Oriol Vinyals: Deep Learning and Artificial General Intelligence

Lex Fridman (01:21.800)

I think partly I would say, do we want that?

Oriol Vinyals (01:24.680)

I really like when we start now with very powerful models,

Lex Fridman (01:29.400)

interacting with them and thinking of them

Oriol Vinyals (01:32.160)

more closer to us.

Lex Fridman (01:34.080)

The question is, if you remove the human side

Lex Fridman (01:37.020)

of the conversation, is that an interesting artifact?

Lex Fridman (01:42.320)

And I would say, probably not.

Oriol Vinyals (01:44.440)

I've seen, for instance, last time we spoke,

Lex Fridman (01:47.400)

like we were talking about StarCraft,

Lex Fridman (01:50.320)

and creating agents that play games involves self play,

Lex Fridman (01:54.920)

but ultimately what people care about was,

Lex Fridman (01:57.660)

how does this agent behave

Lex Fridman (01:59.080)

when the opposite side is a human?

Lex Fridman (02:02.700)

So without a doubt,

Lex Fridman (02:04.720)

we will probably be more empowered by AI.

Oriol Vinyals (02:08.560)

Maybe you can source some questions from an AI system.

Lex Fridman (02:12.480)

I mean, that even today, I would say it's quite plausible

Oriol Vinyals (02:15.020)

that with your creativity,

Lex Fridman (02:17.060)

you might actually find very interesting questions

Oriol Vinyals (02:19.400)

that you can filter.

Lex Fridman (02:20.740)

We call this cherry picking sometimes

Oriol Vinyals (02:22.420)

in the field of language.

Lex Fridman (02:24.400)

And likewise, if I had now the tools on my side,

Oriol Vinyals (02:27.540)

I could say, look, you're asking this interesting question.

Lex Fridman (02:30.660)

From this answer, I like the words chosen

Oriol Vinyals (02:33.240)

by this particular system that created a few words.

Lex Fridman (02:36.600)

Completely replacing it feels not exactly exciting to me.

Oriol Vinyals (02:41.280)

Although in my lifetime, I think way,

Lex Fridman (02:43.780)

I mean, given the trajectory,

Oriol Vinyals (02:45.520)

I think it's possible that perhaps

Lex Fridman (02:48.020)

there could be interesting,

Oriol Vinyals (02:49.880)

maybe self play interviews as you're suggesting

Lex Fridman (02:53.040)

that would look or sound quite interesting

Lex Fridman (02:56.160)

and probably would educate

Lex Fridman (02:57.720)

or you could learn a topic through listening

Oriol Vinyals (03:00.160)

to one of these interviews at a basic level at least.

Lex Fridman (03:03.200)

So you said it doesn't seem exciting to you,

Lex Fridman (03:04.800)

but what if exciting is part of the objective function

Lex Fridman (03:07.520)

the thing is optimized over?

Lex Fridman (03:09.120)

So there's probably a huge amount of data of humans

Lex Fridman (03:12.840)

if you look correctly, of humans communicating online,

Lex Fridman (03:16.080)

and there's probably ways to measure the degree of,

Lex Fridman (03:19.280)

you know, as they talk about engagement.

Lex Fridman (03:21.920)

So you can probably optimize the question

Lex Fridman (03:24.140)

that's most created an engaging conversation in the past.

Lex Fridman (03:28.680)

So actually, if you strictly use the word exciting,

Lex Fridman (03:33.200)

there is probably a way to create

Oriol Vinyals (03:37.240)

a optimally exciting conversations

Lex Fridman (03:40.320)

that involve AI systems.

Oriol Vinyals (03:42.160)

At least one side is AI.

Lex Fridman (03:44.600)

Yeah, that makes sense, I think,

Oriol Vinyals (03:46.560)

maybe looping back a bit to games and the game industry,

Lex Fridman (03:50.240)

when you design algorithms,

Lex Fridman (03:53.040)

you're thinking about winning as the objective, right?

Lex Fridman (03:55.800)

Or the reward function.

Lex Fridman (03:57.320)

But in fact, when we discussed this with Blizzard,

Lex Fridman (04:00.080)

the creators of StarCraft in this case,

Oriol Vinyals (04:02.320)

I think what's exciting, fun,

Lex Fridman (04:05.340)

if you could measure that and optimize for that,

Oriol Vinyals (04:09.160)

that's probably why we play video games

Lex Fridman (04:11.720)

or why we interact or listen or look at cat videos

Oriol Vinyals (04:14.640)

or whatever on the internet.

