Matt Botvinick: Neuroscience, Psychology, and AI at DeepMind
心理与人性生物与进化AI 与机器学习音乐与艺术技术与编程
📋 章节目录
暂无章节信息
🔑 关键词
learningbrainneurosciencehumanhumanspsychologydoingbehaviorneuralresearchmetadoncortexmemoryintelligencesaidprefrontalreinforcementthinkinginteresting
💬 精彩语录
暂无语录
🎙️ 完整对话（2396 条）
Lex Fridman (00:00.000)
The following is a conversation with Matt Botmanek,
以下是与 Matt Botmanek 的对话，
Lex Fridman (00:03.440)
Director of Neuroscience Research at DeepMind.
DeepMind 神经科学研究总监。
Lex Fridman (00:06.680)
He's a brilliant, cross disciplinary mind,
他是一个才华横溢、跨学科的头脑
Lex Fridman (00:09.360)
navigating effortlessly between cognitive psychology,
毫不费力地在认知心理学之间导航，
Lex Fridman (00:12.480)
computational neuroscience, and artificial intelligence.
计算神经科学和人工智能。
Matt Botvinick (00:16.760)
Quick summary of the ads.
广告的快速摘要。
Lex Fridman (00:18.320)
Two sponsors, The Jordan Harbinger Show
两位赞助商，乔丹先驱秀
Lex Fridman (00:21.060)
and Magic Spoon Cereal.
Lex Fridman (00:23.880)
Please consider supporting the podcast
请考虑支持播客
Matt Botvinick (00:25.600)
by going to jordanharbinger.com slash lex
访问 jordanharbinger.com 斜杠 lex
Lex Fridman (00:29.320)
and also going to magicspoon.com slash lex
还可以访问 magicspoon.com 斜线 lex
Lex Fridman (00:33.800)
and using code lex at checkout
并在结帐时使用代码 lex
Lex Fridman (00:36.120)
after you buy all of their cereal.
当你买完他们所有的麦片后。
Matt Botvinick (00:39.080)
Click the links, buy the stuff.
点击链接，购买东西。
Lex Fridman (00:40.920)
It's the best way to support this podcast
这是支持此播客的最佳方式
Lex Fridman (00:43.040)
and the journey I'm on.
以及我正在进行的旅程。
Lex Fridman (00:44.740)
If you enjoy this podcast, subscribe on YouTube,
如果您喜欢这个播客，请在 YouTube 上订阅，
Matt Botvinick (00:47.680)
review it with five stars on Apple Podcast,
在 Apple Podcast 上以五颗星评价它，
Lex Fridman (00:49.920)
follow on Spotify, support on Patreon,
关注 Spotify，支持 Patreon，
Matt Botvinick (00:52.380)
or connect with me on Twitter at lexfriedman,
或者通过 Twitter 上的 lexfriedman 与我联系，
Lex Fridman (00:55.600)
spelled surprisingly without the E,
Matt Botvinick (00:58.920)
just F R I D M A N.
Lex Fridman (01:02.080)
As usual, I'll do a few minutes of ads now
Lex Fridman (01:03.920)
and never any ads in the middle
Lex Fridman (01:05.160)
that can break the flow of the conversation.
Matt Botvinick (01:07.620)
This episode is supported by The Jordan Harbinger Show.
Lex Fridman (01:11.740)
Go to jordanharbinger.com slash lex.
Matt Botvinick (01:15.200)
It's how he knows I sent you.
Lex Fridman (01:16.900)
On that page, subscribe to his podcast
Matt Botvinick (01:19.400)
on Apple Podcast, Spotify, and you know where to look.
Lex Fridman (01:24.320)
I've been binging on his podcast.
Matt Botvinick (01:26.120)
Jordan is a great interviewer
Lex Fridman (01:28.400)
and even a better human being.
Matt Botvinick (01:30.280)
I recently listened to his conversation with Jack Barsky,
Lex Fridman (01:32.760)
former sleeper agent for the KGB in the 80s
Lex Fridman (01:36.120)
and author of Deep Undercover,
Lex Fridman (01:38.880)
which is a memoir that paints yet another
Matt Botvinick (01:40.740)
interesting perspective on the Cold War era.
Lex Fridman (01:43.440)
I've been reading a lot about the Stalin
Lex Fridman (01:46.720)
and then Gorbachev and Putin eras of Russia,
Lex Fridman (01:49.280)
but this conversation made me realize
Matt Botvinick (01:50.800)
that I need to do a deep dive into the Cold War era
Lex Fridman (01:53.680)
to get a complete picture of Russia's recent history.
Matt Botvinick (01:57.120)
Again, go to jordanharbinger.com slash lex.
Lex Fridman (02:01.160)
Subscribe to his podcast.
Matt Botvinick (02:02.880)
It's how he knows I sent you.
Lex Fridman (02:04.440)
It's awesome, you won't regret it.
Matt Botvinick (02:06.740)
This episode is also supported by Magic Spoon,
Lex Fridman (02:10.320)
low carb, keto friendly, super amazingly delicious cereal.
Matt Botvinick (02:15.700)
I've been on a keto or very low carb diet
Lex Fridman (02:18.300)
for a long time now.
Matt Botvinick (02:19.480)
It helps with my mental performance.
Lex Fridman (02:21.300)
It helps with my physical performance,
Matt Botvinick (02:22.840)
even during this crazy push up, pull up challenge I'm doing,
Lex Fridman (02:26.520)
including the running, it just feels great.
Matt Botvinick (02:29.680)
I used to love cereal.
Lex Fridman (02:31.320)
Obviously, I can't have it now
Matt Botvinick (02:33.840)
because most cereals have crazy amounts of sugar,
Lex Fridman (02:36.820)
which is terrible for you, so I quit it years ago.
Lex Fridman (02:40.140)
But Magic Spoon, amazingly, somehow,
Lex Fridman (02:44.260)
is a totally different thing.
Matt Botvinick (02:45.920)
Zero sugar, 11 grams of protein,
Lex Fridman (02:48.340)
and only three net grams of carbs.
Matt Botvinick (02:50.920)
It tastes delicious.
Lex Fridman (02:53.140)
It has a lot of flavors, two new ones,
Matt Botvinick (02:55.200)
including peanut butter.
Lex Fridman (02:56.760)
But if you know what's good for you,
Matt Botvinick (02:58.520)
you'll go with cocoa, my favorite flavor,
Lex Fridman (03:01.560)
and the flavor of champions.
Matt Botvinick (03:04.200)
Click the magicspoon.com slash lex link in the description
Lex Fridman (03:07.880)
and use code lex at checkout for free shipping
Lex Fridman (03:11.040)
and to let them know I sent you.
Lex Fridman (03:13.100)
They have agreed to sponsor this podcast for a long time.
Matt Botvinick (03:16.480)
They're an amazing sponsor and an even better cereal.
Lex Fridman (03:19.920)
I highly recommend it.
Matt Botvinick (03:21.760)
It's delicious, it's good for you, you won't regret it.
Lex Fridman (03:24.720)
And now, here's my conversation with Matt Botpenik.
Lex Fridman (03:29.600)
How much of the human brain do you think we understand?
Lex Fridman (03:33.400)
I think we're at a weird moment
Matt Botvinick (03:36.920)
in the history of neuroscience in the sense that
Lex Fridman (03:45.200)
I feel like we understand a lot about the brain
Matt Botvinick (03:47.320)
at a very high level, but a very coarse level.
Lex Fridman (03:52.600)
When you say high level, what are you thinking?
Lex Fridman (03:54.280)
Are you thinking functional?
Lex Fridman (03:55.440)
Are you thinking structurally?
Lex Fridman (03:56.960)
So in other words, what is the brain for?
Lex Fridman (04:00.960)
What kinds of computation does the brain do?
Lex Fridman (04:05.000)
What kinds of behaviors would we have to explain
Lex Fridman (04:12.320)
if we were gonna look down at the mechanistic level?
Lex Fridman (04:16.560)
And at that level, I feel like we understand
Lex Fridman (04:18.440)
much, much more about the brain
Matt Botvinick (04:19.680)
than we did when I was in high school.
Lex Fridman (04:22.060)
But it's almost like we're seeing it through a fog.
Matt Botvinick (04:25.240)
It's only at a very coarse level.
Lex Fridman (04:26.600)
We don't really understand what the neuronal mechanisms are
Matt Botvinick (04:30.200)
that underlie these computations.
Lex Fridman (04:32.500)
We've gotten better at saying,
Lex Fridman (04:34.600)
what are the functions that the brain is computing
Lex Fridman (04:36.720)
that we would have to understand
Lex Fridman (04:38.400)
if we were gonna get down to the neuronal level?
Lex Fridman (04:40.200)
And at the other end of the spectrum,
Matt Botvinick (04:45.500)
in the last few years, incredible progress has been made
Lex Fridman (04:49.600)
in terms of technologies that allow us to see,
Matt Botvinick (04:54.880)
actually literally see, in some cases,
Lex Fridman (04:57.220)
what's going on at the single unit level,
Matt Botvinick (05:01.040)
even the dendritic level.
Lex Fridman (05:02.640)
And then there's this yawning gap in between.
Matt Botvinick (05:05.800)
Well, that's interesting.
Lex Fridman (05:06.640)
So at the high level,
Lex Fridman (05:07.460)
so that's almost a cognitive science level.
Lex Fridman (05:09.600)
And then at the neuronal level,
Matt Botvinick (05:11.900)
that's neurobiology and neuroscience,
Lex Fridman (05:14.600)
just studying single neurons,
Matt Botvinick (05:16.040)
the synaptic connections and all the dopamine,
Lex Fridman (05:19.800)
all the kind of neurotransmitters.
Matt Botvinick (05:21.560)
One blanket statement I should probably make
Lex Fridman (05:23.360)
is that as I've gotten older,
Matt Botvinick (05:27.760)
I have become more and more reluctant
Lex Fridman (05:30.200)
to make a distinction between psychology and neuroscience.
Matt Botvinick (05:33.400)
To me, the point of neuroscience
Lex Fridman (05:37.240)
is to study what the brain is for.
Matt Botvinick (05:41.780)
If you're a nephrologist
Lex Fridman (05:44.360)
and you wanna learn about the kidney,
Lex Fridman (05:46.560)
you start by saying, what is this thing for?
Lex Fridman (05:50.000)
Well, it seems to be for taking blood on one side
Matt Botvinick (05:55.800)
that has metabolites in it that shouldn't be there,
Lex Fridman (06:01.120)
sucking them out of the blood
Matt Botvinick (06:03.320)
while leaving the good stuff behind,
Lex Fridman (06:05.160)
and then excreting that in the form of urine.
Matt Botvinick (06:07.060)
That's what the kidney is for.
Lex Fridman (06:08.400)
It's like obvious.
Lex Fridman (06:10.240)
So the rest of the work is deciding how it does that.
Lex Fridman (06:13.200)
And this, it seems to me,
Matt Botvinick (06:14.800)
is the right approach to take to the brain.
Lex Fridman (06:17.080)
You say, well, what is the brain for?
Matt Botvinick (06:19.120)
The brain, as far as I can tell, is for producing behavior.
Lex Fridman (06:22.760)
It's for going from perceptual inputs to behavioral outputs,
Lex Fridman (06:27.980)
and the behavioral outputs should be adaptive.
Lex Fridman (06:31.420)
So that's what psychology is about.
Matt Botvinick (06:33.620)
It's about understanding the structure of that function.
Lex Fridman (06:35.920)
And then the rest of neuroscience is about figuring out
Lex Fridman (06:38.920)
how those operations are actually carried out
Lex Fridman (06:41.880)
at a mechanistic level.
Matt Botvinick (06:44.160)
That's really interesting, but so unlike the kidney,
Lex Fridman (06:47.960)
the brain, the gap between the electrical signal
Lex Fridman (06:52.020)
and behavior, so you truly see neuroscience
Lex Fridman (06:57.120)
as the science that touches behavior,
Lex Fridman (07:01.220)
how the brain generates behavior,
Lex Fridman (07:03.260)
or how the brain converts raw visual information
Matt Botvinick (07:07.400)
into understanding.
Lex Fridman (07:08.960)
Like, you basically see cognitive science,
Matt Botvinick (07:12.520)
psychology, and neuroscience as all one science.
Lex Fridman (07:15.860)
Yeah, it's a personal statement.
Lex Fridman (07:19.240)
Is that a hopeful or a realistic statement?
Lex Fridman (07:22.920)
So certainly you will be correct in your feeling
Matt Botvinick (07:26.880)
in some number of years, but that number of years
Lex Fridman (07:29.240)
could be 200, 300 years from now.
Matt Botvinick (07:31.440)
Oh, well, there's a...
Lex Fridman (07:33.400)
Is that aspirational or is that pragmatic engineering
Lex Fridman (07:37.600)
feeling that you have?
Lex Fridman (07:39.360)
It's both in the sense that this is what I hope
Lex Fridman (07:46.520)
and expect will bear fruit over the coming decades,
Lex Fridman (07:53.360)
but it's also pragmatic in the sense that I'm not sure
Lex Fridman (07:57.560)
what we're doing in either psychology or neuroscience
Lex Fridman (08:02.840)
if that's not the framing.
Matt Botvinick (08:04.920)
I don't know what it means to understand the brain
Lex Fridman (08:09.760)
if there's no, if part of the enterprise
Matt Botvinick (08:14.320)
is not about understanding the behavior
Lex Fridman (08:18.520)
that's being produced.
Matt Botvinick (08:20.020)
I mean, yeah, but I would compare it
Lex Fridman (08:23.040)
to maybe astronomers looking at the movement
Matt Botvinick (08:25.880)
of the planets and the stars without any interest
Lex Fridman (08:30.120)
of the underlying physics, right?
Lex Fridman (08:32.360)
And I would argue that at least in the early days,
Lex Fridman (08:35.560)
there is some value to just tracing the movement
Matt Botvinick (08:37.780)
of the planets and the stars without thinking
Lex Fridman (08:41.680)
about the physics too much because it's such a big leap
Matt Botvinick (08:44.100)
to start thinking about the physics
Lex Fridman (08:45.600)
before you even understand even the basic structural
Matt Botvinick (08:48.640)
elements of...
Lex Fridman (08:49.520)
Oh, I agree with that.
Matt Botvinick (08:50.420)
I agree.
Lex Fridman (08:51.260)
But you're saying in the end, the goal should be
Matt Botvinick (08:53.240)
to deeply understand.
Lex Fridman (08:54.760)
Well, right, and I think...
Lex Fridman (08:57.300)
So I thought about this a lot when I was in grad school
Lex Fridman (08:59.240)
because a lot of what I studied in grad school
Matt Botvinick (09:00.600)
was psychology and I found myself a little bit confused
Lex Fridman (09:06.120)
about what it meant to...
Matt Botvinick (09:08.680)
It seems like what we were talking about a lot of the time
Lex Fridman (09:11.500)
were virtual causal mechanisms.
Matt Botvinick (09:14.800)
Like, oh, well, you know, attentional selection
Lex Fridman (09:18.500)
then selects some object in the environment
Lex Fridman (09:22.200)
and that is then passed on to the motor, you know,
Lex Fridman (09:25.600)
information about that is passed on to the motor system.
Lex Fridman (09:27.800)
But these are virtual mechanisms.
Lex Fridman (09:29.760)
These are, you know, they're metaphors.
Matt Botvinick (09:31.480)
They're, you know, there's no reduction going on
Lex Fridman (09:37.040)
in that conversation to some physical mechanism that,
Matt Botvinick (09:40.200)
you know, which is really what it would take
Lex Fridman (09:43.240)
to fully understand, you know, how behavior is rising.
Lex Fridman (09:47.320)
But the causal mechanisms are definitely neurons interacting.
Lex Fridman (09:50.780)
I'm willing to say that at this point in history.
Lex Fridman (09:53.360)
So in psychology, at least for me personally,
Lex Fridman (09:56.240)
there was this strange insecurity about trafficking
Matt Botvinick (10:00.160)
in these metaphors, you know,
Lex Fridman (10:02.680)
which were supposed to explain the function of the mind.
Matt Botvinick (10:07.360)
If you can't ground them in physical mechanisms,
Lex Fridman (10:09.400)
then what is the explanatory validity of these explanations?
Lex Fridman (10:16.120)
And I managed to soothe my own nerves
Lex Fridman (10:21.120)
by thinking about the history of genetics research.
Lex Fridman (10:29.400)
So I'm very far from being an expert
Lex Fridman (10:32.460)
on the history of this field.
Lex Fridman (10:34.660)
But I know enough to say that, you know,
Lex Fridman (10:38.160)
Mendelian genetics preceded, you know, Watson and Crick.
Lex Fridman (10:42.800)
And so there was a significant period of time
Lex Fridman (10:45.520)
during which people were, you know,
Matt Botvinick (10:49.600)
productively investigating the structure of inheritance
Lex Fridman (10:54.760)
using what was essentially a metaphor,
Matt Botvinick (10:56.880)
the notion of a gene, you know.
Lex Fridman (10:58.600)
Oh, genes do this and genes do that.
Lex Fridman (11:00.760)
But, you know, where are the genes?
Lex Fridman (11:02.520)
They're sort of an explanatory thing that we made up.
Lex Fridman (11:06.080)
And we ascribed to them these causal properties.
Lex Fridman (11:08.880)
Oh, there's a dominant, there's a recessive,
Lex Fridman (11:10.640)
and then they recombine it.
Lex Fridman (11:12.800)
And then later, there was a kind of blank there
Matt Botvinick (11:17.460)
that was filled in with a physical mechanism.
Lex Fridman (11:21.620)
That connection was made.
Lex Fridman (11:24.300)
But it was worth having that metaphor
Lex Fridman (11:26.800)
because that gave us a good sense
Matt Botvinick (11:29.360)
of what kind of causal mechanism we were looking for.
Lex Fridman (11:34.280)
And the fundamental metaphor of cognition, you said,
Matt Botvinick (11:38.880)
is the interaction of neurons.
Lex Fridman (11:40.780)
Is that, what is the metaphor?
Matt Botvinick (11:42.680)
No, no, the metaphor,
Lex Fridman (11:44.280)
the metaphors we use in cognitive psychology
Matt Botvinick (11:47.640)
are things like attention, the way that memory works.
Lex Fridman (11:56.040)
I retrieve something from memory, right?
Matt Botvinick (11:59.440)
A memory retrieval occurs.
Lex Fridman (12:01.880)
What is that?
Matt Botvinick (12:02.860)
You know, that's not a physical mechanism
Lex Fridman (12:06.620)
that I can examine in its own right.
Lex Fridman (12:08.960)
But it's still worth having, that metaphorical level.
Lex Fridman (12:13.840)
Yeah, so yeah, I misunderstood actually.
Lex Fridman (12:16.000)
So the higher level of abstractions
Lex Fridman (12:17.640)
is the metaphor that's most useful.
Matt Botvinick (12:19.640)
Yes.
Lex Fridman (12:20.480)
But what about, so how does that connect
Lex Fridman (12:24.420)
to the idea that that arises from interaction of neurons?
Lex Fridman (12:33.000)
Well, even, is the interaction of neurons
Lex Fridman (12:35.940)
also not a metaphor to you?
Lex Fridman (12:38.080)
Or is it literally, like that's no longer a metaphor.
Matt Botvinick (12:42.400)
That's already the lowest level of abstractions
Lex Fridman (12:46.160)
that could actually be directly studied.
Matt Botvinick (12:50.280)
Well, I'm hesitating because I think
Lex Fridman (12:53.840)
what I want to say could end up being controversial.
Lex Fridman (12:57.960)
So what I want to say is, yes,
Lex Fridman (12:59.960)
the interactions of neurons, that's not metaphorical.
Matt Botvinick (13:03.040)
That's a physical fact.
Lex Fridman (13:04.680)
That's where the causal interactions actually occur.
Matt Botvinick (13:08.500)
Now, I suppose you could say,
Lex Fridman (13:09.880)
well, even that is metaphorical relative
Matt Botvinick (13:12.720)
to the quantum events that underlie.
Lex Fridman (13:15.840)
I don't want to go down that rabbit hole.
Matt Botvinick (13:17.320)
It's always turtles on top of turtles.
Lex Fridman (13:18.920)
Yeah, there's turtles all the way down.
Matt Botvinick (13:21.200)
There's a reduction that you can do.
Lex Fridman (13:22.560)
You can say these psychological phenomena
Matt Botvinick (13:25.720)
can be explained through a very different
Lex Fridman (13:28.200)
kind of causal mechanism,
Matt Botvinick (13:29.160)
which has to do with neurotransmitter release.
Lex Fridman (13:31.440)
And so what we're really trying to do
Matt Botvinick (13:33.800)
in neuroscience writ large, as I say,
Lex Fridman (13:37.120)
which for me includes psychology,
Matt Botvinick (13:39.760)
is to take these psychological phenomena
Lex Fridman (13:44.400)
and map them onto neural events.
Matt Botvinick (13:49.980)
I think remaining forever at the level of description
Lex Fridman (13:57.160)
that is natural for psychology,
Matt Botvinick (14:00.520)
for me personally, would be disappointing.
Lex Fridman (14:02.280)
I want to understand how mental activity
Matt Botvinick (14:05.640)
arises from neural activity.
Lex Fridman (14:10.360)
But the converse is also true.
Matt Botvinick (14:13.000)
Studying neural activity without any sense
Lex Fridman (14:15.880)
of what you're trying to explain,
Matt Botvinick (14:19.800)
to me feels like at best groping around at random.
