)
is the thing that's happening at the very base of reality.
Joscha Bach (33:34.740)
Yeah, for something to be real, it needs to be implemented.
Lex Fridman (33:40.020)
So the model that you have of reality
Joscha Bach (33:42.340)
is real in as far as it is a model.
Lex Fridman (33:45.300)
It's an appropriate description of the world
Joscha Bach (33:47.860)
to say that there are models that are being experienced,
Lex Fridman (33:51.500)
but the world that you experience
Joscha Bach (33:54.700)
is not necessarily implemented.
Lex Fridman (33:56.900)
There is a difference between a reality,
Joscha Bach (33:59.380)
a simulation and a simulacrum.
Lex Fridman (34:02.220)
The reality that we're talking about
Joscha Bach (34:04.460)
is something that fully emerges
Lex Fridman (34:06.060)
over a causally closed lowest layer.
Lex Fridman (34:08.620)
And the idea of physicalism is that we are in that layer,
Lex Fridman (34:11.300)
that basically our world emerges over that.
Joscha Bach (34:13.460)
Every alternative to physicalism is a simulation theory,
Lex Fridman (34:16.060)
which basically says that we are
Joscha Bach (34:17.980)
in some kind of simulation universe
Lex Fridman (34:19.460)
and the real world needs to be in a parent universe of that,
Lex Fridman (34:22.100)
where the actual causal structure is, right?
Lex Fridman (34:24.380)
And when you look at the ocean and your own mind,
Joscha Bach (34:27.660)
you are looking at a simulation
Lex Fridman (34:28.900)
that explains what you're going to see next.
Lex Fridman (34:31.460)
So we are living in a simulation.
Lex Fridman (34:32.860)
Yes, but a simulation generated by our own brains.
Joscha Bach (34:35.900)
Yeah.
Lex Fridman (34:36.740)
And this simulation is different from the physical reality
Joscha Bach (34:39.660)
because the causal structure that is being produced,
Lex Fridman (34:42.060)
what you are seeing is different
Joscha Bach (34:43.380)
from the causal structure of physics.
Lex Fridman (34:44.980)
But consistent.
Joscha Bach (34:46.780)
Hopefully, if not, then you are going to end up
Lex Fridman (34:49.780)
in some kind of institution
Joscha Bach (34:51.060)
where people will take care of you
Lex Fridman (34:52.220)
because your behavior will be inconsistent, right?
Joscha Bach (34:54.580)
Your behavior needs to work in such a way
Lex Fridman (34:57.220)
that it's interacting with an accurately predictive
Joscha Bach (35:00.140)
model of reality.
Lex Fridman (35:00.980)
And if your brain is unable to make your model
Joscha Bach (35:03.500)
of reality predictive, you will need help.
Lex Fridman (35:06.180)
So what do you think about Donald Hoffman's argument
Joscha Bach (35:10.260)
that it doesn't have to be consistent,
Lex Fridman (35:12.740)
the dream world to what he calls like the interface
Joscha Bach (35:17.820)
to the actual physical reality,
Lex Fridman (35:19.500)
where there could be evolution?
Joscha Bach (35:20.660)
I think he makes an evolutionary argument,
Lex Fridman (35:23.060)
which is like, it could be an evolutionary advantage
Joscha Bach (35:26.460)
to have the dream world drift away from physical reality.
Lex Fridman (35:30.940)
I think that only works if you have tenure.
Joscha Bach (35:32.780)
As long as you're still interacting with the ground tools,
Lex Fridman (35:35.260)
your model needs to be somewhat predictive.
Joscha Bach (35:38.980)
Well, in some sense, humans have achieved a kind of tenure
Lex Fridman (35:42.740)
in the animal kingdom.
Joscha Bach (35:45.100)
Yeah.
Lex Fridman (35:45.940)
And at some point we became too big to fail,
Lex Fridman (35:47.620)
so we became postmodernist.
Lex Fridman (35:51.420)
It all makes sense now.
Joscha Bach (35:52.660)
We can just change the version of reality that we like.
Lex Fridman (35:54.980)
Oh man.
Joscha Bach (35:56.500)
Okay.
Lex Fridman (35:57.380)
Yeah, but basically you can do magic.
Joscha Bach (36:00.220)
You can change your assessment of reality,
Lex Fridman (36:02.460)
but eventually reality is going to come bite you in the ass
Joscha Bach (36:05.580)
if it's not predictive.
Lex Fridman (36:06.820)
Do you have a sense of what is that base layer
Lex Fridman (36:11.220)
of physical reality?
Lex Fridman (36:12.580)
You have like, so you have these attempts
Joscha Bach (36:15.540)
at the theories of everything,
Lex Fridman (36:17.620)
the very, very small of like strength theory,
Joscha Bach (36:21.140)
or what Stephen Wolfram talks about with the hyper grass.
Lex Fridman (36:25.420)
These are these tiny, tiny, tiny, tiny objects.
Lex Fridman (36:28.540)
And then there is more like quantum mechanics
Lex Fridman (36:32.620)
that's talking about objects that are much larger,
Lex Fridman (36:34.900)
but still very, very, very tiny.
Lex Fridman (36:36.780)
Do you have a sense of where the tiniest thing is
Lex Fridman (36:40.060)
that is like at the lowest level?
Lex Fridman (36:42.900)
The turtle at the very bottom.
Lex Fridman (36:44.740)
Do you have a sense what that turtle is?
Lex Fridman (36:45.980)
I don't think that you can talk about where it is
Joscha Bach (36:48.580)
because space is emerging over the activity of these things.
Lex Fridman (36:51.580)
So space, the coordinates only exist
Joscha Bach (36:55.540)
in relation to the things, other things.
Lex Fridman (36:58.820)
And so you could, in some sense, abstract it into locations
Joscha Bach (37:01.740)
that can hold information and trajectories
Lex Fridman (37:04.300)
that the information can take
Joscha Bach (37:05.540)
between the different locations.
Lex Fridman (37:06.860)
And this is how we construct our notion of space.
Lex Fridman (37:10.380)
And physicists usually have a notion of space
Lex Fridman (37:14.100)
that is continuous.
Lex Fridman (37:15.700)
And this is a point where I tend to agree
Lex Fridman (37:19.140)
with people like Stephen Wolfram
Joscha Bach (37:20.980)
who are very skeptical of the geometric notions.
Lex Fridman (37:23.820)
I think that geometry is the dynamics
Joscha Bach (37:25.980)
of too many parts to count.
Lex Fridman (37:27.300)
And when there are no infinities,
Joscha Bach (37:30.820)
if there were two infinities,
Lex Fridman (37:32.500)
you would be running into contradictions,
Joscha Bach (37:34.220)
which is in some sense what Gödel and Turing discovered
Lex Fridman (37:37.780)
in response to Hilbert's call.
Lex Fridman (37:39.820)
So there are no infinities.
Lex Fridman (37:41.340)
There are no infinities.
Joscha Bach (37:42.180)
Infinities fake.
Lex Fridman (37:43.020)
There is unboundedness, but if you have a language
Joscha Bach (37:45.300)
that talks about infinity, at some point,
Lex Fridman (37:47.340)
the language is going to contradict itself,
Joscha Bach (37:49.580)
which means it's no longer valid.
Lex Fridman (37:51.660)
In order to deal with infinities and mathematics,
Joscha Bach (37:54.020)
you have to postulate the existence initially.
Lex Fridman (37:57.580)
You cannot construct the infinities.
Lex Fridman (37:59.180)
And that's an issue, right?
Lex Fridman (38:00.180)
You cannot build up an infinity from zero.
Lex Fridman (38:02.700)
But in practice, you never do this, right?
Lex Fridman (38:04.700)
When you perform calculations,
Joscha Bach (38:06.020)
you only look at the dynamics of too many parts to count.
Lex Fridman (38:09.060)
And usually these numbers are not that large.
Joscha Bach (38:13.420)
They're not Googles or something.
