Karl Friston: Neuroscience and the Free Energy Principle
生物与进化AI 与机器学习音乐与艺术技术与编程哲学与宗教
📋 章节目录
暂无章节信息
🔑 关键词
brainstatesenergyfreeoilsurfacedropselftermstalkingmodelprinciplestructuretryinggotfunctionexistencelookingactivitymeans
💬 精彩语录
暂无语录
🎙️ 完整对话（1766 条）
Lex Fridman (00:00.000)
The following is a conversation with Carl Friston,
以下是与卡尔·弗里斯顿的对话，
Lex Fridman (00:03.080)
one of the greatest neuroscientists in history.
历史上最伟大的神经科学家之一。
Lex Fridman (00:06.960)
Cited over 245,000 times,
被引用超过 245,000 次，
Lex Fridman (00:10.320)
known for many influential ideas in brain imaging,
以脑成像领域许多有影响力的想法而闻名，
Lex Fridman (00:13.580)
neuroscience, and theoretical neurobiology,
神经科学和理论神经生物学，
Karl Friston (00:16.520)
including especially the fascinating idea
尤其包括令人着迷的想法
Lex Fridman (00:19.820)
of the free energy principle for action and perception.
行动和感知的自由能原理。
Karl Friston (00:24.200)
Carl's mix of humor, brilliance, and kindness,
卡尔集幽默、才华和善良于一身，
Lex Fridman (00:28.040)
to me, are inspiring and captivating.
对我来说，是鼓舞人心和迷人的。
Karl Friston (00:31.160)
This was a huge honor and a pleasure.
这是一项巨大的荣誉和荣幸。
Lex Fridman (00:34.120)
This is the Artificial Intelligence Podcast.
这是人工智能播客。
Karl Friston (00:36.800)
If you enjoy it, subscribe on YouTube,
如果您喜欢，请在 YouTube 上订阅，
Lex Fridman (00:38.920)
review it with five stars on Apple Podcast,
在 Apple Podcast 上以五颗星评价它，
Karl Friston (00:41.280)
support it on Patreon,
在 Patreon 上支持它，
Lex Fridman (00:42.560)
or simply connect with me on Twitter,
或者直接在 Twitter 上与我联系，
Karl Friston (00:44.560)
at Lex Friedman, spelled F R I D M A N.
在 Lex Friedman，拼写为 F R I D M A N。
Lex Fridman (00:48.100)
As usual, I'll do a few minutes of ads now,
像往常一样，我现在会做几分钟的广告，
Lex Fridman (00:50.480)
and never any ads in the middle
中间没有任何广告
Lex Fridman (00:52.040)
that can break the flow of the conversation.
这可能会破坏谈话的流畅性。
Karl Friston (00:54.220)
I hope that works for you,
我希望这对你有用
Lex Fridman (00:55.540)
and doesn't hurt the listening experience.
Karl Friston (00:58.280)
This show is presented by Cash App,
Lex Fridman (01:00.880)
the number one finance app in the App Store.
Karl Friston (01:03.200)
When you get it, use code LEXPODCAST.
Lex Fridman (01:06.280)
Cash App lets you send money to friends by Bitcoin,
Lex Fridman (01:09.400)
and invest in the stock market with as little as $1.
Lex Fridman (01:12.860)
Since Cash App allows you to send
Lex Fridman (01:14.480)
and receive money digitally,
Lex Fridman (01:16.200)
let me mention a surprising fact related to physical money.
Karl Friston (01:20.200)
Of all the currency in the world,
Lex Fridman (01:22.000)
roughly 8% of it is actual physical money.
Karl Friston (01:25.640)
The other 92% of money only exists digitally.
Lex Fridman (01:29.920)
So again, if you get Cash App from the App Store,
Karl Friston (01:32.260)
Google Play, and use the code LEXPODCAST, you get $10,
Lex Fridman (01:37.040)
and Cash App will also donate $10 to FIRST,
Karl Friston (01:39.880)
an organization that is helping to advance robotics
Lex Fridman (01:42.640)
and STEM education for young people around the world.
Lex Fridman (01:45.780)
And now, here's my conversation with Carl Friston.
Lex Fridman (01:50.020)
How much of the human brain do we understand
Karl Friston (01:53.380)
from the low level of neuronal communication
Lex Fridman (01:56.580)
to the functional level to the highest level,
Lex Fridman (02:01.660)
maybe the psychiatric disorder level?
Lex Fridman (02:04.940)
Well, we're certainly in a better position
Karl Friston (02:06.420)
than we were last century.
Lex Fridman (02:08.980)
How far we've got to go, I think,
Karl Friston (02:10.440)
is almost an unanswerable question.
Lex Fridman (02:13.520)
So you'd have to set the parameters,
Karl Friston (02:16.040)
you know, what constitutes understanding, what level
Lex Fridman (02:20.140)
of understanding do you want?
Karl Friston (02:21.740)
I think we've made enormous progress
Lex Fridman (02:25.380)
in terms of broad brush principles.
Karl Friston (02:29.140)
Whether that affords a detailed cartography
Lex Fridman (02:32.460)
of the functional anatomy of the brain and what it does,
Karl Friston (02:35.380)
right down to the microcircuitry and the neurons,
Lex Fridman (02:38.140)
that's probably out of reach at the present time.
Lex Fridman (02:42.220)
So the cartography, so mapping the brain,
Lex Fridman (02:44.900)
do you think mapping of the brain,
Karl Friston (02:47.440)
the detailed, perfect imaging of it,
Lex Fridman (02:50.380)
does that get us closer to understanding
Lex Fridman (02:54.100)
of the mind, of the brain?
Lex Fridman (02:56.060)
So how far does it get us if we have
Lex Fridman (02:59.100)
that perfect cartography of the brain?
Lex Fridman (03:01.580)
I think there are lower bounds on that.
Karl Friston (03:03.060)
It's a really interesting question.
Lex Fridman (03:06.340)
And it would determine the sort of scientific career
Karl Friston (03:09.220)
you'd pursue if you believe that knowing
Lex Fridman (03:11.640)
every dendritic connection, every sort of microscopic,
Karl Friston (03:16.020)
synaptic structure right down to the molecular level
Lex Fridman (03:18.740)
was gonna give you the right kind of information
Karl Friston (03:22.260)
to understand the computational anatomy,
Lex Fridman (03:25.340)
then you'd choose to be a microscopist
Lex Fridman (03:27.180)
and you would study little cubic millimeters of brain
Lex Fridman (03:32.340)
for the rest of your life.
Karl Friston (03:33.500)
If on the other hand you were interested
Lex Fridman (03:35.300)
in holistic functions and a sort of functional anatomy
Karl Friston (03:40.300)
of the sort that a neuropsychologist would understand,
Lex Fridman (03:44.100)
you'd study brain lesions and strokes,
Karl Friston (03:46.380)
just looking at the whole person.
Lex Fridman (03:48.460)
So again, it comes back to at what level
Lex Fridman (03:50.900)
do you want understanding?
Lex Fridman (03:52.980)
I think there are principled reasons not to go too far.
Karl Friston (03:57.900)
If you commit to a view of the brain
Lex Fridman (04:01.660)
as a machine that's performing a form of inference
Lex Fridman (04:06.660)
and representing things, that level of understanding
Lex Fridman (04:15.980)
is necessarily cast in terms of probability densities
Lex Fridman (04:20.540)
and ensemble densities, distributions.
Lex Fridman (04:24.180)
And what that tells you is that you don't really want
Karl Friston (04:27.120)
to look at the atoms to understand the thermodynamics
Lex Fridman (04:30.500)
of probabilistic descriptions of how the brain works.
Lex Fridman (04:34.420)
So I personally wouldn't look at the molecules
Lex Fridman (04:38.620)
or indeed the single neurons in the same way
Karl Friston (04:41.700)
if I wanted to understand the thermodynamics
Lex Fridman (04:44.060)
of some non equilibrium steady state of a gas
Karl Friston (04:47.220)
or an active material, I wouldn't spend my life
Lex Fridman (04:49.980)
looking at the individual molecules
Karl Friston (04:54.200)
that constitute that ensemble.
Lex Fridman (04:55.420)
I'd look at their collective behavior.
Karl Friston (04:57.500)
On the other hand, if you go too coarse grain,
Lex Fridman (05:00.140)
you're gonna miss some basic canonical principles
Karl Friston (05:03.900)
of connectivity and architectures.
Lex Fridman (05:06.300)
I'm thinking here this bit colloquial,
Lex Fridman (05:10.020)
but this current excitement about high field
Lex Fridman (05:13.540)
magnetic resonance imaging at seven Tesla, why?
Karl Friston (05:17.240)
Well, it gives us for the first time the opportunity
Lex Fridman (05:19.460)
to look at the brain in action at the level
Karl Friston (05:22.740)
of a few millimeters that distinguish
Lex Fridman (05:24.700)
between different layers of the cortex
Karl Friston (05:27.260)
that may be very important in terms of evincing
Lex Fridman (05:32.260)
generic principles of conical microcircuitry
Karl Friston (05:35.260)
that are replicated throughout the brain
Lex Fridman (05:37.620)
that may tell us something fundamental
Karl Friston (05:39.120)
about message passing in the brain
Lex Fridman (05:41.140)
and these density dynamics or neuronal
Karl Friston (05:44.260)
or some more population dynamics
Lex Fridman (05:46.820)
that underwrite our brain function.
Lex Fridman (05:49.420)
So somewhere between a millimeter and a meter.
Lex Fridman (05:53.920)
Lingering for a bit on the big questions if you allow me,
Lex Fridman (05:58.340)
what to use the most beautiful or surprising characteristic
Lex Fridman (06:01.940)
of the human brain?
Karl Friston (06:03.420)
I think it's hierarchical and recursive aspect.
Lex Fridman (06:06.780)
It's recurrent aspect.
Karl Friston (06:08.680)
Of the structure or of the actual
Lex Fridman (06:10.820)
representation of power of the brain?
Karl Friston (06:12.920)
Well, I think one speaks to the other.
Lex Fridman (06:15.380)
I was actually answering in a dull minded way
Karl Friston (06:18.420)
from the point of view of purely its anatomy
Lex Fridman (06:20.380)
and its structural aspects.
Karl Friston (06:22.980)
I mean, there are many marvelous organs in the body.
Lex Fridman (06:26.740)
Let's take your liver for example.
Karl Friston (06:28.460)
Without it, you wouldn't be around for very long
Lex Fridman (06:32.660)
and it does some beautiful and delicate biochemistry
Lex Fridman (06:35.980)
and homeostasis and evolved with a finesse
Lex Fridman (06:41.140)
that would easily parallel the brain
Lex Fridman (06:43.020)
but it doesn't have a beautiful anatomy.
Lex Fridman (06:45.140)
It has a simple anatomy which is attractive
Karl Friston (06:47.020)
in a minimalist sense but it doesn't have
Lex Fridman (06:48.940)
that crafted structure of sparse connectivity
Lex Fridman (06:52.820)
and that recurrence and that specialization
Lex Fridman (06:55.380)
that the brain has.
Lex Fridman (06:56.380)
So you said a lot of interesting terms here.
Lex Fridman (06:58.300)
So the recurrence, the sparsity,
Lex Fridman (07:00.700)
but you also started by saying hierarchical.
Lex Fridman (07:03.540)
So I've never thought of our brain as hierarchical.
Karl Friston (07:11.260)
Sort of I always thought it's just like a giant mess,
Lex Fridman (07:14.300)
interconnected mess where it's very difficult
Karl Friston (07:16.740)
to figure anything out.
Lex Fridman (07:18.180)
But in what sense do you see the brain as hierarchical?
Karl Friston (07:21.860)
Well, I see it, it's not a magic soup.
Lex Fridman (07:24.220)
Which of course is what I used to think
Karl Friston (07:28.700)
before I studied medicine and the like.
Lex Fridman (07:34.460)
So a lot of those terms imply each other.
Lex Fridman (07:38.740)
So hierarchies, if you just think about
Lex Fridman (07:41.700)
the nature of a hierarchy,
Lex Fridman (07:43.540)
how would you actually build one?
Lex Fridman (07:46.020)
And what you would have to do is basically
Karl Friston (07:48.300)
carefully remove the right connections
Lex Fridman (07:51.140)
that destroy the completely connected soups
Karl Friston (07:54.740)
that you might have in mind.
Lex Fridman (07:56.100)
So a hierarchy is in and of itself defined
Karl Friston (08:00.140)
by a sparse and particular connectivity structure.
Lex Fridman (08:04.420)
I'm not committing to any particular form of hierarchy.
Lex Fridman (08:08.380)
But your sense is there is some.
Lex Fridman (08:10.140)
Oh, absolutely, yeah.
Karl Friston (08:11.420)
In virtue of the fact that there is a sparsity
Lex Fridman (08:14.220)
of connectivity, not necessarily of a qualitative sort,
Lex Fridman (08:19.220)
but certainly of a quantitative sort.
Lex Fridman (08:20.460)
So it is demonstrably so that the further apart
Karl Friston (08:27.460)
two parts of the brain are,
Lex Fridman (08:29.220)
the less likely they are to be wired,
Karl Friston (08:32.260)
to possess axonal processes, neuronal processes
Lex Fridman (08:35.820)
that directly communicate one message
Karl Friston (08:39.100)
or messages from one part of that brain
Lex Fridman (08:41.140)
to the other part of the brain.
Lex Fridman (08:42.660)
So we know there's a sparse connectivity.
Lex Fridman (08:45.820)
And furthermore, on the basis of anatomical connectivity
Karl Friston (08:50.300)
in traces studies, we know that that sparsity
Lex Fridman (08:57.580)
underwrites a hierarchical and very structured
Karl Friston (09:00.980)
sort of connectivity that might be best understood
Lex Fridman (09:07.140)
like a little bit like an onion.
Karl Friston (09:09.380)
There is a concentric, sometimes referred to as centripetal
Lex Fridman (09:15.540)
by people like Marcel Masulam,
Karl Friston (09:17.500)
hierarchical organization to the brain.
Lex Fridman (09:19.340)
So you can think of the brain as in a rough sense,
Karl Friston (09:23.860)
like an onion, and all the sensory information
Lex Fridman (09:28.060)
and all the afferent outgoing messages
Karl Friston (09:31.260)
that supply commands to your muscles
Lex Fridman (09:33.500)
or to your secretory organs come from the surface.
Lex Fridman (09:37.100)
So there's a massive exchange interface
Lex Fridman (09:40.140)
with the world out there on the surface.
Lex Fridman (09:43.020)
And then underneath, there's a little layer
Lex Fridman (09:45.140)
that sits and looks at the exchange on the surface.
Lex Fridman (09:49.380)
And then underneath that, there's a layer
Lex Fridman (09:51.380)
right the way down to the very center,
Karl Friston (09:53.380)
to the deepest part of the onion.
Lex Fridman (09:55.740)
That's what I mean by a hierarchical organization.
Karl Friston (09:58.700)
There's a discernible structure defined
Lex Fridman (10:02.020)
by the sparsity of connections
Karl Friston (10:04.860)
that lends the architecture a hierarchical structure
Lex Fridman (10:08.620)
that tells one a lot about the kinds of representations
Lex Fridman (10:12.780)
and messages.
Lex Fridman (10:13.620)
Coming back to your earlier question,
Karl Friston (10:15.620)
is this about the representational capacity
Lex Fridman (10:19.140)
or is it about the anatomy?
Karl Friston (10:20.420)
Well, one underwrites the other.