Lex Fridman (14:27.280)
Now, you've kind of talked about this bridging
Matt Botvinick (14:30.280)
of the gap between psychology and neuroscience,
Lex Fridman (14:32.880)
but do you think it's possible,
Matt Botvinick (14:34.040)
like my love is, like I fell in love with psychology
Lex Fridman (14:38.280)
and psychiatry in general with Freud
Lex Fridman (14:40.120)
and when I was really young,
Lex Fridman (14:41.760)
and I hoped to understand the mind.
Lex Fridman (14:43.540)
And for me, understanding the mind,
Lex Fridman (14:45.240)
at least at that young age before I discovered AI
Lex Fridman (14:48.400)
and even neuroscience was to, is psychology.
Lex Fridman (14:52.840)
And do you think it's possible to understand the mind
Lex Fridman (14:55.840)
without getting into all the messy details of neuroscience?
Lex Fridman (14:59.920)
Like you kind of mentioned to you it's appealing
Matt Botvinick (15:03.120)
to try to understand the mechanisms at the lowest level,
Lex Fridman (15:06.040)
but do you think that's needed,
Lex Fridman (15:07.560)
that's required to understand how the mind works?
Lex Fridman (15:11.480)
That's an important part of the whole picture,
Lex Fridman (15:14.760)
but I would be the last person on earth
Lex Fridman (15:18.480)
to suggest that that reality
Matt Botvinick (15:23.440)
renders psychology in its own right unproductive.
Lex Fridman (15:29.440)
I trained as a psychologist.
Matt Botvinick (15:31.160)
I am fond of saying that I have learned much more
Lex Fridman (15:35.000)
from psychology than I have from neuroscience.
Matt Botvinick (15:38.480)
To me, psychology is a hugely important discipline.
Lex Fridman (15:43.740)
And one thing that warms in my heart is that
Matt Botvinick (15:50.360)
ways of investigating behavior
Lex Fridman (15:54.080)
that have been native to cognitive psychology
Matt Botvinick (15:58.000)
since it's dawn in the 60s
Lex Fridman (16:01.600)
are starting to become,
Matt Botvinick (16:03.960)
they're starting to become interesting to AI researchers
Lex Fridman (16:07.680)
for a variety of reasons.
Lex Fridman (16:09.480)
And that's been exciting for me to see.
Lex Fridman (16:11.680)
Can you maybe talk a little bit about what you see
Matt Botvinick (16:14.920)
as beautiful aspects of psychology,
Lex Fridman (16:19.320)
maybe limiting aspects of psychology?
Matt Botvinick (16:21.920)
I mean, maybe just start it off as a science, as a field.
Lex Fridman (16:25.640)
To me, it was when I understood what psychology is,
Matt Botvinick (16:29.760)
analytical psychology,
Lex Fridman (16:30.880)
like the way it's actually carried out,
Matt Botvinick (16:32.760)
it was really disappointing to see two aspects.
Lex Fridman (16:36.240)
One is how small the N is,
Lex Fridman (16:39.200)
how small the number of subject is in the studies.
Lex Fridman (16:43.040)
And two, it was disappointing to see
Lex Fridman (16:45.320)
how controlled the entire,
Lex Fridman (16:47.480)
how much it was in the lab.
Matt Botvinick (16:50.520)
It wasn't studying humans in the wild.
Lex Fridman (16:52.680)
There was no mechanism for studying humans in the wild.
Lex Fridman (16:55.000)
So that's where I became a little bit disillusioned
Lex Fridman (16:57.640)
to psychology.
Lex Fridman (16:59.480)
And then the modern world of the internet
Lex Fridman (17:01.680)
is so exciting to me.
Matt Botvinick (17:02.960)
The Twitter data or YouTube data,
Lex Fridman (17:05.720)
data of human behavior on the internet becomes exciting
Matt Botvinick (17:08.280)
because the N grows and then in the wild grows.
Lex Fridman (17:11.920)
But that's just my narrow sense.
Matt Botvinick (17:13.880)
Like, do you have a optimistic or pessimistic
Lex Fridman (17:16.560)
cynical view of psychology?
Lex Fridman (17:18.160)
How do you see the field broadly?
Lex Fridman (17:21.120)
When I was in graduate school,
Matt Botvinick (17:22.720)
it was early enough that there was still a thrill
Lex Fridman (17:27.800)
in seeing that there were ways of doing,
Matt Botvinick (17:32.960)
there were ways of doing experimental science
Lex Fridman (17:36.560)
that provided insight to the structure of the mind.
Matt Botvinick (17:40.040)
One thing that impressed me most when I was at that stage
Lex Fridman (17:43.720)
in my education was neuropsychology,
Matt Botvinick (17:46.000)
looking at, analyzing the behavior of populations
Lex Fridman (17:51.000)
who had brain damage of different kinds
Lex Fridman (17:55.560)
and trying to understand what the specific deficits were
Lex Fridman (18:02.920)
that arose from a lesion in a particular part of the brain.
Lex Fridman (18:06.760)
And the kind of experimentation that was done
Lex Fridman (18:08.960)
and that's still being done to get answers in that context
Matt Botvinick (18:13.520)
was so creative and it was so deliberate.
Lex Fridman (18:18.160)
It was good science.
Matt Botvinick (18:21.360)
An experiment answered one question but raised another
Lex Fridman (18:24.400)
and somebody would do an experiment
Matt Botvinick (18:25.600)
that answered that question.
Lex Fridman (18:26.600)
And you really felt like you were narrowing in on
Matt Botvinick (18:29.360)
some kind of approximate understanding
Lex Fridman (18:31.760)
of what this part of the brain was for.
Lex Fridman (18:34.840)
Do you have an example from memory
Lex Fridman (18:36.880)
of what kind of aspects of the mind
Lex Fridman (18:39.560)
could be studied in this kind of way?
Lex Fridman (18:41.400)
Oh, sure.
Matt Botvinick (18:42.240)
I mean, the very detailed neuropsychological studies
Lex Fridman (18:45.840)
of language function,
Matt Botvinick (18:49.720)
looking at production and reception
Lex Fridman (18:52.040)
and the relationship between visual function,
Matt Botvinick (18:57.080)
reading and auditory and semantic.
Lex Fridman (19:00.680)
There were these, and still are, these beautiful models
Matt Botvinick (19:03.920)
that came out of that kind of research
Lex Fridman (19:05.560)
that really made you feel like you understood something
Matt Botvinick (19:08.480)
that you hadn't understood before
Lex Fridman (19:10.320)
about how language processing is organized in the brain.
Lex Fridman (19:15.320)
But having said all that,
Lex Fridman (19:20.840)
I think you are, I mean, I agree with you
Matt Botvinick (19:25.400)
that the cost of doing highly controlled experiments
Lex Fridman (19:30.960)
is that you, by construction, miss out on the richness
Lex Fridman (19:36.480)
and complexity of the real world.
Lex Fridman (19:39.160)
One thing that, so I was drawn into science
Matt Botvinick (19:42.360)
by what in those days was called connectionism,
Lex Fridman (19:44.960)
which is, of course, what we now call deep learning.
Lex Fridman (19:49.120)
And at that point in history,
Lex Fridman (19:50.840)
neural networks were primarily being used
Matt Botvinick (19:54.200)
in order to model human cognition.
Lex Fridman (19:56.440)
They weren't yet really useful for industrial applications.
Matt Botvinick (1:00:00.780)
is unlikely to not be deeply meaningful, yeah.
Lex Fridman (1:00:03.340)
Yeah, the circumstantial evidence is overwhelming.
Lex Fridman (1:00:07.140)
But it could be.
Lex Fridman (1:00:07.980)
But you're always open to total flipping at the table.
Matt Botvinick (1:00:10.460)
Hey, of course.
Lex Fridman (1:00:11.620)
So you have coauthored several recent papers
Matt Botvinick (1:00:15.140)
that sort of weave beautifully between the world
Lex Fridman (1:00:17.860)
of neuroscience and artificial intelligence.
Lex Fridman (1:00:20.660)
And maybe if we could, can we just try to dance around
Lex Fridman (1:00:26.380)
and talk about some of them?
Matt Botvinick (1:00:27.500)
Maybe try to pick out interesting ideas
Lex Fridman (1:00:29.740)
that jump to your mind from memory.
Lex Fridman (1:00:32.300)
So maybe looking at, we were talking about
Lex Fridman (1:00:34.300)
the prefrontal cortex, the 2018, I believe, paper
Matt Botvinick (1:00:38.220)
called the Prefrontal Cortex
Lex Fridman (1:00:40.060)
as a Meta Reinforcement Learning System.
Matt Botvinick (1:00:42.140)
What, is there a key idea
Lex Fridman (1:00:44.340)
that you can speak to from that paper?
Matt Botvinick (1:00:47.660)
Yeah, I mean, the key idea is about meta learning.
Lex Fridman (1:00:53.860)
What is meta learning?
Matt Botvinick (1:00:54.860)
Meta learning is, by definition,
Lex Fridman (1:01:00.940)
a situation in which you have a learning algorithm
Lex Fridman (1:01:06.100)
and the learning algorithm operates in such a way
Lex Fridman (1:01:09.780)
that it gives rise to another learning algorithm.
Matt Botvinick (1:01:14.060)
In the earliest applications of this idea,
Lex Fridman (1:01:17.140)
you had one learning algorithm sort of adjusting
Matt Botvinick (1:01:20.340)
the parameters on another learning algorithm.
Lex Fridman (1:01:23.060)
But the case that we're interested in this paper
Matt Botvinick (1:01:25.100)
is one where you start with just one learning algorithm
Lex Fridman (1:01:29.140)
and then another learning algorithm kind of emerges
Matt Botvinick (1:01:33.020)
out of thin air.
Lex Fridman (1:01:35.180)
I can say more about what I mean by that.
Matt Botvinick (1:01:36.700)
I don't mean to be scurrentist,
Lex Fridman (1:01:39.780)
but that's the idea of meta learning.
Matt Botvinick (1:01:44.140)
It relates to the old idea in psychology
Lex Fridman (1:01:46.020)
of learning to learn.
Matt Botvinick (1:01:49.380)
Situations where you have experiences
Lex Fridman (1:01:54.300)
that make you better at learning something new.
Matt Botvinick (1:01:57.980)
A familiar example would be learning a foreign language.
Lex Fridman (1:02:01.380)
The first time you learn a foreign language,
Matt Botvinick (1:02:02.860)
it may be quite laborious and disorienting
Lex Fridman (1:02:06.420)
and novel, but let's say you've learned
Matt Botvinick (1:02:10.300)
two foreign languages.
Lex Fridman (1:02:12.220)
The third foreign language, obviously,
Matt Botvinick (1:02:14.140)
is gonna be much easier to pick up.
Lex Fridman (1:02:15.940)
And why?
Matt Botvinick (1:02:16.780)
Because you've learned how to learn.
Lex Fridman (1:02:18.220)
You know how this goes.
Matt Botvinick (1:02:20.220)
You know, okay, I'm gonna have to learn how to conjugate.
Lex Fridman (1:02:22.140)
I'm gonna have to...
Matt Botvinick (1:02:23.940)
That's a simple form of meta learning
Lex Fridman (1:02:26.340)
in the sense that there's some slow learning mechanism
Matt Botvinick (1:02:30.260)
that's helping you kind of update
Lex Fridman (1:02:33.020)
your fast learning mechanism.
Lex Fridman (1:02:34.300)
Does that make sense?
Lex Fridman (1:02:35.660)
So how from our understanding from the psychology world,
Matt Botvinick (1:02:40.540)
from neuroscience, our understanding
Lex Fridman (1:02:43.180)
how meta learning might work in the human brain,
Lex Fridman (1:02:47.180)
what lessons can we draw from that
Lex Fridman (1:02:49.980)
that we can bring into the artificial intelligence world?
Matt Botvinick (1:02:53.060)
Well, yeah, so the origin of that paper
Lex Fridman (1:02:55.980)
was in AI work that we were doing in my group.
Matt Botvinick (1:03:00.180)
We were looking at what happens
Lex Fridman (1:03:03.700)
when you train a recurrent neural network
Matt Botvinick (1:03:06.260)
using standard reinforcement learning algorithms.
Lex Fridman (1:03:10.180)
But you train that network, not just in one task,
Lex Fridman (1:03:12.660)
but you train it in a bunch of interrelated tasks.
Lex Fridman (1:03:15.140)
And then you ask what happens when you give it
Matt Botvinick (1:03:18.700)
yet another task in that sort of line of interrelated tasks.
Lex Fridman (1:03:23.380)
And what we started to realize is that
Matt Botvinick (1:03:29.380)
a form of meta learning spontaneously happens
Lex Fridman (1:03:31.860)
in recurrent neural networks.
Lex Fridman (1:03:33.780)
And the simplest way to explain it is to say
Lex Fridman (1:03:39.540)
a recurrent neural network has a kind of memory
Matt Botvinick (1:03:43.500)
in its activation patterns.
Lex Fridman (1:03:45.340)
It's recurrent by definition in the sense
Matt Botvinick (1:03:47.540)
that you have units that connect to other units,
Lex Fridman (1:03:50.180)
that connect to other units.
Lex Fridman (1:03:51.060)
So you have sort of loops of connectivity,
Lex Fridman (1:03:53.660)
which allows activity to stick around
Lex Fridman (1:03:55.740)
and be updated over time.
Lex Fridman (1:03:57.380)
In psychology we call, in neuroscience
Matt Botvinick (1:03:59.020)
we call this working memory.
Lex Fridman (1:04:00.100)
It's like actively holding something in mind.
Lex Fridman (1:04:04.260)
And so that memory gives
Lex Fridman (1:04:09.260)
the recurrent neural network a dynamics, right?
Matt Botvinick (1:04:13.100)
The way that the activity pattern evolves over time
Lex Fridman (1:04:17.700)
is inherent to the connectivity
Lex Fridman (1:04:19.980)
of the recurrent neural network, okay?
Lex Fridman (1:04:21.580)
So that's idea number one.
Matt Botvinick (1:04:23.500)
Now, the dynamics of that network are shaped
Lex Fridman (1:04:26.020)
by the connectivity, by the synaptic weights.
Lex Fridman (1:04:29.660)
And those synaptic weights are being shaped
Lex Fridman (1:04:31.660)
by this reinforcement learning algorithm
Matt Botvinick (1:04:33.860)
that you're training the network with.
Lex Fridman (1:04:37.700)
So the punchline is if you train a recurrent neural network
Matt Botvinick (1:04:41.260)
with a reinforcement learning algorithm
Lex Fridman (1:04:43.140)
that's adjusting its weights,
Lex Fridman (1:04:44.180)
and you do that for long enough,
Lex Fridman (1:04:47.060)
the activation dynamics will become very interesting, right?
Lex Fridman (1:04:50.860)
So imagine I give you a task
Lex Fridman (1:04:53.180)
where you have to press one button or another,
Matt Botvinick (1:04:56.060)
left button or right button.
Lex Fridman (1:04:57.580)
And there's some probability
Matt Botvinick (1:05:00.820)
that I'm gonna give you an M&M
Lex Fridman (1:05:02.260)
if you press the left button,
Lex Fridman (1:05:04.220)
and there's some probability I'll give you an M&M
Lex Fridman (1:05:06.220)
if you press the other button.
Lex Fridman (1:05:07.620)
And you have to figure out what those probabilities are
Lex Fridman (1:05:09.340)
just by trying things out.
Lex Fridman (1:05:12.060)
But as I said before,
Lex Fridman (1:05:13.780)
instead of just giving you one of these tasks,
Matt Botvinick (1:05:15.500)
I give you a whole sequence.
Lex Fridman (1:05:17.020)
You know, I give you two buttons
Lex Fridman (1:05:18.700)
and you figure out which one's best.
Lex Fridman (1:05:19.860)
And I go, good job, here's a new box.
Matt Botvinick (1:05:22.180)
Two new buttons, you have to figure out which one's best.
Lex Fridman (1:05:24.100)
Good job, here's a new box.
Lex Fridman (1:05:25.420)
And every box has its own probabilities
Lex Fridman (1:05:27.340)
and you have to figure it out.
Lex Fridman (1:05:28.300)
So if you train a recurrent neural network
Lex Fridman (1:05:30.420)
on that kind of sequence of tasks,
Lex Fridman (1:05:33.700)
what happens, it seemed almost magical to us
Lex Fridman (1:05:37.380)
when we first started kind of realizing what was going on.
Matt Botvinick (1:05:41.180)
The slow learning algorithm that's adjusting
Lex Fridman (1:05:43.620)
the synaptic weights,
Matt Botvinick (1:05:46.980)
those slow synaptic changes give rise to a network dynamics
Lex Fridman (1:05:51.380)
that themselves, that, you know,
Matt Botvinick (1:05:53.020)
the dynamics themselves turn into a learning algorithm.
Lex Fridman (1:05:56.860)
So in other words, you can tell this is happening
Matt Botvinick (1:05:59.060)
by just freezing the synaptic weights saying,
Lex Fridman (1:06:01.020)
okay, no more learning, you're done.
Matt Botvinick (1:06:03.460)
Here's a new box, figure out which button is best.
Lex Fridman (1:06:07.620)
And the recurrent neural network will do this just fine.
Matt Botvinick (1:06:09.620)
There's no, like it figures out which button is best.
Lex Fridman (1:06:13.060)
It kind of transitions from exploring the two buttons
Matt Botvinick (1:06:16.700)
to just pressing the one that it likes best
Lex Fridman (1:06:18.380)
in a very rational way.
Lex Fridman (1:06:20.700)
How is that happening?
Lex Fridman (1:06:21.660)
It's happening because the activity dynamics
Matt Botvinick (1:06:24.700)
of the network have been shaped by the slow learning process
Lex Fridman (1:06:28.460)
that's occurred over many, many boxes.
Lex Fridman (1:06:30.660)
And so what's happened is that this slow learning algorithm
Lex Fridman (1:06:34.660)
that's slowly adjusting the weights
Matt Botvinick (1:06:37.140)
is changing the dynamics of the network,
Lex Fridman (1:06:39.740)
the activity dynamics into its own learning algorithm.
Lex Fridman (1:06:43.460)
And as we were kind of realizing that this is a thing,
Lex Fridman (1:06:51.340)
it just so happened that the group that was working on this
Matt Botvinick (1:06:53.740)
included a bunch of neuroscientists
Lex Fridman (1:06:56.020)
and it started kind of ringing a bell for us,
Matt Botvinick (1:06:59.900)
which is to say that we thought this sounds a lot
Lex Fridman (1:07:02.860)
like the distinction between synaptic learning
Lex Fridman (1:07:06.180)
and activity, synaptic memory
Lex Fridman (1:07:08.460)
and activity based memory in the brain.
Lex Fridman (1:07:11.700)
And it also reminded us of recurrent connectivity
Lex Fridman (1:07:15.900)
that's very characteristic of prefrontal function.
Lex Fridman (1:07:19.620)
So this is kind of why it's good to have people working
Lex Fridman (1:07:22.820)
on AI that know a little bit about neuroscience
Lex Fridman (1:07:26.180)
and vice versa, because we started thinking
Lex Fridman (1:07:29.340)
about whether we could apply this principle to neuroscience.
Lex Fridman (1:07:32.340)
And that's where the paper came from.
Lex Fridman (1:07:33.660)
So the kind of principle of the recurrence
Matt Botvinick (1:07:37.540)
they can see in the prefrontal cortex,
Lex Fridman (1:07:39.540)
then you start to realize that it's possible
Matt Botvinick (1:07:43.660)
for something like an idea of a learning
Lex Fridman (1:07:46.340)
to learn emerging from this learning process
Matt Botvinick (1:07:50.860)
as long as you keep varying the environment sufficiently.
Lex Fridman (1:07:54.500)
Exactly, so the kind of metaphorical transition
Matt Botvinick (1:07:59.300)
we made to neuroscience was to think,
Lex Fridman (1:08:00.740)
okay, well, we know that the prefrontal cortex
Matt Botvinick (1:08:03.660)
is highly recurrent.
Lex Fridman (1:08:04.940)
We know that it's an important locus for working memory
Matt Botvinick (1:08:08.500)
for activation based memory.
Lex Fridman (1:08:11.260)
So maybe the prefrontal cortex
Matt Botvinick (1:08:13.660)
supports reinforcement learning.
Lex Fridman (1:08:15.620)
In other words, what is reinforcement learning?
Matt Botvinick (1:08:19.260)
You take an action, you see how much reward you got,
Lex Fridman (1:08:21.620)
you update your policy of behavior.
Matt Botvinick (1:08:24.580)
Maybe the prefrontal cortex is doing that sort of thing
Lex Fridman (1:08:26.860)
strictly in its activation patterns.
Matt Botvinick (1:08:28.500)
It's keeping around a memory in its activity patterns
Lex Fridman (1:08:31.900)
of what you did, how much reward you got,
Lex Fridman (1:08:35.340)
and it's using that activity based memory
Lex Fridman (1:08:38.980)
as a basis for updating behavior.
Lex Fridman (1:08:41.100)
But then the question is, well,
Lex Fridman (1:08:42.180)
how did the prefrontal cortex get so smart?
Lex Fridman (1:08:44.540)
In other words, where did these activity dynamics come from?
Lex Fridman (1:08:48.020)
How did that program that's implemented
Lex Fridman (1:08:50.780)
in the recurrent dynamics of the prefrontal cortex arise?
Lex Fridman (1:08:54.460)
And one answer that became evident in this work was,
Matt Botvinick (1:08:58.060)
well, maybe the mechanisms that operate
Lex Fridman (1:09:00.940)
on the synaptic level, which we believe are mediated
Matt Botvinick (1:09:05.020)
by dopamine, are responsible for shaping those dynamics.