Lex Fridman (38:14.860)
The infinities that we are dealing with in our universe
Joscha Bach (38:18.540)
are mathematically speaking, relatively small integers.
Lex Fridman (38:23.180)
And still what we're looking at is dynamics
Joscha Bach (38:26.540)
where a trillion things behave similar
Lex Fridman (38:30.660)
to a hundred trillion things
Joscha Bach (38:32.660)
or something that is very, very large
Lex Fridman (38:37.860)
because they're converging.
Lex Fridman (38:39.260)
And these convergent dynamics, these operators,
Lex Fridman (38:41.380)
this is what we deal with when we are doing the geometry.
Joscha Bach (38:45.020)
Geometry is stuff where we can pretend that it's continuous
Lex Fridman (38:48.420)
because if we subdivide the space sufficiently fine grained,
Joscha Bach (38:54.100)
these things approach a certain dynamic.
Lex Fridman (38:56.140)
And this approach dynamic, that is what we mean by it.
Lex Fridman (38:59.220)
But I don't think that infinity would work, so to speak,
Lex Fridman (39:02.860)
that you would know the last digit of pi
Lex Fridman (39:05.060)
and that you have a physical process
Lex Fridman (39:06.540)
that rests on knowing the last digit of pi.
Joscha Bach (39:09.420)
Yeah, that could be just a peculiar quirk
Lex Fridman (39:12.020)
of human cognition that we like discrete.
Joscha Bach (39:15.100)
Discrete makes sense to us.
Lex Fridman (39:16.660)
Infinity doesn't, so in terms of our intuitions.
Joscha Bach (39:19.900)
No, the issue is that everything that we think about
Lex Fridman (39:22.940)
needs to be expressed in some kind of mental language,
Joscha Bach (39:25.660)
not necessarily natural language,
Lex Fridman (39:27.740)
but some kind of mathematical language
Joscha Bach (39:29.860)
that your neurons can speak
Lex Fridman (39:31.700)
that refers to something in the world.
Lex Fridman (39:34.140)
And what we have discovered
Lex Fridman (39:35.460)
is that we cannot construct a notion of infinity
Joscha Bach (39:39.020)
without running into contradictions,
Lex Fridman (39:40.540)
which means that such a language is no longer valid.
Lex Fridman (39:43.620)
And I suspect this is what made Pythagoras so unhappy
Lex Fridman (39:46.780)
when somebody came up with the notion of irrational numbers
Lex Fridman (39:49.380)
before it was time, right?
Lex Fridman (39:50.420)
There's this myth that he had this person killed
Joscha Bach (39:52.700)
when he blabbed out the secret
Lex Fridman (39:54.140)
that not everything can be expressed
Joscha Bach (39:55.740)
as a ratio between two numbers,
Lex Fridman (39:57.300)
but there are numbers between the ratios.
Joscha Bach (39:59.740)
The world was not ready for this.
Lex Fridman (3:00:01.820)
and you really don't know what to do.
Lex Fridman (3:00:03.740)
What should I have done to make you stop screaming?
Lex Fridman (3:00:06.380)
You could have given me candy.
Joscha Bach (3:00:10.020)
I think that's like a cat versus dog discussion.
Lex Fridman (3:00:12.140)
I love it.
Joscha Bach (3:00:13.940)
Cause you said like a fundamental aspect of that is love
Lex Fridman (3:00:19.420)
that makes it all worth it.
Joscha Bach (3:00:21.220)
What, in this monkey riding an elephant in a dream world,
Lex Fridman (3:00:26.740)
what role does love play in the human condition?
Joscha Bach (3:00:31.540)
I think that love is the facilitator
Lex Fridman (3:00:33.540)
of non transactional interaction.
Lex Fridman (3:00:37.140)
And you are observing your own purposes.
Lex Fridman (3:00:40.140)
Some of these purposes go beyond your ego.
Joscha Bach (3:00:42.460)
They go beyond the particular organism
Lex Fridman (3:00:45.140)
that you are and your local interests.
Joscha Bach (3:00:46.780)
That's what you mean by non transactional.
Lex Fridman (3:00:48.540)
Yes, so basically when you are acting
Joscha Bach (3:00:50.180)
in a transactional way, it means that you are respecting
Lex Fridman (3:00:52.860)
something in return for you
Joscha Bach (3:00:55.420)
from the one that you're interacting with.
Lex Fridman (3:00:58.060)
You are interacting with a random stranger,
Joscha Bach (3:00:59.860)
you buy something from them on eBay,
Lex Fridman (3:01:01.340)
you expect a fair value for the money that you sent them
Lex Fridman (3:01:03.900)
and vice versa.
Lex Fridman (3:01:05.420)
Because you don't know that person,
Joscha Bach (3:01:06.580)
you don't have any kind of relationship to them.
Lex Fridman (3:01:09.020)
But when you know this person a little bit better
Lex Fridman (3:01:10.660)
and you know the situation that they're in,
Lex Fridman (3:01:12.740)
you understand what they try to achieve in their life
Lex Fridman (3:01:14.940)
and you approve because you realize that they're
Lex Fridman (3:01:17.700)
in some sense serving the same human sacredness as you are.
Lex Fridman (3:01:22.420)
And they need to think that you have,
Lex Fridman (3:01:23.820)
maybe you give it to them as a present.
Joscha Bach (3:01:26.700)
But, I mean, the feeling itself of joy is a kind of benefit,
Lex Fridman (3:01:32.180)
is a kind of transaction, like...
Joscha Bach (3:01:34.660)
Yes, but the joy is not the point.
Lex Fridman (3:01:36.500)
The joy is the signal that you get.
Joscha Bach (3:01:38.460)
It's the reinforcement signal that your brain sends to you
Lex Fridman (3:01:40.900)
because you are acting on the incentives
Joscha Bach (3:01:43.740)
of the agent that you're a part of.
Lex Fridman (3:01:45.740)
We are meant to be part of something larger.
Joscha Bach (3:01:48.500)
This is the way in which we out competed other hominins.
Lex Fridman (3:01:54.100)
Take that Neanderthals.
Joscha Bach (3:01:56.420)
Yeah, right.
Lex Fridman (3:01:57.420)
And also other humans.
Joscha Bach (3:01:59.620)
There was a population bottleneck for human society
Lex Fridman (3:02:03.100)
that leads to an extreme lack of genetic diversity
Joscha Bach (3:02:06.900)
among humans.
Lex Fridman (3:02:07.740)
If you look at Bushmen in the Kalahari,
Joscha Bach (3:02:11.300)
that basically tribes that are not that far distant
Lex Fridman (3:02:13.900)
to each other have more genetic diversity
Joscha Bach (3:02:15.860)
than exists between Europeans and Chinese.
Lex Fridman (3:02:19.740)
And that's because basically the out of Africa population
Joscha Bach (3:02:23.460)
at some point had a bottleneck
Lex Fridman (3:02:25.060)
of just a few thousand individuals.
Lex Fridman (3:02:27.740)
And what probably happened is not that at any time
Lex Fridman (3:02:30.980)
the number of people shrank below a few hundred thousand.
Lex Fridman (3:02:34.740)
What probably happened is that there was a small group
Lex Fridman (3:02:37.580)
that had a decisive mutation that produced an advantage.
Lex Fridman (3:02:40.460)
And this group multiplied and killed everybody else.
Lex Fridman (3:02:44.100)
And we are descendants of that group.
Joscha Bach (3:02:46.140)
Yeah, I wonder what the peculiar characteristics
Lex Fridman (3:02:50.780)
of that group.
Joscha Bach (3:02:52.140)
Yeah.
Lex Fridman (3:02:53.100)
I mean, we can never know.
Joscha Bach (3:02:53.940)
Me too, and a lot of people do.
Lex Fridman (3:02:55.460)
We can only just listen to the echoes in ours,
Joscha Bach (3:02:58.220)
like the ripples that are still within us.