Lex Fridman (04:16.460)

So it's true that modeling reward

Oriol Vinyals (04:19.500)

beyond the obvious reward functions

Lex Fridman (04:21.320)

we've used to in reinforcement learning

Oriol Vinyals (04:23.720)

is definitely very exciting.

Lex Fridman (04:25.560)

And again, there is some progress actually

Oriol Vinyals (04:28.240)

into a particular aspect of AI, which is quite critical,

Lex Fridman (04:32.140)

which is, for instance, is a conversation

Lex Fridman (04:36.120)

or is the information truthful, right?

Lex Fridman (04:38.200)

So you could start trying to evaluate these

Lex Fridman (04:41.640)

from accepts from the internet, right?

Lex Fridman (04:44.440)

That has lots of information.

Lex Fridman (04:45.840)

And then if you can learn a function automated ideally,

Lex Fridman (04:50.220)

so you can also optimize it more easily,

Oriol Vinyals (04:52.920)

then you could actually have conversations

Lex Fridman (04:54.920)

that optimize for non obvious things such as excitement.

Lex Fridman (04:59.440)

So yeah, that's quite possible.

Lex Fridman (05:01.100)

And then I would say in that case,

Oriol Vinyals (05:03.620)

it would definitely be fun exercise

Lex Fridman (05:05.960)

and quite unique to have at least one site

Oriol Vinyals (05:08.120)

that is fully driven by an excitement reward function.

Lex Fridman (05:12.840)

But obviously, there would be still quite a lot of humanity

Oriol Vinyals (05:16.960)

in the system, both from who is building the system,

Lex Fridman (05:20.800)

of course, and also, ultimately,

Oriol Vinyals (05:23.600)

if we think of labeling for excitement,

Lex Fridman (05:26.040)

that those labels must come from us

Oriol Vinyals (05:28.480)

because it's just hard to have a computational measure

Lex Fridman (05:32.560)

of excitement.

Oriol Vinyals (05:33.520)

As far as I understand, there's no such thing.

Lex Fridman (05:36.160)

Well, as you mentioned truth also,

Oriol Vinyals (05:39.280)

I would actually venture to say that excitement

Lex Fridman (05:41.840)

is easier to label than truth,

Oriol Vinyals (05:44.160)

or is perhaps has lower consequences of failure.

Lex Fridman (05:49.920)

But there is perhaps the humanness that you mentioned,

Oriol Vinyals (05:55.760)

that's perhaps part of a thing that could be labeled.

Lex Fridman (05:58.280)

And that could mean an AI system that's doing dialogue,

Oriol Vinyals (06:02.520)

that's doing conversations should be flawed, for example.

Lex Fridman (06:07.720)

Like that's the thing you optimize for,

Oriol Vinyals (06:09.440)

which is have inherent contradictions by design,

Lex Fridman (06:13.280)

have flaws by design.

Oriol Vinyals (06:15.080)

Maybe it also needs to have a strong sense of identity.

Lex Fridman (06:18.760)

So it has a backstory it told itself that it sticks to.

Oriol Vinyals (06:22.680)

It has memories, not in terms of how the system is designed,

Lex Fridman (06:26.900)

but it's able to tell stories about its past.

Oriol Vinyals (06:30.440)

It's able to have mortality and fear of mortality

Lex Fridman (06:36.040)

in the following way that it has an identity.

Lex Fridman (06:39.080)

And if it says something stupid

Lex Fridman (06:41.200)

and gets canceled on Twitter, that's the end of that system.

Lex Fridman (06:44.680)

So it's not like you get to rebrand yourself.

Lex Fridman (06:47.320)

That system is, that's it.

Lex Fridman (06:49.320)

So maybe the high stakes nature of it,

Lex Fridman (06:52.080)

because you can't say anything stupid now,

Oriol Vinyals (06:54.520)

or because you'd be canceled on Twitter.

Lex Fridman (06:57.680)

And there's stakes to that.

Lex Fridman (06:59.720)

And that I think part of the reason

Lex Fridman (07:01.120)

that makes it interesting.

Lex Fridman (07:03.480)

And then you have a perspective,

Lex Fridman (07:04.680)

like you've built up over time that you stick with,

Lex Fridman (07:07.680)

and then people can disagree with you.

Lex Fridman (07:09.100)

So holding that perspective strongly,

Oriol Vinyals (07:11.760)

holding sort of maybe a controversial,

Lex Fridman (07:14.000)

at least a strong opinion.

Oriol Vinyals (07:16.280)

All of those elements, it feels like they can be learned

Lex Fridman (07:18.800)

because it feels like there's a lot of data

Oriol Vinyals (07:21.720)

on the internet of people having an opinion.