Lex Fridman (1:09:10.180)
So this may be a silly question,
Lex Fridman (1:09:12.420)
but because this kind of several temporal sort of classes
Lex Fridman (1:09:19.340)
of learning are happening and the learning to learnism
Matt Botvinick (1:09:23.020)
emerges, can you keep building stacks of learning
Lex Fridman (1:09:28.660)
to learn to learn, learning to learn to learn
Matt Botvinick (1:09:30.940)
to learn to learn because it keeps,
Lex Fridman (1:09:32.900)
I mean, basically abstractions of more powerful abilities
Matt Botvinick (1:09:37.020)
to generalize of learning complex rules.
Lex Fridman (1:09:41.140)
Yeah, that's overstretching this kind of mechanism.
Matt Botvinick (1:09:46.100)
Well, one of the people in AI who started thinking
Lex Fridman (1:09:51.260)
about meta learning from very early on,
Matt Botvinick (1:09:54.700)
Jürgen Schmidhuber sort of cheekily suggested,
Lex Fridman (1:09:59.780)
I think it may have been in his PhD thesis,
Matt Botvinick (1:10:03.900)
that we should think about meta, meta, meta,
Lex Fridman (1:10:06.900)
meta, meta, meta learning.
Matt Botvinick (1:10:08.740)
That's really what's gonna get us to true intelligence.
Lex Fridman (1:10:13.140)
Certainly there's a poetic aspect to it
Lex Fridman (1:10:15.380)
and it seems interesting and correct
Lex Fridman (1:10:19.260)
that that kind of levels of abstraction would be powerful,
Lex Fridman (1:10:21.660)
but is that something you see in the brain?
Lex Fridman (1:10:23.940)
This kind of, is it useful to think of learning
Lex Fridman (1:10:27.780)
in these meta, meta, meta way or is it just meta learning?
Lex Fridman (1:10:32.100)
Well, one thing that really fascinated me
Matt Botvinick (1:10:35.300)
about this mechanism that we were starting to look at,
Lex Fridman (1:10:39.020)
and other groups started talking
Matt Botvinick (1:10:41.100)
about very similar things at the same time.
Lex Fridman (1:10:44.740)
And then a kind of explosion of interest
Matt Botvinick (1:10:47.020)
in meta learning happened in the AI community
Lex Fridman (1:10:48.980)
shortly after that.
Matt Botvinick (1:10:50.580)
I don't know if we had anything to do with that,
Lex Fridman (1:10:52.060)
but I was gratified to see that a lot of people
Matt Botvinick (1:10:55.620)
started talking about meta learning.
Lex Fridman (1:10:57.780)
One of the things that I liked about the kind of flavor
Matt Botvinick (1:11:01.380)
of meta learning that we were studying was that
Lex Fridman (1:11:04.060)
it didn't require anything special.
Matt Botvinick (1:11:05.940)
It was just, if you took a system that had
Lex Fridman (1:11:08.620)
some form of memory that the function of which
Matt Botvinick (1:11:12.460)
could be shaped by pick URL algorithm,
Lex Fridman (1:11:16.860)
then this would just happen, right?
Matt Botvinick (1:11:19.100)
I mean, there are a lot of forms of,
Lex Fridman (1:11:21.300)
there are a lot of meta learning algorithms
Matt Botvinick (1:11:23.180)
that have been proposed since then
Lex Fridman (1:11:24.500)
that are fascinating and effective
Matt Botvinick (1:11:26.580)
in their domains of application.
Lex Fridman (1:11:29.780)
But they're engineered, they're things that somebody
Matt Botvinick (1:11:32.580)
had to say, well, gee, if we wanted meta learning
Lex Fridman (1:11:34.340)
to happen, how would we do that?
Matt Botvinick (1:11:35.700)
Here's an algorithm that would,
Lex Fridman (1:11:37.060)
but there's something about the kind of meta learning
Matt Botvinick (1:11:39.500)
that we were studying that seemed to me special
Lex Fridman (1:11:42.540)
in the sense that it wasn't an algorithm.
Matt Botvinick (1:11:44.980)
It was just something that automatically happened
Lex Fridman (1:11:48.740)
if you had a system that had memory
Lex Fridman (1:11:51.060)
and it was trained with a reinforcement learning algorithm.
Lex Fridman (1:11:54.020)
And in that sense, it can be as meta as it wants to be.
Matt Botvinick (1:11:59.020)
There's no limit on how abstract the meta learning can get
Lex Fridman (1:12:04.700)
because it's not reliant on a human engineering
Matt Botvinick (1:12:07.980)
a particular meta learning algorithm to get there.
Lex Fridman (1:12:11.540)
And that's, I also, I don't know,
Matt Botvinick (1:12:15.140)
I guess I hope that that's relevant in the brain.
Lex Fridman (1:12:17.820)
I think there's a kind of beauty
Matt Botvinick (1:12:19.180)
in the ability of this emergent.
Lex Fridman (1:12:23.380)
The emergent aspect of it, as opposed to engineered.
Matt Botvinick (1:12:26.460)
Exactly, it's something that just, it just happens
Lex Fridman (1:12:29.020)
in a sense, in a sense, you can't avoid this happening.
Matt Botvinick (1:12:33.620)
If you have a system that has memory
Lex Fridman (1:12:35.820)
and the function of that memory is shaped
Matt Botvinick (1:12:39.660)
by reinforcement learning, and this system is trained
Lex Fridman (1:12:42.740)
in a series of interrelated tasks, this is gonna happen.
Matt Botvinick (1:12:46.900)
You can't stop it.
Lex Fridman (1:12:48.460)
As long as you have certain properties,
Matt Botvinick (1:12:50.140)
maybe like a recurrent structure to.
Lex Fridman (1:12:52.540)
You have to have memory.
Matt Botvinick (1:12:53.380)
It actually doesn't have to be a recurrent neural network.
Lex Fridman (1:12:55.220)
One of, a paper that I was honored to be involved
Matt Botvinick (1:12:58.740)
with even earlier, used a kind of slot based memory.
Lex Fridman (1:13:02.260)
Do you remember the title?
Matt Botvinick (1:13:03.100)
Just for people to understand.
Lex Fridman (1:13:05.060)
It was Memory Augmented Neural Networks.
Matt Botvinick (1:13:08.140)
I think it was, I think the title was
Lex Fridman (1:13:10.180)
Meta Learning in Memory Augmented Neural Networks.
Lex Fridman (1:13:14.660)
And it was the same exact story.
Lex Fridman (1:13:17.940)
If you have a system with memory,
Matt Botvinick (1:13:21.100)
here it was a different kind of memory,
Lex Fridman (1:13:22.780)
but the function of that memory is shaped
Matt Botvinick (1:13:26.860)
by reinforcement learning.
Lex Fridman (1:13:29.900)
Here it was the reads and writes that occurred
Matt Botvinick (1:13:34.300)
on this slot based memory.
Lex Fridman (1:13:36.420)
This will just happen.
Lex Fridman (1:13:39.940)
But this brings us back to something I was saying earlier
Lex Fridman (1:13:42.060)
about the importance of the environment.
Matt Botvinick (1:13:46.340)
This will happen if the system is being trained
Lex Fridman (1:13:49.940)
in a setting where there's like a sequence of tasks
Matt Botvinick (1:13:53.060)
that all share some abstract structure.
Lex Fridman (1:13:56.100)
Sometimes we talk about task distributions.
Lex Fridman (1:13:59.020)
And that's something that's very obviously true
Lex Fridman (1:14:04.180)
of the world that humans inhabit.
Matt Botvinick (1:14:09.500)
Like if you just kind of think about what you do every day,
Lex Fridman (1:14:13.140)
you never do exactly the same thing
Matt Botvinick (1:14:16.280)
that you did the day before.
Lex Fridman (1:14:17.640)
But everything that you do sort of has a family resemblance.
Matt Botvinick (1:14:21.060)
It shares a structure with something that you did before.
Lex Fridman (1:14:23.500)
And so the real world is sort of
Matt Botvinick (1:14:29.260)
saturated with this kind of, this property.
Lex Fridman (1:14:32.700)
It's endless variety with endless redundancy.
Lex Fridman (1:14:37.540)
And that's the setting in which
Lex Fridman (1:14:38.700)
this kind of meta learning happens.
Lex Fridman (1:14:40.540)
And it does seem like we're just so good at finding,
Lex Fridman (1:14:44.980)
just like in this emergent phenomena you described,
Matt Botvinick (1:14:47.820)
we're really good at finding that redundancy,
Lex Fridman (1:14:50.020)
finding those similarities, the family resemblance.
Lex Fridman (1:14:53.480)
Some people call it sort of, what is it?
Lex Fridman (1:14:56.560)
Melanie Mitchell was talking about analogies.
Lex Fridman (1:14:59.180)
So we're able to connect concepts together
Lex Fridman (1:15:01.940)
in this kind of way,
Matt Botvinick (1:15:03.860)
in this same kind of automated emergent way,
Lex Fridman (1:15:06.020)
which there's so many echoes here
Matt Botvinick (1:15:08.620)
of psychology and neuroscience.
Lex Fridman (1:15:10.640)
And obviously now with reinforcement learning
Matt Botvinick (1:15:15.300)
with recurrent neural networks at the core.
Lex Fridman (1:15:18.260)
If we could talk a little bit about dopamine,
Matt Botvinick (1:15:20.180)
you have really, you're a part of coauthoring
Lex Fridman (1:15:23.780)
really exciting recent paper, very recent,
Matt Botvinick (1:15:26.420)
in terms of release on dopamine
Lex Fridman (1:15:28.900)
and temporal difference learning.
Lex Fridman (1:15:31.040)
Can you describe the key ideas of that paper?
Lex Fridman (1:15:34.820)
Sure, yeah.
Matt Botvinick (1:15:35.660)
I mean, one thing I want to pause to do
Lex Fridman (1:15:37.740)
is acknowledge my coauthors
Matt Botvinick (1:15:39.460)
on actually both of the papers we're talking about.
Lex Fridman (1:15:41.540)
So this dopamine paper.
Matt Botvinick (1:15:42.660)
I'll just, I'll certainly post all their names.
Lex Fridman (1:15:45.700)
Okay, wonderful.
Matt Botvinick (1:15:46.540)
Yeah, because I'm sort of abashed
Lex Fridman (1:15:49.300)
to be the spokesperson for these papers
Matt Botvinick (1:15:51.000)
when I had such amazing collaborators on both.
Lex Fridman (1:15:55.180)
So it's a comfort to me to know
Matt Botvinick (1:15:56.980)
that you'll acknowledge them.
Lex Fridman (1:15:58.580)
Yeah, there's an incredible team there, but yeah.
Matt Botvinick (1:16:00.420)
Oh yeah, it's such a, it's so much fun.
Lex Fridman (1:16:03.080)
And in the case of the dopamine paper,
Matt Botvinick (1:16:06.360)
we also collaborated with Naochit at Harvard,
Lex Fridman (1:16:09.020)
who, you know, obviously a paper simply
Matt Botvinick (1:16:11.180)
wouldn't have happened without him.
Lex Fridman (1:16:12.620)
But so you were asking for like a thumbnail sketch of.
Matt Botvinick (1:16:17.540)
Yeah, thumbnail sketch or key ideas or, you know,
Lex Fridman (1:16:20.820)
things, the insights that are, you know,
Matt Botvinick (1:16:22.500)
continuing on our kind of discussion here
Lex Fridman (1:16:24.780)
between neuroscience and AI.
Matt Botvinick (1:16:26.900)
Yeah, I mean, this was another,
Lex Fridman (1:16:28.860)
a lot of the work that we've done so far
Matt Botvinick (1:16:30.620)
is taking ideas that have bubbled up in AI
Lex Fridman (1:16:35.380)
and, you know, asking the question of whether the brain
Matt Botvinick (1:16:39.660)
might be doing something related,
Lex Fridman (1:16:41.460)
which I think on the surface sounds like something
Matt Botvinick (1:16:45.420)
that's really mainly of use to neuroscience.
Lex Fridman (1:16:49.380)
We see it also as a way of validating
Lex Fridman (1:16:53.600)
what we're doing on the AI side.
Lex Fridman (1:16:55.320)
If we can gain some evidence that the brain
Matt Botvinick (1:16:57.940)
is using some technique that we've been trying out
Lex Fridman (1:17:01.760)
in our AI work, that gives us confidence
Matt Botvinick (1:17:05.500)
that, you know, it may be a good idea,
Lex Fridman (1:17:07.780)
that it'll, you know, scale to rich, complex tasks,
Matt Botvinick (1:17:11.560)
that it'll interface well with other mechanisms.
Lex Fridman (1:17:14.840)
So you see it as a two way road.
Matt Botvinick (1:17:16.860)
Yeah, for sure. Just because a particular paper
Lex Fridman (1:17:18.520)
is a little bit focused on from one to the,
Matt Botvinick (1:17:21.140)
from AI, from neural networks to neuroscience.
Lex Fridman (1:17:25.620)
Ultimately the discussion, the thinking,
Matt Botvinick (1:17:28.380)
the productive longterm aspect of it
Lex Fridman (1:17:30.840)
is the two way road nature of the whole interaction.
Matt Botvinick (1:17:33.220)
Yeah, I mean, we've talked about the notion
Lex Fridman (1:17:36.260)
of a virtuous circle between AI and neuroscience.
Matt Botvinick (1:17:39.300)
And, you know, the way I see it,
Lex Fridman (1:17:42.660)
that's always been there since the two fields,
Matt Botvinick (1:17:47.460)
you know, jointly existed.
Lex Fridman (1:17:50.100)
There have been some phases in that history
Matt Botvinick (1:17:52.140)
when AI was sort of ahead.
Lex Fridman (1:17:53.540)
There are some phases when neuroscience was sort of ahead.
Matt Botvinick (1:17:56.340)
I feel like given the burst of innovation
Lex Fridman (1:18:00.660)
that's happened recently on the AI side,
Matt Botvinick (1:18:03.780)
AI is kind of ahead in the sense that
Lex Fridman (1:18:06.320)
there are all of these ideas that we, you know,
Matt Botvinick (1:18:10.620)
for which it's exciting to consider
Lex Fridman (1:18:12.660)
that there might be neural analogs.
Lex Fridman (1:18:16.100)
And neuroscience, you know,
Lex Fridman (1:18:19.620)
in a sense has been focusing on approaches
Matt Botvinick (1:18:22.420)
to studying behavior that come from, you know,
Lex Fridman (1:18:24.860)
that are kind of derived from this earlier era
Matt Botvinick (1:18:27.540)
of cognitive psychology.
Lex Fridman (1:18:29.620)
And, you know, so in some ways fail to connect
Matt Botvinick (1:18:33.540)
with some of the issues that we're grappling with in AI.
Lex Fridman (1:18:36.700)
Like how do we deal with, you know,
Matt Botvinick (1:18:37.940)
large, you know, complex environments.
Lex Fridman (1:18:41.560)
But, you know, I think it's inevitable
Matt Botvinick (1:18:45.220)
that this circle will keep turning
Lex Fridman (1:18:47.920)
and there will be a moment
Matt Botvinick (1:18:49.540)
in the not too different distant future
Lex Fridman (1:18:51.300)
when neuroscience is pelting AI researchers
Matt Botvinick (1:18:54.640)
with insights that may change the direction of our work.
Lex Fridman (1:18:58.260)
Just a quick human question.
Matt Botvinick (1:19:00.940)
Is it, you have parts of your brain,
Lex Fridman (1:19:05.460)
this is very meta, but they're able to both think
Matt Botvinick (1:19:08.260)
about neuroscience and AI.
Lex Fridman (1:19:10.300)
You know, I don't often meet people like that.
Lex Fridman (1:19:14.220)
So do you think, let me ask a meta plasticity question.
Lex Fridman (1:19:19.780)
Do you think a human being can be both good at AI
Lex Fridman (1:19:22.660)
and neuroscience?
Lex Fridman (1:19:23.580)
It's like what, on the team at DeepMind,
Lex Fridman (1:19:26.500)
what kind of human can occupy these two realms?
Lex Fridman (1:19:30.180)
And is that something you see everybody should be doing,
Matt Botvinick (1:19:33.340)
can be doing, or is that a very special few
Lex Fridman (1:19:36.620)
can kind of jump?
Matt Botvinick (1:19:37.460)
Just like we talk about art history,
Lex Fridman (1:19:39.180)
I would think it's a special person
Matt Botvinick (1:19:41.020)
that can major in art history
Lex Fridman (1:19:43.620)
and also consider being a surgeon.
Matt Botvinick (1:19:46.860)
Otherwise known as a dilettante.
Lex Fridman (1:19:48.380)
A dilettante, yeah.
Matt Botvinick (1:19:50.140)
Easily distracted.
Lex Fridman (1:19:52.100)
No, I think it does take a special kind of person
Matt Botvinick (1:19:58.620)
to be truly world class at both AI and neuroscience.
Lex Fridman (1:20:02.660)
And I am not on that list.
Matt Botvinick (1:20:05.940)
I happen to be someone whose interest in neuroscience
Lex Fridman (1:20:10.300)
and psychology involved using the kinds
Matt Botvinick (1:20:15.940)
of modeling techniques that are now very central in AI.
Lex Fridman (1:20:20.940)
And that sort of, I guess, bought me a ticket
Matt Botvinick (1:20:24.140)
to be involved in all of the amazing things
Lex Fridman (1:20:26.500)
that are going on in AI research right now.
Matt Botvinick (1:20:29.500)
I do know a few people who I would consider
Lex Fridman (1:20:32.660)
pretty expert on both fronts,
Lex Fridman (1:20:34.780)
and I won't embarrass them by naming them,
Lex Fridman (1:20:36.260)
but there are exceptional people out there
Matt Botvinick (1:20:40.540)
who are like this.
Lex Fridman (1:20:41.380)
The one thing that I find is a barrier
Matt Botvinick (1:20:45.900)
to being truly world class on both fronts
Lex Fridman (1:20:49.300)
is just the complexity of the technology
Matt Botvinick (1:20:54.980)
that's involved in both disciplines now.
Lex Fridman (1:20:58.180)
So the engineering expertise that it takes
Matt Botvinick (1:21:02.980)
to do truly frontline, hands on AI research
Lex Fridman (1:21:07.860)
is really, really considerable.
Matt Botvinick (1:21:10.620)
The learning curve of the tools,
Lex Fridman (1:21:11.940)
just like the specifics of just whether it's programming
Matt Botvinick (1:21:15.260)
or the kind of tools necessary to collect the data,
Lex Fridman (1:21:17.500)
to manage the data, to distribute, to compute,
Matt Botvinick (1:21:19.780)
all that kind of stuff.
Lex Fridman (1:21:20.780)
And on the neuroscience, I guess, side,
Matt Botvinick (1:21:22.380)
there'll be all different sets of tools.
Lex Fridman (1:21:24.580)
Exactly, especially with the recent explosion
Matt Botvinick (1:21:26.820)
in neuroscience methods.
Lex Fridman (1:21:28.980)
So having said all that,
Matt Botvinick (1:21:32.100)
I think the best scenario for both neuroscience
Lex Fridman (1:21:39.860)
and AI is to have people interacting
Matt Botvinick (1:21:44.860)
who live at every point on this spectrum
Lex Fridman (1:21:48.140)
from exclusively focused on neuroscience
Matt Botvinick (1:21:51.900)
to exclusively focused on the engineering side of AI.
Lex Fridman (1:21:55.540)
But to have those people inhabiting a community
Matt Botvinick (1:22:01.060)
where they're talking to people who live elsewhere
Lex Fridman (1:22:03.740)
on the spectrum.
Lex Fridman (1:22:04.820)
And I may be someone who's very close to the center
Lex Fridman (1:22:08.660)
in the sense that I have one foot in the neuroscience world
Lex Fridman (1:22:12.180)
and one foot in the AI world,
Lex Fridman (1:22:14.020)
and that central position, I will admit,
Matt Botvinick (1:22:17.220)
prevents me, at least someone
Lex Fridman (1:22:19.060)
with my limited cognitive capacity,
Matt Botvinick (1:22:21.300)
from having true technical expertise in either domain.
Lex Fridman (1:22:26.820)
But at the same time, I at least hope
Matt Botvinick (1:22:30.140)
that it's worthwhile having people around
Lex Fridman (1:22:32.340)
who can kind of see the connections.
Matt Botvinick (1:22:34.980)
Yeah, the community, the emergent intelligence
Lex Fridman (1:22:39.100)
of the community when it's nicely distributed is useful.
Matt Botvinick (1:22:43.300)
Exactly, yeah.
Lex Fridman (1:22:44.580)
So hopefully that, I mean, I've seen that work,
Matt Botvinick (1:22:46.620)
I've seen that work out well at DeepMind.
Lex Fridman (1:22:48.420)
There are people who, I mean, even if you just focus
Matt Botvinick (1:22:52.860)
on the AI work that happens at DeepMind,
Lex Fridman (1:22:55.820)
it's been a good thing to have some people around
Matt Botvinick (1:22:59.540)
doing that kind of work whose PhDs are in neuroscience
Lex Fridman (1:23:03.260)
or psychology.
Matt Botvinick (1:23:04.780)
Every academic discipline has its kind of blind spots
Lex Fridman (1:23:09.780)
and kind of unfortunate obsessions and its metaphors
Lex Fridman (1:23:16.820)
and its reference points,
Lex Fridman (1:23:18.260)
and having some intellectual diversity is really healthy.
Matt Botvinick (1:23:24.020)
People get each other unstuck, I think.
Lex Fridman (1:23:28.420)
I see it all the time at DeepMind.
Lex Fridman (1:23:30.620)
And I like to think that the people
Lex Fridman (1:23:33.060)
who bring some neuroscience background to the table
Matt Botvinick (1:23:35.940)
are helping with that.