Lex Fridman (10:24.580)
If one simply thinks of the brain
Karl Friston (10:26.380)
as a message passing machine,
Lex Fridman (10:29.100)
a process that is in the service of doing something,
Karl Friston (10:33.460)
then the circuitry and the connectivity
Lex Fridman (10:37.500)
that shape that message passing also dictate its function.
Lex Fridman (10:42.500)
So you've done a lot of amazing work
Lex Fridman (10:44.980)
in a lot of directions.
Lex Fridman (10:46.660)
So let's look at one aspect of that,
Lex Fridman (10:49.260)
of looking into the brain
Lex Fridman (10:51.060)
and trying to study this onion structure.
Lex Fridman (10:54.980)
What can we learn about the brain by imaging it?
Karl Friston (10:57.660)
Which is one way to sort of look at the anatomy of it.
Lex Fridman (11:00.860)
Broadly speaking, what are the methods of imaging,
Lex Fridman (11:04.700)
but even bigger, what can we learn about it?
Lex Fridman (11:07.620)
Right, so well, most human neuroimaging
Karl Friston (11:13.820)
that you might see in science journals
Lex Fridman (11:18.980)
that speaks to the way the brain works,
Karl Friston (11:22.540)
measures brain activity over time.
Lex Fridman (11:24.340)
So that's the first thing to say,
Karl Friston (11:26.020)
that we're effectively looking at fluctuations
Lex Fridman (11:30.020)
in neuronal responses,
Karl Friston (11:32.580)
usually in response to some sensory input
Lex Fridman (11:36.140)
or some instruction, some task.
Karl Friston (11:40.660)
Not necessarily, there's a lot of interest
Lex Fridman (11:42.260)
in just looking at the brain
Karl Friston (11:44.740)
in terms of resting state, endogenous,
Lex Fridman (11:46.860)
or intrinsic activity.
Lex Fridman (11:48.940)
But crucially, at every point,
Lex Fridman (11:50.780)
looking at these fluctuations,
Karl Friston (11:52.460)
either induced or intrinsic in the neural activity,
Lex Fridman (11:57.340)
and understanding them at two levels.
Lex Fridman (11:59.020)
So normally, people would recourse
Lex Fridman (12:03.540)
to two principles of brain organization
Karl Friston (12:06.260)
that are complementary.
Lex Fridman (12:07.420)
One, functional specialization or segregation.
Lex Fridman (12:10.460)
So what does that mean?
Lex Fridman (12:11.860)
It simply means that there are certain parts of the brain
Karl Friston (12:16.300)
that may be specialized for certain kinds of processing.
Lex Fridman (12:19.580)
For example, visual motion,
Karl Friston (12:21.900)
our ability to recognize or to perceive movement
Lex Fridman (12:26.060)
in the visual world.
Lex Fridman (12:27.780)
And furthermore, that specialized processing
Lex Fridman (12:31.100)
may be spatially or anatomically segregated,
Karl Friston (12:34.540)
leading to functional segregation.
Lex Fridman (12:37.060)
Which means that if I were to compare your brain activity
Karl Friston (12:40.900)
during a period of viewing a static image,
Lex Fridman (12:45.220)
and then compare that to the responses of fluctuations
Karl Friston (12:49.260)
in the brain when you were exposed to a moving image,
Lex Fridman (12:52.660)
say a flying bird,
Karl Friston (12:54.900)
we'd expect to see
Lex Fridman (12:56.300)
restricted, segregated differences in activity.
Lex Fridman (13:01.620)
And those are basically the hotspots
Lex Fridman (13:03.780)
that you see in the statistical parametric maps
Karl Friston (13:06.780)
that test for the significance of the responses
Lex Fridman (13:08.900)
that are circumscribed.
Lex Fridman (13:11.060)
So now, basically, we're talking about
Lex Fridman (13:13.740)
some people have perhaps unkindly called a neocartography.
Karl Friston (13:19.300)
This is a phrenology augmented by modern day neuroimaging,
Lex Fridman (13:24.420)
basically finding blobs or bumps on the brain
Karl Friston (13:28.980)
that do this or do that,
Lex Fridman (13:30.860)
and trying to understand the cartography
Karl Friston (13:33.100)
of that functional specialization.
Lex Fridman (13:35.700)
So how much is there such,
Karl Friston (13:38.860)
this is such a beautiful sort of ideal to strive for.
Lex Fridman (13:43.220)
We humans, scientists, would like this,
Karl Friston (13:46.860)
to hope that there's a beautiful structure to this
Lex Fridman (13:48.980)
where it's, like you said, there's segregated regions
Karl Friston (13:51.420)
that are responsible for the different function.
Lex Fridman (13:54.100)
How much hope is there to find such regions
Lex Fridman (13:57.100)
in terms of looking at the progress of studying the brain?
Lex Fridman (14:00.420)
Oh, I think enormous progress has been made
Karl Friston (14:02.740)
in the past 20 or 30 years.
Lex Fridman (14:06.860)
So this is beyond incremental.
Karl Friston (14:08.940)
At the advent of brain imaging,
Lex Fridman (14:11.940)
the very notion of functional segregation
Karl Friston (14:14.820)
was just a hypothesis based upon a century,
Lex Fridman (14:18.900)
if not more, of careful neuropsychology,
Karl Friston (14:21.420)
looking at people who had lost via insult
Lex Fridman (14:24.660)
or traumatic brain injury particular parts of the brain,
Lex Fridman (14:29.300)
and then saying, well, they can't do this
Lex Fridman (14:30.780)
or they can't do that.
Karl Friston (14:32.060)
For example, losing the visual cortex
Lex Fridman (14:34.100)
and not being able to see,
Karl Friston (14:35.620)
or losing particular parts of the visual cortex
Lex Fridman (14:39.580)
or regions known as V5
Karl Friston (14:44.060)
or the middle temporal region, MT,
Lex Fridman (14:47.980)
and noticing that they selectively
Karl Friston (14:49.580)
could not see moving things.
Lex Fridman (14:51.460)
And so that created the hypothesis
Karl Friston (14:55.140)
that perhaps visual movement processing
Lex Fridman (14:59.700)
was located in this functionally segregated area.
Lex Fridman (15:02.740)
And you could then go and put invasive electrodes
Lex Fridman (15:05.700)
in animal models and say, yes, indeed,
Karl Friston (15:08.260)
we can excite activity here.
Lex Fridman (15:10.500)
We can form receptive fields that are sensitive to
Karl Friston (15:13.980)
or defined in terms of visual motion.
Lex Fridman (15:16.220)
But at no point could you exclude the possibility
Karl Friston (15:18.340)
that everywhere else in the brain
Lex Fridman (15:20.020)
was also very interested in visual motion.
Karl Friston (15:23.140)
By the way, I apologize to interrupt,
Lex Fridman (15:24.900)
but a tiny little tangent.
Karl Friston (15:26.980)
You said animal models, just out of curiosity,
Lex Fridman (15:31.180)
from your perspective, how different is the human brain
Karl Friston (15:34.460)
versus the other animals
Lex Fridman (15:35.460)
in terms of our ability to study the brain?
Karl Friston (15:38.980)
Well, clearly, the further away you go from a human brain,
Lex Fridman (15:43.620)
the greater the differences,
Lex Fridman (15:45.740)
but not as remarkable as you might think.
Lex Fridman (15:48.940)
So people will choose their level of approximation
Karl Friston (15:53.180)
to the human brain,
Lex Fridman (15:54.300)
depending upon the kinds of questions
Karl Friston (15:56.540)
that they want to answer.
Lex Fridman (15:57.380)
So if you're talking about sort of canonical principles
Karl Friston (16:00.620)
of microcircuitry, it might be perfectly okay
Lex Fridman (16:02.780)
to look at a mouse, indeed.
Karl Friston (16:04.540)
You could even look at flies, worms.
Lex Fridman (16:08.100)
If, on the other hand, you wanted to look at
Karl Friston (16:09.740)
the finer details of organization of visual cortex
Lex Fridman (16:13.100)
and V1, V2, these are designated patches of cortex
Karl Friston (16:17.420)
that may do different things, indeed, do.
Lex Fridman (16:20.660)
You'd probably want to use a primate
Karl Friston (16:23.340)
that looked a little bit more like a human,
Lex Fridman (16:26.940)
because there are lots of ethical issues
Karl Friston (16:28.260)
in terms of the use of nonhuman primates
Lex Fridman (16:32.780)
to answer questions about human anatomy.
Lex Fridman (16:37.500)
But I think most people assume
Lex Fridman (16:39.060)
that most of the important principles are conserved
Karl Friston (16:43.300)
in a continuous way, right from, well, yes,
Lex Fridman (16:48.020)
worms right through to you and me.
Lex Fridman (16:53.260)
So now returning to, so that was the early sort of ideas
Lex Fridman (16:56.900)
of studying the functional regions of the brain
Karl Friston (17:00.820)
by if there's some damage to it,
Lex Fridman (17:02.860)
to try to infer that that part of the brain
Karl Friston (17:06.140)
might be somewhat responsible for this type of function.
Lex Fridman (17:09.100)
So where does that lead us?
Lex Fridman (17:11.140)
What are the next steps beyond that?
Lex Fridman (17:12.980)
Right, well, I'll just actually just reverse a bit,
Karl Friston (17:16.140)
come back to your sort of notion
Lex Fridman (17:17.860)
that the brain is a magic soup.
Karl Friston (17:19.580)
That was actually a very prominent idea at one point,
Lex Fridman (17:23.780)
notions such as Lashley's law of mass action
Karl Friston (17:28.140)
inherited from the observation that for certain animals,
Lex Fridman (17:33.140)
if you just took out spoonfuls of the brain,
Karl Friston (17:36.020)
it didn't matter where you took these spoonfuls out,
Lex Fridman (17:38.220)
they always showed the same kinds of deficits.
Lex Fridman (17:40.420)
So it was very difficult to infer functional specialization
Lex Fridman (17:44.540)
purely on the basis of lesion deficit studies.
Lex Fridman (17:49.260)
But once we had the opportunity
Lex Fridman (17:50.820)
to look at the brain lighting up
Lex Fridman (17:52.380)
and it's literally it's sort of excitement, neuronal excitement
Lex Fridman (17:57.740)
when looking at this versus that,
Karl Friston (18:01.100)
one was able to say, yes, indeed,
Lex Fridman (18:03.260)
these functionally specialized responses
Karl Friston (18:05.020)
are very restricted and they're here or they're over there.
Lex Fridman (18:08.780)
If I do this, then this part of the brain lights up.
Lex Fridman (18:11.340)
And that became doable in the early 90s.
Lex Fridman (18:16.900)
In fact, shortly before with the advent
Karl Friston (18:19.100)
of positron emission tomography.
Lex Fridman (18:21.060)
And then functional magnetic resonance imaging
Karl Friston (18:23.420)
came along in the early 90s.
Lex Fridman (18:26.700)
And since that time, there has been an explosion
Karl Friston (18:29.780)
of discovery, refinement, confirmation.
Lex Fridman (18:36.300)
There are people who believe that it's all in the anatomy.
Karl Friston (18:38.940)
If you understand the anatomy,
Lex Fridman (18:40.260)
then you understand the function at some level.
Lex Fridman (18:43.260)
And many, many hypotheses were predicated
Lex Fridman (18:45.780)
on a deep understanding of the anatomy and the connectivity,
Lex Fridman (18:49.980)
but they were all confirmed
Lex Fridman (18:51.260)
and taken much further with neuroimaging.
Lex Fridman (18:55.140)
So that's what I meant by we've made an enormous amount
Lex Fridman (18:57.580)
of progress in this century indeed,
Lex Fridman (19:01.180)
and in relation to the previous century,
Lex Fridman (19:03.940)
by looking at these functionally selective responses.
Lex Fridman (19:09.620)
But that wasn't the whole story.
Lex Fridman (19:11.100)
So there's this sort of near phrenology,
Lex Fridman (19:13.580)
but finding bumps and hot spots in the brain
Lex Fridman (19:15.620)
that did this or that.
Karl Friston (19:17.620)
The bigger question was, of course,
Lex Fridman (19:20.020)
the functional integration.
Lex Fridman (19:22.140)
How all of these regionally specific responses
Lex Fridman (19:26.980)
were orchestrated, how they were distributed,
Lex Fridman (19:29.860)
how did they relate to distributed processing
Lex Fridman (19:32.700)
and indeed representations in the brain.
Lex Fridman (19:35.580)
So then you turn to the more challenging issue
Lex Fridman (19:39.540)
of the integration, the connectivity.
Lex Fridman (19:42.660)
And then we come back to this beautiful,
Lex Fridman (19:44.980)
sparse, recurrent, hierarchical connectivity
Karl Friston (19:49.060)
that seems characteristic of the brain
Lex Fridman (19:51.140)
and probably not many other organs.
Lex Fridman (19:53.620)
But nevertheless, we come back to this challenge
Lex Fridman (19:58.260)
of trying to figure out how everything is integrated.
Karl Friston (1:00:01.720)
into different sets of states to exist,
Lex Fridman (1:00:04.440)
the active states, the Markov blanket
Karl Friston (1:00:06.800)
cannot be influenced by the external states.
Lex Fridman (1:00:09.800)
And we already know that the internal states
Karl Friston (1:00:11.560)
can't be influenced by the external states
Lex Fridman (1:00:13.640)
because the blanket separates them.
Lex Fridman (1:00:15.720)
So what does that mean?
Lex Fridman (1:00:16.560)
Well, it means the active states, the internal states
Karl Friston (1:00:19.280)
are now jointly not influenced by external states.
Lex Fridman (1:00:23.440)
They only have autonomous dynamics.
Lex Fridman (1:00:26.120)
So now you've got a picture of an oil drop
Lex Fridman (1:00:30.000)
that has autonomy, it has autonomous states,
Karl Friston (1:00:34.040)
it has autonomous states in the sense
Lex Fridman (1:00:35.440)
that there must be some parts of the surface of the oil drop
Karl Friston (1:00:38.320)
that are not influenced by the external states
Lex Fridman (1:00:40.400)
and all the interior.
Lex Fridman (1:00:41.880)
And together, those two states endow
Lex Fridman (1:00:44.520)
even a little oil drop with autonomous states
Karl Friston (1:00:47.640)
that look as if they are optimizing
Lex Fridman (1:00:51.360)
their variational free energy or their negative elbow,
Karl Friston (1:00:56.200)
their moral evidence.
Lex Fridman (1:00:59.400)
And that would be an interesting intellectual exercise.
Lex Fridman (1:01:03.240)
And you could say, you could even go into the realms
Lex Fridman (1:01:05.600)
of panpsychism, that everything that exists
Karl Friston (1:01:08.160)
is implicitly making inferences on self evidencing.
Lex Fridman (1:01:13.000)
Now we make the next move, but what about living things?
Karl Friston (1:01:17.040)
I mean, so let me ask you,
Lex Fridman (1:01:19.200)
what's the difference between an oil drop
Lex Fridman (1:01:21.600)
and a little tadpole or a little lava or a plankton?
Lex Fridman (1:01:27.200)
The picture was just painted of an oil drop.
Karl Friston (1:01:30.720)
Just immediately in a matter of minutes
Lex Fridman (1:01:32.840)
took me into the world of panpsychism,
Karl Friston (1:01:35.200)
where you've just convinced me,
Lex Fridman (1:01:38.040)
made me feel like an oil drop is a living,
Karl Friston (1:01:41.240)
it's certainly an autonomous system,
Lex Fridman (1:01:43.400)
but almost a living system.
Lex Fridman (1:01:44.720)
So it has sensory capabilities and acting capabilities
Lex Fridman (1:01:48.960)
and it maintains something.
Lex Fridman (1:01:50.600)
So what is the difference between that
Lex Fridman (1:01:53.920)
and something that we traditionally
Lex Fridman (1:01:56.120)
think of as a living system?