Lex Fridman (3:03:01.660)
So I suspect what eventually made a big difference
Joscha Bach (3:03:04.420)
was the ability to organize at scale,
Lex Fridman (3:03:07.580)
to program each other.
Joscha Bach (3:03:09.260)
With ideas.
Lex Fridman (3:03:11.380)
That we became programmable,
Joscha Bach (3:03:12.620)
that we were willing to work in lockstep,
Lex Fridman (3:03:14.500)
that we went above the tribal level,
Joscha Bach (3:03:17.420)
that we no longer were groups of a few hundred individuals
Lex Fridman (3:03:20.700)
and acted on direct reputation systems transactionally,
Lex Fridman (3:03:24.460)
but that we basically evolved an adaptation
Lex Fridman (3:03:27.420)
to become state building.
Joscha Bach (3:03:28.980)
Yeah.
Lex Fridman (3:03:31.740)
To form collectives outside of the direct collectives.
Joscha Bach (3:03:35.700)
Yes, and that's basically a part of us became committed
Lex Fridman (3:03:38.580)
to serving something outside of what we know.
Joscha Bach (3:03:41.940)
Yeah, then that's kind of what love is.
Lex Fridman (3:03:44.140)
And it's terrifying because it meant
Joscha Bach (3:03:45.820)
that we eradicated the others.
Lex Fridman (3:03:48.900)
Right, it's a force.
Joscha Bach (3:03:49.820)
It's an adaptive force that gets us ahead in evolution,
Lex Fridman (3:03:52.940)
which means we displace something else
Joscha Bach (3:03:54.540)
that doesn't have that.
Lex Fridman (3:03:56.780)
Oh, so we had to murder a lot of people
Joscha Bach (3:03:58.740)
that weren't about love.
Lex Fridman (3:04:00.380)
So love led to destruction.
Joscha Bach (3:04:01.660)
They didn't have the same strong love as we did.
Lex Fridman (3:04:04.020)
Right, that's why I mentioned this thing with fascism.
Lex Fridman (3:04:07.420)
When you see these speeches, do you want total war?
Lex Fridman (3:04:12.220)
And everybody says, yes, right?
Joscha Bach (3:04:14.180)
This is this big, oh my God, we are part of something
Lex Fridman (3:04:17.620)
that is more important than me
Joscha Bach (3:04:18.660)
that gives meaning to my existence.
Lex Fridman (3:04:22.980)
Fair enough.
Lex Fridman (3:04:27.020)
Do you have advice for young people today
Lex Fridman (3:04:30.980)
in high school, in college,
Joscha Bach (3:04:33.140)
that are thinking about what to do with their career,
Lex Fridman (3:04:37.260)
with their life, so that at the end of the whole thing,
Lex Fridman (3:04:40.420)
they can be proud of what they did?
Lex Fridman (3:04:43.820)
Don't cheat.
Joscha Bach (3:04:45.860)
Have integrity, aim for integrity.
Lex Fridman (3:04:48.540)
So what does integrity look like when you're at the river
Lex Fridman (3:04:50.860)
or the leaf or the fat frog in a lake?
Lex Fridman (3:04:54.580)
It basically means that you try to figure out
Lex Fridman (3:04:57.700)
what the thing is that is the most right.
Lex Fridman (3:05:02.060)
And this doesn't mean that you have to look
Joscha Bach (3:05:04.620)
for what other people tell you what's right,
Lex Fridman (3:05:07.140)
but you have to aim for moral autonomy.
Lex Fridman (3:05:09.740)
So things need to be right independently
Lex Fridman (3:05:12.220)
of what other people say.
Joscha Bach (3:05:14.100)
I always felt that when people told me
Lex Fridman (3:05:17.620)
to listen to what others say, like read the room,
Joscha Bach (3:05:22.940)
build your ideas of what's true
Lex Fridman (3:05:25.060)
based on the high status people of your in group,
Joscha Bach (3:05:27.020)
that does not protect me from fascism.
Lex Fridman (3:05:29.780)
The only way to protect yourself from fascism
Joscha Bach (3:05:31.940)
is to decide it's the world that is being built here,
Lex Fridman (3:05:35.580)
the world that I want to be in.
Lex Fridman (3:05:37.620)
And so in some sense, try to make your behavior sustainable,
Lex Fridman (3:05:41.740)
act in such a way that you would feel comfortable
Joscha Bach (3:05:44.540)
on all sides of the transaction.
Lex Fridman (3:05:46.420)
Realize that everybody is you in a different timeline,
Lex Fridman (3:05:48.900)
but is seeing things differently
Lex Fridman (3:05:51.140)
and has reasons to do so.
Joscha Bach (3:05:53.940)
Yeah, I've come to realize this recently,
Lex Fridman (3:05:58.100)
that there is an inner voice
Joscha Bach (3:05:59.340)
that tells you what's right and wrong.
Lex Fridman (3:06:02.820)
And speaking of reading the room,
Joscha Bach (3:06:06.180)
there's times what integrity looks like
Lex Fridman (3:06:08.060)
is there's times when a lot of people
Joscha Bach (3:06:10.460)
are doing something wrong.
Lex Fridman (3:06:12.180)
And what integrity looks like
Joscha Bach (3:06:13.740)
is not going on Twitter and tweeting about it,
Lex Fridman (3:06:16.500)
but not participating quietly, not doing.
Lex Fridman (3:06:20.260)
So it's not like signaling or not all this kind of stuff,
Lex Fridman (3:06:24.100)
but actually living your, what you think is right.
Joscha Bach (3:06:28.020)
Like living it, not signaling.
Lex Fridman (3:06:28.860)
There's also sometimes this expectation
Joscha Bach (3:06:30.980)
that others are like us.
Lex Fridman (3:06:32.260)
So imagine the possibility
Joscha Bach (3:06:34.380)
that some of the people around you are space aliens
Lex Fridman (3:06:37.060)
that only look human, right?
Lex Fridman (3:06:39.500)
So they don't have the same prayers as you do.
Lex Fridman (3:06:41.620)
They don't have the same impulses
Joscha Bach (3:06:44.100)
that's what's right and wrong.
Lex Fridman (3:06:45.180)
There's a large diversity in these basic impulses
Joscha Bach (3:06:48.900)
that people can have in a given situation.
Lex Fridman (3:06:51.820)
And now realize that you are a space alien, right?
Joscha Bach (3:06:54.660)
You are not actually human.
Lex Fridman (3:06:55.900)
You think that you are human,
Lex Fridman (3:06:57.220)
but you don't know what it means,
Lex Fridman (3:06:58.820)
like what it's like to be human.
Joscha Bach (3:07:00.780)
You just make it up as you go along like everybody else.
Lex Fridman (3:07:04.020)
And you have to figure that out,
Lex Fridman (3:07:05.740)
what it means that you are a full human being,
Lex Fridman (3:07:09.620)
what it means to be human in the world
Lex Fridman (3:07:11.180)
and how to connect with others on that.
Lex Fridman (3:07:13.540)
And there is also something, don't be afraid
Joscha Bach (3:07:17.340)
in the sense that if you do this, you're not good enough.
Lex Fridman (3:07:20.980)
Because if you are acting on these incentives of integrity,
Joscha Bach (3:07:23.580)
you become trustworthy.
Lex Fridman (3:07:25.140)
That's the way in which you can recognize each other.
Joscha Bach (3:07:28.420)
There is a particular place where you can meet.
Lex Fridman (3:07:30.700)
You can figure out what that place is,
Joscha Bach (3:07:33.060)
where you will give support to people
Lex Fridman (3:07:35.420)
because you realize that they act with integrity
Lex Fridman (3:07:38.420)
and they will also do that.
Lex Fridman (3:07:40.300)
So in some sense, you are safe if you do that.
Joscha Bach (3:07:43.860)
You're not always protected.