Lex Fridman (1:23:37.460)
So one of my probably the deepest passion for me,
Lex Fridman (1:23:41.420)
what I would say, maybe we kind of spoke off mic
Lex Fridman (1:23:44.140)
a little bit about it, but that I think is a blind spot
Matt Botvinick (1:23:49.460)
for at least robotics and AI folks
Lex Fridman (1:23:51.380)
is human robot interaction, human agent interaction.
Matt Botvinick (1:23:55.540)
Maybe do you have thoughts about how we reduce the size
Lex Fridman (1:24:01.860)
of that blind spot?
Lex Fridman (1:24:02.980)
Do you also share the feeling that not enough folks
Lex Fridman (1:24:07.460)
are studying this aspect of interaction?
Matt Botvinick (1:24:10.260)
Well, I'm actually pretty intensively interested
Lex Fridman (1:24:14.540)
in this issue now, and there are people in my group
Matt Botvinick (1:24:17.060)
who've actually pivoted pretty hard over the last few years
Lex Fridman (1:24:20.940)
from doing more traditional cognitive psychology
Lex Fridman (1:24:24.180)
and cognitive neuroscience to doing experimental work
Lex Fridman (1:24:28.060)
on human agent interaction.
Lex Fridman (1:24:30.220)
And there are a couple of reasons that I'm
Lex Fridman (1:24:33.700)
pretty passionately interested in this.
Matt Botvinick (1:24:35.500)
One is it's kind of the outcome of having thought
Lex Fridman (1:24:42.460)
for a few years now about what we're up to.
Lex Fridman (1:24:46.900)
Like what are we doing?
Lex Fridman (1:24:49.340)
Like what is this AI research for?
Lex Fridman (1:24:53.420)
So what does it mean to make the world a better place?
Lex Fridman (1:24:57.020)
I think I'm pretty sure that means making life better
Matt Botvinick (1:24:59.740)
for humans.
Lex Fridman (1:25:02.620)
And so how do you make life better for humans?
Matt Botvinick (1:25:05.820)
That's a proposition that when you look at it carefully
Lex Fridman (1:25:10.540)
and honestly is rather horrendously complicated,
Matt Botvinick (1:25:15.860)
especially when the AI systems
Lex Fridman (1:25:18.820)
that you're building are learning systems.
Matt Botvinick (1:25:25.220)
They're not, you're not programming something
Lex Fridman (1:25:29.060)
that you then introduce to the world
Lex Fridman (1:25:31.420)
and it just works as programmed,
Lex Fridman (1:25:33.140)
like Google Maps or something.
Matt Botvinick (1:25:36.500)
We're building systems that learn from experience.
Lex Fridman (1:25:39.700)
So that typically leads to AI safety questions.
Lex Fridman (1:25:43.500)
How do we keep these things from getting out of control?
Lex Fridman (1:25:45.420)
How do we keep them from doing things that harm humans?
Lex Fridman (1:25:49.060)
And I mean, I hasten to say,
Lex Fridman (1:25:51.820)
I consider those hugely important issues.
Lex Fridman (1:25:54.500)
And there are large sectors of the research community
Lex Fridman (1:25:58.900)
at DeepMind and of course elsewhere
Matt Botvinick (1:26:00.780)
who are dedicated to thinking hard all day,
Lex Fridman (1:26:03.460)
every day about that.
Lex Fridman (1:26:04.980)
But there's, I guess I would say a positive side to this too
Lex Fridman (1:26:09.620)
which is to say, well, what would it mean
Lex Fridman (1:26:13.300)
to make human life better?
Lex Fridman (1:26:15.900)
And how can we imagine learning systems doing that?
Lex Fridman (1:26:21.180)
And in talking to my colleagues about that,
Lex Fridman (1:26:23.500)
we reached the initial conclusion
Matt Botvinick (1:26:25.700)
that it's not sufficient to philosophize about that.
Lex Fridman (1:26:30.100)
You actually have to take into account
Lex Fridman (1:26:32.060)
how humans actually work and what humans want
Lex Fridman (1:26:37.860)
and the difficulties of knowing what humans want
Lex Fridman (1:26:41.740)
and the difficulties that arise
Lex Fridman (1:26:43.780)
when humans want different things.
Lex Fridman (1:26:47.380)
And so human agent interaction has become,
Lex Fridman (1:26:50.900)
a quite intensive focus of my group lately.
Matt Botvinick (1:26:56.460)
If for no other reason that,
Lex Fridman (1:26:59.020)
in order to really address that issue in an adequate way,
Matt Botvinick (1:27:04.660)
you have to, I mean, psychology becomes part of the picture.
Lex Fridman (1:27:07.340)
Yeah, and so there's a few elements there.
Lex Fridman (1:27:10.380)
So if you focus on solving like the,
Lex Fridman (1:27:12.900)
if you focus on the robotics problem,
Matt Botvinick (1:27:14.700)
let's say AGI without humans in the picture
Lex Fridman (1:27:18.140)
is you're missing fundamentally the final step.
Matt Botvinick (1:27:22.300)
When you do want to help human civilization,
Lex Fridman (1:27:24.580)
you eventually have to interact with humans.
Lex Fridman (1:27:27.340)
And when you create a learning system, just as you said,
Lex Fridman (1:27:31.380)
that will eventually have to interact with humans,
Matt Botvinick (1:27:34.340)
the interaction itself has to be become,
Lex Fridman (1:27:37.900)
has to become part of the learning process.
Lex Fridman (1:27:40.780)
So you can't just watch, well, my sense is,
Lex Fridman (1:27:43.820)
it sounds like your sense is you can't just watch humans
Matt Botvinick (1:27:46.580)
to learn about humans.
Lex Fridman (1:27:48.260)
You have to also be part of the human world.
Matt Botvinick (1:27:50.220)
You have to interact with humans.
Lex Fridman (1:27:51.420)
Yeah, exactly.
Lex Fridman (1:27:52.260)
And I mean, then questions arise that start imperceptibly,
Lex Fridman (1:27:57.380)
but inevitably to slip beyond the realm of engineering.
Lex Fridman (1:28:02.380)
So questions like, if you have an agent
Lex Fridman (1:28:05.940)
that can do something that you can't do,
Lex Fridman (1:28:10.900)
under what conditions do you want that agent to do it?
Lex Fridman (1:28:13.780)
So if I have a robot that can play Beethoven sonatas
Matt Botvinick (1:28:24.700)
better than any human, in the sense that the sensitivity,
Lex Fridman (1:28:30.740)
the expression is just beyond what any human,
Lex Fridman (1:28:33.940)
do I want to listen to that?
Lex Fridman (1:28:36.300)
Do I want to go to a concert and hear a robot play?
Matt Botvinick (1:28:38.780)
These aren't engineering questions.
Lex Fridman (1:28:41.340)
These are questions about human preference
Lex Fridman (1:28:44.340)
and human culture.
Lex Fridman (1:28:45.980)
Psychology bordering on philosophy.
Matt Botvinick (1:28:47.940)
Yeah, and then you start asking,
Lex Fridman (1:28:50.260)
well, even if we knew the answer to that,
Matt Botvinick (1:28:54.660)
is it our place as AI engineers
Lex Fridman (1:28:57.060)
to build that into these agents?
Matt Botvinick (1:28:59.180)
Probably the agents should interact with humans
Lex Fridman (1:29:03.500)
beyond the population of AI engineers
Lex Fridman (1:29:05.620)
and figure out what those humans want.
Lex Fridman (1:29:08.780)
And then when you start,
Matt Botvinick (1:29:10.620)
I referred this the moment ago,
Lex Fridman (1:29:11.780)
but even that becomes complicated.
Lex Fridman (1:29:14.340)
Be quote, what if two humans want different things?
Lex Fridman (1:29:19.100)
And you have only one agent that's able to interact with them
Lex Fridman (1:29:22.380)
and try to satisfy their preferences.
Lex Fridman (1:29:24.620)
Then you're into the realm of economics
Lex Fridman (1:29:30.340)
and social choice theory and even politics.
Lex Fridman (1:29:33.660)
So there's a sense in which,
Matt Botvinick (1:29:35.540)
if you kind of follow what we're doing
Lex Fridman (1:29:37.980)
to its logical conclusion,
Matt Botvinick (1:29:39.940)
then it goes beyond questions of engineering and technology
Lex Fridman (1:29:45.060)
and starts to shade imperceptibly into questions
Lex Fridman (1:29:48.420)
about what kind of society do you want?
Lex Fridman (1:29:51.660)
And actually, once that dawned on me,
Matt Botvinick (1:29:55.740)
I actually felt,
Lex Fridman (1:29:58.620)
I don't know what the right word is,
Matt Botvinick (1:29:59.860)
quite refreshed in my involvement in AI research.
Lex Fridman (1:30:03.020)
It was almost like building this kind of stuff
Matt Botvinick (1:30:06.300)
is gonna lead us back to asking really fundamental questions
Lex Fridman (1:30:10.220)
about what is this,
Matt Botvinick (1:30:13.860)
what's the good life and who gets to decide
Lex Fridman (1:30:16.700)
and bringing in viewpoints from multiple sub communities
Matt Botvinick (1:30:23.780)
to help us shape the way that we live.
Lex Fridman (1:30:27.460)
There's something, it started making me feel like
Matt Botvinick (1:30:30.820)
doing AI research in a fully responsible way, would,
Lex Fridman (1:30:38.300)
could potentially lead to a kind of like cultural renewal.
Matt Botvinick (1:30:42.820)
Yeah, it's the way to understand human beings
Lex Fridman (1:30:48.180)
at the individual, at the societal level.
Matt Botvinick (1:30:50.340)
It may become a way to answer all the silly human questions
Lex Fridman (1:30:54.020)
of the meaning of life and all those kinds of things.
Matt Botvinick (1:30:57.060)
Even if it doesn't give us a way
Lex Fridman (1:30:58.060)
of answering those questions,
Matt Botvinick (1:30:59.220)
it may force us back to thinking about them.
Lex Fridman (1:31:03.660)
And it might bring, it might restore a certain,
Matt Botvinick (1:31:06.940)
I don't know, a certain depth to,
Lex Fridman (1:31:10.460)
or even dare I say spirituality to the way that,
Matt Botvinick (1:31:16.380)
to the world, I don't know.
Lex Fridman (1:31:18.060)
Maybe that's too grandiose.
Matt Botvinick (1:31:19.380)
Well, I'm with you.
Lex Fridman (1:31:21.020)
I think it's AI will be the philosophy of the 21st century,
Matt Botvinick (1:31:27.620)
the way which will open the door.
Lex Fridman (1:31:29.020)
I think a lot of AI researchers are afraid to open that door
Matt Botvinick (1:31:32.500)
of exploring the beautiful richness
Lex Fridman (1:31:35.660)
of the human agent interaction, human AI interaction.
Matt Botvinick (1:31:39.540)
I'm really happy that somebody like you
Lex Fridman (1:31:42.380)
have opened that door.
Lex Fridman (1:31:43.700)
And one thing I often think about is the usual schema
Lex Fridman (1:31:49.500)
for thinking about human agent interaction
Matt Botvinick (1:31:54.500)
as this kind of dystopian, oh, our robot overlords.
Lex Fridman (1:32:00.460)
And again, I hasten to say AI safety is hugely important.
Lex Fridman (1:32:03.500)
And I'm not saying we shouldn't be thinking
Lex Fridman (1:32:06.420)
about those risks, totally on board for that.
Lex Fridman (1:32:09.540)
But there's, having said that,
Lex Fridman (1:32:17.060)
what often follows for me is the thought
Matt Botvinick (1:32:18.860)
that there's another kind of narrative
Lex Fridman (1:32:22.980)
that might be relevant, which is,
Matt Botvinick (1:32:24.780)
when we think of humans gaining more and more information
Lex Fridman (1:32:31.020)
about human life, the narrative there is usually
Matt Botvinick (1:32:36.380)
that they gain more and more wisdom
Lex Fridman (1:32:38.540)
and they get closer to enlightenment
Lex Fridman (1:32:40.700)
and they become more benevolent.
Lex Fridman (1:32:43.260)
And the Buddha is like, that's a totally different narrative.
Lex Fridman (1:32:47.300)
And why isn't it the case that we imagine
Lex Fridman (1:32:50.380)
that the AI systems that we're creating
Matt Botvinick (1:32:52.460)
are just gonna, like, they're gonna figure out
Lex Fridman (1:32:53.980)
more and more about the way the world works
Lex Fridman (1:32:55.660)
and the way that humans interact
Lex Fridman (1:32:56.820)
and they'll become beneficent.
Matt Botvinick (1:32:59.180)
I'm not saying that will happen.
Lex Fridman (1:33:00.500)
I don't honestly expect that to happen
Matt Botvinick (1:33:05.420)
without some careful, setting things up very carefully.
Lex Fridman (1:33:08.820)
But it's another way things could go, right?
Lex Fridman (1:33:11.340)
And yeah, and I would even push back on that.
Lex Fridman (1:33:13.820)
I personally believe that the most trajectories,
Matt Botvinick (1:33:18.820)
natural human trajectories will lead us towards progress.
Lex Fridman (1:33:25.460)
So for me, there is a kind of sense
Matt Botvinick (1:33:28.420)
that most trajectories in AI development
Lex Fridman (1:33:30.820)
will lead us into trouble.
Matt Botvinick (1:33:32.540)
To me, and we over focus on the worst case.
Lex Fridman (1:33:37.140)
It's like in computer science,
Matt Botvinick (1:33:38.500)
theoretical computer science has been this focus
Lex Fridman (1:33:40.860)
on worst case analysis.
Matt Botvinick (1:33:42.060)
There's something appealing to our human mind
Lex Fridman (1:33:45.180)
at some lowest level to be good.
Matt Botvinick (1:33:47.660)
I mean, we don't wanna be eaten by the tiger, I guess.
Lex Fridman (1:33:50.220)
So we wanna do the worst case analysis.
Lex Fridman (1:33:52.300)
But the reality is that shouldn't stop us
Lex Fridman (1:33:55.660)
from actually building out all the other trajectories
Matt Botvinick (1:33:58.620)
which are potentially leading to all the positive worlds,
Lex Fridman (1:34:01.900)
all the enlightenment.
Matt Botvinick (1:34:04.540)
There's a book, Enlightenment Now,
Lex Fridman (1:34:05.700)
with Steven Pinker and so on.
Matt Botvinick (1:34:06.980)
This is looking generally at human progress.
Lex Fridman (1:34:09.660)
And there's so many ways that human progress
Matt Botvinick (1:34:12.300)
can happen with AI.
Lex Fridman (1:34:13.900)
And I think you have to do that research.
Matt Botvinick (1:34:16.300)
You have to do that work.
Lex Fridman (1:34:17.380)
You have to do the, not just the AI safety work
Matt Botvinick (1:34:20.700)
of the one worst case analysis.
Lex Fridman (1:34:22.500)
How do we prevent that?
Lex Fridman (1:34:23.500)
But the actual tools and the glue
Lex Fridman (1:34:27.540)
and the mechanisms of human AI interaction
Matt Botvinick (1:34:31.340)
that would lead to all the positive actions that can go.
Lex Fridman (1:34:34.180)
It's a super exciting area, right?
Matt Botvinick (1:34:36.540)
Yeah, we should be spending,
Lex Fridman (1:34:38.340)
we should be spending a lot of our time saying
Lex Fridman (1:34:40.820)
what can go wrong.
Lex Fridman (1:34:42.860)
I think it's harder to see that there's work to be done
Matt Botvinick (1:34:47.860)
to bring into focus the question of what it would look like
Lex Fridman (1:34:51.540)
for things to go right.
Matt Botvinick (1:34:54.420)
That's not obvious.
Lex Fridman (1:34:57.660)
And we wouldn't be doing this if we didn't have the sense
Lex Fridman (1:34:59.620)
there was huge potential, right?
Lex Fridman (1:35:01.980)
We're not doing this for no reason.
Matt Botvinick (1:35:05.100)
We have a sense that AGI would be a major boom to humanity.
Lex Fridman (1:35:10.100)
But I think it's worth starting now,
Matt Botvinick (1:35:13.700)
even when our technology is quite primitive,
Lex Fridman (1:35:15.620)
asking exactly what would that mean?
Matt Botvinick (1:35:19.420)
We can start now with applications
Lex Fridman (1:35:21.060)
that are already gonna make the world a better place,
Matt Botvinick (1:35:22.580)
like solving protein folding.
Lex Fridman (1:35:25.060)
I think DeepMind has gotten heavy
Matt Botvinick (1:35:27.860)
into science applications lately,
Lex Fridman (1:35:30.060)
which I think is a wonderful, wonderful move
Matt Botvinick (1:35:34.380)
for us to be making.
Lex Fridman (1:35:36.060)
But when we think about AGI,
Matt Botvinick (1:35:37.260)
when we think about building fully intelligent
Lex Fridman (1:35:39.860)
agents that are gonna be able to, in a sense,
Matt Botvinick (1:35:42.460)
do whatever they want,
Lex Fridman (1:35:45.540)
we should start thinking about
Lex Fridman (1:35:46.740)
what do we want them to want, right?
Lex Fridman (1:35:48.940)
What kind of world do we wanna live in?
Matt Botvinick (1:35:52.300)
That's not an easy question.
Lex Fridman (1:35:54.300)
And I think we just need to start working on it.
Lex Fridman (1:35:56.700)
And even on the path to,
Lex Fridman (1:35:58.620)
it doesn't have to be AGI,
Lex Fridman (1:35:59.900)
but just intelligent agents that interact with us
Lex Fridman (1:36:02.300)
and help us enrich our own existence on social networks,
Matt Botvinick (1:36:06.220)
for example, on recommender systems of various intelligence.
Lex Fridman (1:36:08.820)
And there's so much interesting interaction
Matt Botvinick (1:36:10.540)
that's yet to be understood and studied.
Lex Fridman (1:36:12.300)
And how do you create,
Matt Botvinick (1:36:15.540)
I mean, Twitter is struggling with this very idea,
Lex Fridman (1:36:19.460)
how do you create AI systems
Lex Fridman (1:36:21.420)
that increase the quality and the health of a conversation?
Lex Fridman (1:36:24.380)
For sure.
Matt Botvinick (1:36:25.220)
That's a beautiful human psychology question.
Lex Fridman (1:36:28.500)
And how do you do that
Matt Botvinick (1:36:29.740)
without deception being involved,
Lex Fridman (1:36:34.740)
without manipulation being involved,
Lex Fridman (1:36:38.100)
maximizing human autonomy?
Lex Fridman (1:36:42.420)
And how do you make these choices in a democratic way?
Lex Fridman (1:36:45.820)
How do we face the,
Lex Fridman (1:36:50.180)
again, I'm speaking for myself here.
Lex Fridman (1:36:52.740)
How do we face the fact that
Lex Fridman (1:36:55.700)
it's a small group of people
Matt Botvinick (1:36:57.740)
who have the skillset to build these kinds of systems,
Lex Fridman (1:37:01.340)
but what it means to make the world a better place
Matt Botvinick (1:37:05.860)
is something that we all have to be talking about.
Lex Fridman (1:37:09.020)
Yeah, the world that we're trying to make a better place
Matt Botvinick (1:37:14.020)
includes a huge variety of different kinds of people.
Lex Fridman (1:37:18.020)
Yeah, how do we cope with that?
Matt Botvinick (1:37:19.420)
This is a problem that has been discussed
Lex Fridman (1:37:22.820)
in gory, extensive detail in social choice theory.
Matt Botvinick (1:37:28.500)
One thing I'm really interested in
Lex Fridman (1:37:29.900)
and one thing I'm really enjoying
Matt Botvinick (1:37:32.900)
about the recent direction work has taken
Lex Fridman (1:37:35.180)
in some parts of my team is that,
Matt Botvinick (1:37:36.900)
yeah, we're reading the AI literature,
Lex Fridman (1:37:38.620)
we're reading the neuroscience literature,
Lex Fridman (1:37:39.940)
but we've also started reading economics
Lex Fridman (1:37:42.940)
and, as I mentioned, social choice theory,
Matt Botvinick (1:37:44.820)
even some political theory,
Lex Fridman (1:37:45.940)
because it turns out that it all becomes relevant.
Matt Botvinick (1:37:50.380)
It all becomes relevant.
Lex Fridman (1:37:53.540)
But at the same time,
Matt Botvinick (1:37:55.660)
we've been trying not to write philosophy papers,
Lex Fridman (1:38:00.140)
we've been trying not to write physician papers.
Matt Botvinick (1:38:01.980)
We're trying to figure out ways
Lex Fridman (1:38:03.780)
of doing actual empirical research
Matt Botvinick (1:38:05.740)
that kind of take the first small steps
Lex Fridman (1:38:07.780)
to thinking about what it really means
Matt Botvinick (1:38:10.820)
for humans with all of their complexity
Lex Fridman (1:38:13.580)
and contradiction and paradox
Matt Botvinick (1:38:18.540)
to be brought into contact with these AI systems
Lex Fridman (1:38:22.340)
in a way that really makes the world a better place.
Matt Botvinick (1:38:25.540)
Often, reinforcement learning frameworks
Lex Fridman (1:38:27.540)
actually kind of allow you to do that,
Matt Botvinick (1:38:30.860)
machine learning, and so that's the exciting thing about AI
Lex Fridman (1:38:33.580)
is it allows you to reduce the unsolvable problem,
Matt Botvinick (1:38:37.260)
philosophical problem, into something more concrete
Lex Fridman (1:38:40.380)
that you can get ahold of.
Matt Botvinick (1:38:41.700)
Yeah, and it allows you to kind of define the problem
Lex Fridman (1:38:43.900)
in some way that allows for growth in the system
Matt Botvinick (1:38:49.980)
that's sort of, you know,
Lex Fridman (1:38:51.140)
you're not responsible for the details, right?