Lex Fridman (1:01:59.880)
That it could die or it can't,
Karl Friston (1:02:02.200)
I mean, yeah, mortality, I'm not exactly sure.
Lex Fridman (1:02:05.240)
I'm not sure what the right answer there is
Karl Friston (1:02:08.640)
because they can move,
Lex Fridman (1:02:09.640)
like movement seems like an essential element
Karl Friston (1:02:11.840)
to being able to act in the environment,
Lex Fridman (1:02:13.520)
but the oil drop is doing that.
Lex Fridman (1:02:15.800)
So I don't know.
Lex Fridman (1:02:16.640)
Is it?
Karl Friston (1:02:18.120)
The oil drop will be moved,
Lex Fridman (1:02:19.720)
but does it in and of itself move autonomously?
Karl Friston (1:02:22.640)
Well, the surface is performing actions
Lex Fridman (1:02:26.880)
that maintain its structure.
Karl Friston (1:02:29.680)
Yeah, you're being too clever.
Lex Fridman (1:02:30.800)
I was, I had in mind a passive little oil drop
Karl Friston (1:02:34.560)
that's sitting there at the bottom
Lex Fridman (1:02:37.680)
on the top of a glass of water.
Karl Friston (1:02:39.240)
Sure, I guess.
Lex Fridman (1:02:40.600)
What I'm trying to say is you're absolutely right.
Karl Friston (1:02:42.200)
You've nailed it.
Lex Fridman (1:02:44.400)
It's movement.
Lex Fridman (1:02:45.880)
So where does that movement come from?
Lex Fridman (1:02:47.200)
If it comes from the inside,
Karl Friston (1:02:49.400)
then you've got, I think, something that's living.
Lex Fridman (1:02:53.320)
What do you mean from the inside?
Lex Fridman (1:02:54.680)
What I mean is that the internal states
Lex Fridman (1:02:58.840)
that can influence the active states,
Karl Friston (1:03:01.040)
where the active states can influence,
Lex Fridman (1:03:02.600)
but they're not influenced by the external states,
Karl Friston (1:03:05.120)
can cause movement.
Lex Fridman (1:03:07.160)
So there are two types of oil drops, if you like.
Karl Friston (1:03:10.440)
There are oil drops where the internal states
Lex Fridman (1:03:12.800)
are so random that they average themselves away,
Lex Fridman (1:03:20.320)
and the thing cannot, on average,
Lex Fridman (1:03:23.800)
when you do the averaging, move.
Lex Fridman (1:03:25.960)
So a nice example of that would be the Sun.
Lex Fridman (1:03:29.400)
The Sun certainly has internal states.
Karl Friston (1:03:31.160)
There's lots of intrinsic autonomous activity going on,
Lex Fridman (1:03:34.360)
but because it's not coordinated,
Karl Friston (1:03:35.840)
because it doesn't have the deep, in the millennial sense,
Lex Fridman (1:03:38.080)
the hierarchical structure that the brain does,
Karl Friston (1:03:40.960)
there is no overall mode or pattern or organization
Lex Fridman (1:03:45.800)
that expresses itself on the surface
Karl Friston (1:03:48.160)
that allows it to actually swim.
Lex Fridman (1:03:51.400)
It can certainly have a very active surface,
Lex Fridman (1:03:54.080)
but en masse, at the scale of the actual surface of the Sun,
Lex Fridman (1:03:58.280)
the average position of that surface cannot, in itself, move,
Karl Friston (1:04:02.920)
because the internal dynamics are more like a hot gas.
Lex Fridman (1:04:06.680)
They are literally like a hot gas,
Karl Friston (1:04:08.480)
whereas your internal dynamics are much more structured
Lex Fridman (1:04:11.440)
and deeply structured,
Lex Fridman (1:04:12.880)
and now you can express on your active states
Lex Fridman (1:04:16.440)
with your muscles and your secretory organs,
Karl Friston (1:04:19.720)
your autonomic nervous system and its effectors,
Lex Fridman (1:04:22.920)
you can actually move, and that's all you can do.
Lex Fridman (1:04:26.760)
And that's something which,
Lex Fridman (1:04:28.240)
if you haven't thought of it like this before,
Karl Friston (1:04:30.440)
I think it's nice to just realize
Lex Fridman (1:04:32.440)
there is no other way that you can change the universe
Karl Friston (1:04:37.040)
other than simply moving.
Lex Fridman (1:04:39.240)
Whether that moving is articulating with my voice box
Karl Friston (1:04:43.840)
or walking around or squeezing juices
Lex Fridman (1:04:46.600)
out of my secretory organs,
Karl Friston (1:04:48.720)
there's only one way you can change the universe.
Lex Fridman (1:04:51.960)
It's moving.
Lex Fridman (1:04:53.240)
And the fact that you do so nonrandomly makes you alive.
Lex Fridman (1:04:58.600)
Yeah, so it's that nonrandomness.
Lex Fridman (1:05:00.600)
And that would be manifested,
Lex Fridman (1:05:04.840)
we realize, in terms of essentially swimming,
Karl Friston (1:05:07.800)
essentially moving, changing one's shape,
Lex Fridman (1:05:10.480)
a morphogenesis that is dynamic and possibly adaptive.
Lex Fridman (1:05:15.760)
So that's what I was trying to get at
Lex Fridman (1:05:17.920)
between the difference between the oil drop
Lex Fridman (1:05:19.160)
and the little tadpole.
Lex Fridman (1:05:21.000)
The tadpole is moving around.
Karl Friston (1:05:23.000)
Its active states are actually changing the external states.
Lex Fridman (1:05:26.920)
And there's now a cycle,
Karl Friston (1:05:28.480)
an action perception cycle, if you like,
Lex Fridman (1:05:30.400)
a recurrent dynamic that's going on
Karl Friston (1:05:34.040)
that depends upon this deeply structured autonomous behavior
Lex Fridman (1:05:39.360)
that rests upon internal dynamics
Karl Friston (1:05:44.600)
that are not only modeling
Lex Fridman (1:05:48.880)
the data impressed upon their surface or the blanket states,
Lex Fridman (1:05:53.800)
but they are actively resampling those data by moving.
Lex Fridman (1:05:58.800)
They're moving towards chemical gradients and chemotaxis.
Lex Fridman (1:06:03.920)
So they've gone beyond just being good little models
Lex Fridman (1:06:08.280)
of the kind of world they live in.
Karl Friston (1:06:11.120)
For example, an oil droplet could, in a panpsychic sense,
Lex Fridman (1:06:15.960)
be construed as a little being
Karl Friston (1:06:18.440)
that has now perfectly inferred.
Lex Fridman (1:06:20.720)
It's a passive, nonliving oil drop
Karl Friston (1:06:23.800)
living in a bowl of water.
Lex Fridman (1:06:25.600)
No problem.
Lex Fridman (1:06:27.960)
But to now equip that oil drop with the ability to go out
Lex Fridman (1:06:31.040)
and test that hypothesis about different states of beings.
Lex Fridman (1:06:34.040)
So it can actually push its surface over there, over there,
Lex Fridman (1:06:36.920)
and test for chemical gradients,
Karl Friston (1:06:38.680)
or then you start to move to a much more lifelike form.
Lex Fridman (1:06:42.880)
This is all fun, theoretically interesting,
Lex Fridman (1:06:44.920)
but it actually is quite important
Lex Fridman (1:06:47.000)
in terms of reflecting what I have seen
Karl Friston (1:06:49.800)
since the turn of the millennium,
Lex Fridman (1:06:53.120)
which is this move towards an inactive
Lex Fridman (1:06:56.600)
and embodied understanding of intelligence.
Lex Fridman (1:07:00.760)
And you say you're from machine learning.
Lex Fridman (1:07:03.760)
So what that means,
Lex Fridman (1:07:05.720)
the central importance of movement,
Karl Friston (1:07:10.680)
I think has yet to really hit machine learning.
Lex Fridman (1:07:14.000)
It certainly has now diffused itself throughout robotics.
Lex Fridman (1:07:20.400)
And perhaps you could say certain problems in active vision
Lex Fridman (1:07:23.200)
where you actually have to move the camera
Karl Friston (1:07:25.360)
to sample this and that.
Lex Fridman (1:07:27.240)
But machine learning of the data mining deep learning sort
Karl Friston (1:07:31.600)
simply hasn't contended with this issue.
Lex Fridman (1:07:34.000)
What it's done, instead of dealing with the movement problem
Lex Fridman (1:07:37.240)
and the active sampling of data,
Lex Fridman (1:07:39.160)
it's just said, we don't need to worry about,
Karl Friston (1:07:40.720)
we can see all the data because we've got big data.
Lex Fridman (1:07:43.120)
So we can ignore movement.
Lex Fridman (1:07:45.120)
So that for me is an important omission
Lex Fridman (1:07:50.920)
in current machine learning.
Karl Friston (1:07:52.200)
The current machine learning is much more like the oil drop.
Lex Fridman (1:07:54.800)
Yes.
Lex Fridman (1:07:55.960)
But an oil drop that enjoys exposure
Lex Fridman (1:07:59.480)
to nearly all the data that it will ever need to be exposed to,
Karl Friston (1:08:03.600)
as opposed to the tadpoles swimming out
Lex Fridman (1:08:05.760)
to find the right data.
Karl Friston (1:08:07.360)
For example, it likes food.
Lex Fridman (1:08:10.320)
That's a good hypothesis.
Karl Friston (1:08:11.240)
Let's test it out.
Lex Fridman (1:08:12.080)
Let's go and move and ingest food, for example,
Lex Fridman (1:08:15.720)
and see is that evidence that I'm the kind of thing
Lex Fridman (1:08:18.680)
that likes this kind of food.
Lex Fridman (1:08:20.400)
So the next natural question, and forgive this question,
Lex Fridman (1:08:24.000)
but if we think of sort of even artificial intelligence
Karl Friston (1:08:27.120)
systems, which I just painted a beautiful picture
Lex Fridman (1:08:29.440)
of existence and life.
Lex Fridman (1:08:32.840)
So do you ascribe, do you find within this framework
Lex Fridman (1:08:39.920)
a possibility of defining consciousness
Lex Fridman (1:08:45.240)
or exploring the idea of consciousness?
Lex Fridman (1:08:47.360)
Like what, you know, self awareness
Lex Fridman (1:08:52.760)
and expand it to consciousness?
Lex Fridman (1:08:55.160)
Yeah.
Lex Fridman (1:08:56.160)
How can we start to think about consciousness
Lex Fridman (1:08:58.840)
within this framework?
Lex Fridman (1:08:59.720)
Is it possible?
Lex Fridman (1:09:00.920)
Well, yeah, I think it's possible to think about it,
Karl Friston (1:09:03.160)
whether you'll get it.
Lex Fridman (1:09:04.320)
Get anywhere is another question.
Lex Fridman (1:09:06.400)
And again, I'm not sure that I'm licensed
Lex Fridman (1:09:10.320)
to answer that question.
Karl Friston (1:09:12.760)
I think you'd have to speak to a qualified philosopher
Lex Fridman (1:09:15.120)
to get a definitive answer to that.
Lex Fridman (1:09:17.320)
But certainly, there's a lot of interest
Lex Fridman (1:09:19.680)
in using not just these ideas, but related ideas
Karl Friston (1:09:23.400)
from information theory to try and tie down
Lex Fridman (1:09:27.960)
the maths and the calculus and the geometry of consciousness,
Karl Friston (1:09:34.080)
either in terms of sort of a minimal consciousness,
Lex Fridman (1:09:39.720)
even less than a minimal selfhood.
Lex Fridman (1:09:42.440)
And what I'm talking about is the ability, effectively,
Lex Fridman (1:09:48.320)
to plan, to have agency.
Lex Fridman (1:09:52.440)
So you could argue that a virus does have a form of agency
Lex Fridman (1:09:57.760)
in virtue of the way that it selectively
Karl Friston (1:10:00.360)
finds hosts and cells to live in and moves around.
Lex Fridman (1:10:05.880)
But you wouldn't endow it with the capacity
Karl Friston (1:10:09.280)
to think about planning and moving in a purposeful way
Lex Fridman (1:10:14.880)
where it countenances the future.
Karl Friston (1:10:17.280)
Whereas you might an ant.
Lex Fridman (1:10:18.680)
You might think an ant's not quite as unconscious
Karl Friston (1:10:22.040)
as a virus.
Lex Fridman (1:10:24.280)
It certainly seems to have a purpose.
Karl Friston (1:10:26.160)
It talks to its friends en route during its foraging.
Lex Fridman (1:10:29.680)
It has a different kind of autonomy, which is biotic,
Lex Fridman (1:10:37.160)
but beyond a virus.
Lex Fridman (1:10:38.720)
So there's something about, so there's
Karl Friston (1:10:41.400)
some line that has to do with the complexity of planning
Lex Fridman (1:10:45.680)
that may contain an answer.
Karl Friston (1:10:48.000)
I mean, it would be beautiful if we
Lex Fridman (1:10:49.960)
can find a line beyond which we could say a being is conscious.
Karl Friston (1:10:55.480)
Yes, it will be.
Lex Fridman (1:10:56.560)
These are wonderful lines that we've drawn with existence,
Karl Friston (1:11:00.800)
life, and consciousness.
Lex Fridman (1:11:02.680)
Yes, it will be very nice.
Karl Friston (1:11:05.320)
One little wrinkle there, and this
Lex Fridman (1:11:07.280)
is something I've only learned in the past few months,
Karl Friston (1:11:09.400)
is the philosophical notion of vagueness.
Lex Fridman (1:11:12.400)
So you're saying it would be wonderful to draw a line.
Karl Friston (1:11:14.840)
I had always assumed that that line at some point
Lex Fridman (1:11:18.080)
would be drawn until about four months ago,
Lex Fridman (1:11:22.720)
and the philosopher taught me about vagueness.
Lex Fridman (1:11:24.880)
So I don't know if you've come across this,
Lex Fridman (1:11:26.720)
but it's a technical concept.
Lex Fridman (1:11:28.440)
And I think most revealingly illustrated with,
Lex Fridman (1:11:33.480)
at what point does a pile of sand become a pile?
Lex Fridman (1:11:37.120)
Is it one grain, two grains, three grains, or four grains?
Lex Fridman (1:11:41.620)
So at what point would you draw the line
Lex Fridman (1:11:44.200)
between being a pile of sand and a collection of grains of sand?
Karl Friston (1:11:51.320)
In the same way, is it right to ask,
Lex Fridman (1:11:53.520)
where would I draw the line between conscious
Lex Fridman (1:11:55.560)
and unconscious?
Lex Fridman (1:11:56.740)
And it might be a vague concept.
Karl Friston (1:11:59.680)
Having said that, I agree with you entirely.
Lex Fridman (1:12:02.920)
Systems that have the ability to plan.
Lex Fridman (1:12:06.480)
So just technically, what that means
Lex Fridman (1:12:08.440)
is your inferential self evidencing,
Karl Friston (1:12:13.880)
by which I simply mean the thermodynamics and gradient
Lex Fridman (1:12:19.880)
flows that underwrite the preservation of your oil
Karl Friston (1:12:22.960)
droplet like form, can be described
Lex Fridman (1:12:29.080)
as an optimization of log Bayesian model
Karl Friston (1:12:32.560)
evidence, your elbow.
Lex Fridman (1:12:36.680)
That self evidencing must be evidence
Karl Friston (1:12:39.760)
for a model of what's causing the sensory impressions
Lex Fridman (1:12:44.000)
on the sensory part of your surface or your Markov
Karl Friston (1:12:47.040)
blanket.