Lex Fridman (3:07:44.940)
There are people which will abuse you
Lex Fridman (3:07:47.100)
and that are bad actors in a way
Lex Fridman (3:07:49.940)
that it's hard to imagine before you meet them.
Lex Fridman (3:07:52.780)
But there is also people which will try to protect you.
Lex Fridman (3:07:57.780)
Yeah, that's such a, thank you for saying that.
Joscha Bach (3:08:00.820)
That's such a hopeful message
Lex Fridman (3:08:03.820)
that no matter what happens to you,
Joscha Bach (3:08:05.380)
there'll be a place, there's people you'll meet
Lex Fridman (3:08:11.740)
that also have what you have
Lex Fridman (3:08:15.620)
and you will find happiness there and safety there.
Lex Fridman (3:08:20.180)
Yeah, but it doesn't need to end well.
Joscha Bach (3:08:21.700)
It can also all go wrong.
Lex Fridman (3:08:23.500)
So there's no guarantees in this life.
Lex Fridman (3:08:26.380)
So you can do everything right and you still can fail
Lex Fridman (3:08:29.380)
and you can see horrible things happening to you
Joscha Bach (3:08:32.500)
that traumatize you and mutilate you
Lex Fridman (3:08:35.060)
and you have to be grateful if it doesn't happen.
Lex Fridman (3:08:40.300)
And ultimately be grateful no matter what happens
Lex Fridman (3:08:42.940)
because even just being alive is pretty damn nice.
Joscha Bach (3:08:46.860)
Yeah, even that, you know.
Lex Fridman (3:08:49.580)
The gratefulness in some sense is also just generated
Joscha Bach (3:08:52.260)
by your brain to keep you going, it's all the trick.
Lex Fridman (3:08:58.900)
Speaking of which, Camus said,
Joscha Bach (3:09:02.900)
I see many people die because they judge
Lex Fridman (3:09:05.540)
that life is not worth living.
Joscha Bach (3:09:08.020)
I see others paradoxically getting killed
Lex Fridman (3:09:10.820)
for the ideas or illusions that give them
Joscha Bach (3:09:12.860)
a reason for living.
Lex Fridman (3:09:15.020)
What is called the reason for living
Joscha Bach (3:09:16.420)
is also an excellent reason for dying.
Lex Fridman (3:09:19.420)
I therefore conclude that the meaning of life
Joscha Bach (3:09:22.020)
is the most urgent of questions.
Lex Fridman (3:09:24.660)
So I have to ask what Jascha Bach is the meaning of life?
Joscha Bach (3:09:31.500)
It is an urgent question according to Camus.
Lex Fridman (3:09:35.260)
I don't think that there's a single answer to this.
Joscha Bach (3:09:37.940)
Nothing makes sense unless the mind makes it so.
Lex Fridman (3:09:41.340)
So you basically have to project a purpose.
Lex Fridman (3:09:44.820)
And if you zoom out far enough,
Lex Fridman (3:09:47.380)
there's the heat test of the universe
Lex Fridman (3:09:49.060)
and everything is meaningless,
Lex Fridman (3:09:50.500)
everything is just a blip in between.
Lex Fridman (3:09:52.100)
And the question is, do you find meaning
Lex Fridman (3:09:54.020)
in this blip in between?
Lex Fridman (3:09:55.820)
Do you find meaning in observing squirrels?
Lex Fridman (3:09:59.780)
Do you find meaning in raising children
Lex Fridman (3:10:01.740)
and projecting a multi generational organism
Lex Fridman (3:10:04.420)
into the future?
Lex Fridman (3:10:05.660)
Do you find meaning in projecting an aesthetic
Lex Fridman (3:10:08.260)
of the world that you like to the future
Lex Fridman (3:10:10.620)
and trying to serve that aesthetic?
Lex Fridman (3:10:12.340)
And if you do, then life has that meaning.
Lex Fridman (3:10:15.300)
And if you don't, then it doesn't.
Lex Fridman (3:10:17.140)
I kind of enjoy the idea that you just create
Joscha Bach (3:10:21.780)
the most vibrant, the most weird,
Lex Fridman (3:10:25.660)
the most unique kind of blip you can,
Joscha Bach (3:10:28.740)
given your environment, given your set of skills,
Lex Fridman (3:10:32.020)
just be the most weird set of,
Joscha Bach (3:10:38.740)
like local pocket of complexity you can be.
Lex Fridman (3:10:41.740)
So that like, when people study the universe,
Joscha Bach (3:10:44.500)
they'll pause and be like, oh, that's weird.
Lex Fridman (3:10:47.340)
It looks like a useful strategy,
Lex Fridman (3:10:50.580)
but of course it's still motivated reasoning.
Lex Fridman (3:10:52.780)
You're obviously acting on your incentives here.
Joscha Bach (3:10:57.780)
It's still a story we tell ourselves within a dream
Lex Fridman (3:11:00.700)
that's hardly in touch with the reality.
Joscha Bach (3:11:03.860)
It's definitely a good strategy if you are a podcaster.
Lex Fridman (3:11:10.180)
And a human, which I'm still trying to figure out if I am.
Joscha Bach (3:11:13.060)
It has a mutual relationship somehow.
Lex Fridman (3:11:15.020)
Somehow.
Joscha Bach (3:11:16.060)
Josh, you're one of the most incredible people I know.
Lex Fridman (3:11:20.860)
I really love talking to you.
Joscha Bach (3:11:22.380)
I love talking to you again,
Lex Fridman (3:11:23.500)
and it's really an honor that you spend
Joscha Bach (3:11:26.060)
your valuable time with me.
Lex Fridman (3:11:27.100)
I hope we get to talk many times
Joscha Bach (3:11:28.580)
through our short and meaningless lives.
Lex Fridman (3:11:33.580)
Or meaningful.
Joscha Bach (3:11:34.620)
Or meaningful.
Lex Fridman (3:11:35.900)
Thank you, Alex.
Joscha Bach (3:11:36.740)
I enjoyed this conversation very much.
Lex Fridman (3:11:39.020)
Thanks for listening to this conversation with Josche Bach.
Joscha Bach (3:11:41.700)
A thank you to Coinbase, Codecademy, Linode,
Lex Fridman (3:11:45.900)
NetSuite, and ExpressVPN.
Joscha Bach (3:11:48.500)
Check them out in the description to support this podcast.
Lex Fridman (3:11:52.020)
Now, let me leave you with some words from Carl Jung.
Joscha Bach (3:11:55.780)
People will do anything, no matter how absurd,
Lex Fridman (3:11:59.020)
in order to avoid facing their own souls.
Joscha Bach (3:12:01.780)
One does not become enlightened
Lex Fridman (3:12:03.580)
by imagining figures of light,
Lex Fridman (3:12:05.780)
but by making the darkness conscious.
Lex Fridman (3:12:09.260)
Thank you for listening, and hope to see you next time.
Lex Fridman (40:01.060)
And I think he was right.
Lex Fridman (40:02.380)
That has confused mathematicians very seriously
Joscha Bach (40:06.060)
because these numbers are not values, they are functions.
Lex Fridman (40:09.660)
And so you can calculate these functions
Joscha Bach (40:11.580)
to a certain degree of approximation,
Lex Fridman (40:13.260)
but you cannot pretend that pi has actually a value.
Joscha Bach (40:17.060)
Pi is a function that would approach this value
Lex Fridman (40:20.020)
to some degree,
Lex Fridman (40:21.500)
but nothing in the world rests on knowing pi.
Lex Fridman (40:26.980)
How important is this distinction
Lex Fridman (40:28.620)
between discrete and continuous for you to get to the book?
Lex Fridman (40:32.180)
Because there's a, I mean, in discussion of your favorite
Joscha Bach (40:36.580)
flavor of the theory of everything,
Lex Fridman (40:39.180)
there's a few on the table.
Lex Fridman (40:41.140)
So there's string theory, there's a particular,
Lex Fridman (40:45.260)
there's a little quantum gravity,
Joscha Bach (40:48.180)
which focused on one particular unification.