Matt Botvinick (1:38:54.100)
You say, this is generally what I want you to do,
Lex Fridman (1:38:56.700)
and then learning takes care of the rest.
Matt Botvinick (1:38:59.580)
Of course, the safety issues arise in that context,
Lex Fridman (1:39:04.100)
but I think also some of these positive issues
Matt Botvinick (1:39:05.980)
arise in that context.
Lex Fridman (1:39:06.940)
What would it mean for an AI system
Lex Fridman (1:39:09.180)
to really come to understand what humans want?
Lex Fridman (1:39:14.780)
And with all of the subtleties of that, right?
Matt Botvinick (1:39:18.940)
You know, humans want help with certain things,
Lex Fridman (1:39:24.660)
but they don't want everything done for them, right?
Matt Botvinick (1:39:27.420)
There is, part of the satisfaction
Lex Fridman (1:39:29.660)
that humans get from life is in accomplishing things.
Lex Fridman (1:39:32.700)
So if there were devices around that did everything for,
Lex Fridman (1:39:34.660)
you know, I often think of the movie WALLI, right?
Matt Botvinick (1:39:37.500)
That's like dystopian in a totally different way.
Lex Fridman (1:39:39.380)
It's like, the machines are doing everything for us.
Matt Botvinick (1:39:41.340)
That's not what we wanted.
Lex Fridman (1:39:43.780)
You know, anyway, I find this, you know,
Matt Botvinick (1:39:46.700)
this opens up a whole landscape of research
Lex Fridman (1:39:50.500)
that feels affirmative and exciting.
Matt Botvinick (1:39:52.740)
To me, it's one of the most exciting, and it's wide open.
Lex Fridman (1:39:56.020)
We have to, because it's a cool paper,
Matt Botvinick (1:39:58.260)
talk about dopamine.
Lex Fridman (1:39:59.300)
Oh yeah, okay, so I can.
Matt Botvinick (1:40:01.100)
We were gonna, I was gonna give you a quick summary.
Lex Fridman (1:40:04.980)
Yeah, a quick summary of, what's the title of the paper?
Matt Botvinick (1:40:09.900)
I think we called it a distributional code for value
Lex Fridman (1:40:14.900)
in dopamine based reinforcement learning, yes.
Lex Fridman (1:40:19.020)
So that's another project that grew out of pure AI research.
Lex Fridman (1:40:25.740)
A number of people at DeepMind and a few other places
Matt Botvinick (1:40:29.620)
had started working on a new version
Lex Fridman (1:40:32.340)
of reinforcement learning,
Matt Botvinick (1:40:35.740)
which was defined by taking something
Lex Fridman (1:40:38.940)
in traditional reinforcement learning and just tweaking it.
Lex Fridman (1:40:41.420)
So the thing that they took
Lex Fridman (1:40:42.740)
from traditional reinforcement learning was a value signal.
Lex Fridman (1:40:46.860)
So at the center of reinforcement learning,
Lex Fridman (1:40:49.540)
at least most algorithms, is some representation
Matt Botvinick (1:40:52.580)
of how well things are going,
Lex Fridman (1:40:54.140)
your expected cumulative future reward.
Lex Fridman (1:40:57.660)
And that's usually represented as a single number.
Lex Fridman (1:41:01.220)
So if you imagine a gambler in a casino
Lex Fridman (1:41:04.260)
and the gambler's thinking, well, I have this probability
Lex Fridman (1:41:07.980)
of winning such and such an amount of money,
Lex Fridman (1:41:09.540)
and I have this probability of losing such and such
Lex Fridman (1:41:11.260)
an amount of money, that situation would be represented
Matt Botvinick (1:41:14.860)
as a single number, which is like the expected,
Lex Fridman (1:41:17.260)
the weighted average of all those outcomes.
Lex Fridman (1:41:20.580)
And this new form of reinforcement learning said,
Lex Fridman (1:41:23.740)
well, what if we generalize that
Lex Fridman (1:41:26.460)
to a distributional representation?
Lex Fridman (1:41:28.140)
So now we think of the gambler as literally thinking,
Matt Botvinick (1:41:30.820)
well, there's this probability
Lex Fridman (1:41:32.260)
that I'll win this amount of money,
Lex Fridman (1:41:33.620)
and there's this probability
Lex Fridman (1:41:34.580)
that I'll lose that amount of money,
Lex Fridman (1:41:35.700)
and we don't reduce that to a single number.
Lex Fridman (1:41:37.820)
And it had been observed through experiments,
Matt Botvinick (1:41:40.580)
through just trying this out,
Lex Fridman (1:41:42.420)
that that kind of distributional representation
Matt Botvinick (1:41:45.900)
really accelerated reinforcement learning
Lex Fridman (1:41:49.620)
and led to better policies.
Matt Botvinick (1:41:52.380)
What's your intuition about,
Lex Fridman (1:41:53.620)
so we're talking about rewards.
Matt Botvinick (1:41:55.260)
Yeah.
Lex Fridman (1:41:56.100)
So what's your intuition why that is, why does it do that?
Matt Botvinick (1:41:58.420)
Well, it's kind of a surprising historical note,
Lex Fridman (1:42:02.620)
at least surprised me when I learned it,
Matt Botvinick (1:42:04.460)
that this had been proven to be true.
Lex Fridman (1:42:07.260)
This had been tried out in a kind of heuristic way.
Lex Fridman (1:42:09.820)
People thought, well, gee, what would happen if we tried?
Lex Fridman (1:42:12.500)
And then it had this, empirically,
Matt Botvinick (1:42:14.580)
it had this striking effect.
Lex Fridman (1:42:17.300)
And it was only then that people started thinking,
Lex Fridman (1:42:19.300)
well, gee, wait, why?
Lex Fridman (1:42:21.380)
Wait, why?
Lex Fridman (1:42:22.220)
Why is this working?
Lex Fridman (1:42:23.420)
And that's led to a series of studies
Matt Botvinick (1:42:26.180)
just trying to figure out why it works, which is ongoing.
Lex Fridman (1:42:29.740)
But one thing that's already clear from that research
Matt Botvinick (1:42:31.780)
is that one reason that it helps
Lex Fridman (1:42:34.340)
is that it drives richer representation learning.
Lex Fridman (1:42:39.420)
So if you imagine two situations
Lex Fridman (1:42:43.060)
that have the same expected value,
Matt Botvinick (1:42:45.300)
the same kind of weighted average value,
Lex Fridman (1:42:48.980)
standard deep reinforcement learning algorithms
Matt Botvinick (1:42:51.300)
are going to take those two situations
Lex Fridman (1:42:53.500)
and kind of, in terms of the way
Matt Botvinick (1:42:55.020)
they're represented internally,
Lex Fridman (1:42:56.460)
they're gonna squeeze them together
Matt Botvinick (1:42:58.180)
because the thing that you're trying to represent,
Lex Fridman (1:43:02.580)
which is their expected value, is the same.
Lex Fridman (1:43:04.180)
So all the way through the system,
Lex Fridman (1:43:06.260)
things are gonna be mushed together.
Lex Fridman (1:43:08.420)
But what if those two situations
Lex Fridman (1:43:11.060)
actually have different value distributions?
Matt Botvinick (1:43:13.940)
They have the same average value,
Lex Fridman (1:43:16.900)
but they have different distributions of value.
Matt Botvinick (1:43:19.900)
In that situation, distributional learning
Lex Fridman (1:43:22.300)
will maintain the distinction between these two things.
Lex Fridman (1:43:25.100)
So to make a long story short,
Lex Fridman (1:43:26.820)
distributional learning can keep things separate
Matt Botvinick (1:43:30.020)
in the internal representation
Lex Fridman (1:43:32.180)
that might otherwise be conflated or squished together.
Lex Fridman (1:43:35.140)
And maintaining those distinctions
Lex Fridman (1:43:36.380)
can be useful when the system is now faced
Matt Botvinick (1:43:40.180)
with some other task where the distinction is important.
Lex Fridman (1:43:43.260)
If we look at the optimistic
Lex Fridman (1:43:44.540)
and pessimistic dopamine neurons.
Lex Fridman (1:43:46.580)
So first of all, what is dopamine?
Matt Botvinick (1:43:50.900)
Oh, God.
Lex Fridman (1:43:51.740)
Why is this at all useful
Lex Fridman (1:43:58.220)
to think about in the artificial intelligence sense?
Lex Fridman (1:44:00.740)
But what do we know about dopamine in the human brain?
Lex Fridman (1:44:04.180)
What is it?
Lex Fridman (1:44:05.620)
Why is it useful?
Lex Fridman (1:44:06.460)
Why is it interesting?
Lex Fridman (1:44:07.460)
What does it have to do with the prefrontal cortex
Lex Fridman (1:44:09.380)
and learning in general?
Lex Fridman (1:44:10.260)
Yeah, so, well, this is also a case
Matt Botvinick (1:44:15.540)
where there's a huge amount of detail and debate.
Lex Fridman (1:44:19.660)
But one currently prevailing idea
Matt Botvinick (1:44:24.740)
is that the function of this neurotransmitter dopamine
Lex Fridman (1:44:29.060)
resembles a particular component
Matt Botvinick (1:44:33.460)
of standard reinforcement learning algorithms,
Lex Fridman (1:44:36.860)
which is called the reward prediction error.
Lex Fridman (1:44:39.860)
So I was talking a moment ago
Lex Fridman (1:44:41.580)
about these value representations.
Lex Fridman (1:44:44.220)
How do you learn them?
Lex Fridman (1:44:45.180)
How do you update them based on experience?
Matt Botvinick (1:44:46.900)
Well, if you made some prediction about a future reward
Lex Fridman (1:44:51.820)
and then you get more reward than you were expecting,
Matt Botvinick (1:44:54.460)
then probably retrospectively,
Lex Fridman (1:44:56.020)
you want to go back and increase the value representation
Matt Botvinick (1:45:00.740)
that you attached to that earlier situation.
Lex Fridman (1:45:03.820)
If you got less reward than you were expecting,
Matt Botvinick (1:45:06.180)
you should probably decrement that estimate.
Lex Fridman (1:45:08.540)
And that's the process of temporal difference.
Matt Botvinick (1:45:10.300)
Exactly, this is the central mechanism
Lex Fridman (1:45:12.020)
of temporal difference learning,
Matt Botvinick (1:45:12.860)
which is one of several sort of the backbone
Lex Fridman (1:45:17.660)
of our momentarium in NRL.
Lex Fridman (1:45:20.420)
And this connection between the reward prediction error
Lex Fridman (1:45:25.020)
and dopamine was made in the 1990s.
Lex Fridman (1:45:31.940)
And there's been a huge amount of research
Lex Fridman (1:45:33.420)
that seems to back it up.
Matt Botvinick (1:45:35.860)
Dopamine may be doing other things,
Lex Fridman (1:45:37.340)
but this is clearly, at least roughly,
Matt Botvinick (1:45:39.860)
one of the things that it's doing.
Lex Fridman (1:45:42.460)
But the usual idea was that dopamine
Matt Botvinick (1:45:45.100)
was representing these reward prediction errors,
Lex Fridman (1:45:48.060)
again, in this like kind of single number way
Matt Botvinick (1:45:51.340)
that representing your surprise with a single number.
Lex Fridman (1:45:56.700)
And in distributional reinforcement learning,
Matt Botvinick (1:45:58.500)
this kind of new elaboration of the standard approach,
Lex Fridman (1:46:03.660)
it's not only the value function
Matt Botvinick (1:46:06.060)
that's represented as a single number,
Lex Fridman (1:46:08.460)
it's also the reward prediction error.
Lex Fridman (1:46:10.940)
And so what happened was that Will Dabney,
Lex Fridman (1:46:16.180)
one of my collaborators who was one of the first people
Matt Botvinick (1:46:18.980)
to work on distributional temporal difference learning,
Lex Fridman (1:46:22.300)
talked to a guy in my group, Zeb Kurt Nelson,
Matt Botvinick (1:46:25.740)
who's a computational neuroscientist,
Lex Fridman (1:46:27.660)
and said, gee, you know, is it possible
Matt Botvinick (1:46:29.580)
that dopamine might be doing something
Lex Fridman (1:46:31.740)
like this distributional coding thing?
Lex Fridman (1:46:33.420)
And they started looking at what was in the literature,
Lex Fridman (1:46:35.980)
and then they brought me in,
Lex Fridman (1:46:36.820)
and we started talking to Nao Uchida,
Lex Fridman (1:46:39.220)
and we came up with some specific predictions
Matt Botvinick (1:46:41.300)
about if the brain is using
Lex Fridman (1:46:43.500)
this kind of distributional coding,
Matt Botvinick (1:46:45.140)
then in the tasks that Nao has studied,
Lex Fridman (1:46:47.340)
you should see this, this, this, and this,
Lex Fridman (1:46:49.300)
and that's where the paper came from.
Lex Fridman (1:46:50.620)
We kind of enumerated a set of predictions,
Matt Botvinick (1:46:53.540)
all of which ended up being fairly clearly confirmed,
Lex Fridman (1:46:57.260)
and all of which leads to at least some initial indication
Matt Botvinick (1:47:00.740)
that the brain might be doing something
Lex Fridman (1:47:02.180)
like this distributional coding,
Matt Botvinick (1:47:03.420)
that dopamine might be representing surprise signals
Lex Fridman (1:47:06.780)
in a way that is not just collapsing everything
Matt Botvinick (1:47:09.980)
to a single number, but instead is kind of respecting
Lex Fridman (1:47:12.180)
the variety of future outcomes, if that makes sense.
Lex Fridman (1:47:16.620)
So yeah, so that's showing, suggesting possibly
Lex Fridman (1:47:19.580)
that dopamine has a really interesting
Matt Botvinick (1:47:21.900)
representation scheme in the human brain
Lex Fridman (1:47:25.940)
for its reward signal.
Matt Botvinick (1:47:27.660)
Exactly. That's fascinating.
Lex Fridman (1:47:29.660)
That's another beautiful example of AI
Matt Botvinick (1:47:32.140)
revealing something nice about neuroscience,
Lex Fridman (1:47:34.460)
potentially suggesting possibilities.
Matt Botvinick (1:47:36.260)
Well, you never know.
Lex Fridman (1:47:37.100)
So the minute you publish a paper like that,
Matt Botvinick (1:47:39.260)
the next thing you think is, I hope that replicates.
Lex Fridman (1:47:42.620)
Like, I hope we see that same thing in other data sets,
Lex Fridman (1:47:44.940)
but of course, several labs now
Lex Fridman (1:47:47.380)
are doing the followup experiments, so we'll know soon.
Lex Fridman (1:47:50.180)
But it has been a lot of fun for us
Lex Fridman (1:47:52.580)
to take these ideas from AI
Lex Fridman (1:47:54.780)
and kind of bring them into neuroscience
Lex Fridman (1:47:56.820)
and see how far we can get.
Lex Fridman (1:47:58.980)
So we kind of talked about it a little bit,
Lex Fridman (1:48:01.300)
but where do you see the field of neuroscience
Lex Fridman (1:48:04.020)
and artificial intelligence heading broadly?
Lex Fridman (1:48:07.740)
Like, what are the possible exciting areas
Matt Botvinick (1:48:12.580)
that you can see breakthroughs in the next,
Lex Fridman (1:48:15.300)
let's get crazy, not just three or five years,
Lex Fridman (1:48:17.980)
but the next 10, 20, 30 years
Lex Fridman (1:48:22.340)
that would make you excited
Lex Fridman (1:48:26.100)
and perhaps you'd be part of?
Lex Fridman (1:48:29.020)
On the neuroscience side,
Matt Botvinick (1:48:32.980)
there's a great deal of interest now
Lex Fridman (1:48:34.420)
in what's going on in AI.
Lex Fridman (1:48:36.780)
And at the same time,
Lex Fridman (1:48:41.500)
I feel like, so neuroscience,
Matt Botvinick (1:48:45.900)
especially the part of neuroscience
Lex Fridman (1:48:50.100)
that's focused on circuits and systems,
Matt Botvinick (1:48:54.180)
kind of like really mechanism focused,
Lex Fridman (1:48:57.780)
there's been this explosion in new technology.
Lex Fridman (1:49:01.980)
And up until recently,
Lex Fridman (1:49:05.100)
the experiments that have exploited this technology
Matt Botvinick (1:49:08.940)
have not involved a lot of interesting behavior.
Lex Fridman (1:49:13.340)
And this is for a variety of reasons,
Matt Botvinick (1:49:16.300)
one of which is in order to employ
Lex Fridman (1:49:18.700)
some of these technologies,
Matt Botvinick (1:49:19.860)
you actually have to, if you're studying a mouse,
Lex Fridman (1:49:22.260)
you have to head fix the mouse.
Matt Botvinick (1:49:23.620)
In other words, you have to like immobilize the mouse.
Lex Fridman (1:49:26.260)
And so it's been tricky to come up
Matt Botvinick (1:49:28.700)
with ways of eliciting interesting behavior
Lex Fridman (1:49:30.860)
from a mouse that's restrained in this way,
Lex Fridman (1:49:33.460)
but people have begun to create
Lex Fridman (1:49:35.660)
very interesting solutions to this,
Matt Botvinick (1:49:39.460)
like virtual reality environments
Lex Fridman (1:49:41.300)
where the animal can kind of move a track ball.
Lex Fridman (1:49:43.860)
And as people have kind of begun to explore
Lex Fridman (1:49:48.780)
what you can do with these technologies,
Matt Botvinick (1:49:50.260)
I feel like more and more people are asking,
Lex Fridman (1:49:52.820)
well, let's try to bring behavior into the picture.
Matt Botvinick (1:49:55.740)
Let's try to like reintroduce behavior,
Lex Fridman (1:49:58.220)
which was supposed to be what this whole thing was about.
Lex Fridman (1:50:01.020)
And I'm hoping that those two trends,
Lex Fridman (1:50:05.700)
the kind of growing interest in behavior
Lex Fridman (1:50:09.180)
and the widespread interest in what's going on in AI,
Lex Fridman (1:50:14.180)
will come together to kind of open a new chapter
Matt Botvinick (1:50:17.580)
in neuroscience research where there's a kind of
Lex Fridman (1:50:22.580)
a rebirth of interest in the structure of behavior
Lex Fridman (1:50:25.820)
and its underlying substrates,
Lex Fridman (1:50:27.540)
but that that research is being informed
Matt Botvinick (1:50:31.340)
by computational mechanisms
Lex Fridman (1:50:33.700)
that we're coming to understand in AI.
Matt Botvinick (1:50:36.740)
If we can do that, then we might be taking a step closer
Lex Fridman (1:50:39.580)
to this utopian future that we were talking about earlier
Matt Botvinick (1:50:43.260)
where there's really no distinction
Lex Fridman (1:50:44.860)
between psychology and neuroscience.
Matt Botvinick (1:50:46.940)
Neuroscience is about studying the mechanisms
Lex Fridman (1:50:50.900)
that underlie whatever it is the brain is for,
Lex Fridman (1:50:54.660)
and what is the brain for?
Lex Fridman (1:50:56.340)
What is the brain for? It's for behavior.
Matt Botvinick (1:50:58.460)
I feel like we could maybe take a step toward that now
Lex Fridman (1:51:03.100)
if people are motivated in the right way.
Matt Botvinick (1:51:06.780)
You also asked about AI.
Lex Fridman (1:51:08.780)
So that was a neuroscience question.
Matt Botvinick (1:51:10.340)
You said neuroscience, that's right.
Lex Fridman (1:51:12.180)
And especially places like DeepMind
Matt Botvinick (1:51:13.740)
are interested in both branches.
Lex Fridman (1:51:15.260)
So what about the engineering of intelligence systems?
Matt Botvinick (1:51:20.820)
I think one of the key challenges
Lex Fridman (1:51:24.900)
that a lot of people are seeing now in AI
Matt Botvinick (1:51:28.700)
is to build systems that have the kind of flexibility
Lex Fridman (1:51:34.300)
and the kind of flexibility that humans have in two senses.
Matt Botvinick (1:51:38.580)
One is that humans can be good at many things.
Lex Fridman (1:51:41.860)
They're not just expert at one thing.
Lex Fridman (1:51:44.300)
And they're also flexible in the sense
Lex Fridman (1:51:45.620)
that they can switch between things very easily
Lex Fridman (1:51:49.660)
and they can pick up new things very quickly
Lex Fridman (1:51:52.060)
because they very ably see what a new task has in common
Matt Botvinick (1:51:57.620)
with other things that they've done.
Lex Fridman (1:52:01.860)
And that's something that our AI systems
Matt Botvinick (1:52:05.340)
just blatantly do not have.
Lex Fridman (1:52:09.100)
There are some people who like to argue
Matt Botvinick (1:52:11.380)
that deep learning and deep RL
Lex Fridman (1:52:13.740)
are simply wrong for getting that kind of flexibility.
Matt Botvinick (1:52:17.080)
I don't share that belief,
Lex Fridman (1:52:20.060)
but the simpler fact of the matter
Matt Botvinick (1:52:22.620)
is we're not building things yet
Lex Fridman (1:52:23.860)
that do have that kind of flexibility.
Lex Fridman (1:52:25.500)
And I think the attention of a large part
Lex Fridman (1:52:28.700)
of the AI community is starting to pivot to that question.
Lex Fridman (1:52:31.500)
How do we get that?
Lex Fridman (1:52:33.460)
That's gonna lead to a focus on abstraction.
Matt Botvinick (1:52:38.060)
It's gonna lead to a focus on
Lex Fridman (1:52:40.460)
what in psychology we call cognitive control,
Matt Botvinick (1:52:43.620)
which is the ability to switch between tasks,
Lex Fridman (1:52:45.900)
the ability to quickly put together a program of behavior
Matt Botvinick (1:52:49.300)
that you've never executed before,
Lex Fridman (1:52:51.740)
but you know makes sense for a particular set of demands.