Lex Fridman (1:12:48.520)
If that model is capable of planning,
Karl Friston (1:12:51.160)
it must include a model of the future consequences
Lex Fridman (1:12:53.800)
of your active states or your action, just planning.
Lex Fridman (1:12:56.800)
So we're now in the game of planning as inference.
Lex Fridman (1:12:59.200)
Now notice what we've made, though.
Karl Friston (1:13:00.620)
We've made quite a big move away from big data and machine
Lex Fridman (1:13:04.400)
learning, because again, it's the consequences of moving.
Karl Friston (1:13:08.520)
It's the consequences of selecting those data or those
Lex Fridman (1:13:11.600)
data or looking over there.
Lex Fridman (1:13:14.240)
And that tells you immediately that even
Lex Fridman (1:13:17.420)
to be a contender for a conscious artifact or a strong
Karl Friston (1:13:22.760)
AI or generalized, I don't know what that's called nowadays,
Lex Fridman (1:13:26.760)
then you've got to have movement in the game.
Lex Fridman (1:13:29.280)
And furthermore, you've got to have a generative model
Lex Fridman (1:13:32.560)
of the sort you might find in, say, a variational auto
Karl Friston (1:13:34.920)
encoder that is thinking about the future conditioned
Lex Fridman (1:13:39.840)
upon different courses of action.
Karl Friston (1:13:41.900)
Now that brings a number of things to the table, which
Lex Fridman (1:13:44.400)
now you start to think, well, those
Karl Friston (1:13:46.280)
have got all the right ingredients
Lex Fridman (1:13:47.400)
to talk about consciousness.
Karl Friston (1:13:48.640)
I've now got to select among a number of different courses
Lex Fridman (1:13:51.600)
of action into the future as part of planning.
Karl Friston (1:13:54.920)
I've now got free will.
Lex Fridman (1:13:56.440)
The act of selecting this course of action or that policy
Karl Friston (1:13:59.440)
or that policy or that action suddenly
Lex Fridman (1:14:02.480)
makes me into an inference machine,
Karl Friston (1:14:04.720)
a self evidencing artifact that now
Lex Fridman (1:14:09.480)
looks as if it's selecting amongst different alternative
Karl Friston (1:14:12.480)
ways forward, as I actively swim here or swim there
Lex Fridman (1:14:15.480)
or look over here, look over there.
Lex Fridman (1:14:17.920)
So I think you've now got to a situation
Lex Fridman (1:14:19.920)
if there is planning in the mix.
Karl Friston (1:14:22.240)
You're now getting much closer to that line
Lex Fridman (1:14:25.200)
if that line were ever to exist.
Karl Friston (1:14:27.360)
I don't think it gets you quite as far as self aware, though.
Lex Fridman (1:14:32.400)
And then you have to, I think, grapple with the question,
Lex Fridman (1:14:39.680)
how would formally write down a calculus or a maths
Lex Fridman (1:14:43.160)
of self awareness?
Karl Friston (1:14:44.620)
I don't think it's impossible to do.
Lex Fridman (1:14:47.840)
But I think there would be pressure on you
Karl Friston (1:14:51.400)
to actually commit to a formal definition of what
Lex Fridman (1:14:53.360)
you mean by self awareness.
Karl Friston (1:14:55.920)
I think most people that I know would probably
Lex Fridman (1:15:00.480)
say that a goldfish, a pet fish, was not self aware.
Karl Friston (1:15:07.000)
They would probably argue about their favorite cat,
Lex Fridman (1:15:10.600)
but would be quite happy to say that their mom was self aware.
Karl Friston (1:15:14.000)
So.
Lex Fridman (1:15:15.480)
I mean, but that might very well connect
Karl Friston (1:15:17.400)
to some level of complexity with planning.
Lex Fridman (1:15:20.960)
It seems like self awareness is essential for complex planning.
Karl Friston (1:15:26.520)
Yeah.
Lex Fridman (1:15:27.240)
Do you want to take that further?
Karl Friston (1:15:28.080)
Because I think you're absolutely right.
Lex Fridman (1:15:29.880)
Again, the line is unclear, but it
Karl Friston (1:15:31.400)
seems like integrating yourself into the world,
Lex Fridman (1:15:36.400)
into your planning, is essential for constructing complex plans.
Karl Friston (1:15:42.320)
Yes.
Lex Fridman (1:15:43.080)
Yeah.
Lex Fridman (1:15:43.720)
So mathematically describing that in the same elegant way
Lex Fridman (1:15:47.440)
as you have with the free energy principle might be difficult.
Karl Friston (1:15:51.120)
Well, yes and no.
Lex Fridman (1:15:53.360)
I don't think that, well, perhaps we should just,
Lex Fridman (1:15:55.360)
can we just go back?
Lex Fridman (1:15:57.000)
That's a very important answer you gave.
Lex Fridman (1:15:58.720)
And I think if I just unpacked it,
Lex Fridman (1:16:01.880)
you'd see the truisms that you've just exposed for us.
Lex Fridman (1:16:06.480)
But let me, sorry, I'm mindful that I didn't answer
Lex Fridman (1:16:10.400)
your question before.
Lex Fridman (1:16:11.360)
Well, what's the free energy principle good for?
Lex Fridman (1:16:13.880)
Is it just a pretty theoretical exercise
Lex Fridman (1:16:15.680)
to explain nonequilibrium steady states?
Lex Fridman (1:16:17.880)
Yes, it is.
Karl Friston (1:16:19.400)
It does nothing more for you than that.
Lex Fridman (1:16:21.360)
It can be regarded, it's going to sound very arrogant,
Lex Fridman (1:16:24.040)
but it is of the sort of theory of natural selection,
Lex Fridman (1:16:27.880)
or a hypothesis of natural selection.
Karl Friston (1:16:32.800)
Beautiful, undeniably true, but tells you
Lex Fridman (1:16:37.200)
absolutely nothing about why you have legs and eyes.
Karl Friston (1:16:42.040)
It tells you nothing about the actual phenotype,
Lex Fridman (1:16:44.720)
and it wouldn't allow you to build something.
Lex Fridman (1:16:48.080)
So the free energy principle by itself
Lex Fridman (1:16:51.120)
is as vacuous as most tautological theories.
Lex Fridman (1:16:54.880)
And by tautological, of course,
Lex Fridman (1:16:56.200)
I'm talking to the theory of natural,
Karl Friston (1:16:58.800)
the survival of the fittest.
Lex Fridman (1:17:00.080)
What's the fittest of those that survive?
Lex Fridman (1:17:01.760)
Why do they cycle?
Lex Fridman (1:17:02.600)
It's the fitter.
Karl Friston (1:17:03.440)
It just goes around in circles.
Lex Fridman (1:17:05.640)
In a sense, the free energy principle has that same
Karl Friston (1:17:08.560)
deflationary tautology under the hood.
Lex Fridman (1:17:15.000)
It's a characteristic of things that exist.
Lex Fridman (1:17:17.680)
Why do they exist?
Lex Fridman (1:17:18.520)
Because they minimize their free energy.
Lex Fridman (1:17:19.720)
Why do they minimize their free energy?
Lex Fridman (1:17:21.400)
Because they exist.
Lex Fridman (1:17:22.440)
And you just keep on going round and round and round.
Lex Fridman (1:17:24.720)
But the practical thing,
Karl Friston (1:17:28.080)
which you don't get from natural selection,
Lex Fridman (1:17:32.680)
but you could say has now manifest in things
Karl Friston (1:17:35.680)
like differential evolution or genetic algorithms
Lex Fridman (1:17:38.160)
and MCMC, for example, in machine learning.
Karl Friston (1:17:41.320)
The practical thing you can get is,
Lex Fridman (1:17:43.240)
if it looks as if things that exist
Karl Friston (1:17:45.440)
are trying to have density dynamics
Lex Fridman (1:17:49.400)
and look as though they're optimizing
Karl Friston (1:17:51.560)
a variational free energy,
Lex Fridman (1:17:53.320)
and a variational free energy has to be
Karl Friston (1:17:55.200)
a functional of a generative model,
Lex Fridman (1:17:57.280)
a probabilistic description of causes and consequences,
Karl Friston (1:18:01.720)
causes out there, consequences in the sensorium
Lex Fridman (1:18:04.560)
on the sensory parts of the Markov blanket,
Karl Friston (1:18:07.080)
then it should, in theory, be possible
Lex Fridman (1:18:08.680)
to write down the generative model,
Karl Friston (1:18:10.400)
work out the gradients,
Lex Fridman (1:18:11.840)
and then cause it to autonomously self evidence.
Lex Fridman (1:18:15.920)
So you should be able to write down oil droplets.
Lex Fridman (1:18:18.160)
You should be able to create artifacts
Karl Friston (1:18:20.080)
where you have supplied the objective function
Lex Fridman (1:18:24.240)
that supplies the gradients,
Karl Friston (1:18:25.560)
that supplies the self organizing dynamics
Lex Fridman (1:18:28.320)
to non equilibrium steady state.
Lex Fridman (1:18:30.280)
So there is actually a practical application
Lex Fridman (1:18:32.680)
of the free energy principle
Karl Friston (1:18:34.120)
when you can write down your required evidence
Lex Fridman (1:18:37.800)
in terms of, well, when you can write down
Karl Friston (1:18:40.400)
the generative model that is the thing
Lex Fridman (1:18:43.720)
that has the evidence.
Karl Friston (1:18:44.800)
The probability of these sensory data
Lex Fridman (1:18:46.760)
or this data, given that model,
Karl Friston (1:18:50.880)
is effectively the thing that the elbow
Lex Fridman (1:18:54.200)
or the variational free energy bounds or approximates.
Karl Friston (1:18:57.400)
That means that you can actually write down the model
Lex Fridman (1:19:01.720)
and the kind of thing that you want to engineer,
Karl Friston (1:19:04.640)
the kind of AGI or artificial general intelligence
Lex Fridman (1:19:10.440)
that you want to manifest probabilistically,
Lex Fridman (1:19:14.480)
and then you engineer, a lot of hard work,
Lex Fridman (1:19:16.720)
but you would engineer a robot and a computer
Karl Friston (1:19:19.760)
to perform a gradient descent on that objective function.
Lex Fridman (1:19:23.400)
So it does have a practical implication.
Lex Fridman (1:19:26.160)
Now, why am I wittering on about that?
Lex Fridman (1:19:27.440)
It did seem relevant to, yes.
Lex Fridman (1:19:28.880)
So what kinds of, so the answer to,
Lex Fridman (1:19:32.920)
would it be easier or would it be hard?
Karl Friston (1:19:34.240)
Well, mathematically, it's easy.
Lex Fridman (1:19:36.200)
I've just told you all you need to do
Karl Friston (1:19:38.000)
is write down your perfect artifact,
Lex Fridman (1:19:43.880)
probabilistically, in the form
Karl Friston (1:19:45.280)
of a probabilistic generative model,
Lex Fridman (1:19:46.480)
a probability distribution over the causes
Lex Fridman (1:19:48.800)
and consequences of the world
Lex Fridman (1:19:52.520)
in which this thing is immersed.
Lex Fridman (1:19:54.680)
And then you just engineer a computer and a robot
Lex Fridman (1:19:58.040)
to perform a gradient descent on that objective function.
Karl Friston (1:20:00.800)
No problem.
Lex Fridman (1:20:02.720)
But of course, the big problem
Karl Friston (1:20:04.160)
is writing down the generative model.
Lex Fridman (1:20:05.960)
So that's where the heavy lifting comes in.
Lex Fridman (1:20:08.040)
So it's the form and the structure of that generative model
Lex Fridman (1:20:12.160)
which basically defines the artifact that you will create
Karl Friston (1:20:15.640)
or, indeed, the kind of artifact that has self awareness.
Lex Fridman (1:20:19.600)
So that's where all the hard work comes,
Karl Friston (1:20:22.080)
very much like natural selection doesn't tell you
Lex Fridman (1:20:24.640)
in the slightest why you have eyes.
Lex Fridman (1:20:27.000)
So you have to drill down on the actual phenotype,
Lex Fridman (1:20:29.480)
the actual generative model.
Lex Fridman (1:20:31.480)
So with that in mind, what did you tell me
Lex Fridman (1:20:36.360)
that tells me immediately the kinds of generative models
Lex Fridman (1:20:40.720)
I would have to write down in order to have self awareness?
Lex Fridman (1:20:43.480)
What you said to me was I have to have a model
Karl Friston (1:20:48.200)
that is effectively fit for purpose
Lex Fridman (1:20:50.440)
for this kind of world in which I operate.
Lex Fridman (1:20:53.680)
And if I now make the observation
Lex Fridman (1:20:55.880)
that this kind of world is effectively largely populated
Karl Friston (1:20:59.120)
by other things like me, i.e. you,
Lex Fridman (1:21:02.360)
then it makes enormous sense
Karl Friston (1:21:04.200)
that if I can develop a hypothesis
Lex Fridman (1:21:07.280)
that we are similar kinds of creatures,
Karl Friston (1:21:11.520)
in fact, the same kind of creature,
Lex Fridman (1:21:13.640)
but I am me and you are you,
Karl Friston (1:21:16.320)
then it becomes, again, mandated to have a sense of self.
Lex Fridman (1:21:21.440)
So if I live in a world
Karl Friston (1:21:23.640)
that is constituted by things like me,
Lex Fridman (1:21:26.560)
basically a social world, a community,
Karl Friston (1:21:29.520)
then it becomes necessary now for me to infer
Lex Fridman (1:21:32.320)
that it's me talking and not you talking.
Karl Friston (1:21:34.400)
I wouldn't need that if I was on Mars by myself
Lex Fridman (1:21:37.280)
or if I was in the jungle as a feral child.
Karl Friston (1:21:40.040)
If there was nothing like me around,
Lex Fridman (1:21:43.120)
there would be no need to have an inference
Karl Friston (1:21:46.200)
at a hypothesis, oh yes, it is me
Lex Fridman (1:21:49.160)
that is experiencing or causing these sounds
Lex Fridman (1:21:51.240)
and it is not you.
Lex Fridman (1:21:52.360)
It's only when there's ambiguity in play
Karl Friston (1:21:54.680)
induced by the fact that there are others in that world.
Lex Fridman (1:21:58.240)
So I think that the special thing about self aware artifacts
Karl Friston (1:22:03.520)
is that they have learned to, or they have acquired,
Lex Fridman (1:22:08.280)
or at least are equipped with, possibly by evolution,
Karl Friston (1:22:11.680)
generative models that allow for the fact
Lex Fridman (1:22:14.560)
there are lots of copies of things like them around,
Lex Fridman (1:22:17.360)
and therefore they have to work out it's you and not me.
Lex Fridman (1:22:20.600)
That's brilliant.
Karl Friston (1:22:23.280)
I've never thought of that.
Lex Fridman (1:22:24.560)
I never thought of that, that the purpose
Karl Friston (1:22:28.160)
of the really usefulness of consciousness
Lex Fridman (1:22:31.360)
or self awareness in the context of planning
Karl Friston (1:22:34.200)
existing in the world is so you can operate
Lex Fridman (1:22:36.920)
with other things like you, and like you could,
Karl Friston (1:22:39.240)
it doesn't have to necessarily be human.
Lex Fridman (1:22:40.800)
It could be other kind of similar creatures.
Karl Friston (1:22:43.480)
Absolutely, well, we view a lot of our attributes
Lex Fridman (1:22:46.080)
into our pets, don't we?
Karl Friston (1:22:47.880)
Or we try to make our robots humanoid.