Lex Fridman (40:53.180)
There's just a bunch of favorite flavors
Joscha Bach (40:56.140)
of different people trying to propose
Lex Fridman (40:59.460)
a theory of everything.
Joscha Bach (41:01.260)
Eric Weinstein and a bunch of people throughout history.
Lex Fridman (41:04.780)
And then of course, Stephen Wolfram,
Joscha Bach (41:06.660)
who I think is one of the only people doing a discrete.
Lex Fridman (41:10.860)
No, no, there's a bunch of physicists
Joscha Bach (41:12.580)
who do this right now.
Lex Fridman (41:13.700)
And like Toffoli and Tomasello.
Lex Fridman (41:17.660)
And digital physics is something
Lex Fridman (41:21.940)
that is, I think, growing in popularity.
Lex Fridman (41:24.460)
But the main reason why this is interesting
Lex Fridman (41:29.460)
is because it's important sometimes to settle disagreements.
Joscha Bach (41:34.700)
I don't think that you need infinities at all,
Lex Fridman (41:36.980)
and you never needed them.
Joscha Bach (41:38.940)
You can always deal with very large numbers
Lex Fridman (41:40.900)
and you can deal with limits, right?
Joscha Bach (41:42.260)
We are fine with doing that.
Lex Fridman (41:43.780)
You don't need any kind of infinity.
Joscha Bach (41:45.300)
You can build your computer algebra systems just as well
Lex Fridman (41:48.340)
without believing in infinity in the first place.
Lex Fridman (41:50.300)
So you're okay with limits?
Lex Fridman (41:51.940)
Yeah, so basically a limit means that something
Joscha Bach (41:54.420)
is behaving pretty much the same
Lex Fridman (41:57.460)
if you make the number large.
Joscha Bach (41:59.100)
Right, because it's converging to a certain value.
Lex Fridman (42:02.420)
And at some point the difference becomes negligible
Lex Fridman (42:04.780)
and you can no longer measure it.
Lex Fridman (42:06.620)
And in this sense, you have things
Joscha Bach (42:08.660)
that if you have an ngon which has enough corners,
Lex Fridman (42:12.820)
then it's going to behave like a circle at some point, right?
Lex Fridman (42:15.180)
And it's only going to be in some kind of esoteric thing
Lex Fridman (42:18.380)
that cannot exist in the physical universe
Joscha Bach (42:21.060)
that you would be talking about this perfect circle.
Lex Fridman (42:23.820)
And now it turns out that it also wouldn't work
Joscha Bach (42:25.900)
in mathematics because you cannot construct mathematics
Lex Fridman (42:28.380)
that has infinite resolution
Joscha Bach (42:30.020)
without running into contradictions.
Lex Fridman (42:32.820)
So that is itself not that important
Lex Fridman (42:35.020)
because we never did that, right?
Lex Fridman (42:36.220)
It's just a thing that some people thought we could.
Lex Fridman (42:39.020)
And this leads to confusion.
Lex Fridman (42:40.780)
So for instance, Roger Penrose uses this as an argument
Joscha Bach (42:43.580)
to say that there are certain things
Lex Fridman (42:46.140)
that mathematicians can do dealing with infinities
Lex Fridman (42:50.580)
and by extension our mind can do
Lex Fridman (42:53.220)
that computers cannot do.
Joscha Bach (42:55.180)
Yeah, he talks about that the human mind
Lex Fridman (42:58.420)
can do certain mathematical things
Joscha Bach (43:00.780)
that the computer as defined
Lex Fridman (43:02.900)
by the universal Turing machine cannot.
Joscha Bach (43:06.140)
Yes.
Lex Fridman (43:07.180)
So that it has to do with infinity.
Joscha Bach (43:08.900)
Yes, it's one of the things.
Lex Fridman (43:10.260)
So he is basically pointing at the fact
Joscha Bach (43:13.100)
that there are things that are possible
Lex Fridman (43:15.580)
in the mathematical mind and in pure mathematics
Joscha Bach (43:21.420)
that are not possible in machines
Lex Fridman (43:24.100)
that can be constructed in the physical universe.
Lex Fridman (43:27.060)
And because he's an honest guy,
Lex Fridman (43:29.140)
he thinks this means that present physics
Joscha Bach (43:31.660)
cannot explain operations that happen in our mind.
Lex Fridman (43:34.860)
Do you think he's right?
Lex Fridman (43:35.780)
And so let's leave his discussion
Lex Fridman (43:38.700)
of consciousness aside for the moment.
Lex Fridman (43:40.780)
Do you think he's right about just
Lex Fridman (43:42.780)
what he's basically referring to as intelligence?
Lex Fridman (43:46.060)
So is the human mind fundamentally more capable
Lex Fridman (43:50.780)
as a thinking machine than a universal Turing machine?
Joscha Bach (43:53.940)
No.
Lex Fridman (43:55.460)
But so he's suggesting that, right?
Lex Fridman (43:58.740)
So our mind is actually less than a Turing machine.
Lex Fridman (44:01.020)
There can be no Turing machine
Joscha Bach (44:02.100)
because it's defined as having an infinite tape.
Lex Fridman (44:05.100)
And we always only have a finite tape.
Lex Fridman (44:07.260)
But he's saying it's better.
Lex Fridman (44:08.100)
Our minds can only perform finitely many operations.
Joscha Bach (44:10.140)
Yes, he thinks so.
Lex Fridman (44:10.980)
He's saying it can do the kind of computation
Joscha Bach (44:13.100)
that the Turing machine cannot.
Lex Fridman (44:14.620)
And that's because he thinks that our minds
Joscha Bach (44:16.660)
can do operations that have infinite resolution
Lex Fridman (44:19.500)
in some sense.
Lex Fridman (44:21.020)
And I don't think that's the case.
Lex Fridman (44:23.260)
Our minds are just able to discover these limit operators
Joscha Bach (44:26.340)
over too many parts to count.
Lex Fridman (44:27.700)
I see.
Lex Fridman (44:30.300)
What about his idea that consciousness
Lex Fridman (44:32.740)
is more than a computation?
Lex Fridman (44:37.460)
So it's more than something that a Turing machine can do.
Lex Fridman (44:42.100)
So again, saying that there's something special
Joscha Bach (44:44.540)
about our mind that cannot be replicated in a machine.
Lex Fridman (44:49.820)
The issue is that I don't even know
Lex Fridman (44:51.380)
how to construct a language to express
Lex Fridman (44:54.300)
this statement correctly.
Joscha Bach (44:56.420)
Well,
Lex Fridman (45:01.100)
the basic statement is there's a human experience
Joscha Bach (45:06.900)
that includes intelligence, that includes self awareness,
Lex Fridman (45:09.420)
that includes the hard problem of consciousness.
Lex Fridman (45:12.980)
And the question is, can that be fully simulated
Lex Fridman (45:16.860)
in the computer, in the mathematical model of the computer
Lex Fridman (45:20.940)
as we understand it today?
Lex Fridman (45:23.620)
Roger Penrose says no.
Lex Fridman (45:25.900)
So the universe of Turing machine
Lex Fridman (45:30.220)
cannot simulate the universe.
Lex Fridman (45:32.460)
So the interesting question is,
Lex Fridman (45:34.420)
and you have to ask him this is, why not?
Lex Fridman (45:36.500)
What is this specific thing that cannot be modeled?
Lex Fridman (45:39.900)
And when I looked at his writings
Lex Fridman (45:42.340)
and I haven't read all of it,
Lex Fridman (45:43.540)
but when I read, for instance,
Joscha Bach (45:45.940)
the section that he writes in the introduction
Lex Fridman (45:49.060)
to a road to infinity,
Joscha Bach (45:51.020)
the thing that he specifically refers to
Lex Fridman (45:53.260)
is the way in which human minds deal with infinities.