Matt Botvinick (1:52:55.260)
It's very closely related to what the prefrontal cortex does
Lex Fridman (1:52:59.140)
on the neuroscience side.
Lex Fridman (1:53:01.060)
So I think it's gonna be an interesting new chapter.
Lex Fridman (1:53:05.380)
So that's the reasoning side and cognition side,
Lex Fridman (1:53:07.420)
but let me ask the over romanticized question.
Lex Fridman (1:53:10.540)
Do you think we'll ever engineer an AGI system
Matt Botvinick (1:53:13.700)
that we humans would be able to love
Lex Fridman (1:53:17.140)
and that would love us back?
Lex Fridman (1:53:19.580)
So have that level and depth of connection?
Lex Fridman (1:53:26.220)
I love that question.
Lex Fridman (1:53:27.860)
And it relates closely to things
Lex Fridman (1:53:31.980)
that I've been thinking about a lot lately,
Matt Botvinick (1:53:33.900)
in the context of this human AI research.
Lex Fridman (1:53:36.620)
There's social psychology research
Matt Botvinick (1:53:41.140)
in particular by Susan Fisk at Princeton
Lex Fridman (1:53:44.940)
the department where I used to work,
Matt Botvinick (1:53:48.420)
where she dissects human attitudes toward other humans
Lex Fridman (1:53:54.500)
into a sort of two dimensional scheme.
Lex Fridman (1:53:59.900)
And one dimension is about ability.
Lex Fridman (1:54:03.940)
How able, how capable is this other person?
Lex Fridman (1:54:10.100)
But the other dimension is warmth.
Lex Fridman (1:54:11.780)
So you can imagine another person who's very skilled
Lex Fridman (1:54:15.580)
and capable, but is very cold.
Lex Fridman (1:54:19.540)
And you wouldn't really like highly,
Matt Botvinick (1:54:22.500)
you might have some reservations about that other person.
Lex Fridman (1:54:26.660)
But there's also a kind of reservation
Matt Botvinick (1:54:28.980)
that we might have about another person
Lex Fridman (1:54:31.020)
who elicits in us or displays a lot of human warmth,
Lex Fridman (1:54:34.860)
but is not good at getting things done.
Lex Fridman (1:54:37.940)
We reserve our greatest esteem really
Matt Botvinick (1:54:40.940)
for people who are both highly capable
Lex Fridman (1:54:43.820)
and also quite warm.
Matt Botvinick (1:54:47.300)
That's like the best of the best.
Lex Fridman (1:54:49.820)
This isn't a normative statement I'm making.
Matt Botvinick (1:54:53.300)
This is just an empirical statement.
Lex Fridman (1:54:55.780)
This is what humans seem...
Matt Botvinick (1:54:57.180)
These are the two dimensions that people seem to kind of like
Lex Fridman (1:54:59.740)
along which people size other people up.
Lex Fridman (1:55:02.660)
And in AI research,
Lex Fridman (1:55:03.980)
there's a lot of people who think that humans are
Matt Botvinick (1:55:06.580)
very capable, and in AI research,
Lex Fridman (1:55:08.700)
we really focus on this capability thing.
Matt Botvinick (1:55:11.420)
We want our agents to be able to do stuff.
Lex Fridman (1:55:13.420)
This thing can play go at a superhuman level.
Matt Botvinick (1:55:15.460)
That's awesome.
Lex Fridman (1:55:16.860)
But that's only one dimension.
Lex Fridman (1:55:18.700)
What about the other dimension?
Lex Fridman (1:55:20.060)
What would it mean for an AI system to be warm?
Lex Fridman (1:55:25.060)
And I don't know, maybe there are easy solutions here.
Lex Fridman (1:55:27.620)
Like we can put a face on our AI systems.
Matt Botvinick (1:55:30.620)
It's cute, it has big ears.
Lex Fridman (1:55:32.020)
I mean, that's probably part of it.
Lex Fridman (1:55:33.820)
But I think it also has to do with a pattern of behavior.
Lex Fridman (1:55:36.540)
A pattern of what would it mean for an AI system
Matt Botvinick (1:55:40.180)
to display caring, compassionate behavior
Lex Fridman (1:55:43.460)
in a way that actually made us feel like it was for real?
Matt Botvinick (1:55:47.740)
That we didn't feel like it was simulated.
Lex Fridman (1:55:49.940)
We didn't feel like we were being duped.
Matt Botvinick (1:55:53.100)
To me, people talk about the Turing test
Lex Fridman (1:55:55.740)
or some descendant of it.
Matt Botvinick (1:55:57.860)
I feel like that's the ultimate Turing test.
Lex Fridman (1:56:01.140)
Is there an AI system that can not only convince us
Matt Botvinick (1:56:05.460)
that it knows how to reason
Lex Fridman (1:56:07.180)
and it knows how to interpret language,
Lex Fridman (1:56:09.100)
but that we're comfortable saying,
Lex Fridman (1:56:12.700)
yeah, that AI system's a good guy.
Matt Botvinick (1:56:15.980)
On the warmth scale, whatever warmth is,
Lex Fridman (1:56:18.700)
we kind of intuitively understand it,
Lex Fridman (1:56:20.860)
but we also wanna be able to, yeah,
Lex Fridman (1:56:25.060)
we don't understand it explicitly enough yet
Matt Botvinick (1:56:29.180)
to be able to engineer it.
Lex Fridman (1:56:30.940)
Exactly.
Lex Fridman (1:56:31.780)
And that's an open scientific question.
Lex Fridman (1:56:33.620)
You kind of alluded it several times
Matt Botvinick (1:56:35.340)
in the human AI interaction.
Lex Fridman (1:56:37.220)
That's a question that should be studied
Lex Fridman (1:56:38.900)
and probably one of the most important questions
Lex Fridman (1:56:42.300)
as we move to AGI.
Matt Botvinick (1:56:43.540)
We humans are so good at it.
Lex Fridman (1:56:46.020)
Yeah.
Matt Botvinick (1:56:46.860)
It's not just that we're born warm.
Lex Fridman (1:56:50.140)
I suppose some people are warmer than others
Matt Botvinick (1:56:53.060)
given whatever genes they manage to inherit.
Lex Fridman (1:56:55.700)
But there are also learned skills involved.
Matt Botvinick (1:57:01.620)
There are ways of communicating to other people
Lex Fridman (1:57:04.740)
that you care, that they matter to you,
Lex Fridman (1:57:07.740)
that you're enjoying interacting with them, right?
Lex Fridman (1:57:11.100)
And we learn these skills from one another.
Lex Fridman (1:57:14.140)
And it's not out of the question
Lex Fridman (1:57:16.740)
that we could build engineered systems.
Matt Botvinick (1:57:20.020)
I think it's hopeless, as you say,
Lex Fridman (1:57:21.460)
that we could somehow hand design
Matt Botvinick (1:57:23.580)
these sorts of behaviors.
Lex Fridman (1:57:26.100)
But it's not out of the question
Matt Botvinick (1:57:27.060)
that we could build systems that kind of,
Lex Fridman (1:57:30.060)
we instill in them something that sets them out
Matt Botvinick (1:57:34.460)
in the right direction,
Lex Fridman (1:57:35.980)
so that they end up learning what it is
Matt Botvinick (1:57:39.580)
to interact with humans
Lex Fridman (1:57:40.540)
in a way that's gratifying to humans.
Matt Botvinick (1:57:44.180)
I mean, honestly, if that's not where we're headed,
Lex Fridman (1:57:49.220)
I want out.
Matt Botvinick (1:57:50.340)
I think it's exciting as a scientific problem,
Lex Fridman (1:57:54.940)
just as you described.
Matt Botvinick (1:57:56.820)
I honestly don't see a better way to end it
Lex Fridman (1:57:59.500)
than talking about warmth and love.
Lex Fridman (1:58:01.180)
And Matt, I don't think I've ever had such a wonderful
Lex Fridman (1:58:05.380)
conversation where my questions were so bad
Lex Fridman (1:58:07.540)
and your answers were so beautiful.
Lex Fridman (1:58:09.380)
So I deeply appreciate it.
Matt Botvinick (1:58:10.740)
I really enjoyed it.
Lex Fridman (1:58:11.580)
Thanks for talking to me.
Matt Botvinick (1:58:12.420)
Well, it's been very fun.
Lex Fridman (1:58:13.260)
As you can probably tell,
Matt Botvinick (1:58:17.140)
there's something I like about kind of thinking
Lex Fridman (1:58:19.020)
outside the box and like,
Lex Fridman (1:58:21.060)
so it's good having an opportunity to do that.
Lex Fridman (1:58:22.940)
Awesome.
Matt Botvinick (1:58:23.780)
Thanks so much for doing it.
Lex Fridman (1:58:25.620)
Thanks for listening to this conversation
Matt Botvinick (1:58:27.180)
with Matt Bopenik.
Lex Fridman (1:58:28.420)
And thank you to our sponsors,
Matt Botvinick (1:58:30.540)
The Jordan Harbinger Show
Lex Fridman (1:58:32.300)
and Magic Spoon Low Carb Keto Cereal.
Matt Botvinick (1:58:36.140)
Please consider supporting this podcast
Lex Fridman (1:58:38.020)
by going to jordanharbinger.com slash lex
Lex Fridman (1:58:41.020)
and also going to magicspoon.com slash lex
Lex Fridman (1:58:44.940)
and using code lex at checkout.
Matt Botvinick (1:58:48.220)
Click the links, buy all the stuff.
Lex Fridman (1:58:50.900)
It's the best way to support this podcast
Lex Fridman (1:58:52.860)
and the journey I'm on in my research and the startup.
Lex Fridman (1:58:57.260)
If you enjoy this thing, subscribe on YouTube,
Matt Botvinick (1:58:59.580)
review it with the five stars in Apple Podcasts,
Lex Fridman (1:59:02.380)
support it on Patreon, follow on Spotify
Matt Botvinick (1:59:05.380)
or connect with me on Twitter at lexfreedman.
Lex Fridman (1:59:08.220)
Again, spelled miraculously without the E,
Matt Botvinick (1:59:12.220)
just F R I D M A N.
Lex Fridman (1:59:15.060)
And now let me leave you with some words
Matt Botvinick (1:59:17.100)
from neurologist V.S. Amarachandran.
Lex Fridman (1:59:20.820)
How can a three pound mass of jelly
Matt Botvinick (1:59:23.340)
that you can hold in your palm imagine angels,
Lex Fridman (1:59:26.620)
contemplate the meaning of an infinity
Lex Fridman (1:59:28.700)
and even question its own place in the cosmos?
Lex Fridman (1:59:31.740)
Especially awe inspiring is the fact that any single brain,
Matt Botvinick (1:59:35.660)
including yours, is made up of atoms
Lex Fridman (1:59:38.580)
that were forged in the hearts
Matt Botvinick (1:59:40.060)
of countless far flung stars billions of years ago.
Lex Fridman (1:59:45.500)
These particles drifted for eons and light years
Matt Botvinick (1:59:48.340)
until gravity and change brought them together here now.
Lex Fridman (1:59:53.180)
These atoms now form a conglomerate, your brain,
Matt Botvinick (1:59:57.540)
that can not only ponder the very stars they gave at birth,
Lex Fridman (20:00.200)
So you always found neural networks
Matt Botvinick (20:02.080)
in biological form beautiful.
Lex Fridman (20:04.080)
Oh, neural networks were very concretely the thing
Matt Botvinick (20:07.160)
that drew me into science.
Lex Fridman (20:09.160)
I was handed, are you familiar with the PDP books
Matt Botvinick (20:13.320)
from the 80s when I was in,
Lex Fridman (20:15.720)
I went to medical school before I went into science.
Matt Botvinick (20:18.240)
And, yeah.
Lex Fridman (20:19.160)
Really, interesting.
Matt Botvinick (20:20.800)
Wow.
Lex Fridman (20:21.960)
I also did a graduate degree in art history,
Lex Fridman (20:23.920)
so I'm kind of exploring.
Lex Fridman (20:26.480)
Well, art history, I understand.
Matt Botvinick (20:28.560)
That's just a curious, creative mind.
Lex Fridman (20:31.280)
But medical school, with the dream of what,
Lex Fridman (20:33.960)
if we take that slight tangent?
Lex Fridman (20:36.560)
What, did you want to be a surgeon?
Matt Botvinick (20:39.120)
I actually was quite interested in surgery.
Lex Fridman (20:41.680)
I was interested in surgery and psychiatry.
Lex Fridman (20:44.200)
And I thought, I must be the only person on the planet
Lex Fridman (20:49.520)
who was torn between those two fields.
Lex Fridman (20:52.680)
And I said exactly that to my advisor in medical school,
Lex Fridman (20:56.840)
who turned out, I found out later,
Matt Botvinick (20:59.440)
to be a famous psychoanalyst.
Lex Fridman (21:01.920)
And he said to me, no, no, it's actually not so uncommon
Matt Botvinick (21:05.160)
to be interested in surgery and psychiatry.
Lex Fridman (21:07.520)
And he conjectured that the reason
Matt Botvinick (21:10.480)
that people develop these two interests
Lex Fridman (21:12.600)
is that both fields are about going beneath the surface
Lex Fridman (21:16.360)
and kind of getting into the kind of secret.
Lex Fridman (21:19.120)
I mean, maybe you understand this as someone
Matt Botvinick (21:21.040)
who was interested in psychoanalysis.
Lex Fridman (21:23.440)
There's sort of a, there's a cliche phrase
Matt Botvinick (21:26.200)
that people use now, like in NPR,
Lex Fridman (21:28.400)
the secret life of blankety blank, right?
Lex Fridman (21:31.400)
And that was part of the thrill of surgery,
Lex Fridman (21:33.560)
was seeing the secret activity
Matt Botvinick (21:38.120)
that's inside everybody's abdomen and thorax.
Lex Fridman (21:40.560)
That's a very poetic way to connect it to disciplines
Matt Botvinick (21:43.880)
that are very, practically speaking,
Lex Fridman (21:45.560)
different from each other.
Matt Botvinick (21:46.520)
That's for sure, that's for sure, yes.
Lex Fridman (21:48.480)
So how did we get onto medical school?
Lex Fridman (21:52.480)
So I was in medical school
Lex Fridman (21:53.720)
and I was doing a psychiatry rotation
Lex Fridman (21:57.360)
and my kind of advisor in that rotation
Lex Fridman (22:02.280)
asked me what I was interested in.
Lex Fridman (22:04.720)
And I said, well, maybe psychiatry.
Lex Fridman (22:07.800)
He said, why?
Lex Fridman (22:09.280)
And I said, well, I've always been interested
Lex Fridman (22:11.120)
in how the brain works.
Matt Botvinick (22:13.080)
I'm pretty sure that nobody's doing scientific research
Lex Fridman (22:16.160)
that addresses my interests,
Matt Botvinick (22:19.160)
which are, I didn't have a word for it then,
Lex Fridman (22:21.880)
but I would have said about cognition.
Lex Fridman (22:25.200)
And he said, well, you know, I'm not sure that's true.
Lex Fridman (22:27.680)
You might be interested in these books.
Lex Fridman (22:29.600)
And he pulled down the PDB books from his shelf
Lex Fridman (22:32.440)
and they were still shrink wrapped.
Matt Botvinick (22:33.960)
He hadn't read them, but he handed them to me.
Lex Fridman (22:36.920)
He said, you feel free to borrow these.
Lex Fridman (22:38.680)
And that was, you know, I went back to my dorm room
Lex Fridman (22:41.440)
and I just, you know, read them cover to cover.
Lex Fridman (22:43.400)
And what's PDB?
Lex Fridman (22:44.960)
Parallel distributed processing,
Matt Botvinick (22:46.520)
which was one of the original names for deep learning.
Lex Fridman (22:50.840)
And so I apologize for the romanticized question,
Lex Fridman (22:55.000)
but what idea in the space of neuroscience
Lex Fridman (22:58.360)
and the space of the human brain is to you
Lex Fridman (23:00.840)
the most beautiful, mysterious, surprising?
Lex Fridman (23:03.880)
What had always fascinated me,
Matt Botvinick (23:08.480)
even when I was a pretty young kid, I think,
Lex Fridman (23:12.320)
was the paradox that lies in the fact
Matt Botvinick (23:21.360)
that the brain is so mysterious
Lex Fridman (23:25.640)
and seems so distant.
Lex Fridman (23:30.640)
But at the same time,
Lex Fridman (23:32.520)
it's responsible for the full transparency
Matt Botvinick (23:37.360)
of everyday life.
Lex Fridman (23:39.040)
The brain is literally what makes everything obvious
Lex Fridman (23:41.520)
and familiar.
Lex Fridman (23:43.080)
And there's always one in the room with you.
Matt Botvinick (23:47.280)
Yeah.
Lex Fridman (23:48.120)
I used to teach, when I taught at Princeton,
Matt Botvinick (23:50.520)
I used to teach a cognitive neuroscience course.
Lex Fridman (23:53.000)
And the very last thing I would say to the students was,
Matt Botvinick (23:56.720)
you know, people often,
Lex Fridman (24:00.160)
when people think of scientific inspiration,
Matt Botvinick (24:04.200)
the metaphor is often, well, look to the stars.
Lex Fridman (24:08.120)
The stars will inspire you to wonder at the universe
Lex Fridman (24:12.360)
and think about your place in it and how things work.
Lex Fridman (24:15.800)
And I'm all for looking at the stars,
Lex Fridman (24:18.360)
but I've always been much more inspired.
Lex Fridman (24:21.600)
And my sense of wonder comes from the,
Matt Botvinick (24:25.360)
not from the distant, mysterious stars,
Lex Fridman (24:28.560)
but from the extremely intimately close brain.
Matt Botvinick (24:34.440)
Yeah.
Lex Fridman (24:35.280)
There's something just endlessly fascinating
Matt Botvinick (24:38.680)
to me about that.
Lex Fridman (24:40.000)
The, like, just like you said,
Matt Botvinick (24:41.360)
the one that's close and yet distant
Lex Fridman (24:45.500)
in terms of our understanding of it.
Lex Fridman (24:48.000)
Do you, are you also captivated by the fact
Lex Fridman (24:53.640)
that this very conversation is happening
Lex Fridman (24:56.040)
because two brains are communicating so that?
Lex Fridman (24:57.560)
Yes, exactly.
Matt Botvinick (24:59.120)
The, I guess what I mean is the subjective nature
Lex Fridman (25:03.800)
of the experience, if it can take a small attention
Matt Botvinick (25:06.320)
into the mystical of it, the consciousness,
Lex Fridman (25:10.240)
or when you were saying you're captivated
Matt Botvinick (25:13.320)
by the idea of the brain,
Lex Fridman (25:14.920)
are you talking about specifically
Lex Fridman (25:16.320)
the mechanism of cognition?
Lex Fridman (25:18.200)
Or are you also just, like, at least for me,
Matt Botvinick (25:23.080)
it's almost like paralyzing the beauty and the mystery
Lex Fridman (25:26.600)
of the fact that it creates the entirety of the experience,
Matt Botvinick (25:29.480)
not just the reasoning capability, but the experience.
Lex Fridman (25:32.880)
Well, I definitely resonate with that latter thought.
Lex Fridman (25:38.920)
And I often find discussions of artificial intelligence
Lex Fridman (25:45.280)
to be disappointingly narrow.
Matt Botvinick (25:50.720)
Speaking as someone who has always had an interest in art.
Lex Fridman (25:55.720)
Right.
Matt Botvinick (25:56.560)
I was just gonna go there
Lex Fridman (25:57.400)
because it sounds like somebody who has an interest in art.
Matt Botvinick (26:00.200)
Yeah, I mean, there are many layers
Lex Fridman (26:04.000)
to full bore human experience.
Lex Fridman (26:08.200)
And in some ways it's not enough to say,
Lex Fridman (26:12.040)
oh, well, don't worry, we're talking about cognition,
Lex Fridman (26:15.020)
but we'll add emotion, you know?
Lex Fridman (26:17.240)
There's an incredible scope
Matt Botvinick (26:21.200)
to what humans go through in every moment.
Lex Fridman (26:25.280)
And yes, so that's part of what fascinates me,
Matt Botvinick (26:33.320)
is that our brains are producing that.
Lex Fridman (26:40.040)
But at the same time, it's so mysterious to us.
Lex Fridman (26:43.040)
How?
Lex Fridman (26:46.240)
Our brains are literally in our heads
Matt Botvinick (26:49.120)
producing this experience.
Lex Fridman (26:50.600)
Producing the experience.
Lex Fridman (26:52.120)
And yet it's so mysterious to us.
Lex Fridman (26:55.100)
And so, and the scientific challenge
Matt Botvinick (26:57.000)
of getting at the actual explanation for that
Lex Fridman (27:00.880)
is so overwhelming.
Matt Botvinick (27:03.360)
That's just, I don't know.
Lex Fridman (27:05.600)
Certain people have fixations on particular questions
Lex Fridman (27:08.440)
and that's always, that's just always been mine.
Lex Fridman (27:11.680)
Yeah, I would say the poetry of that is fascinating.
Lex Fridman (27:14.020)
And I'm really interested in natural language as well.
Lex Fridman (27:16.740)
And when you look at artificial intelligence community,
Matt Botvinick (27:19.440)
it always saddens me how much
Lex Fridman (27:23.880)
when you try to create a benchmark
Matt Botvinick (27:25.720)
for the community to gather around,
Lex Fridman (27:28.200)
how much of the magic of language is lost
Matt Botvinick (27:30.920)
when you create that benchmark.