Lex Fridman (1:22:49.920)
And I think there's a deep reason for that,
Karl Friston (1:22:51.840)
that it's just much easier to read the world
Lex Fridman (1:22:54.680)
if you can make the simplifying assumption
Karl Friston (1:22:56.240)
that basically you're me, and it's just your turn to talk.
Lex Fridman (1:23:00.040)
I mean, when we talk about planning,
Karl Friston (1:23:01.520)
when you talk specifically about planning,
Lex Fridman (1:23:04.200)
the highest, if you like, manifestation or realization
Karl Friston (1:23:07.800)
of that planning is what we're doing now.
Lex Fridman (1:23:09.600)
I mean, the human condition doesn't get any higher
Karl Friston (1:23:12.560)
than this talking about the philosophy of existence
Lex Fridman (1:23:16.800)
and the conversation.
Lex Fridman (1:23:17.920)
But in that conversation, there is a beautiful art
Lex Fridman (1:23:23.760)
of turn taking and mutual inference, theory of mind.
Karl Friston (1:23:28.120)
I have to know when you wanna listen.
Lex Fridman (1:23:29.680)
I have to know when you want to interrupt.
Karl Friston (1:23:31.120)
I have to make sure that you're online.
Lex Fridman (1:23:32.520)
I have to have a model in my head
Karl Friston (1:23:34.360)
of your model in your head.
Lex Fridman (1:23:35.800)
That's the highest, the most sophisticated form
Karl Friston (1:23:38.320)
of generative model, where the generative model
Lex Fridman (1:23:40.160)
actually has a generative model
Karl Friston (1:23:41.360)
of somebody else's generative model.
Lex Fridman (1:23:42.800)
And I think that, and what we are doing now evinces
Karl Friston (1:23:47.080)
the kinds of generative models
Lex Fridman (1:23:49.160)
that would support self awareness,
Karl Friston (1:23:51.280)
because without that, we'd both be talking over each other,
Lex Fridman (1:23:54.640)
or we'd be singing together in a choir.
Karl Friston (1:23:58.320)
That's not a brilliant analogy for what I'm trying to say,
Lex Fridman (1:24:01.280)
but yeah, we wouldn't have this discourse.
Karl Friston (1:24:05.160)
We wouldn't have this.
Lex Fridman (1:24:06.000)
Yeah, the dance of it.
Karl Friston (1:24:06.840)
Yeah, that's right.
Lex Fridman (1:24:07.680)
As I interrupt, I mean, that's beautifully put.
Karl Friston (1:24:12.680)
I'll re listen to this conversation many times.
Lex Fridman (1:24:17.360)
There's so much poetry in this, and mathematics.
Karl Friston (1:24:21.600)
Let me ask the silliest, or perhaps the biggest question
Lex Fridman (1:24:26.240)
as a last kind of question.
Karl Friston (1:24:29.840)
We've talked about living in existence
Lex Fridman (1:24:33.360)
and the objective function under which
Karl Friston (1:24:35.120)
these objects would operate.
Lex Fridman (1:24:37.480)
What do you think is the objective function
Lex Fridman (1:24:39.720)
of our existence?
Lex Fridman (1:24:41.560)
What's the meaning of life?
Lex Fridman (1:24:44.160)
What do you think is the, for you, perhaps,
Lex Fridman (1:24:47.160)
the purpose, the source of fulfillment,
Karl Friston (1:24:50.200)
the source of meaning for your existence,
Lex Fridman (1:24:53.080)
as one blob in this soup?
Karl Friston (1:24:57.640)
I'm tempted to answer that, again, as a physicist,
Lex Fridman (1:25:00.480)
until it's the free energy I expect
Karl Friston (1:25:03.040)
consequent upon my behavior.
Lex Fridman (1:25:05.480)
So technically, we could get a really interesting
Karl Friston (1:25:08.400)
conversation about what that comprises
Lex Fridman (1:25:10.720)
in terms of searching for information,
Karl Friston (1:25:13.160)
resolving uncertainty about the kind of thing that I am.
Lex Fridman (1:25:16.560)
But I suspect that you want a slightly more personal
Lex Fridman (1:25:20.240)
and fun answer, but which can be consistent with that.
Lex Fridman (1:25:25.120)
And I think it's reassuringly simple
Lex Fridman (1:25:30.320)
and hops back to what you were taught as a child,
Lex Fridman (1:25:36.520)
that you have certain beliefs about the kind of creature
Lex Fridman (1:25:39.520)
and the kind of person you are.
Lex Fridman (1:25:41.840)
And all that self evidencing means,
Karl Friston (1:25:44.840)
all that minimizing variational free energy
Lex Fridman (1:25:46.840)
in an inactive and embodied way,
Karl Friston (1:25:50.000)
means is fulfilling the beliefs about
Lex Fridman (1:25:53.640)
what kind of thing you are.
Lex Fridman (1:25:55.720)
And of course, we're all given those scripts,
Lex Fridman (1:25:58.040)
those narratives, at a very early age,
Karl Friston (1:26:01.240)
usually in the form of bedtime stories or fairy stories
Lex Fridman (1:26:04.280)
that I'm a princess and I'm gonna meet a beast
Karl Friston (1:26:07.160)
who's gonna transform and he's gonna be a prince.
Lex Fridman (1:26:09.360)
And so the narratives are all around you
Karl Friston (1:26:11.880)
from your parents to the friends
Lex Fridman (1:26:14.560)
to the society feeds these stories.
Lex Fridman (1:26:17.680)
And then your objective function is to fulfill.
Lex Fridman (1:26:21.000)
Exactly, that narrative that has been encultured
Karl Friston (1:26:24.240)
by your immediate family, but as you say,
Lex Fridman (1:26:27.160)
also the sort of the culture in which you grew up
Lex Fridman (1:26:29.600)
and you create for yourself.
Lex Fridman (1:26:30.880)
I mean, again, because of this active inference,
Karl Friston (1:26:33.440)
this inactive aspect of self evidencing,
Lex Fridman (1:26:36.680)
not only am I modeling my environment,
Karl Friston (1:26:40.960)
my eco niche, my external states out there,
Lex Fridman (1:26:44.040)
but I'm actively changing them all the time
Lex Fridman (1:26:46.520)
and doing the same back, we're doing it together.
Lex Fridman (1:26:49.840)
So there's a synchrony that means that I'm creating
Karl Friston (1:26:53.680)
my own culture over different timescales.
Lex Fridman (1:26:56.800)
So the question now is for me being very selfish,
Lex Fridman (1:27:00.800)
what scripts were I given?
Lex Fridman (1:27:02.240)
It basically was a mixture between Einstein and shark homes.
Lex Fridman (1:27:06.160)
So I smoke as heavily as possible,
Lex Fridman (1:27:09.880)
try to avoid too much interpersonal contact,
Karl Friston (1:27:15.280)
enjoy the fantasy that you're a popular scientist
Lex Fridman (1:27:21.000)
who's gonna make a difference in a slightly quirky way.
Lex Fridman (1:27:23.400)
So that's what I grew up on.
Lex Fridman (1:27:25.200)
My father was an engineer and loved science
Lex Fridman (1:27:28.320)
and he loved sort of things like Sir Arthur Edmonds,
Lex Fridman (1:27:33.320)
Spacetime and Gravitation, which was the first
Karl Friston (1:27:37.440)
understandable version of general relativity.
Lex Fridman (1:27:41.800)
So all the fairy stories I was told as I was growing up
Karl Friston (1:27:45.840)
were all about these characters.
Lex Fridman (1:27:48.640)
I'm keeping the Hobbit out of this
Karl Friston (1:27:50.600)
because that doesn't quite fit my narrative.
Lex Fridman (1:27:53.280)
There's a journey of exploration, I suppose, of sorts.
Lex Fridman (1:27:56.240)
So yeah, I've just grown up to be what I imagine
Lex Fridman (1:28:01.240)
a mild mannered Sherlock Holmes slash Albert Einstein
Karl Friston (1:28:05.720)
would do in my shoes.
Lex Fridman (1:28:07.960)
And you did it elegantly and beautifully.
Karl Friston (1:28:10.080)
Carl was a huge honor talking today, it was fun.
Lex Fridman (1:28:12.440)
Thank you so much for your time.
Karl Friston (1:28:13.560)
No, thank you. Appreciate it.
Lex Fridman (1:28:15.560)
Thank you for listening to this conversation
Karl Friston (1:28:17.240)
with Carl Friston and thank you
Lex Fridman (1:28:19.280)
to our presenting sponsor, Cash App.
Karl Friston (1:28:21.320)
Please consider supporting the podcast
Lex Fridman (1:28:23.080)
by downloading Cash App and using code LexPodcast.
Karl Friston (1:28:27.080)
If you enjoy this podcast, subscribe on YouTube,
Lex Fridman (1:28:29.720)
review it with five stars on Apple Podcast,
Karl Friston (1:28:32.040)
support on Patreon, or simply connect with me on Twitter
Lex Fridman (1:28:35.440)
at LexFriedman.
Lex Fridman (1:28:37.480)
And now let me leave you with some words from Carl Friston.
Lex Fridman (1:28:41.520)
Your arm moves because you predict it will
Lex Fridman (1:28:44.600)
and your motor system seeks to minimize prediction error.
Lex Fridman (1:28:48.040)
Thank you for listening and hope to see you next time.
Lex Fridman (20:01.100)
But what's your feeling?
Lex Fridman (20:02.820)
What's the general consensus?
Lex Fridman (20:04.220)
Have we moved away from the magic soup view of the brain?
Lex Fridman (20:07.820)
So there is a deep structure to it.
Lex Fridman (20:11.940)
And then maybe a further question.
Lex Fridman (20:14.460)
You said some people believe that the structure
Karl Friston (20:16.820)
is most of it, that you can really get
Lex Fridman (20:19.180)
at the core of the function
Karl Friston (20:20.220)
by just deeply understanding the structure.
Lex Fridman (20:22.540)
Where do you sit on that, do you?
Karl Friston (20:25.180)
I think it's got some mileage to it, yes, yeah.
Lex Fridman (20:28.260)
So it's a worthy pursuit of going,
Karl Friston (20:31.180)
of studying through imaging and all the different methods
Lex Fridman (20:34.780)
to actually study the structure.
Karl Friston (20:36.260)
No, absolutely, yeah, yeah.
Lex Fridman (20:38.340)
Sorry, I'm just noting, you were accusing me
Karl Friston (20:41.180)
of using lots of long words
Lex Fridman (20:42.460)
and then you introduced one there, which is deep,
Karl Friston (20:44.300)
which is interesting.
Lex Fridman (20:46.340)
Because deep is the sort of millennial equivalent
Karl Friston (20:50.420)
of hierarchical.
Lex Fridman (20:51.660)
So if you've put deep in front of anything,
Karl Friston (20:53.860)
not only are you very millennial and very trending,
Lex Fridman (20:57.620)
but you're also implying a hierarchical architecture.
Lex Fridman (21:01.540)
So it is a depth, which is, for me, the beautiful thing.
Lex Fridman (21:05.260)
That's right, the word deep kind of,
Karl Friston (21:07.380)
yeah, exactly, it implies hierarchy.
Lex Fridman (21:10.260)
I didn't even think about that.
Karl Friston (21:11.500)
That indeed, the implicit meaning
Lex Fridman (21:14.620)
of the word deep is hierarchy.
Karl Friston (21:16.820)
Yep. Yeah.
Lex Fridman (21:18.340)
So deep inside the onion is the center of your soul.
Karl Friston (21:22.380)
Beautifully put.
Lex Fridman (21:23.500)
Maybe briefly, if you could paint a picture
Karl Friston (21:26.780)
of the kind of methods of neuroimaging,
Lex Fridman (21:30.980)
maybe the history which you were a part of,
Karl Friston (21:33.420)
from statistical parametric mapping.
Lex Fridman (21:35.180)
I mean, just what's out there that's interesting
Karl Friston (21:37.940)
for people maybe outside the field
Lex Fridman (21:40.540)
to understand of what are the actual methodologies
Lex Fridman (21:43.460)
of looking inside the human brain?
Lex Fridman (21:45.340)
Right, well, you can answer that question
Karl Friston (21:47.420)
from two perspectives.
Lex Fridman (21:48.300)
Basically, it's the modality.
Lex Fridman (21:49.900)
What kind of signal are you measuring?
Lex Fridman (21:52.580)
And they can range from,
Lex Fridman (21:55.460)
and let's limit ourselves
Lex Fridman (21:56.860)
to sort of imaging based noninvasive techniques.
Lex Fridman (22:01.060)
So you've essentially got brain scanners,
Lex Fridman (22:03.020)
and brain scanners can either measure
Karl Friston (22:05.340)
the structural attributes, the amount of water,
Lex Fridman (22:07.660)
the amount of fat, or the amount of iron
Karl Friston (22:09.300)
in different parts of the brain,
Lex Fridman (22:10.380)
and you can make lots of inferences
Karl Friston (22:11.620)
about the structure of the organ of the sort
Lex Fridman (22:15.460)
that you might have produced from an X ray,
Lex Fridman (22:17.780)
but a very nuanced X ray that is looking
Lex Fridman (22:21.340)
at this kind of property or that kind of property.
Lex Fridman (22:24.380)
So looking at the anatomy noninvasively
Lex Fridman (22:27.860)
would be the first sort of neuroimaging
Karl Friston (22:30.140)
that people might want to employ.
Lex Fridman (22:32.180)
Then you move on to the kinds of measurements
Karl Friston (22:34.940)
that reflect dynamic function,
Lex Fridman (22:38.100)
and the most prevalent of those fall into two camps.
Karl Friston (22:42.020)
You've got these metabolic, sometimes hemodynamic,
Lex Fridman (22:46.460)
blood related signals.
Lex Fridman (22:48.940)
So these metabolic and or hemodynamic signals
Lex Fridman (22:53.460)
are basically proxies for elevated activity
Lex Fridman (22:58.380)
and message passing and neuronal dynamics
Lex Fridman (23:03.460)
in particular parts of the brain.
Karl Friston (23:05.340)
Characteristically though, the time constants
Lex Fridman (23:07.660)
of these hemodynamic or metabolic responses
Karl Friston (23:11.420)
to neural activity are much longer
Lex Fridman (23:14.300)
than the neural activity itself.
Lex Fridman (23:15.820)
And this is referring,
Lex Fridman (23:19.020)
forgive me for the dumb questions,
Lex Fridman (23:20.580)
but this would be referring to blood,
Lex Fridman (23:22.940)
like the flow of blood.
Karl Friston (23:24.260)
Absolutely, absolutely.
Lex Fridman (23:25.100)
So there's a ton of,
Karl Friston (23:26.500)
it seems like there's a ton of blood vessels in the brain.
Lex Fridman (23:29.420)
Yeah.
Lex Fridman (23:30.260)
So what's the interaction between the flow of blood
Lex Fridman (23:33.700)
and the function of the neurons?
Lex Fridman (23:35.980)
Is there an interplay there or?
Lex Fridman (23:37.420)
Yup, yup, and that interplay accounts for several careers
Karl Friston (23:42.780)
of world renowned scientists, yes, absolutely.
Lex Fridman (23:47.220)
So this is known as neurovascular coupling,
Karl Friston (23:49.140)
is exactly what you said.
Lex Fridman (23:50.180)
It's how does the neural activity,
Karl Friston (23:52.300)
the neuronal infrastructure, the actual message passing
Lex Fridman (23:54.620)
that we think underlies our capacity to perceive and act,
Lex Fridman (24:01.060)
how is that coupled to the vascular responses
Lex Fridman (24:06.060)
that supply the energy for that neural processing?