Lex Fridman (45:57.660)
And that itself can, I think, easily be deconstructed.
Lex Fridman (46:03.060)
A lot of people feel that our experience
Joscha Bach (46:05.580)
cannot be explained in a mechanical way.
Lex Fridman (46:08.660)
And therefore it needs to be different.
Lex Fridman (46:11.060)
And I concur, our experience is not mechanical.
Lex Fridman (46:14.500)
Our experience is simulated.
Joscha Bach (46:16.700)
It exists only in a simulation.
Lex Fridman (46:18.420)
The only simulation can be conscious.
Joscha Bach (46:19.980)
Physical systems cannot be conscious
Lex Fridman (46:21.580)
because they're only mechanical.
Joscha Bach (46:23.020)
Cells cannot be conscious.
Lex Fridman (46:25.100)
Neurons cannot be conscious.
Joscha Bach (46:26.300)
Brains cannot be conscious.
Lex Fridman (46:27.460)
People cannot be conscious
Joscha Bach (46:28.660)
as far as if you understand them as physical systems.
Lex Fridman (46:31.620)
What can be conscious is the story of the system
Joscha Bach (46:36.220)
in the world where you write all these things
Lex Fridman (46:37.980)
into the story.
Joscha Bach (46:39.420)
You have experiences for the same reason
Lex Fridman (46:41.420)
that a character novel has experiences
Joscha Bach (46:43.260)
because it's written into the story.
Lex Fridman (46:45.780)
And now the system is acting on that story.
Lex Fridman (46:48.220)
And it's not a story that is written in a natural language.
Lex Fridman (46:50.660)
It's written in a perceptual language,
Joscha Bach (46:52.500)
in this multimedia language of the game engine.
Lex Fridman (46:55.380)
And in there, you write in what kind of experience you have
Lex Fridman (46:59.340)
and what this means for the behavior of the system,
Lex Fridman (47:01.460)
for your behavior tendencies, for your focus,
Joscha Bach (47:03.700)
for your attention, for your experience of valence
Lex Fridman (47:05.460)
and so on.
Lex Fridman (47:06.420)
And this is being used to inform the behavior of the system
Lex Fridman (47:09.620)
in the next step.
Lex Fridman (47:10.740)
And then the story updates with the reactions of the system
Lex Fridman (47:15.780)
and the changes in the world and so on.
Lex Fridman (47:17.780)
And you live inside of that model.
Lex Fridman (47:19.340)
You don't live inside of the physical reality.
Lex Fridman (47:23.420)
And I mean, just to linger on it, like you say, okay,
Lex Fridman (47:28.820)
it's in the perceptual language,
Joscha Bach (47:30.860)
the multimodal perceptual language.
Lex Fridman (47:33.300)
That's the experience.
Joscha Bach (47:34.900)
That's what consciousness is within that model,
Lex Fridman (47:38.900)
within that story.
Lex Fridman (47:40.860)
But do you have agency?
Lex Fridman (47:43.980)
When you play a video game, you can turn left
Lex Fridman (47:46.020)
and you can turn right in that story.
Lex Fridman (47:49.620)
So in that dream world, how much control do you have?
Lex Fridman (47:54.220)
Is there such a thing as you in that story?
Lex Fridman (47:57.620)
Like, is it right to say the main character,
Joscha Bach (48:00.980)
you know, everybody's NPCs,
Lex Fridman (48:02.540)
and then there's the main character
Lex Fridman (48:04.380)
and you're controlling the main character?
Lex Fridman (48:07.020)
Or is that an illusion?
Lex Fridman (48:08.700)
Is there a main character that you're controlling?
Lex Fridman (48:10.900)
I'm getting to the point of like the free will point.
Joscha Bach (48:14.540)
Imagine that you are building a robot that plays soccer.
Lex Fridman (48:17.780)
And you've been to MIT computer science,
Lex Fridman (48:19.900)
you basically know how to do that, right?
Lex Fridman (48:22.060)
And so you would say the robot is an agent
Joscha Bach (48:25.300)
that solves a control problem,
Lex Fridman (48:27.780)
how to get the ball into the goal.
Lex Fridman (48:29.300)
And it needs to perceive the world
Lex Fridman (48:30.740)
and the world is disturbing him in trying to do this, right?
Lex Fridman (48:33.260)
So he has to control many variables to make that happen
Lex Fridman (48:35.660)
and to project itself and the ball into the future
Lex Fridman (48:38.820)
and understand its position on the field
Lex Fridman (48:40.700)
relative to the ball and so on,
Lex Fridman (48:42.140)
and the position of its limbs
Lex Fridman (48:44.620)
or in the space around it and so on.
Lex Fridman (48:46.940)
So it needs to have an adequate model
Lex Fridman (48:48.460)
that abstracting reality in a useful way.
Lex Fridman (48:51.380)
And you could say that this robot does have agency
Lex Fridman (48:55.900)
over what it's doing in some sense.
Lex Fridman (48:58.420)
And the model is going to be a control model.
Lex Fridman (49:01.500)
And inside of that control model,
Joscha Bach (49:03.060)
you can possibly get to a point
Lex Fridman (49:05.780)
where this thing is sufficiently abstract
Joscha Bach (49:07.820)
to discover its own agency.
Lex Fridman (49:09.540)
Our current robots don't do that.
Joscha Bach (49:10.860)
They don't have a unified model of the universe,
Lex Fridman (49:13.140)
but there's not a reason why we shouldn't be getting there
Joscha Bach (49:16.140)
at some point in the not too distant future.
Lex Fridman (49:18.660)
And once that happens,
Joscha Bach (49:20.060)
you will notice that the robot tells a story
Lex Fridman (49:23.220)
about a robot playing soccer.
Lex Fridman (49:25.980)
So the robot will experience itself playing soccer
Lex Fridman (49:29.420)
in a simulation of the world that it uses
Joscha Bach (49:32.060)
to construct a model of the locations of its legs
Lex Fridman (49:35.340)
and limbs in space on the field
Joscha Bach (49:38.180)
with relationship to the ball.
Lex Fridman (49:39.380)
And it's not going to be at the level of the molecules.
Joscha Bach (49:42.220)
It will be an abstraction that is exactly at the level
Lex Fridman (49:45.300)
that is most suitable for past planning
Joscha Bach (49:47.420)
of the movements of the robot.
Lex Fridman (49:49.940)
It's going to be a high level abstraction,
Lex Fridman (49:51.420)
but a very useful one that is as predictive
Lex Fridman (49:53.700)
as we can make it.
Lex Fridman (49:55.180)
And in that side of that story,
Lex Fridman (49:56.580)
there is a model of the agency of that system.
Lex Fridman (49:58.780)
So this model can accurately predict
Lex Fridman (50:03.060)
that the contents of the model
Joscha Bach (50:04.740)
are going to be driving the behavior of the robot
Lex Fridman (50:07.380)
in the immediate future.
Lex Fridman (50:08.900)
But there's the hard problem of consciousness,
Lex Fridman (50:12.340)
which I would also,
Joscha Bach (50:14.580)
there's a subjective experience of free will as well
Lex Fridman (50:18.060)
that I'm not sure where the robot gets that,
Joscha Bach (50:20.740)
where that little leap is.
Lex Fridman (50:22.660)
Because for me right now,
Joscha Bach (50:24.260)
everything I imagine with that robot,
Lex Fridman (50:26.260)
as it gets more and more and more sophisticated,
Joscha Bach (50:29.020)
the agency comes from the programmer of the robot still,
Lex Fridman (50:33.540)
of what was programmed in.
Joscha Bach (50:35.820)
You could probably do an end to end learning system.
Lex Fridman (50:38.500)
You maybe need to give it a few priors.
Lex Fridman (50:40.300)
So you nudge the architecture in the right direction
Lex Fridman (50:42.460)
that it converges more quickly,
Lex Fridman (50:44.340)
but ultimately discovering the suitable hyperparameters
Lex Fridman (50:47.980)
of the architecture is also only a search process.