Lex Fridman (27:33.240)
That there's something, we talk about experience,
Matt Botvinick (27:35.920)
the music of the language, the wit,
Lex Fridman (27:38.600)
the something that makes a rich experience,
Matt Botvinick (27:41.080)
something that would be required to pass
Lex Fridman (27:43.800)
the spirit of the Turing test is lost in these benchmarks.
Lex Fridman (27:47.660)
And I wonder how to get it back in
Lex Fridman (27:50.240)
because it's very difficult.
Matt Botvinick (27:51.920)
The moment you try to do like real good rigorous science,
Lex Fridman (27:55.160)
you lose some of that magic.
Matt Botvinick (27:56.960)
When you try to study cognition
Lex Fridman (28:00.160)
in a rigorous scientific way,
Matt Botvinick (28:01.560)
it feels like you're losing some of the magic.
Lex Fridman (28:03.800)
The seeing cognition in a mechanistic way
Matt Botvinick (28:07.520)
that AI folk at this stage in our history.
Lex Fridman (28:10.060)
Well, I agree with you, but at the same time,
Matt Botvinick (28:13.040)
one thing that I found really exciting
Lex Fridman (28:18.040)
about that first wave of deep learning models in cognition
Matt Botvinick (28:22.960)
was the fact that the people who were building these models
Lex Fridman (28:29.640)
were focused on the richness and complexity
Matt Botvinick (28:32.960)
of human cognition.
Lex Fridman (28:34.800)
So an early debate in cognitive science,
Matt Botvinick (28:40.080)
which I sort of witnessed as a grad student
Lex Fridman (28:41.820)
was about something that sounds very dry,
Matt Botvinick (28:44.200)
which is the formation of the past tense.
Lex Fridman (28:47.180)
But there were these two camps.
Matt Botvinick (28:49.200)
One said, well, the mind encodes certain rules
Lex Fridman (28:54.400)
and it also has a list of exceptions
Matt Botvinick (28:57.900)
because of course, the rule is add ED,
Lex Fridman (29:00.380)
but that's not always what you do.
Lex Fridman (29:01.820)
So you have to have a list of exceptions.
Lex Fridman (29:05.000)
And then there were the connectionists
Matt Botvinick (29:06.960)
who evolved into the deep learning people who said,
Lex Fridman (29:10.700)
well, if you look carefully at the data,
Matt Botvinick (29:13.820)
if you actually look at corpora, like language corpora,
Lex Fridman (29:18.280)
it turns out to be very rich
Matt Botvinick (29:20.080)
because yes, there are most verbs
Lex Fridman (29:25.080)
that you just tack on ED, and then there are exceptions,
Lex Fridman (29:28.640)
but there are rules that the exceptions aren't just random.
Lex Fridman (29:36.040)
There are certain clues to which verbs
Matt Botvinick (29:39.560)
should be exceptional.
Lex Fridman (29:41.040)
And then there are exceptions to the exceptions.
Lex Fridman (29:44.120)
And there was a word that was kind of deployed
Lex Fridman (29:47.760)
in order to capture this, which was quasi regular.
Matt Botvinick (29:51.760)
In other words, there are rules, but it's messy.
Lex Fridman (29:54.740)
And there's either structure even among the exceptions.
Lex Fridman (29:58.760)
And it would be, yeah, you could try to write down,
Lex Fridman (2:00:00.860)
but can also think about its own ability to think
Lex Fridman (2:00:04.180)
and wonder about its own ability to wander.
Lex Fridman (2:00:07.820)
With the arrival of humans, it has been said,
Matt Botvinick (2:00:10.660)
the universe has suddenly become conscious of itself.
Lex Fridman (2:00:14.580)
This truly is the greatest mystery of all.
Matt Botvinick (2:00:18.620)
Thank you for listening and hope to see you next time.
Lex Fridman (30:01.280)
we could try to write down the structure
Matt Botvinick (30:03.820)
in some sort of closed form,
Lex Fridman (30:04.840)
but really the right way to understand
Lex Fridman (30:07.560)
how the brain is handling all this,
Lex Fridman (30:09.080)
and by the way, producing all of this,
Matt Botvinick (30:11.440)
is to build a deep neural network
Lex Fridman (30:14.000)
and train it on this data
Lex Fridman (30:15.200)
and see how it ends up representing all of this richness.
Lex Fridman (30:18.520)
So the way that deep learning
Matt Botvinick (30:21.420)
was deployed in cognitive psychology
Lex Fridman (30:23.720)
was that was the spirit of it.
Matt Botvinick (30:25.960)
It was about that richness.
Lex Fridman (30:29.560)
And that's something that I always found very compelling,
Matt Botvinick (30:31.960)
still do.
Lex Fridman (30:33.160)
Is there something especially interesting
Lex Fridman (30:36.200)
and profound to you
Lex Fridman (30:37.520)
in terms of our current deep learning neural network,
Matt Botvinick (30:40.480)
artificial neural network approaches,
Lex Fridman (30:42.640)
and whatever we do understand
Lex Fridman (30:46.300)
about the biological neural networks in our brain?
Lex Fridman (30:49.000)
Is there, there's quite a few differences.
Matt Botvinick (30:52.440)
Are some of them to you,
Lex Fridman (30:54.680)
either interesting or perhaps profound
Matt Botvinick (30:58.040)
in terms of the gap we might want to try to close
Lex Fridman (31:03.040)
in trying to create a human level intelligence?
Lex Fridman (31:07.560)
What I would say here is something
Lex Fridman (31:08.840)
that a lot of people are saying,
Matt Botvinick (31:10.720)
which is that one seeming limitation
Lex Fridman (31:16.580)
of the systems that we're building now
Matt Botvinick (31:18.960)
is that they lack the kind of flexibility,
Lex Fridman (31:22.900)
the readiness to sort of turn on a dime
Matt Botvinick (31:25.960)
when the context calls for it
Lex Fridman (31:28.200)
that is so characteristic of human behavior.
Lex Fridman (31:32.200)
So is that connected to you to the,
Lex Fridman (31:34.920)
like which aspect of the neural networks in our brain
Lex Fridman (31:37.720)
is that connected to?
Lex Fridman (31:39.160)
Is that closer to the cognitive science level of,
Matt Botvinick (31:45.080)
now again, see like my natural inclination
Lex Fridman (31:47.320)
is to separate into three disciplines of neuroscience,
Matt Botvinick (31:51.640)
cognitive science and psychology.
Lex Fridman (31:54.280)
And you've already kind of shut that down
Matt Botvinick (31:56.380)
by saying you're kind of see them as separate,
Lex Fridman (31:58.360)
but just to look at those layers,
Matt Botvinick (32:01.500)
I guess where is there something about the lowest layer
Lex Fridman (32:05.320)
of the way the neural neurons interact
Matt Botvinick (32:09.160)
that is profound to you in terms of this difference
Lex Fridman (32:13.320)
to the artificial neural networks,
Matt Botvinick (32:15.480)
or is all the key differences
Lex Fridman (32:17.220)
at a higher level of abstraction?
Matt Botvinick (32:20.720)
One thing I often think about is that,
Lex Fridman (32:24.440)
if you take an introductory computer science course
Lex Fridman (32:27.140)
and they are introducing you to the notion
Lex Fridman (32:29.600)
of Turing machines,
Matt Botvinick (32:31.480)
one way of articulating
Lex Fridman (32:36.000)
what the significance of a Turing machine is,
Matt Botvinick (32:39.320)
is that it's a machine emulator.
Lex Fridman (32:42.760)
It can emulate any other machine.
Lex Fridman (32:47.540)
And that to me,
Lex Fridman (32:52.960)
that way of looking at a Turing machine
Matt Botvinick (32:56.200)
really sticks with me.
Lex Fridman (32:57.640)
I think of humans as maybe sharing
Matt Botvinick (33:01.960)
in some of that character.
Lex Fridman (33:05.000)
We're capacity limited,
Matt Botvinick (33:06.160)
we're not Turing machines obviously,
Lex Fridman (33:07.540)
but we have the ability to adapt behaviors
Matt Botvinick (33:11.040)
that are very much unlike anything we've done before,
Lex Fridman (33:15.420)
but there's some basic mechanism
Matt Botvinick (33:17.720)
that's implemented in our brain
Lex Fridman (33:18.960)
that allows us to run software.
Lex Fridman (33:22.400)
But just on that point, you mentioned Turing machine,
Lex Fridman (33:24.600)
but nevertheless, it's fundamentally
Matt Botvinick (33:26.840)
our brains are just computational devices in your view.
Lex Fridman (33:29.720)
Is that what you're getting at?
Matt Botvinick (33:31.160)
It was a little bit unclear to this line you drew.
Lex Fridman (33:35.680)
Is there any magic in there
Lex Fridman (33:37.800)
or is it just basic computation?
Lex Fridman (33:40.720)
I'm happy to think of it as just basic computation,
Lex Fridman (33:43.320)
but mind you, I won't be satisfied
Lex Fridman (33:46.120)
until somebody explains to me
Lex Fridman (33:48.280)
what the basic computations are
Lex Fridman (33:49.840)
that are leading to the full richness of human cognition.
Matt Botvinick (33:54.760)
It's not gonna be enough for me
Lex Fridman (33:56.680)
to understand what the computations are
Matt Botvinick (33:58.880)
that allow people to do arithmetic or play chess.
Lex Fridman (34:02.160)
I want the whole thing.
Lex Fridman (34:06.360)
And a small tangent,
Lex Fridman (34:07.780)
because you kind of mentioned coronavirus,
Matt Botvinick (34:10.480)
there's group behavior.
Lex Fridman (34:12.400)
Oh, sure.
Matt Botvinick (34:13.480)
Is there something interesting
Lex Fridman (34:14.960)
to your search of understanding the human mind
Matt Botvinick (34:18.720)
where behavior of large groups
Lex Fridman (34:21.520)
or just behavior of groups is interesting,
Matt Botvinick (34:24.240)
seeing that as a collective mind,
Lex Fridman (34:25.640)
as a collective intelligence,
Matt Botvinick (34:27.120)
perhaps seeing the groups of people
Lex Fridman (34:28.880)
as a single intelligent organisms,
Matt Botvinick (34:31.080)
especially looking at the reinforcement learning work
Lex Fridman (34:34.200)
you've done recently.
Matt Botvinick (34:35.600)
Well, yeah, I can't.
Lex Fridman (34:36.920)
I mean, I have the honor of working
Matt Botvinick (34:41.760)
with a lot of incredibly smart people
Lex Fridman (34:43.640)
and I wouldn't wanna take any credit
Matt Botvinick (34:45.480)
for leading the way on the multiagent work
Lex Fridman (34:48.820)
that's come out of my group or DeepMind lately,
Lex Fridman (34:51.360)
but I do find it fascinating.
Lex Fridman (34:53.840)
And I mean, I think it can't be debated.
Matt Botvinick (35:00.760)
You know, human behavior arises within communities.
Lex Fridman (35:06.000)
That just seems to me self evident.
Lex Fridman (35:08.960)
But to me, it is self evident,
Lex Fridman (35:11.400)
but that seems to be a profound aspects
Matt Botvinick (35:14.720)
of something that created.
Lex Fridman (35:16.040)
That was like, if you look at like 2001 Space Odyssey
Matt Botvinick (35:19.160)
when the monkeys touched the...
Lex Fridman (35:21.360)
Yeah.
Matt Botvinick (35:22.200)
That's the magical moment I think Yuval Harari argues
Lex Fridman (35:25.320)
that the ability of our large numbers of humans
Matt Botvinick (35:29.400)
to hold an idea, to converge towards idea together,
Lex Fridman (35:31.880)
like you said, shaking hands versus bumping elbows,
Matt Botvinick (35:34.360)
somehow converge without being in a room altogether,
Lex Fridman (35:40.880)
just kind of this like distributed convergence
Matt Botvinick (35:43.380)
towards an idea over a particular period of time
Lex Fridman (35:46.720)
seems to be fundamental to just every aspect
Matt Botvinick (35:51.520)
of our cognition, of our intelligence,
Lex Fridman (35:53.400)
because humans, I will talk about reward,
Lex Fridman (35:56.720)
but it seems like we don't really have
Lex Fridman (35:58.720)
a clear objective function under which we operate,
Lex Fridman (36:01.320)
but we all kind of converge towards one somehow.
Lex Fridman (36:04.160)
And that to me has always been a mystery
Matt Botvinick (36:07.600)
that I think is somehow productive
Lex Fridman (36:09.840)
for also understanding AI systems.
Lex Fridman (36:13.620)
But I guess that's the next step.
Lex Fridman (36:16.520)
The first step is try to understand the mind.
Matt Botvinick (36:18.780)
Well, I don't know.
Lex Fridman (36:19.700)
I mean, I think there's something to the argument
Matt Botvinick (36:22.520)
that that kind of like strictly bottom up approach
Lex Fridman (36:27.520)
is wrongheaded.
Matt Botvinick (36:29.920)
In other words, there are basic phenomena,
Lex Fridman (36:34.880)
basic aspects of human intelligence
Matt Botvinick (36:36.860)
that can only be understood in the context of groups.
Lex Fridman (36:43.280)
I'm perfectly open to that.
Matt Botvinick (36:44.680)
I've never been particularly convinced by the notion
Lex Fridman (36:48.680)
that we should consider intelligence
Matt Botvinick (36:52.360)
to inhere at the level of communities.
Lex Fridman (36:55.600)
I don't know why, I'm sort of stuck on the notion
Matt Botvinick (36:58.720)
that the basic unit that we want to understand
Lex Fridman (37:01.380)
is individual humans.
Lex Fridman (37:02.720)
And if we have to understand that
Lex Fridman (37:05.880)
in the context of other humans, fine.
Lex Fridman (37:08.560)
But for me, intelligence is just,
Lex Fridman (37:11.320)
I stubbornly define it as something
Matt Botvinick (37:14.640)
that is an aspect of an individual human.
Lex Fridman (37:18.800)
That's just my, I don't know if that's a matter of taste.
Matt Botvinick (37:20.200)
I'm with you, but that could be the reductionist dream
Lex Fridman (37:22.880)
of a scientist because you can understand a single human.
Matt Botvinick (37:26.400)
It also is very possible that intelligence can only arise
Lex Fridman (37:30.760)
when there's multiple intelligences.
Matt Botvinick (37:33.040)
When there's multiple sort of, it's a sad thing,
Lex Fridman (37:37.480)
if that's true, because it's very difficult to study.
Lex Fridman (37:39.880)
But if it's just one human,
Lex Fridman (37:42.440)
that one human would not be homosapien,
Matt Botvinick (37:44.880)
would not become that intelligent.
Lex Fridman (37:46.520)
That's a possibility.
Matt Botvinick (37:48.500)
I'm with you.
Lex Fridman (37:50.040)
One thing I will say along these lines
Matt Botvinick (37:52.800)
is that I think a serious effort
Lex Fridman (38:01.280)
to understand human intelligence
Lex Fridman (38:05.600)
and maybe to build humanlike intelligence
Lex Fridman (38:09.680)
needs to pay just as much attention
Matt Botvinick (38:11.840)
to the structure of the environment
Lex Fridman (38:14.000)
as to the structure of the cognizing system,
Matt Botvinick (38:20.040)
whether it's a brain or an AI system.
Lex Fridman (38:23.260)
That's one thing I took away actually
Matt Botvinick (38:24.640)
from my early studies with the pioneers
Lex Fridman (38:27.920)
of neural network research,
Matt Botvinick (38:29.900)
people like Jay McClelland and John Cohen.
Lex Fridman (38:34.080)
The structure of cognition is really,
Matt Botvinick (38:38.600)
it's only partly a function of the architecture of the brain
Lex Fridman (38:44.480)
and the learning algorithms that it implements.
Lex Fridman (38:46.980)
What really shapes it is the interaction of those things
Lex Fridman (38:51.520)
with the structure of the world
Matt Botvinick (38:54.460)
in which those things are embedded.
Lex Fridman (38:56.680)
And that's especially important for,
Matt Botvinick (38:58.280)
that's made most clear in reinforcement learning
Lex Fridman (39:00.880)
where the simulated environment is,
Matt Botvinick (39:03.720)
you can only learn as much as you can simulate.
Lex Fridman (39:05.800)
And that's what DeepMind made very clear
Matt Botvinick (39:09.360)
with the other aspect of the environment,
Lex Fridman (39:11.080)
which is the self play mechanism of the other agent,
Matt Botvinick (39:15.600)
of the competitive behavior,
Lex Fridman (39:16.840)
which the other agent becomes the environment essentially.
Lex Fridman (39:20.000)
And that's, I mean, one of the most exciting ideas in AI
Lex Fridman (39:24.080)
is the self play mechanism that's able to learn successfully.
Lex Fridman (39:27.960)
So there you go.
Lex Fridman (39:28.800)
There's a thing where competition is essential
Matt Botvinick (39:31.600)
for learning, at least in that context.
Lex Fridman (39:35.040)
So if we can step back into another sort of beautiful world,
Matt Botvinick (39:37.960)
which is the actual mechanics,
Lex Fridman (39:42.040)
the dirty mess of it of the human brain,
Lex Fridman (39:44.680)
is there something for people who might not know?
Lex Fridman (39:49.440)
Is there something you can comment on
Matt Botvinick (39:51.120)
or describe the key parts of the brain
Lex Fridman (39:53.960)
that are important for intelligence or just in general,
Lex Fridman (39:56.840)
what are the different parts of the brain
Lex Fridman (39:58.620)
that you're curious about that you've studied
Lex Fridman (40:01.120)
and that are just good to know about
Lex Fridman (40:03.880)
when you're thinking about cognition?
Matt Botvinick (40:06.240)
Well, my area of expertise, if I have one,
Lex Fridman (40:11.200)
is prefrontal cortex.
Lex Fridman (40:14.200)
So, you know. What's that?
Lex Fridman (40:16.560)
Where do we?
Matt Botvinick (40:18.200)
It depends on who you ask.
Lex Fridman (40:21.520)
The technical definition is anatomical.
Matt Botvinick (40:25.640)
There are parts of your brain
Lex Fridman (40:30.680)
that are responsible for motor behavior
Lex Fridman (40:32.480)
and they're very easy to identify.
Lex Fridman (40:35.740)
And the region of your cerebral cortex,
Matt Botvinick (40:40.760)
the sort of outer crust of your brain
Lex Fridman (40:43.960)
that lies in front of those
Matt Botvinick (40:46.440)
is defined as the prefrontal cortex.
Lex Fridman (40:49.360)
And when you say anatomical, sorry to interrupt,
Lex Fridman (40:51.960)
so that's referring to sort of the geographic region
Lex Fridman (40:57.160)
as opposed to some kind of functional definition.
Matt Botvinick (41:00.160)
Exactly, so this is kind of the coward's way out.
Lex Fridman (41:04.400)
I'm telling you what the prefrontal cortex is
Matt Botvinick (41:06.000)
just in terms of what part of the real estate it occupies.
Lex Fridman (41:09.640)
It's the thing in the front of the brain.
Matt Botvinick (41:10.720)
Yeah, exactly.
Lex Fridman (41:11.680)
And in fact, the early history
Matt Botvinick (41:14.960)
of neuroscientific investigation
Lex Fridman (41:20.840)
of what this front part of the brain does
Matt Botvinick (41:23.480)
is sort of funny to read
Lex Fridman (41:25.760)
because it was really World War I
Matt Botvinick (41:32.280)
that started people down this road
Lex Fridman (41:34.580)
of trying to figure out what different parts of the brain,
Matt Botvinick (41:37.280)
the human brain do in the sense
Lex Fridman (41:39.440)
that there were a lot of people with brain damage
Matt Botvinick (41:42.560)
who came back from the war with brain damage.
Lex Fridman (41:44.800)
And that provided, as tragic as that was,
Matt Botvinick (41:47.740)
it provided an opportunity for scientists
Lex Fridman (41:49.900)
to try to identify the functions of different brain regions.
Lex Fridman (41:53.440)
And that was actually incredibly productive,
Lex Fridman (41:56.200)
but one of the frustrations that neuropsychologists faced
Matt Botvinick (41:59.480)
was they couldn't really identify exactly
Lex Fridman (42:02.160)
what the deficit was that arose from damage
Matt Botvinick (42:05.040)
to these most kind of frontal parts of the brain.
Lex Fridman (42:08.440)
It was just a very difficult thing to pin down.
Matt Botvinick (42:13.680)
There were a couple of neuropsychologists
Lex Fridman (42:16.080)
who identified through a large amount
Matt Botvinick (42:20.600)
of clinical experience and close observation,
Lex Fridman (42:23.000)
they started to put their finger on a syndrome
Matt Botvinick (42:26.240)
that was associated with frontal damage.
Lex Fridman (42:27.680)
Actually, one of them was a Russian neuropsychologist
Matt Botvinick (42:30.480)
named Luria, who students of cognitive psychology still read.
Lex Fridman (42:36.160)
And what he started to figure out was that
Matt Botvinick (42:41.360)
the frontal cortex was somehow involved in flexibility,
Lex Fridman (42:48.060)
in guiding behaviors that required someone
Matt Botvinick (42:52.320)
to override a habit, or to do something unusual,
Lex Fridman (42:57.600)
or to change what they were doing in a very flexible way
Matt Botvinick (43:01.040)
from one moment to another.
Lex Fridman (43:02.560)
So focused on like new experiences.
Lex Fridman (43:05.080)
And so the way your brain processes
Lex Fridman (43:08.800)
and acts in new experiences.
Matt Botvinick (43:10.960)
Yeah, what later helped bring this function
Lex Fridman (43:14.760)
into better focus was a distinction
Matt Botvinick (43:17.240)
between controlled and automatic behavior,
Lex Fridman (43:19.880)
or in other literatures, this is referred to
Matt Botvinick (43:23.680)
as habitual behavior versus goal directed behavior.