Lex Fridman (07:24.520)

And then combine that with a metric of excitement,

Oriol Vinyals (07:27.800)

you can start to create something that,

Lex Fridman (07:30.020)

as opposed to trying to optimize

Oriol Vinyals (07:31.680)

for sort of grammatical clarity and truthfulness,

Lex Fridman (07:38.120)

the factual consistency over many sentences,

Oriol Vinyals (07:42.000)

you optimize for the humanness.

Lex Fridman (07:45.320)

And there's obviously data for humanness on the internet.

Lex Fridman (07:48.860)

So I wonder if there's a future where that's part,

Lex Fridman (07:53.760)

or I mean, I sometimes wonder that about myself.

Oriol Vinyals (07:56.400)

I'm a huge fan of podcasts,

Lex Fridman (07:58.120)

and I listen to some podcasts,

Lex Fridman (08:00.760)

and I think like, what is interesting about this?

Lex Fridman (08:03.240)

What is compelling?

Oriol Vinyals (08:05.960)

The same way you watch other games.

Lex Fridman (08:07.440)

Like you said, watch, play StarCraft,

Oriol Vinyals (08:09.160)

or have Magnus Carlsen play chess.

Lex Fridman (08:13.040)

So I'm not a chess player,

Lex Fridman (08:14.920)

but it's still interesting to me.

Lex Fridman (08:16.120)

What is that?

Oriol Vinyals (08:16.960)

That's the stakes of it,

Lex Fridman (08:19.440)

maybe the end of a domination of a series of wins.

Oriol Vinyals (08:23.400)

I don't know, there's all those elements

Lex Fridman (08:25.440)

somehow connect to a compelling conversation.

Lex Fridman (08:28.000)

And I wonder how hard is that to replace,

Lex Fridman (08:30.200)

because ultimately all of that connects

Oriol Vinyals (08:31.840)

to the initial proposition of how to test,

Lex Fridman (08:35.480)

whether an AI is intelligent or not with the Turing test,

Oriol Vinyals (08:38.640)

which I guess my question comes from a place

Lex Fridman (08:41.760)

of the spirit of that test.

Oriol Vinyals (08:43.680)

Yes, I actually recall,

Lex Fridman (08:45.440)

I was just listening to our first podcast

Oriol Vinyals (08:47.920)

where we discussed Turing test.

Lex Fridman (08:50.380)

So I would say from a neural network,

Oriol Vinyals (08:54.760)

AI builder perspective,

Lex Fridman (08:57.640)

there's usually you try to map

Oriol Vinyals (09:01.360)

many of these interesting topics you discuss to benchmarks,

Lex Fridman (09:05.200)

and then also to actual architectures

Oriol Vinyals (09:08.140)

on the how these systems are currently built,

Lex Fridman (09:10.640)

how they learn, what data they learn from,

Lex Fridman (09:13.080)

what are they learning, right?

Lex Fridman (09:14.300)

We're talking about weights of a mathematical function,

Lex Fridman (09:17.800)

and then looking at the current state of the game,

Lex Fridman (09:21.560)

maybe what do we need leaps forward

Oriol Vinyals (09:26.000)

to get to the ultimate stage of all these experiences,

Lex Fridman (09:30.660)

lifetime experience, fears,

Oriol Vinyals (09:32.880)

like words that currently,

Lex Fridman (09:34.800)

barely we're seeing progress

Oriol Vinyals (09:38.020)

just because what's happening today

Lex Fridman (09:40.120)

is you take all these human interactions,

Oriol Vinyals (09:44.020)

it's a large vast variety of human interactions online,

Lex Fridman (09:47.960)

and then you're distilling these sequences, right?

Oriol Vinyals (09:51.640)

Going back to my passion,

Lex Fridman (09:53.000)

like sequences of words, letters, images, sound,

Oriol Vinyals (09:56.920)

there's more modalities here to be at play.

Lex Fridman (09:59.840)

And then you're trying to just learn a function

Oriol Vinyals (10:03.360)

that will be happy,

Lex Fridman (10:04.400)

that maximizes the likelihood of seeing all these

Oriol Vinyals (10:08.840)

through a neural network.

Lex Fridman (10:10.880)

Now, I think there's a few places

Oriol Vinyals (10:14.200)

where the way currently we train these models

Lex Fridman (10:17.240)

would clearly lack to be able to develop

Oriol Vinyals (10:20.000)

the kinds of capabilities you save.

Lex Fridman (10:22.120)

I'll tell you maybe a couple.

Oriol Vinyals (10:23.560)

One is the lifetime of an agent or a model.

Lex Fridman (10:27.640)

So you learn from this data offline, right?