Lex Fridman (24:09.900)
So there's a delicate web of large vessels,
Lex Fridman (24:13.380)
arteries and veins, that gets progressively finer
Lex Fridman (24:16.420)
and finer in detail until it perfuses
Lex Fridman (24:18.860)
at a microscopic level,
Karl Friston (24:20.420)
the machinery where little neurons lie.
Lex Fridman (24:23.780)
So coming back to this sort of onion perspective,
Karl Friston (24:27.420)
we were talking before using the onion as a metaphor
Lex Fridman (24:30.500)
for a deep hierarchical structure,
Lex Fridman (24:32.460)
but also I think it's just anatomically quite
Lex Fridman (24:36.220)
a useful metaphor.
Karl Friston (24:37.780)
All the action, all the heavy lifting
Lex Fridman (24:40.140)
in terms of neural computation is done
Karl Friston (24:41.660)
on the surface of the brain,
Lex Fridman (24:43.940)
and then the interior of the brain is constituted
Karl Friston (24:47.380)
by fatty wires, essentially, axonal processes
Lex Fridman (24:52.980)
that are enshrouded by myelin sheaths.
Lex Fridman (24:55.940)
And these, when you dissect them, they look fatty and white,
Lex Fridman (24:59.620)
and so it's called white matter,
Karl Friston (25:01.240)
as opposed to the actual neuro pill,
Lex Fridman (25:04.100)
which does the computation constituted largely by neurons,
Lex Fridman (25:07.220)
and that's known as gray matter.
Lex Fridman (25:08.700)
So the gray matter is a surface or a skin
Karl Friston (25:13.260)
that sits on top of this big ball,
Lex Fridman (25:16.260)
now we are talking magic soup,
Lex Fridman (25:17.780)
but a big ball of connections like spaghetti,
Lex Fridman (25:20.860)
very carefully structured with sparse connectivity
Karl Friston (25:23.100)
that preserves this deep hierarchical structure,
Lex Fridman (25:25.760)
but all the action takes place on the surface,
Karl Friston (25:28.300)
on the cortex of the onion, and that means
Lex Fridman (25:34.180)
that you have to supply the right amount of blood flow,
Karl Friston (25:38.560)
the right amount of nutrient,
Lex Fridman (25:41.100)
which is rapidly absorbed and used by neural cells
Karl Friston (25:45.240)
that don't have the same capacity
Lex Fridman (25:46.820)
that your leg muscles would have
Karl Friston (25:48.780)
to basically spend their energy budget
Lex Fridman (25:52.500)
and then claim it back later.
Lex Fridman (25:55.000)
So one peculiar thing about cerebral metabolism,
Lex Fridman (25:58.420)
brain metabolism, is it really needs to be driven
Karl Friston (26:01.440)
in the moment, which means you basically
Lex Fridman (26:03.060)
have to turn on the taps.
Lex Fridman (26:04.860)
So if there's lots of neural activity
Lex Fridman (26:08.700)
in one part of the brain, a little patch
Karl Friston (26:10.780)
of a few millimeters, even less possibly,
Lex Fridman (26:14.100)
you really do have to water that piece
Karl Friston (26:16.060)
of the garden now and quickly,
Lex Fridman (26:18.420)
and by quickly I mean within a couple of seconds.
Lex Fridman (26:21.540)
So that contains a lot of, hence the imaging
Lex Fridman (26:26.060)
could tell you a story of what's happening.
Karl Friston (26:28.260)
Absolutely, but it is slightly compromised
Lex Fridman (26:32.240)
in terms of the resolution.
Lex Fridman (26:33.460)
So the deployment of these little microvessels
Lex Fridman (26:37.580)
that water the garden to enable the neural activity
Karl Friston (26:42.380)
to play out, the spatial resolution
Lex Fridman (26:45.320)
is in order of a few millimeters,
Lex Fridman (26:48.020)
and crucially, the temporal resolution
Lex Fridman (26:50.380)
is the order of a few seconds.
Lex Fridman (26:52.220)
So you can't get right down and dirty
Lex Fridman (26:54.420)
into the actual spatial and temporal scale
Karl Friston (26:57.400)
of neural activity in and of itself.
Lex Fridman (26:59.920)
To do that, you'd have to turn
Karl Friston (27:00.860)
to the other big imaging modality,
Lex Fridman (27:02.580)
which is the recording of electromagnetic signals
Karl Friston (27:05.580)
as they're generated in real time.
Lex Fridman (27:07.740)
So here, the temporal bandwidth, if you like,
Karl Friston (27:10.320)
or the low limit on the temporal resolution
Lex Fridman (27:12.960)
is incredibly small, talking about milliseconds.
Lex Fridman (27:17.960)
And then you can get into the phasic fast responses
Lex Fridman (27:23.520)
that is in and of itself the neural activity,
Lex Fridman (27:27.440)
and start to see the succession or cascade
Lex Fridman (27:32.780)
of hierarchical recurrent message passing
Karl Friston (27:35.220)
evoked by a particular stimulus.
Lex Fridman (27:37.140)
But the problem is you're looking
Karl Friston (27:39.400)
at electromagnetic signals that have passed
Lex Fridman (27:42.440)
through an enormous amount of magic soup
Karl Friston (27:45.640)
or spaghetti of collectivity, and through the scalp
Lex Fridman (27:49.480)
and the skull, and it's become spatially very diffuse.
Lex Fridman (27:52.920)
So it's very difficult to know where you are.
Lex Fridman (27:54.940)
So you've got this sort of catch 22.
Karl Friston (27:58.600)
You can either use an imaging modality
Lex Fridman (28:00.280)
that tells you within millimeters
Karl Friston (28:02.840)
which part of the brain is activated,
Lex Fridman (28:04.440)
but you don't know when,
Karl Friston (28:05.800)
or you've got these electromagnetic EEG, MEG setups
Lex Fridman (28:10.800)
that tell you to within a few milliseconds
Karl Friston (28:15.840)
when something has responded, but you're not aware.
Lex Fridman (28:19.240)
So you've got these two complementary measures,
Karl Friston (28:22.520)
either indirect via the blood flow,
Lex Fridman (28:25.840)
or direct via the electromagnetic signals
Karl Friston (28:28.780)
caused by neural activity.
Lex Fridman (28:31.080)
These are the two big imaging devices.
Lex Fridman (28:33.360)
And then the second level of responding to your question,
Lex Fridman (28:36.960)
what are the, from the outside,
Lex Fridman (28:39.440)
what are the big ways of using this technology?
Lex Fridman (28:44.240)
So once you've chosen the kind of neural imaging
Karl Friston (28:47.160)
that you want to use to answer your set questions,
Lex Fridman (28:50.240)
and sometimes it would have to be both,
Karl Friston (28:53.360)
then you've got a whole raft of analyses,
Lex Fridman (28:57.540)
time series analyses usually, that you can bring to bear
Karl Friston (29:01.920)
in order to answer your questions
Lex Fridman (29:04.360)
or address your hypothesis about those data.
Lex Fridman (29:07.000)
And interestingly, they both fall
Lex Fridman (29:08.920)
into the same two camps we were talking about before,
Karl Friston (29:11.320)
this dialectic between specialization and integration,
Lex Fridman (29:14.800)
differentiation and integration.
Lex Fridman (29:17.120)
So it's the cartography, the blobology analyses.
Lex Fridman (29:20.880)
I apologize, I probably shouldn't interrupt so much,
Lex Fridman (29:23.220)
but just heard a fun word, the blah.
Lex Fridman (29:27.080)
Blobology.
Karl Friston (29:27.920)
Blobology.
Lex Fridman (29:29.160)
It's a neologism, which means the study of blobs.
Lex Fridman (29:33.000)
So nothing bob.
Lex Fridman (29:34.720)
Are you being witty and humorous,
Karl Friston (29:36.640)
or does the word blobology ever appear
Lex Fridman (29:39.640)
in a textbook somewhere?
Karl Friston (29:40.760)
It would appear in a popular book.
Lex Fridman (29:43.320)
It would not appear in a worthy specialist journal.
Karl Friston (29:47.960)
Yeah, I thought so.
Lex Fridman (29:48.960)
It's the fond word for the study of literally little blobs
Karl Friston (29:53.560)
on brain maps showing activations.
Lex Fridman (29:56.160)
So the kind of thing that you'd see in the newspapers
Karl Friston (29:59.500)
on ABC or BBC reporting the latest finding
Lex Fridman (30:04.080)
from brain imaging.
Karl Friston (30:05.320)
Interestingly though, the maths involved
Lex Fridman (30:10.080)
in that stream of analysis does actually call upon
Karl Friston (30:15.340)
the mathematics of blobs.
Lex Fridman (30:17.640)
So seriously, they're actually called Euler characteristics
Lex Fridman (30:21.800)
and they have a lot of fancy names in mathematics.
Lex Fridman (30:27.500)
We'll talk about it, about your ideas
Karl Friston (30:28.920)
in free energy principle.
Lex Fridman (30:30.400)
I mean, there's echoes of blobs there
Karl Friston (30:33.320)
when you consider sort of entities,
Lex Fridman (30:36.720)
mathematically speaking.
Karl Friston (30:38.000)
Yes, absolutely.
Lex Fridman (30:40.800)
Well, circumscribed, well defined,
Karl Friston (30:43.080)
you entities of, well, from the free energy point of view,
Lex Fridman (30:48.120)
entities of anything, but from the point of view
Karl Friston (30:50.280)
of the analysis, the cartography of the brain,
Lex Fridman (30:55.800)
these are the entities that constitute the evidence
Karl Friston (30:59.200)
for this functional segregation.
Lex Fridman (31:01.680)
You have segregated this function in this blob
Lex Fridman (31:04.500)
and it is not outside of the blob.
Lex Fridman (31:06.760)
And that's basically the, if you were a map maker
Karl Friston (31:10.200)
of America and you did not know its structure,
Lex Fridman (31:14.080)
the first thing were you doing constituting
Karl Friston (31:16.400)
or creating a map would be to identify the cities,
Lex Fridman (31:19.140)
for example, or the mountains or the rivers.
Karl Friston (31:22.000)
All of these uniquely spatially localizable features,
Lex Fridman (31:26.920)
possibly topological features have to be placed somewhere
Karl Friston (31:30.680)
because that requires a mathematics of identifying
Lex Fridman (31:33.520)
what does a city look like on a satellite image
Karl Friston (31:36.080)
or what does a river look like
Lex Fridman (31:37.400)
or what does a mountain look like?
Lex Fridman (31:39.120)
What would it, you know, what data features
Lex Fridman (31:42.240)
would evidence that particular top,
Karl Friston (31:46.520)
you know, that particular thing
Lex Fridman (31:48.560)
that you wanted to put on the map?
Lex Fridman (31:50.400)
And they normally are characterized
Lex Fridman (31:52.760)
in terms of literally these blobs
Karl Friston (31:54.320)
or these sort of, another way of looking at this
Lex Fridman (31:57.000)
is that a certain statistical measure
Karl Friston (32:01.560)
of the degree of activation crosses a threshold
Lex Fridman (32:04.000)
and in crossing that threshold
Karl Friston (32:06.520)
in the spatially restricted part of the brain,
Lex Fridman (32:09.600)
it creates a blob.
Lex Fridman (32:11.040)
And that's basically what statistical parametric mapping does.
Lex Fridman (32:14.000)
It's basically mathematically finessed blobology.
Karl Friston (32:19.840)
Okay, so those are the,
Lex Fridman (32:20.920)
you kind of described these two methodologies for,
Karl Friston (32:24.000)
one is temporally noisy, one is spatially noisy
Lex Fridman (32:26.920)
and you kind of have to play and figure out
Lex Fridman (32:28.360)
what can be useful.
Lex Fridman (32:31.400)
It'd be great if you can sort of comment.
Karl Friston (32:33.000)
I got a chance recently to spend a day
Lex Fridman (32:34.920)
at a company called Neuralink
Karl Friston (32:37.200)
that uses brain computer interfaces
Lex Fridman (32:39.420)
and their dream is to, well,
Karl Friston (32:42.720)
there's a bunch of sort of dreams,
Lex Fridman (32:45.200)
but one of them is to understand the brain
Karl Friston (32:47.380)
by sort of, you know, getting in there
Lex Fridman (32:51.520)
past the so called sort of factory wall,
Karl Friston (32:53.800)
getting in there and be able to listen,
Lex Fridman (32:55.200)
communicate both directions.
Lex Fridman (32:57.160)
What are your thoughts about this,
Lex Fridman (32:59.360)
the future of this kind of technology
Karl Friston (33:01.040)
of brain computer interfaces
Lex Fridman (33:02.360)
to be able to now have a window
Karl Friston (33:06.920)
or direct contact within the brain
Lex Fridman (33:08.500)
to be able to measure some of the signals,
Karl Friston (33:10.040)
to be able to sense signals,
Lex Fridman (33:11.320)
to understand some of the functionality of the brain?
Karl Friston (33:15.040)
Ambivalent, my sense is ambivalent.
Lex Fridman (33:17.760)
So it's a mixture of good and bad
Lex Fridman (33:19.920)
and I acknowledge that freely.
Lex Fridman (33:22.400)
So the good bits, if you just look at the legacy
Karl Friston (33:24.760)
of that kind of reciprocal but invasive
Lex Fridman (33:29.240)
your brain stimulation,
Karl Friston (33:31.160)
I didn't paint a complete picture
Lex Fridman (33:33.160)
when I was talking about sort of the ways
Karl Friston (33:34.680)
we understand the brain prior to neuroimaging.
Lex Fridman (33:37.080)
It wasn't just lesion deficit studies.
Karl Friston (33:39.680)
Some of the early work, in fact,
Lex Fridman (33:41.380)
literally 100 years from where we're sitting
Karl Friston (33:43.500)
at the institution of neurology,
Lex Fridman (33:45.240)
was done by stimulating the brain of say dogs
Lex Fridman (33:50.000)
and looking at how they responded
Lex Fridman (33:51.880)
with their muscles or with their salivation
Lex Fridman (33:56.400)
and imputing what that part of the brain must be doing.
Lex Fridman (34:00.920)
If I stimulate it and I vote this kind of response,
Karl Friston (34:06.000)
then that tells me quite a lot
Lex Fridman (34:07.240)
about the functional specialization.
Lex Fridman (34:09.080)
So there's a long history of brain stimulation
Lex Fridman (34:12.280)
which continues to enjoy a lot of attention nowadays.
Karl Friston (34:16.840)
Positive attention.
Lex Fridman (34:17.720)
Oh yes, absolutely.
Karl Friston (34:19.600)
You know, deep brain stimulation for Parkinson's disease
Lex Fridman (34:22.160)
is now a standard treatment
Lex Fridman (34:23.600)
and also a wonderful vehicle
Lex Fridman (34:25.560)
to try and understand the neuronal dynamics
Karl Friston (34:29.000)
underlying movement disorders like Parkinson's disease.
Lex Fridman (34:33.120)
Even interest in magnetic stimulation,
Karl Friston (34:37.920)
stimulating the magnetic fields
Lex Fridman (34:39.240)
and will it work in people who are depressed, for example.
Karl Friston (34:43.320)
Quite a crude level of understanding what you're doing,
Lex Fridman (34:45.700)
but there is historical evidence
Karl Friston (34:49.080)
that these kinds of brute force interventions
Lex Fridman (34:51.800)
do change things.
Karl Friston (34:53.380)
They, you know, it's a little bit like banging the TV
Lex Fridman (34:56.000)
when the valves aren't working properly,
Lex Fridman (34:58.240)
but it's still, it works.
Lex Fridman (35:00.720)
So, you know, there is a long history.
Karl Friston (35:04.480)
Brain computer interfacing or BCI,
Lex Fridman (35:08.700)
I think is a beautiful example of that.