Lex Fridman (43:28.280)
So it's very, very clear that the human brain
Matt Botvinick (43:33.440)
has pathways that are dedicated to habits,
Lex Fridman (43:36.600)
to things that you do all the time,
Lex Fridman (43:39.360)
and they need to be automatized
Lex Fridman (43:42.440)
so that they don't require you to concentrate too much.
Lex Fridman (43:45.140)
So that leaves your cognitive capacity
Lex Fridman (43:47.840)
free to do other things.
Matt Botvinick (43:49.800)
Just think about the difference
Lex Fridman (43:51.640)
between driving when you're learning to drive
Matt Botvinick (43:55.960)
versus driving after you're a fairly expert.
Lex Fridman (43:59.160)
There are brain pathways that slowly absorb
Matt Botvinick (44:03.560)
those frequently performed behaviors
Lex Fridman (44:07.840)
so that they can be habits, so that they can be automatic.
Matt Botvinick (44:12.360)
That's kind of like the purest form of learning.
Lex Fridman (44:14.900)
I guess it's happening there, which is why,
Matt Botvinick (44:18.360)
I mean, this is kind of jumping ahead,
Lex Fridman (44:20.000)
which is why that perhaps is the most useful for us
Matt Botvinick (44:22.480)
to focusing on and trying to see
Lex Fridman (44:24.080)
how artificial intelligence systems can learn.
Lex Fridman (44:27.340)
Is that the way you think?
Lex Fridman (44:28.180)
It's interesting.
Matt Botvinick (44:29.000)
I do think about this distinction
Lex Fridman (44:30.040)
between controlled and automatic,
Matt Botvinick (44:31.440)
or goal directed and habitual behavior a lot
Lex Fridman (44:34.600)
in thinking about where we are in AI research.
Lex Fridman (44:42.960)
But just to finish the kind of dissertation here,
Lex Fridman (44:46.480)
the role of the prefrontal cortex
Matt Botvinick (44:51.380)
is generally understood these days
Lex Fridman (44:54.600)
sort of in contradistinction to that habitual domain.
Matt Botvinick (45:00.440)
In other words, the prefrontal cortex
Lex Fridman (45:02.320)
is what helps you override those habits.
Matt Botvinick (45:05.840)
It's what allows you to say,
Lex Fridman (45:07.440)
well, what I usually do in this situation is X,
Lex Fridman (45:10.800)
but given the context, I probably should do Y.
Lex Fridman (45:14.160)
I mean, the elbow bump is a great example, right?
Matt Botvinick (45:18.080)
Reaching out and shaking hands
Lex Fridman (45:19.300)
is probably a habitual behavior,
Lex Fridman (45:22.520)
and it's the prefrontal cortex that allows us
Lex Fridman (45:26.000)
to bear in mind that there's something unusual
Matt Botvinick (45:28.760)
going on right now, and in this situation,
Lex Fridman (45:31.360)
I need to not do the usual thing.
Matt Botvinick (45:34.720)
The kind of behaviors that Luria reported,
Lex Fridman (45:38.560)
and he built tests for detecting these kinds of things,
Matt Botvinick (45:42.040)
were exactly like this.
Lex Fridman (45:43.460)
So in other words, when I stick out my hand,
Matt Botvinick (45:47.540)
I want you instead to present your elbow.
Lex Fridman (45:49.760)
A patient with frontal damage
Matt Botvinick (45:51.080)
would have a great deal of trouble with that.
Lex Fridman (45:53.520)
Somebody proffering their hand would elicit a handshake.
Matt Botvinick (45:58.800)
The prefrontal cortex is what allows us to say,
Lex Fridman (46:00.920)
hold on, hold on, that's the usual thing,
Lex Fridman (46:03.840)
but I have the ability to bear in mind
Lex Fridman (46:07.120)
even very unusual contexts and to reason about
Lex Fridman (46:10.520)
what behavior is appropriate there.
Lex Fridman (46:13.240)
Just to get a sense, are us humans special
Lex Fridman (46:17.560)
in the presence of the prefrontal cortex?
Lex Fridman (46:20.680)
Do mice have a prefrontal cortex?
Lex Fridman (46:22.640)
Do other mammals that we can study?
Lex Fridman (46:25.900)
If no, then how do they integrate new experiences?
Matt Botvinick (46:30.040)
Yeah, that's a really tricky question
Lex Fridman (46:33.760)
and a very timely question
Matt Botvinick (46:35.840)
because we have revolutionary new technologies
Lex Fridman (46:44.040)
for monitoring, measuring,
Lex Fridman (46:48.280)
and also causally influencing neural behavior
Lex Fridman (46:52.040)
in mice and fruit flies.
Lex Fridman (46:57.000)
And these techniques are not fully available
Lex Fridman (47:00.640)
even for studying brain function in monkeys,
Matt Botvinick (47:06.080)
let alone humans.
Lex Fridman (47:08.160)
And so it's a very sort of, for me at least,
Matt Botvinick (47:12.920)
a very urgent question whether the kinds of things
Lex Fridman (47:16.160)
that we wanna understand about human intelligence
Matt Botvinick (47:18.000)
can be pursued in these other organisms.
Lex Fridman (47:22.000)
And to put it briefly, there's disagreement.
Matt Botvinick (47:26.500)
People who study fruit flies will often tell you,
Lex Fridman (47:32.960)
hey, fruit flies are smarter than you think.
Lex Fridman (47:35.520)
And they'll point to experiments where fruit flies
Lex Fridman (47:37.600)
were able to learn new behaviors,
Matt Botvinick (47:40.320)
were able to generalize from one stimulus to another
Lex Fridman (47:44.180)
in a way that suggests that they have abstractions
Matt Botvinick (47:47.500)
that guide their generalization.
Lex Fridman (47:51.880)
I've had many conversations in which
Matt Botvinick (47:53.840)
I will start by observing,
Lex Fridman (47:58.160)
recounting some observation about mouse behavior
Matt Botvinick (48:05.200)
where it seemed like mice were taking an awfully long time
Lex Fridman (48:09.060)
to learn a task that for a human would be profoundly trivial.
Lex Fridman (48:13.660)
And I will conclude from that,
Lex Fridman (48:16.460)
that mice really don't have the cognitive flexibility
Matt Botvinick (48:18.800)
that we want to explain.
Lex Fridman (48:20.100)
And then a mouse researcher will say to me,
Matt Botvinick (48:21.760)
well, hold on, that experiment may not have worked
Lex Fridman (48:26.360)
because you asked a mouse to deal with stimuli
Lex Fridman (48:31.280)
and behaviors that were very unnatural for the mouse.
Lex Fridman (48:34.300)
If instead you kept the logic of the experiment the same,
Lex Fridman (48:38.760)
but presented the information in a way
Lex Fridman (48:44.440)
that aligns with what mice are used to dealing with
Matt Botvinick (48:46.880)
in their natural habitats,
Lex Fridman (48:48.480)
you might find that a mouse actually has more intelligence
Matt Botvinick (48:51.080)
than you think.
Lex Fridman (48:52.440)
And then they'll go on to show you videos
Matt Botvinick (48:54.920)
of mice doing things in their natural habitat,
Lex Fridman (48:57.440)
which seem strikingly intelligent,
Matt Botvinick (49:00.000)
dealing with physical problems.
Lex Fridman (49:02.920)
I have to drag this piece of food back to my lair,
Lex Fridman (49:07.180)
but there's something in my way
Lex Fridman (49:08.560)
and how do I get rid of that thing?
Lex Fridman (49:10.400)
So I think these are open questions
Lex Fridman (49:13.160)
to put it, to sum that up.
Lex Fridman (49:15.400)
And then taking a small step back related to that
Lex Fridman (49:18.520)
is you kind of mentioned we're taking a little shortcut
Matt Botvinick (49:21.440)
by saying it's a geographic part of the prefrontal cortex
Lex Fridman (49:26.600)
is a region of the brain.
Lex Fridman (49:28.280)
But if we, what's your sense in a bigger philosophical view,
Lex Fridman (49:33.720)
prefrontal cortex and the brain in general,
Lex Fridman (49:36.260)
do you have a sense that it's a set of subsystems
Lex Fridman (49:38.840)
in the way we've kind of implied
Matt Botvinick (49:41.180)
that are pretty distinct or to what degree is it that
Lex Fridman (49:46.180)
or to what degree is it a giant interconnected mess
Matt Botvinick (49:49.460)
where everything kind of does everything
Lex Fridman (49:51.380)
and it's impossible to disentangle them?
Matt Botvinick (49:54.920)
I think there's overwhelming evidence
Lex Fridman (49:57.020)
that there's functional differentiation,
Matt Botvinick (50:00.060)
that it's clearly not the case
Lex Fridman (50:03.460)
that all parts of the brain are doing the same thing.
Matt Botvinick (50:07.100)
This follows immediately from the kinds of studies
Lex Fridman (50:11.100)
of brain damage that we were chatting about before.
Matt Botvinick (50:14.620)
It's obvious from what you see
Lex Fridman (50:18.060)
if you stick an electrode in the brain
Lex Fridman (50:19.620)
and measure what's going on at the level of neural activity.
Lex Fridman (50:25.960)
Having said that, there are two other things to add,
Matt Botvinick (50:30.680)
which kind of, I don't know,
Lex Fridman (50:32.740)
maybe tug in the other direction.
Matt Botvinick (50:34.340)
One is that it's when you look carefully
Lex Fridman (50:39.740)
at functional differentiation in the brain,
Lex Fridman (50:42.220)
what you usually end up concluding,
Lex Fridman (50:44.900)
at least this is my observation of the literature,
Matt Botvinick (50:48.140)
is that the differences between regions are graded
Lex Fridman (50:52.780)
rather than being discreet.
Lex Fridman (50:55.180)
So it doesn't seem like it's easy
Lex Fridman (50:57.460)
to divide the brain up into true modules
Matt Botvinick (51:03.300)
that have clear boundaries and that have
Lex Fridman (51:07.460)
you know, clear channels of communication between them.
Lex Fridman (51:16.020)
And this applies to the prefrontal cortex?
Lex Fridman (51:18.020)
Yeah, oh yeah.
Matt Botvinick (51:18.860)
The prefrontal cortex is made up
Lex Fridman (51:20.200)
of a bunch of different subregions,
Matt Botvinick (51:23.140)
the functions of which are not clearly defined
Lex Fridman (51:27.380)
and the borders of which seem to be quite vague.
Lex Fridman (51:32.300)
And then there's another thing that's popping up
Lex Fridman (51:34.420)
in very recent research, which, you know, which,
Matt Botvinick (51:40.280)
involves application of these new techniques,
Lex Fridman (51:44.940)
which there are a number of studies that suggest that
Matt Botvinick (51:48.820)
parts of the brain that we would have previously thought
Lex Fridman (51:51.540)
were quite focused in their function
Matt Botvinick (51:57.740)
are actually carrying signals
Lex Fridman (51:59.100)
that we wouldn't have thought would be there.
Matt Botvinick (52:01.340)
For example, looking in the primary visual cortex,
Lex Fridman (52:04.500)
which is classically thought of as basically
Matt Botvinick (52:07.900)
the first cortical way station
Lex Fridman (52:09.380)
for processing visual information.
Matt Botvinick (52:10.900)
Basically what it should care about is, you know,
Lex Fridman (52:12.980)
where are the edges in this scene that I'm viewing?
Matt Botvinick (52:17.460)
It turns out that if you have enough data,
Lex Fridman (52:19.460)
you can recover information from primary visual cortex
Matt Botvinick (52:22.220)
about all sorts of things.
Lex Fridman (52:23.220)
Like, you know, what behavior the animal is engaged
Matt Botvinick (52:26.180)
in right now and how much reward is on offer
Lex Fridman (52:29.340)
in the task that it's pursuing.
Lex Fridman (52:31.340)
So it's clear that even regions whose function
Lex Fridman (52:36.740)
is pretty well defined at a core screen
Matt Botvinick (52:40.540)
are nonetheless carrying some information
Lex Fridman (52:42.860)
about information from very different domains.
Matt Botvinick (52:47.060)
So, you know, the history of neuroscience
Lex Fridman (52:49.780)
is sort of this oscillation between the two views
Matt Botvinick (52:52.660)
that you articulated, you know, the kind of modular view
Lex Fridman (52:55.460)
and then the big, you know, mush view.
Matt Botvinick (52:57.740)
And, you know, I think, I guess we're gonna end up
Lex Fridman (53:01.580)
somewhere in the middle.
Matt Botvinick (53:02.800)
Which is unfortunate for our understanding
Lex Fridman (53:05.580)
because there's something about our, you know,
Matt Botvinick (53:08.880)
conceptual system that finds it's easy to think about
Lex Fridman (53:11.380)
a modularized system and easy to think about
Matt Botvinick (53:13.680)
a completely undifferentiated system.
Lex Fridman (53:15.500)
But something that kind of lies in between is confusing.
Lex Fridman (53:19.980)
But we're gonna have to get used to it, I think.
Lex Fridman (53:21.860)
Unless we can understand deeply the lower level mechanism
Matt Botvinick (53:24.660)
of neuronal communication.
Lex Fridman (53:25.860)
Yeah, yeah.
Lex Fridman (53:26.760)
But on that topic, you kind of mentioned information.
Lex Fridman (53:29.660)
Just to get a sense, I imagine something
Matt Botvinick (53:31.860)
that there's still mystery and disagreement on
Lex Fridman (53:34.620)
is how does the brain carry information and signal?
Matt Botvinick (53:38.060)
Like what in your sense is the basic mechanism
Lex Fridman (53:43.380)
of communication in the brain?
Matt Botvinick (53:46.420)
Well, I guess I'm old fashioned in that I consider
Lex Fridman (53:52.020)
the networks that we use in deep learning research
Matt Botvinick (53:54.340)
to be a reasonable approximation to, you know,
Lex Fridman (53:59.080)
the mechanisms that carry information in the brain.
Lex Fridman (54:02.500)
So the usual way of articulating that is to say,
Lex Fridman (54:06.180)
what really matters is a rate code.
Lex Fridman (54:08.540)
What matters is how quickly is an individual neuron spiking?
Lex Fridman (54:14.580)
You know, what's the frequency at which it's spiking?
Lex Fridman (54:16.380)
Is it right?
Lex Fridman (54:17.200)
So the timing of the spike.
Lex Fridman (54:18.040)
Yeah, is it firing fast or slow?
Lex Fridman (54:20.340)
Let's, you know, let's put a number on that.
Lex Fridman (54:22.740)
And that number is enough to capture
Lex Fridman (54:24.380)
what neurons are doing.
Matt Botvinick (54:26.140)
There's, you know, there's still uncertainty
Lex Fridman (54:30.620)
about whether that's an adequate description
Matt Botvinick (54:34.500)
of how information is transmitted within the brain.
Lex Fridman (54:39.880)
There, you know, there are studies that suggest
Matt Botvinick (54:42.820)
that the precise timing of spikes matters.
Lex Fridman (54:46.060)
There are studies that suggest that there are computations
Matt Botvinick (54:50.660)
that go on within the dendritic tree, within a neuron,
Lex Fridman (54:54.520)
that are quite rich and structured
Lex Fridman (54:57.100)
and that really don't equate to anything that we're doing
Lex Fridman (54:59.980)
in our artificial neural networks.
Matt Botvinick (55:02.820)
Having said that, I feel like we can get,
Lex Fridman (55:05.360)
I feel like we're getting somewhere
Matt Botvinick (55:08.260)
by sticking to this high level of abstraction.
Lex Fridman (55:11.620)
Just the rate, and by the way,
Matt Botvinick (55:13.380)
we're talking about the electrical signal.
Lex Fridman (55:16.220)
I remember reading some vague paper somewhere recently
Matt Botvinick (55:20.060)
where the mechanical signal, like the vibrations
Lex Fridman (55:23.420)
or something of the neurons, also communicates information.
Matt Botvinick (55:28.820)
I haven't seen that, but.
Lex Fridman (55:30.260)
There's somebody who was arguing
Matt Botvinick (55:32.100)
that the electrical signal, this is in a nature paper,
Lex Fridman (55:36.840)
something like that, where the electrical signal
Matt Botvinick (55:38.780)
is actually a side effect of the mechanical signal.
Lex Fridman (55:43.740)
But I don't think that changes the story.
Lex Fridman (55:46.100)
But it's almost an interesting idea
Lex Fridman (55:49.060)
that there could be a deeper, it's always like in physics
Matt Botvinick (55:52.420)
with quantum mechanics, there's always a deeper story
Lex Fridman (55:55.740)
that could be underlying the whole thing.
Lex Fridman (55:57.500)
But you think it's basically the rate of spiking
Lex Fridman (56:00.540)
that gets us, that's like the lowest hanging fruit
Matt Botvinick (56:02.820)
that can get us really far.
Lex Fridman (56:04.060)
This is a classical view.
Matt Botvinick (56:06.580)
I mean, this is not, the only way in which this stance
Lex Fridman (56:10.700)
would be controversial is in the sense
Matt Botvinick (56:13.580)
that there are members of the neuroscience community
Lex Fridman (56:17.100)
who are interested in alternatives.
Lex Fridman (56:18.820)
But this is really a very mainstream view.
Lex Fridman (56:21.400)
The way that neurons communicate
Matt Botvinick (56:22.940)
is that neurotransmitters arrive,
Lex Fridman (56:30.180)
they wash up on a neuron, the neuron has receptors
Matt Botvinick (56:34.500)
for those transmitters, the meeting of the transmitter
Lex Fridman (56:39.040)
with these receptors changes the voltage of the neuron.
Lex Fridman (56:42.340)
And if enough voltage change occurs, then a spike occurs,
Lex Fridman (56:46.860)
one of these like discrete events.
Lex Fridman (56:48.660)
And it's that spike that is conducted down the axon
Lex Fridman (56:52.300)
and leads to neurotransmitter release.
Matt Botvinick (56:54.580)
This is just like neuroscience 101.
Lex Fridman (56:56.860)
This is like the way the brain is supposed to work.
Matt Botvinick (56:59.300)
Now, what we do when we build artificial neural networks
Lex Fridman (57:03.660)
of the kind that are now popular in the AI community
Matt Botvinick (57:08.060)
is that we don't worry about those individual spikes.
Lex Fridman (57:11.780)
We just worry about the frequency
Matt Botvinick (57:14.220)
at which those spikes are being generated.
Lex Fridman (57:16.980)
And people talk about that as the activity of a neuron.
Lex Fridman (57:22.340)
And so the activity of units in a deep learning system
Lex Fridman (57:27.180)
is broadly analogous to the spike rate of a neuron.
Matt Botvinick (57:32.900)
There are people who believe that there are other forms
Lex Fridman (57:38.020)
of communication in the brain.
Matt Botvinick (57:39.180)
In fact, I've been involved in some research recently
Lex Fridman (57:41.260)
that suggests that the voltage fluctuations
Matt Botvinick (57:46.260)
that occur in populations of neurons
Lex Fridman (57:49.260)
that are sort of below the level of spike production
Matt Botvinick (57:54.860)
may be important for communication.
Lex Fridman (57:57.220)
But I'm still pretty old school in the sense
Matt Botvinick (58:00.220)
that I think that the things that we're building
Lex Fridman (58:02.700)
in AI research constitute reasonable models
Matt Botvinick (58:06.980)
of how a brain would work.
Lex Fridman (58:10.300)
Let me ask just for fun a crazy question, because I can.
Lex Fridman (58:14.220)
Do you think it's possible we're completely wrong
Lex Fridman (58:17.020)
about the way this basic mechanism
Matt Botvinick (58:20.060)
of neuronal communication, that the information
Lex Fridman (58:23.700)
is stored in some very different kind of way in the brain?
Matt Botvinick (58:26.340)
Oh, heck yes.
Lex Fridman (58:27.580)
I mean, look, I wouldn't be a scientist
Matt Botvinick (58:29.900)
if I didn't think there was any chance we were wrong.
Lex Fridman (58:32.500)
But I mean, if you look at the history
Matt Botvinick (58:36.420)
of deep learning research as it's been applied
Lex Fridman (58:39.900)
to neuroscience, of course the vast majority
Matt Botvinick (58:42.620)
of deep learning research these days isn't about neuroscience.
Lex Fridman (58:45.380)
But if you go back to the 1980s,
Matt Botvinick (58:49.060)
there's sort of an unbroken chain of research
Lex Fridman (58:52.740)
in which a particular strategy is taken,
Matt Botvinick (58:54.940)
which is, hey, let's train a deep learning system.
Lex Fridman (59:00.180)
Let's train a multi layer neural network
Matt Botvinick (59:04.060)
on this task that we trained our rat on,
Lex Fridman (59:09.260)
or our monkey on, or this human being on.
Lex Fridman (59:12.300)
And then let's look at what the units
Lex Fridman (59:15.700)
deep in the system are doing.
Lex Fridman (59:17.700)
And let's ask whether what they're doing
Lex Fridman (59:20.780)
resembles what we know about what neurons
Matt Botvinick (59:23.260)
deep in the brain are doing.
Lex Fridman (59:24.620)
And over and over and over and over,
Matt Botvinick (59:28.540)
that strategy works in the sense that
Lex Fridman (59:32.020)
the learning algorithms that we have access to,
Matt Botvinick (59:34.340)
which typically center on back propagation,
Lex Fridman (59:37.740)
they give rise to patterns of activity,
Matt Botvinick (59:42.060)
patterns of response,
Lex Fridman (59:45.220)
patterns of neuronal behavior in these artificial models
Matt Botvinick (59:48.740)
that look hauntingly similar to what you see in the brain.
Lex Fridman (59:53.660)
And is that a coincidence?
Matt Botvinick (59:57.380)
At a certain point, it starts looking like such coincidence
🔗 相关节目