Lex Fridman (50:50.340)
And as the search process was evolution,
Lex Fridman (50:52.740)
that has informed our brain architecture
Lex Fridman (50:55.300)
so we can converge in a single lifetime
Lex Fridman (50:57.380)
on useful interaction with the world
Lex Fridman (50:59.500)
and the formation of a self model.
Lex Fridman (51:00.340)
The problem is if we define hyperparameters broadly,
Lex Fridman (51:03.500)
so it's not just the parameters that control
Lex Fridman (51:06.820)
this end to end learning system,
Lex Fridman (51:08.700)
but the entirety of the design of the robot.
Lex Fridman (51:11.180)
Like there's, you have to remove the human completely
Joscha Bach (51:15.060)
from the picture.
Lex Fridman (51:15.900)
And then in order to build the robot,
Joscha Bach (51:17.300)
you have to create an entire universe.
Lex Fridman (51:20.340)
Cause you have to go, you can't just shortcut evolution.
Joscha Bach (51:22.620)
You have to go from the very beginning
Lex Fridman (51:24.620)
in order for it to have,
Joscha Bach (51:25.860)
cause I feel like there's always a human
Lex Fridman (51:28.020)
pulling the strings and that makes it seem like
Joscha Bach (51:32.620)
the robot is cheating.
Lex Fridman (51:33.900)
It's getting a shortcut to consciousness.
Lex Fridman (51:35.940)
And you are looking at the current Boston Dynamics robots.
Lex Fridman (51:38.300)
It doesn't look as if there is somebody
Joscha Bach (51:40.140)
pulling the strings.
Lex Fridman (51:40.980)
It doesn't look like cheating anymore.
Joscha Bach (51:42.420)
Okay, so let's go there.
Lex Fridman (51:43.420)
Cause I got to talk to you about this.
Lex Fridman (51:44.860)
So obviously with the case of Boston Dynamics,
Lex Fridman (51:47.740)
as you may or may not know,
Joscha Bach (51:49.780)
it's always either hard coded or remote controlled.
Lex Fridman (51:54.100)
There's no intelligence.
Joscha Bach (51:55.220)
I don't know how the current generation
Lex Fridman (51:57.460)
of Boston Dynamics robots works,
Lex Fridman (51:59.060)
but what I've been told about the previous ones
Lex Fridman (52:02.020)
was that it's basically all cybernetic control,
Joscha Bach (52:05.260)
which means you still have feedback mechanisms and so on,
Lex Fridman (52:08.620)
but it's not deep learning for the most part
Joscha Bach (52:11.340)
as it's currently done.
Lex Fridman (52:13.220)
It's for the most part,
Joscha Bach (52:14.700)
just identifying a control hierarchy
Lex Fridman (52:16.940)
that is congruent to the limbs that exist
Lex Fridman (52:19.820)
and the parameters that need to be optimized
Lex Fridman (52:21.460)
for the movement of these limbs.
Lex Fridman (52:22.580)
And then there is a convergence progress.
Lex Fridman (52:24.500)
So it's basically just regression
Joscha Bach (52:26.220)
that you would need to control this.
Lex Fridman (52:27.900)
But again, I don't know whether that's true.
Joscha Bach (52:29.420)
That's just what I've been told about how they work.
Lex Fridman (52:31.420)
We have to separate several levels of discussion here.
Lex Fridman (52:35.020)
So the only thing they do is pretty sophisticated control
Lex Fridman (52:39.300)
with no machine learning
Joscha Bach (52:40.900)
in order to maintain balance or to right itself.
Lex Fridman (52:45.980)
It's a control problem in terms of using the actuators
Joscha Bach (52:49.380)
to when it's pushed or when it steps on a thing
Lex Fridman (52:52.420)
that's uneven, how to always maintain balance.
Lex Fridman (52:55.420)
And there's a tricky set of heuristics around that,
Lex Fridman (52:57.940)
but that's the only goal.
Joscha Bach (53:00.460)
Everything you see Boston Dynamics doing
Lex Fridman (53:02.660)
in terms of that to us humans is compelling,
Joscha Bach (53:06.140)
which is any kind of higher order movement,
Lex Fridman (53:09.460)
like turning, wiggling its butt,
Joscha Bach (53:13.220)
like jumping back on its two feet, dancing.
Lex Fridman (53:18.740)
Dancing is even worse because dancing is hard coded in.
Joscha Bach (53:22.460)
It's choreographed by humans.
Lex Fridman (53:25.300)
There's choreography software.
Lex Fridman (53:27.420)
So there is no, of all that high level movement,
Lex Fridman (53:30.900)
there's no anything that you can call,
Joscha Bach (53:34.220)
certainly can't call AI,
Lex Fridman (53:35.940)
but there's no even like basic heuristics.
Joscha Bach (53:39.500)
It's all hard coded in.
Lex Fridman (53:41.060)
And yet we humans immediately project agency onto them,
Joscha Bach (53:47.660)
which is fascinating.
Lex Fridman (53:48.900)
So the gap here doesn't necessarily have agency.
Lex Fridman (53:53.140)
What it has is cybernetic control.
Lex Fridman (53:55.340)
And the cybernetic control means you have a hierarchy
Joscha Bach (53:57.420)
of feedback loops that keep the behavior
Lex Fridman (53:59.740)
in certain boundaries so the robot doesn't fall over
Lex Fridman (54:02.340)
and it's able to perform the movements.
Lex Fridman (54:04.140)
And the choreography cannot really happen
Joscha Bach (54:06.660)
with motion capture because the robot would fall over
Lex Fridman (54:09.220)
because the physics of the robot,
Joscha Bach (54:10.620)
the weight distribution and so on is different
Lex Fridman (54:12.780)
from the weight distribution in the human body.
Lex Fridman (54:15.340)
So if you were using the directly motion captured movements
Lex Fridman (54:19.580)
of a human body to project it into this robot,
Joscha Bach (54:21.740)
it wouldn't work.
Lex Fridman (54:22.580)
You can do this with a computer animation.
Lex Fridman (54:24.100)
It will look a little bit off, but who cares?
Lex Fridman (54:26.140)
But if you want to correct for the physics,
Joscha Bach (54:29.100)
you need to basically tell the robot
Lex Fridman (54:31.500)
where it should move its limbs.
Lex Fridman (54:33.740)
And then the control algorithm is going
Lex Fridman (54:35.860)
to approximate a solution that makes it possible
Joscha Bach (54:38.980)
within the physics of the robot.
Lex Fridman (54:41.020)
And you have to find the basic solution
Joscha Bach (54:43.900)
for making that happen.
Lex Fridman (54:44.780)
And there's probably going to be some regression necessary
Joscha Bach (54:47.580)
to get the control architecture to make these movements.
Lex Fridman (54:51.220)
But those two layers are separate.
Lex Fridman (54:52.660)
So the thing, the higher level instruction
Lex Fridman (54:56.180)
of how you should move and where you should move
Joscha Bach (54:59.060)
is a higher level.
Lex Fridman (54:59.900)
Yeah, so I expect that the control level
Joscha Bach (55:01.700)
of these robots at some level is dumb.
Lex Fridman (55:03.620)
This is just the physical control movement,
Joscha Bach (55:06.180)
the motor architecture.
Lex Fridman (55:07.860)
But it's a relatively smart motor architecture.
Joscha Bach (55:10.340)
It's just that there is no high level deliberation
Lex Fridman (55:12.500)
about what decisions to make necessarily, right?
Lex Fridman (55:14.620)
But see, it doesn't feel like free will or consciousness.
Lex Fridman (55:17.900)
No, no, that was not where I was trying to get to.
Joscha Bach (55:20.580)
I think that in our own body, we have that too.
Lex Fridman (55:24.540)
So we have a certain thing that is basically
Joscha Bach (55:26.900)
just a cybernetic control architecture
Lex Fridman (55:29.540)
that is moving our limbs.