Karl Friston (35:11.000)
It's sort of carved out its own niche
Lex Fridman (35:12.720)
and its own aspirations
Lex Fridman (35:14.420)
and there've been enormous advances within limits.
Lex Fridman (35:20.800)
Advances in terms of our ability to understand
Lex Fridman (35:25.480)
how the brain, the embodied brain,
Lex Fridman (35:29.720)
engages with the world.
Karl Friston (35:32.480)
I'm thinking here of sensory substitution,
Lex Fridman (35:34.760)
the augmenting our sensory capacities
Karl Friston (35:37.240)
by giving ourselves extra ways of sensing
Lex Fridman (35:40.820)
and sampling the world,
Karl Friston (35:42.260)
ranging from sort of trying to replace lost visual signals
Lex Fridman (35:48.300)
through to giving people completely new signals.
Karl Friston (35:50.380)
So, one of the, I think, most engaging examples of this
Lex Fridman (35:57.100)
is equipping people with a sense of magnetic fields.
Lex Fridman (36:00.660)
So you can actually give them magnetic sensors
Lex Fridman (36:03.620)
that enable them to feel,
Karl Friston (36:05.460)
should we say, tactile pressure around their tummy,
Lex Fridman (36:08.980)
where they are in relation to the magnetic field of the Earth.
Lex Fridman (36:13.100)
And after a few weeks, they take it for granted.
Lex Fridman (36:17.660)
They integrate it, they embody this,
Karl Friston (36:19.340)
simulate this new sensory information
Lex Fridman (36:22.340)
into the way that they literally feel their world,
Lex Fridman (36:25.420)
but now equipped with this sense of magnetic direction.
Lex Fridman (36:29.300)
So that tells you something
Karl Friston (36:31.020)
about the brain's plastic potential
Lex Fridman (36:32.980)
to remodel and its plastic capacity
Karl Friston (36:37.980)
to suddenly try to explain the sensory data at hand
Lex Fridman (36:43.640)
by augmenting the sensory sphere
Lex Fridman (36:48.440)
and the kinds of things that you can measure.
Lex Fridman (36:51.600)
Clearly, that's purely for entertainment
Lex Fridman (36:54.720)
and understanding the nature and the power of our brains.
Lex Fridman (37:00.120)
I would imagine that most BCI is pitched
Karl Friston (37:03.400)
at solving clinical and human problems
Lex Fridman (37:08.600)
such as locked in syndrome, such as paraplegia,
Karl Friston (37:12.240)
or replacing lost sensory capacities
Lex Fridman (37:16.120)
like blindness and deafness.
Lex Fridman (37:18.960)
So then we come to the negative part of my ambivalence,
Lex Fridman (37:24.240)
the other side of it.
Lex Fridman (37:26.960)
So I don't want to be deflationary
Lex Fridman (37:30.960)
because much of my deflationary comments
Karl Friston (37:33.400)
is probably large out of ignorance than anything else.
Lex Fridman (37:37.120)
But generally speaking, the bandwidth
Lex Fridman (37:42.480)
and the bit rates that you get
Lex Fridman (37:44.360)
from brain computer interfaces as we currently know them,
Karl Friston (37:49.160)
we're talking about bits per second.
Lex Fridman (37:51.460)
So that would be like me only being able to communicate
Karl Friston (37:55.600)
with any world or with you using very, very, very slow Morse code.
Lex Fridman (38:06.560)
And it is not even within an order of magnitude
Karl Friston (38:13.440)
near what we actually need for an inactive realization
Lex Fridman (38:18.600)
of what people aspire to when they think about
Karl Friston (38:21.280)
sort of curing people with paraplegia or replacing sight
Lex Fridman (38:28.760)
despite heroic efforts.
Lex Fridman (38:30.280)
So one has to ask, is there a lower bound
Lex Fridman (38:33.760)
on the kinds of recurrent information exchange
Lex Fridman (38:41.040)
between a brain and some augmented or artificial interface?
Lex Fridman (38:46.040)
And then we come back to, interestingly,
Lex Fridman (38:51.620)
what I was talking about before,
Lex Fridman (38:52.820)
which is if you're talking about function
Karl Friston (38:56.460)
in terms of inference, and I presume we'll get to that
Lex Fridman (39:00.220)
later on in terms of the free energy principle,
Karl Friston (39:01.900)
then at the moment, there may be fundamental reasons
Lex Fridman (39:05.140)
to assume that is the case.
Karl Friston (39:06.400)
We're talking about ensemble activity.
Lex Fridman (39:08.540)
We're talking about basically, for example,
Karl Friston (39:13.540)
let's paint the challenge facing brain computer interfacing
Lex Fridman (39:20.800)
in terms of controlling another system
Karl Friston (39:24.520)
that is highly and deeply structured,
Lex Fridman (39:27.080)
very relevant to our lives, very nonlinear,
Karl Friston (39:30.560)
that rests upon the kind of nonequilibrium
Lex Fridman (39:34.360)
steady states and dynamics that the brain does,
Lex Fridman (39:37.400)
the weather, all right?
Lex Fridman (39:39.540)
So imagine you had some very aggressive satellites
Karl Friston (39:45.840)
that could produce signals that could perturb
Lex Fridman (39:48.280)
some little parts of the weather system.
Lex Fridman (39:53.000)
And then what you're asking now is,
Lex Fridman (39:55.140)
can I meaningfully get into the weather
Lex Fridman (39:58.320)
and change it meaningfully and make the weather respond
Lex Fridman (40:01.300)
in a way that I want it to?
Karl Friston (40:03.360)
You're talking about chaos control on a scale
Lex Fridman (40:06.600)
which is almost unimaginable.
Lex Fridman (40:08.860)
So there may be fundamental reasons
Lex Fridman (40:11.120)
why BCI, as you might read about it in a science fiction novel,
Karl Friston (40:18.480)
aspirational BCI may never actually work
Lex Fridman (40:22.920)
in the sense that to really be integrated
Lex Fridman (40:26.920)
and be part of the system is a requirement
Lex Fridman (40:32.160)
that requires you to have evolved with that system,
Karl Friston (40:35.240)
that you have to be part of a very delicately structured,
Lex Fridman (40:43.320)
deeply structured, dynamic, ensemble activity
Karl Friston (40:48.000)
that is not like rewiring a broken computer
Lex Fridman (40:51.600)
or plugging in a peripheral interface adapter.
Karl Friston (40:54.660)
It is much more like getting into the weather patterns
Lex Fridman (40:58.140)
or, come back to your magic soup,
Karl Friston (41:00.680)
getting into the active matter
Lex Fridman (41:02.420)
and meaningfully relate that to the outside world.
Lex Fridman (41:07.140)
So I think there are enormous challenges there.
Lex Fridman (41:09.920)
So I think the example of the weather is a brilliant one.
Lex Fridman (41:13.260)
And I think you paint a really interesting picture
Lex Fridman (41:15.300)
and it wasn't as negative as I thought.
Karl Friston (41:17.420)
It's essentially saying that it might be
Lex Fridman (41:19.920)
incredibly challenging, including the low bound
Karl Friston (41:22.140)
of the bandwidth and so on.
Lex Fridman (41:23.640)
I kind of, so just to full disclosure,
Karl Friston (41:26.900)
I come from the machine learning world.
Lex Fridman (41:28.680)
So my natural thought is the hardest part
Karl Friston (41:32.760)
is the engineering challenge of controlling the weather,
Lex Fridman (41:34.820)
of getting those satellites up and running and so on.
Lex Fridman (41:37.560)
And once they are, then the rest is fundamentally
Lex Fridman (41:42.220)
the same approaches that allow you to win in the game of Go
Karl Friston (41:46.860)
will allow you to potentially play in this soup,
Lex Fridman (41:49.580)
in this chaos.
Lex Fridman (41:51.000)
So I have a hope that sort of machine learning methods
Lex Fridman (41:54.500)
will help us play in this soup.
Lex Fridman (41:58.820)
But perhaps you're right that it is a biology
Lex Fridman (42:04.220)
and the brain is just an incredible system
Karl Friston (42:08.660)
that may be almost impossible to get in.
Lex Fridman (42:12.220)
But for me, what seems impossible
Karl Friston (42:15.780)
is the incredible mess of blood vessels
Lex Fridman (42:19.780)
that you also described without,
Karl Friston (42:22.300)
we also value the brain.
Lex Fridman (42:24.620)
You can't make any mistakes, you can't damage things.
Lex Fridman (42:27.100)
So to me, that engineering challenge seems nearly impossible.
Lex Fridman (42:31.340)
One of the things I was really impressed by at Neuralink
Karl Friston (42:35.900)
is just talking to brilliant neurosurgeons
Lex Fridman (42:39.660)
and the roboticists that made me realize
Karl Friston (42:43.340)
that even though it seems impossible,
Lex Fridman (42:45.860)
if anyone can do it, it's some of these world class
Karl Friston (42:48.620)
engineers that are trying to take it on.
Lex Fridman (42:50.800)
So I think the conclusion of our discussion here
Karl Friston (42:55.860)
of this part is basically that the problem is really hard
Lex Fridman (43:00.980)
but hopefully not impossible.
Karl Friston (43:02.580)
Absolutely.
Lex Fridman (43:03.800)
So if it's okay, let's start with the basics.
Lex Fridman (43:07.260)
So you've also formulated a fascinating principle,
Lex Fridman (43:12.180)
the free energy principle.
Karl Friston (43:13.500)
Could we maybe start at the basics
Lex Fridman (43:15.340)
and what is the free energy principle?
Karl Friston (43:19.660)
Well, in fact, the free energy principle
Lex Fridman (43:23.700)
inherits a lot from the building
Karl Friston (43:29.220)
of these data analytic approaches
Lex Fridman (43:31.240)
to these very high dimensional time series
Karl Friston (43:34.180)
you get from the brain.
Lex Fridman (43:35.960)
So I think it's interesting to acknowledge that.
Lex Fridman (43:37.940)
And in particular, the analysis tools
Lex Fridman (43:39.980)
that try to address the other side,
Karl Friston (43:43.100)
which is a functional integration,
Lex Fridman (43:44.360)
so the connectivity analyses.
Lex Fridman (43:46.060)
So on the one hand, but I should also acknowledge
Lex Fridman (43:51.920)
it inherits an awful lot from machine learning as well.
Lex Fridman (43:55.360)
So the free energy principle is just a formal statement
Lex Fridman (44:02.600)
that the existential imperatives for any system
Karl Friston (44:08.500)
that manages to survive in a changing world
Lex Fridman (44:11.380)
can be cast as an inference problem
Karl Friston (44:18.900)
in the sense that you can interpret
Lex Fridman (44:21.240)
the probability of existing as the evidence that you exist.
Lex Fridman (44:25.720)
And if you can write down that problem of existence
Lex Fridman (44:29.460)
as a statistical problem,
Karl Friston (44:30.900)
then you can use all the maths that has been developed
Lex Fridman (44:33.940)
for inference to understand and characterize
Karl Friston (44:38.940)
the ensemble dynamics that must be in play
Lex Fridman (44:42.960)
in the service of that inference.
Lex Fridman (44:45.560)
So technically, what that means is
Lex Fridman (44:48.240)
you can always interpret anything that exists
Karl Friston (44:52.160)
in virtue of being separate from the environment
Lex Fridman (44:55.660)
in which it exists as trying to minimize
Karl Friston (45:02.100)
variational free energy.
Lex Fridman (45:03.600)
And if you're from the machine learning community,
Karl Friston (45:05.560)
you will know that as a negative evidence lower bound
Lex Fridman (45:09.200)
or a negative elbow, which is the same as saying
Karl Friston (45:13.120)
you're trying to maximize or it will look as if
Lex Fridman (45:16.400)
all your dynamics are trying to maximize
Karl Friston (45:19.720)
the complement of that which is the marginal likelihood
Lex Fridman (45:24.000)
or the evidence for your own existence.
Lex Fridman (45:26.480)
So that's basically the free energy principle.
Lex Fridman (45:30.120)
But to even take a sort of a small step backwards,
Karl Friston (45:34.120)
you said the existential imperative.
Lex Fridman (45:38.360)
There's a lot of beautiful poetic words here,
Lex Fridman (45:40.120)
but to put it crudely, it's a fascinating idea
Lex Fridman (45:46.840)
of basically just of trying to describe
Karl Friston (45:49.520)
if you're looking at a blob,
Lex Fridman (45:51.800)
how do you know this thing is alive?
Lex Fridman (45:54.320)
What does it mean to be alive?
Lex Fridman (45:55.680)
What does it mean to exist?
Lex Fridman (45:57.520)
And so you can look at the brain,
Lex Fridman (45:59.440)
you can look at parts of the brain,
Karl Friston (46:00.720)
or this is just a general principle
Lex Fridman (46:02.840)
that applies to almost any system.
Karl Friston (46:07.240)
That's just a fascinating sort of philosophically
Lex Fridman (46:10.160)
at every level question and a methodology
Karl Friston (46:13.120)
to try to answer that question.
Lex Fridman (46:14.360)
What does it mean to be alive?
Lex Fridman (46:16.040)
So that's a huge endeavor and it's nice
Lex Fridman (46:21.480)
that there's at least some,
Karl Friston (46:23.200)
from some perspective, a clean answer.
Lex Fridman (46:25.480)
So maybe can you talk about that optimization view of it?
Lex Fridman (46:30.280)
So what's trying to be minimized, maximized?
Lex Fridman (46:33.520)
A system that's alive, what is it trying to minimize?
Karl Friston (46:36.840)
Right, you've made a big move there.
Lex Fridman (46:40.600)
First of all, it's good to make big moves.
Lex Fridman (46:45.960)
But you've assumed that the thing exists
Lex Fridman (46:52.000)
in a state that could be living or nonliving.
Lex Fridman (46:54.680)
So I may ask you, what licenses you
Lex Fridman (46:57.960)
to say that something exists?
Karl Friston (47:00.120)
That's why I use the word existential.
Lex Fridman (47:02.280)
It's beyond living, it's just existence.
Lex Fridman (47:05.440)
So if you drill down onto the definition
Lex Fridman (47:08.000)
of things that exist, then they have certain properties
Karl Friston (47:13.520)
if you borrow the maths
Lex Fridman (47:16.400)
from nonequilibrium steady state physics
Karl Friston (47:19.360)
that enable you to interpret their existence
Lex Fridman (47:26.240)
in terms of this optimization procedure.
Lex Fridman (47:29.280)
So it's good you introduced the word optimization.
Lex Fridman (47:32.160)
So what the free energy principle
Karl Friston (47:36.000)
in its sort of most ambitious,
Lex Fridman (47:39.800)
but also most deflationary and simplest, says
Karl Friston (47:44.720)
is that if something exists,
Lex Fridman (47:47.200)
then it must, by the mathematics
Karl Friston (47:51.440)
of nonequilibrium steady state,
Lex Fridman (47:55.120)
exhibit properties that make it look
Karl Friston (47:59.600)
as if it is optimizing a particular quantity.
Lex Fridman (48:03.680)
And it turns out that particular quantity
Karl Friston (48:06.120)
happens to be exactly the same
Lex Fridman (48:08.520)
as the evidence lower bound in machine learning
Karl Friston (48:11.360)
or Bayesian model evidence in Bayesian statistics.
Lex Fridman (48:15.280)
Or, and then I can list a whole other list
Karl Friston (48:18.800)
of ways of understanding this key quantity,
Lex Fridman (48:23.480)
which is a bound on surprise or self information
Karl Friston (48:29.120)
if you have information theory.
Lex Fridman (48:31.000)
There are a number of different perspectives
Karl Friston (48:34.080)
on this quantity.