Lex Fridman (10:30.820)

So you're just passively observing and maximizing these,

Lex Fridman (10:33.880)

it's almost like a mountains,

Oriol Vinyals (10:35.360)

like a landscape of mountains,

Lex Fridman (10:37.600)

and then everywhere there's data

Oriol Vinyals (10:39.140)

that humans interacted in this way,

Lex Fridman (10:41.040)

you're trying to make that higher

Lex Fridman (10:43.000)

and then lower where there's no data.

Lex Fridman (10:45.720)

And then these models generally

Oriol Vinyals (10:48.480)

don't then experience themselves.

Lex Fridman (10:51.160)

They just are observers, right?

Oriol Vinyals (10:52.520)

They're passive observers of the data.

Lex Fridman (10:54.600)

And then we're putting them to then generate data

Oriol Vinyals (10:57.440)

when we interact with them,

Lex Fridman (10:59.180)

but that's very limiting.

Oriol Vinyals (11:00.900)

The experience they actually experience

Lex Fridman (11:03.480)

when they could maybe be optimizing

Oriol Vinyals (11:05.680)

or further optimizing the weights,

Lex Fridman (11:07.440)

we're not even doing that.

Lex Fridman (11:08.640)

So to be clear, and again, mapping to AlphaGo, AlphaStar,

Lex Fridman (11:14.080)

we train the model.

Lex Fridman (11:15.280)

And when we deploy it to play against humans,

Lex Fridman (11:18.260)

or in this case interact with humans,

Oriol Vinyals (11:20.400)

like language models,

Lex Fridman (11:21.840)

they don't even keep training, right?

Oriol Vinyals (11:23.560)

They're not learning in the sense of the weights

Lex Fridman (11:26.220)

that you've learned from the data,

Oriol Vinyals (11:28.240)

they don't keep changing.

Lex Fridman (11:29.820)

Now, there's something a bit more feels magical,

Lex Fridman (11:33.540)

but it's understandable if you're into Neuronet,

Lex Fridman (11:36.240)

which is, well, they might not learn

Oriol Vinyals (11:39.180)

in the strict sense of the words,

Lex Fridman (11:40.520)

the weights changing,

Oriol Vinyals (11:41.520)

maybe that's mapping to how neurons interconnect

Lex Fridman (11:44.400)

and how we learn over our lifetime.

Lex Fridman (11:46.680)

But it's true that the context of the conversation

Lex Fridman (11:50.320)

that takes place when you talk to these systems,

Lex Fridman (11:55.020)

it's held in their working memory, right?

Lex Fridman (11:57.280)

It's almost like you start the computer,

Oriol Vinyals (12:00.160)

it has a hard drive that has a lot of information,

Lex Fridman (12:02.880)

you have access to the internet,

Oriol Vinyals (12:04.040)

which has probably all the information,

Lex Fridman (12:06.360)

but there's also a working memory

Oriol Vinyals (12:08.520)

where these agents, as we call them,

Lex Fridman (12:11.120)

or start calling them, build upon.

Oriol Vinyals (12:13.880)

Now, this memory is very limited.

Lex Fridman (12:16.640)

I mean, right now we're talking, to be concrete,

Oriol Vinyals (12:19.240)

about 2,000 words that we hold,

Lex Fridman (12:21.780)

and then beyond that, we start forgetting what we've seen.

Lex Fridman (12:24.880)

So you can see that there's some short term coherence

Lex Fridman (12:28.080)

already, right, with what you said.

Oriol Vinyals (12:29.880)

I mean, it's a very interesting topic.

Lex Fridman (12:32.340)

Having sort of a mapping, an agent to have consistency,

Oriol Vinyals (12:37.440)

then if you say, oh, what's your name,

Lex Fridman (12:40.800)

it could remember that,

Lex Fridman (12:42.280)

but then it might forget beyond 2,000 words,

Lex Fridman (12:45.020)

which is not that long of context

Oriol Vinyals (12:47.520)

if we think even of these podcast books are much longer.

Lex Fridman (12:51.800)

So technically speaking, there's a limitation there,

Oriol Vinyals (12:55.160)

super exciting from people that work on deep learning

Lex Fridman (12:58.220)

to be working on, but I would say we lack maybe benchmarks

Lex Fridman (13:03.080)

and the technology to have this lifetime like experience

Lex Fridman (13:07.880)

of memory that keeps building up.

Oriol Vinyals (13:10.900)

However, the way it learns offline

Lex Fridman (13:13.200)

is clearly very powerful, right?

Lex Fridman (13:14.920)

So you asked me three years ago, I would say,

Lex Fridman (13:17.840)

oh, we're very far.