Lex Fridman (55:31.300)
And deep learning can help in discovering
Lex Fridman (55:34.300)
such an architecture if you don't have it
Joscha Bach (55:35.940)
in the first place.
Lex Fridman (55:37.220)
If you already know your hardware,
Joscha Bach (55:38.620)
you can maybe handcraft it.
Lex Fridman (55:40.700)
But if you don't know your hardware,
Joscha Bach (55:41.900)
you can search for such an architecture.
Lex Fridman (55:43.740)
And this work already existed in the 80s and 90s.
Joscha Bach (55:46.980)
People were starting to search for control architectures
Lex Fridman (55:49.820)
by motor babbling and so on,
Lex Fridman (55:51.140)
and just use reinforcement learning architectures
Lex Fridman (55:53.900)
to discover such a thing.
Lex Fridman (55:55.580)
And now imagine that you have
Lex Fridman (55:57.740)
the cybernetic control architecture already inside of you.
Lex Fridman (56:01.540)
And you extend this a little bit.
Lex Fridman (56:03.700)
So you are seeking out food, for instance,
Joscha Bach (56:06.460)
or rest or and so on.
Lex Fridman (56:08.300)
And you get to have a baby at some point.
Lex Fridman (56:11.820)
And now you add more and more control layers to this.
Lex Fridman (56:15.740)
And the system is reverse engineering
Joscha Bach (56:17.780)
its own control architecture
Lex Fridman (56:19.620)
and builds a high level model to synchronize
Joscha Bach (56:22.460)
the pursuit of very different conflicting goals.
Lex Fridman (56:26.340)
And this is how I think you get to purposes.
Joscha Bach (56:28.180)
Purposes are models of your goals.
Lex Fridman (56:30.060)
The goals may be intrinsic
Joscha Bach (56:31.540)
as the result of the different set point violations
Lex Fridman (56:33.820)
that you have,
Joscha Bach (56:34.660)
hunger and thirst for very different things,
Lex Fridman (56:37.140)
and rest and pain avoidance and so on.
Lex Fridman (56:39.380)
And you put all these things together
Lex Fridman (56:41.100)
and eventually you need to come up with a strategy
Joscha Bach (56:44.180)
to synchronize them all.
Lex Fridman (56:46.020)
And you don't need just to do this alone by yourself
Joscha Bach (56:49.340)
because we are state building organisms.
Lex Fridman (56:51.340)
We cannot function as isolation
Joscha Bach (56:53.700)
the way that homo sapiens is set up.
Lex Fridman (56:55.820)
So our own behavior only makes sense
Joscha Bach (56:58.100)
when you zoom out very far into a society
Lex Fridman (57:00.980)
or even into ecosystemic intelligence on the planet
Lex Fridman (57:04.900)
and our place in it.
Lex Fridman (57:06.500)
So the individual behavior only makes sense
Joscha Bach (57:08.460)
in these larger contexts.
Lex Fridman (57:09.980)
And we have a number of priors built into us.
Lex Fridman (57:11.820)
So we are behaving as if we were acting
Lex Fridman (57:14.660)
on these high level goals pretty much right from the start.
Lex Fridman (57:17.900)
And eventually in the course of our life,
Lex Fridman (57:19.820)
we can reverse engineer the goals that we're acting on,
Lex Fridman (57:22.700)
what actually are our higher level purposes.
Lex Fridman (57:25.820)
And the more we understand that,
Joscha Bach (57:27.100)
the more our behavior makes sense.
Lex Fridman (57:28.660)
But this is all at this point,
Joscha Bach (57:30.380)
complex stories within stories
Lex Fridman (57:32.420)
that are driving our behavior.
Joscha Bach (57:34.580)
Yeah, I just don't know how big of a leap it is
Lex Fridman (57:38.500)
to start create a system
Joscha Bach (57:40.940)
that's able to tell stories within stories.
Lex Fridman (57:44.340)
Like how big of a leap that is
Joscha Bach (57:45.580)
from where currently Boston Dynamics is
Lex Fridman (57:48.260)
or any robot that's operating in the physical space.
Lex Fridman (57:53.820)
And that leap might be big
Lex Fridman (57:56.220)
if it requires to solve the hard problem of consciousness,
Joscha Bach (57:59.380)
which is telling a hell of a good story.
Lex Fridman (58:01.620)
I suspect that consciousness itself is relatively simple.
Joscha Bach (58:05.220)
What's hard is perception
Lex Fridman (58:07.300)
and the interface between perception and reasoning.
Joscha Bach (58:11.100)
That's for instance, the idea of the consciousness prior
Lex Fridman (58:14.700)
that would be built into such a system by Yoshua Bengio.
Lex Fridman (58:18.740)
And what he describes, and I think that's accurate,
Lex Fridman (58:22.260)
is that our own model of the world
Joscha Bach (58:27.260)
can be described through something like an energy function.
Lex Fridman (58:29.820)
The energy function is modeling the contradictions
Joscha Bach (58:32.700)
that exist within the model at any given point.
Lex Fridman (58:34.820)
And you try to minimize these contradictions,
Joscha Bach (58:36.620)
the tangents in the model.
Lex Fridman (58:38.340)
And to do this, you need to sometimes test things.
Joscha Bach (58:41.380)
You need to conditionally disambiguate figure and ground.
Lex Fridman (58:43.740)
You need to distinguish whether this is true
Joscha Bach (58:46.500)
or that is true, and so on.
Lex Fridman (58:47.940)
Eventually you get to an interpretation,
Lex Fridman (58:49.500)
but you will need to manually depress a few points
Lex Fridman (58:52.300)
in your model to let it snap into a state that makes sense.
Lex Fridman (58:55.580)
And this function that tries to get the biggest dip
Lex Fridman (58:57.740)
in the energy function in your model,
Joscha Bach (58:59.620)
according to Yoshua Bengio, is related to consciousness.
Lex Fridman (59:02.340)
It's a low dimensional discrete function
Joscha Bach (59:04.620)
that tries to maximize this dip in the energy function.
Lex Fridman (59:09.580)
Yeah, I think I would need to dig into details
Joscha Bach (59:13.340)
because I think the way he uses the word consciousness
Lex Fridman (59:15.580)
is more akin to like self awareness,
Joscha Bach (59:17.780)
like modeling yourself within the world,
Lex Fridman (59:20.860)
as opposed to the subjective experience, the hard problem.
Joscha Bach (59:23.660)
No, it's not even the self is in the world.
Lex Fridman (59:26.580)
The self is the agent and you don't need to be aware
Joscha Bach (59:28.820)
of yourself in order to be conscious.
Lex Fridman (59:31.100)
The self is just a particular content that you can have,
Lex Fridman (59:34.380)
but you don't have to have.
Lex Fridman (59:35.980)
But you can be conscious in, for instance, a dream at night
Joscha Bach (59:39.700)
or during a meditation state where you don't have a self.
Lex Fridman (59:42.940)
Right.
Joscha Bach (59:43.780)
Where you're just aware of the fact that you are aware.
Lex Fridman (59:45.620)
And what we mean by consciousness in the colloquial sense
Joscha Bach (59:49.900)
is largely this reflexive self awareness,
Lex Fridman (59:53.820)
that we become aware of the fact
Joscha Bach (59:55.220)
that you're paying attention,
Lex Fridman (59:57.300)
that we are the thing that pays attention.
Joscha Bach (59:59.220)
We are the thing that pays attention, right.
🔗 相关节目
Joscha Bach: Artificial Consciousness and the Nature of Reality
Joscha Bach
Nick Bostrom: Simulation and Superintelligence
Nick Bostrom
Wojciech Zaremba: OpenAI Codex, GPT-3, Robotics, and the Future of AI
Wojciech Zaremba
Alex Garland: Ex Machina, Devs, Annihilation, and the Poetry of Science
Alex Garland