Lex Fridman (48:34.920)
It's just basically the log probability
Karl Friston (48:36.920)
of being in a particular state.
Lex Fridman (48:40.120)
I'm telling this story as an honest,
Karl Friston (48:42.840)
an attempt to answer your question.
Lex Fridman (48:45.640)
And I'm answering it as if I was pretending
Karl Friston (48:49.280)
to be a physicist who was trying to understand
Lex Fridman (48:52.400)
the fundaments of nonequilibrium steady state.
Lex Fridman (48:58.040)
And I shouldn't really be doing that
Lex Fridman (48:59.640)
because the last time I was taught physics,
Karl Friston (49:02.240)
I was in my 20s.
Lex Fridman (49:03.760)
What kind of systems,
Karl Friston (49:04.720)
when you think about the free energy principle,
Lex Fridman (49:06.400)
what kind of systems are you imagining
Lex Fridman (49:08.640)
as a sort of more specific kind of case study?
Lex Fridman (49:11.640)
Yeah, I'm imagining a range of systems,
Lex Fridman (49:15.720)
but at its simplest, a single celled organism
Lex Fridman (49:23.400)
that can be identified from its eco niche
Karl Friston (49:26.120)
or its environment.
Lex Fridman (49:27.680)
So at its simplest, that's basically
Lex Fridman (49:31.480)
what I always imagined in my head.
Lex Fridman (49:33.840)
And you may ask, well, is there any,
Lex Fridman (49:36.680)
how on earth can you even elaborate questions
Lex Fridman (49:41.680)
about the existence of a single drop of oil, for example?
Lex Fridman (49:48.000)
But there are deep questions there.
Lex Fridman (49:49.440)
Why doesn't the oil, why doesn't the thing,
Karl Friston (49:52.560)
the interface between the drop of oil
Lex Fridman (49:55.160)
that contains an interior
Lex Fridman (49:57.480)
and the thing that is not the drop of oil,
Lex Fridman (50:00.320)
which is the solvent in which it is immersed,
Lex Fridman (50:03.880)
how does that interface persist over time?
Lex Fridman (50:07.080)
Why doesn't the oil just dissolve into solvent?
Lex Fridman (50:10.400)
So what special properties of the exchange
Lex Fridman (50:16.160)
between the surface of the oil drop
Lex Fridman (50:18.200)
and the external states in which it's immersed,
Lex Fridman (50:22.000)
if you're a physicist, say it would be the heat bath.
Karl Friston (50:24.040)
You've got a physical system, an ensemble again,
Lex Fridman (50:28.280)
we're talking about density dynamics, ensemble dynamics,
Karl Friston (50:30.320)
an ensemble of atoms or molecules immersed in the heat bath.
Lex Fridman (50:35.840)
But the question is, how did the heat bath get there?
Lex Fridman (50:38.760)
And why does it not dissolve?
Lex Fridman (50:41.360)
How is it maintaining itself?
Karl Friston (50:42.800)
Exactly.
Lex Fridman (50:43.640)
What actions is it?
Karl Friston (50:44.480)
I mean, it's such a fascinating idea of a drop of oil
Lex Fridman (50:47.520)
and I guess it would dissolve in water,
Karl Friston (50:50.000)
it wouldn't dissolve in water.
Lex Fridman (50:51.680)
So what?
Lex Fridman (50:52.520)
Precisely, so why not?
Lex Fridman (50:53.840)
So why not?
Lex Fridman (50:54.680)
Why not?
Lex Fridman (50:55.520)
And how do you mathematically describe,
Karl Friston (50:57.000)
I mean, it's such a beautiful idea.
Lex Fridman (50:58.680)
And also the idea of like, where does the thing,
Lex Fridman (51:02.120)
where does the drop of oil end and where does it begin?
Lex Fridman (51:07.120)
Right, so I mean, you're asking deep questions,
Karl Friston (51:10.400)
deep in a nonmillennial sense here.
Lex Fridman (51:12.520)
In a hierarchical sense.
Lex Fridman (51:16.280)
But what you can do, so this is the deflationary part of it.
Lex Fridman (51:21.720)
Can I just qualify my answer by saying that normally
Karl Friston (51:23.920)
when I'm asked this question,
Lex Fridman (51:24.760)
I answer from the point of view of a psychologist,
Karl Friston (51:26.640)
we talk about predictive processing and predictive coding
Lex Fridman (51:29.080)
and the brain as an inference machine,
Lex Fridman (51:31.640)
but you haven't asked me from that perspective,
Lex Fridman (51:33.920)
I'm answering from the point of view of a physicist.
Lex Fridman (51:36.240)
So the question is not so much why,
Lex Fridman (51:41.000)
but if it exists, what properties must it display?
Lex Fridman (51:44.440)
So that's the deflationary part of the free energy principle.
Lex Fridman (51:46.880)
The free energy principle does not supply an answer
Karl Friston (51:50.800)
as to why, it's saying if something exists,
Lex Fridman (51:54.520)
then it must display these properties.
Karl Friston (51:57.720)
That's the sort of thing that's on offer.
Lex Fridman (52:01.560)
And it so happens that these properties it must display
Karl Friston (52:05.240)
are actually intriguing and have this inferential gloss,
Lex Fridman (52:10.880)
this sort of self evidencing gloss that inherits on the fact
Karl Friston (52:14.920)
that the very preservation of the boundary
Lex Fridman (52:19.720)
between the oil drop and the not oil drop
Karl Friston (52:22.640)
requires an optimization of a particular function
Lex Fridman (52:26.160)
or a functional that defines the presence
Karl Friston (52:30.600)
or the existence of this oil drop,
Lex Fridman (52:33.040)
which is why I started with existential imperatives.
Karl Friston (52:36.160)
It is a necessary condition for existence
Lex Fridman (52:39.520)
that this must occur because the boundary
Karl Friston (52:42.880)
basically defines the thing that's existing.
Lex Fridman (52:46.040)
So it is that self assembly aspect
Karl Friston (52:47.960)
it's that you were hinting at in biology,
Lex Fridman (52:53.080)
sometimes known as autopoiesis
Karl Friston (52:56.880)
in computational chemistry with self assembly.
Lex Fridman (53:00.160)
It's the, what does it look like?
Karl Friston (53:03.600)
Sorry, how would you describe things
Lex Fridman (53:06.040)
that configure themselves out of nothing?
Karl Friston (53:08.640)
The way they clearly demarcate themselves
Lex Fridman (53:12.000)
from the states or the soup in which they are immersed.
Lex Fridman (53:18.240)
So from the point of view of computational chemistry,
Lex Fridman (53:20.920)
for example, you would just understand that
Karl Friston (53:23.440)
as a configuration of a macro molecule
Lex Fridman (53:25.440)
to minimize its free energy, its thermodynamic free energy.
Karl Friston (53:28.920)
It's exactly the same principle that we've been talking about
Lex Fridman (53:31.480)
that thermodynamic free energy is just the negative elbow.
Karl Friston (53:35.040)
It's the same mathematical construct.
Lex Fridman (53:38.200)
So the very emergence of existence, of structure, of form
Karl Friston (53:42.520)
that can be distinguished from the environment
Lex Fridman (53:45.040)
or the thing that is not the thing
Karl Friston (53:49.240)
necessitates the existence of an objective function
Lex Fridman (53:56.120)
that it looks as if it is minimizing.
Karl Friston (53:58.160)
It's finding a free energy minima.
Lex Fridman (54:00.280)
And so just to clarify, I'm trying to wrap my head around.
Lex Fridman (54:04.880)
So the free energy principle says that if something exists,
Lex Fridman (54:09.560)
these are the properties it should display.
Karl Friston (54:11.520)
Yeah.
Lex Fridman (54:12.520)
So what that means is we can't just look,
Karl Friston (54:17.520)
we can't just go into a soup and there's no mechanism.
Lex Fridman (54:21.320)
Free energy principle doesn't give us a mechanism
Karl Friston (54:23.480)
to find the things that exist.
Lex Fridman (54:25.680)
Is that what's implying, is being implied
Karl Friston (54:28.680)
that you can kind of use it to reason,
Lex Fridman (54:33.200)
to think about like, study a particular system
Lex Fridman (54:36.040)
and say, does this exhibit these qualities?
Lex Fridman (54:40.600)
That's an excellent question.
Lex Fridman (54:42.400)
But to answer that, I'd have to return
Lex Fridman (54:44.680)
to your previous question about what's the difference
Karl Friston (54:46.240)
between living and nonliving things.
Lex Fridman (54:48.520)
Yes, well, actually, sorry.
Lex Fridman (54:51.880)
So yeah, maybe we can go there.
Lex Fridman (54:54.080)
Maybe we can go there, you kind of drew a line
Lex Fridman (54:57.160)
and forgive me for the stupid questions,
Lex Fridman (54:58.960)
but you kind of drew a line between living and existing.
Lex Fridman (55:02.560)
Is there an interesting sort of distinction?
Lex Fridman (55:07.200)
Yeah, I think there is.
Lex Fridman (55:09.600)
So things do exist, grains of sand,
Lex Fridman (55:15.240)
rocks on the moon, trees, you.
Lex Fridman (55:19.480)
So all of these things can be separated from the environment
Lex Fridman (55:24.960)
in which they are immersed.
Lex Fridman (55:26.320)
And therefore, they must at some level
Lex Fridman (55:28.200)
be optimizing their free energy,
Karl Friston (55:32.320)
taking this sort of model evidence interpretation
Lex Fridman (55:36.200)
of this quantity that basically means
Karl Friston (55:38.160)
they're self evidencing.
Lex Fridman (55:39.640)
Another nice little twist of phrase here
Karl Friston (55:42.840)
is that you are your own existence proof,
Lex Fridman (55:45.520)
statistically speaking, which I don't think
Karl Friston (55:48.880)
I said that, somebody did, but I love that phrase.
Lex Fridman (55:53.480)
You are your own existence proof.
Lex Fridman (55:55.600)
Yeah, so it's so existential, isn't it?
Lex Fridman (55:59.280)
I'm gonna have to think about that for a few days.
Karl Friston (56:01.440)
That's a beautiful line.
Lex Fridman (56:06.080)
So the step through to answer your question
Karl Friston (56:09.760)
about what's it good for,
Lex Fridman (56:13.840)
we'll go along the following lines.
Karl Friston (56:15.760)
First of all, you have to define what it means
Lex Fridman (56:18.720)
to exist, which now, as you've rightly pointed out,
Karl Friston (56:22.040)
you have to define what probabilistic properties
Lex Fridman (56:25.000)
must the states of something possess
Lex Fridman (56:27.440)
so it knows where it finishes.
Lex Fridman (56:30.600)
And then you write that down in terms
Karl Friston (56:32.720)
of statistical dependencies, again, sparsity.
Lex Fridman (56:36.000)
Again, it's not what's connected or what's correlated
Karl Friston (56:39.680)
or what depends upon, it's what's not correlated
Lex Fridman (56:43.720)
and what doesn't depend upon something.
Karl Friston (56:45.960)
Again, it comes down to the deep structures,
Lex Fridman (56:49.680)
not in this instance, hierarchical,
Lex Fridman (56:50.920)
but the structures that emerge
Lex Fridman (56:54.040)
from removing connectivity and dependency.
Lex Fridman (56:56.920)
And in this instance, basically being able
Lex Fridman (57:00.160)
to identify the surface of the oil drop
Karl Friston (57:02.720)
from the water in which it is immersed.
Lex Fridman (57:06.480)
And when you do that, you start to realize,
Karl Friston (57:09.120)
well, there are actually four kinds of states
Lex Fridman (57:12.720)
in any given universe that contains anything.
Karl Friston (57:15.680)
The things that are internal to the surface,
Lex Fridman (57:18.840)
the things that are external to the surface
Lex Fridman (57:20.680)
and the surface in and of itself,
Lex Fridman (57:22.560)
which is why I use a metaphor,
Karl Friston (57:24.040)
a little single celled organism
Lex Fridman (57:25.440)
that has an interior and exterior
Lex Fridman (57:27.080)
and then the surface of the cell.
Lex Fridman (57:29.520)
And that's mathematically a Markov blanket.
Karl Friston (57:32.720)
Just to pause, I'm in awe of this concept
Lex Fridman (57:34.960)
that there's the stuff outside the surface,
Karl Friston (57:36.560)
stuff inside the surface and the surface itself,
Lex Fridman (57:38.960)
the Markov blanket.
Karl Friston (57:40.240)
It's just the most beautiful kind of notion
Lex Fridman (57:43.680)
about trying to explore what it means
Karl Friston (57:46.160)
to exist mathematically.
Lex Fridman (57:48.240)
I apologize, it's just a beautiful idea.
Lex Fridman (57:50.720)
But it came out of California, so that's.
Lex Fridman (57:53.080)
I changed my mind.
Karl Friston (57:54.000)
I take it all back.
Lex Fridman (57:55.080)
So anyway, so you were just talking
Karl Friston (57:59.440)
about the surface, about the Markov blanket.
Lex Fridman (58:01.280)
So this surface or these blanket states
Karl Friston (58:04.680)
that are the, because they are now defined
Lex Fridman (58:09.680)
in relation to these independencies
Lex Fridman (58:17.560)
and what different states internal blanket
Lex Fridman (58:21.560)
or external states can,
Karl Friston (58:23.840)
which ones can influence each other
Lex Fridman (58:25.280)
and which cannot influence each other.
Karl Friston (58:27.760)
You can now apply standard results
Lex Fridman (58:30.960)
that you would find in non equilibrium physics
Karl Friston (58:33.600)
or steady state or thermodynamics or hydrodynamics,
Lex Fridman (58:37.880)
usually out of equilibrium solutions
Lex Fridman (58:41.640)
and apply them to this partition.
Lex Fridman (58:43.200)
And what it looks like is if all the normal gradient flows
Karl Friston (58:48.120)
that you would associate with any non equilibrium system
Lex Fridman (58:52.080)
apply in such a way that part of the Markov blanket
Lex Fridman (58:57.600)
and the internal states seem to be hill climbing
Lex Fridman (59:01.560)
or doing a gradient descent on the same quantity.
Lex Fridman (59:05.840)
And that means that you can now describe
Lex Fridman (59:09.360)
the very existence of this oil drop.
Karl Friston (59:13.120)
You can write down the existence of this oil drop
Lex Fridman (59:16.000)
in terms of flows, dynamics, equations of motion,
Karl Friston (59:20.600)
where the blanket states or part of them,
Lex Fridman (59:24.160)
we call them active states and the internal states
Karl Friston (59:28.040)
now seem to be and must be trying to look
Lex Fridman (59:32.680)
as if they're minimizing the same function,
Karl Friston (59:35.480)
which is a low probability of occupying these states.
Lex Fridman (59:39.280)
Interesting thing is that what would they be called
Lex Fridman (59:44.040)
if you were trying to describe these things?
Lex Fridman (59:45.680)
So what we're talking about are internal states,
Karl Friston (59:50.080)
external states and blanket states.
Lex Fridman (59:52.040)
Now let's carve the blanket states
Karl Friston (59:54.080)
into two sensory states and active states.
Lex Fridman (59:57.200)
Operationally, it has to be the case
Karl Friston (59:59.520)
that in order for this carving up
🔗 相关节目
Alex Garland: Ex Machina, Devs, Annihilation, and the Poetry of Science
Alex Garland
Rob Reid: The Existential Threat of Engineered Viruses and Lab Leaks
Rob Reid
Andrew Huberman: Sleep, Dreams, Creativity & the Limits of the Human Mind
Andrew Huberman
Andrew Huberman: Neuroscience of Optimal Performance
Andrew Huberman