Oriol Vinyals (13:18.680)

I think we've seen the power of this imitation,

Lex Fridman (13:22.240)

again, on the internet scale that has enabled this

Oriol Vinyals (13:26.280)

to feel like at least the knowledge,

Lex Fridman (13:28.800)

the basic knowledge about the world now

Oriol Vinyals (13:30.720)

is incorporated into the weights,

Lex Fridman (13:33.160)

but then this experience is lacking.

Lex Fridman (13:36.600)

And in fact, as I said, we don't even train them

Lex Fridman (13:39.360)

when we're talking to them,

Oriol Vinyals (13:41.200)

other than their working memory, of course, is affected.

Lex Fridman (13:44.800)

So that's the dynamic part,

Lex Fridman (13:46.600)

but they don't learn in the same way

Lex Fridman (13:48.300)

that you and I have learned, right?

Oriol Vinyals (13:50.640)

From basically when we were born and probably before.

Lex Fridman (13:54.080)

So lots of fascinating, interesting questions you asked there.

Oriol Vinyals (13:57.440)

I think the one I mentioned is this idea of memory

Lex Fridman (14:01.720)

and experience versus just kind of observe the world

Lex Fridman (14:05.540)

and learn its knowledge, which I think for that,

Lex Fridman (14:08.040)

I would argue lots of recent advancements

Oriol Vinyals (14:10.400)

that make me very excited about the field.

Lex Fridman (14:13.480)

And then the second maybe issue that I see is

Oriol Vinyals (14:18.240)

all these models, we train them from scratch.

Lex Fridman (14:21.320)

That's something I would have complained three years ago

Oriol Vinyals (14:24.100)

or six years ago or 10 years ago.

Lex Fridman (14:26.480)

And it feels if we take inspiration from how we got here,

Lex Fridman (14:31.440)

how the universe evolved us and we keep evolving,

Lex Fridman (14:35.340)

it feels that is a missing piece,

Oriol Vinyals (14:37.920)

that we should not be training models from scratch

Lex Fridman (14:41.400)

every few months,

Oriol Vinyals (14:42.560)

that there should be some sort of way

Lex Fridman (14:45.320)

in which we can grow models much like as a species

Lex Fridman (14:49.040)

and many other elements in the universe

Lex Fridman (14:51.560)

is building from the previous sort of iterations.

Lex Fridman (14:55.080)

And that from a just purely neural network perspective,

Lex Fridman (14:59.600)

even though we would like to make it work,

Oriol Vinyals (15:02.360)

it's proven very hard to not throw away

Lex Fridman (15:06.300)

the previous weights, right?

Oriol Vinyals (15:07.720)

This landscape we learn from the data

Lex Fridman (15:09.720)

and refresh it with a brand new set of weights,

Oriol Vinyals (15:13.400)

given maybe a recent snapshot of these data sets

Lex Fridman (15:17.020)

we train on, et cetera, or even a new game we're learning.

Lex Fridman (15:20.000)

So that feels like something is missing fundamentally.

Lex Fridman (15:24.200)

We might find it, but it's not very clear

Lex Fridman (15:27.480)

how it will look like.

Lex Fridman (15:28.460)

There's many ideas and it's super exciting as well.

Oriol Vinyals (15:30.860)

Yes, just for people who don't know,

Lex Fridman (15:32.480)

when you're approaching a new problem in machine learning,

Oriol Vinyals (15:35.760)

you're going to come up with an architecture

Lex Fridman (15:38.240)

that has a bunch of weights

Lex Fridman (15:41.000)

and then you initialize them somehow,

Lex Fridman (15:43.400)

which in most cases is some version of random.

Lex Fridman (15:47.320)

So that's what you mean by starting from scratch.

Lex Fridman (15:49.020)

And it seems like it's a waste every time you solve

Oriol Vinyals (15:54.480)

the game of Go and chess, StarCraft, protein folding,

Lex Fridman (15:59.720)

like surely there's some way to reuse the weights

Oriol Vinyals (16:03.200)

as we grow this giant database of neural networks

Lex Fridman (16:08.400)

that have solved some of the toughest problems in the world.

Lex Fridman (16:10.760)

And so some of that is, what is that?

Lex Fridman (16:15.240)

Methods, how to reuse weights,

Lex Fridman (16:19.080)

how to learn, extract what's generalizable

Lex Fridman (16:22.480)

or at least has a chance to be

Lex Fridman (16:25.160)

and throw away the other stuff.

Lex Fridman (16:27.840)

And maybe the neural network itself

Oriol Vinyals (16:29.580)

should be able to tell you that.

Lex Fridman (16:31.640)

Like what, yeah, how do you,

Lex Fridman (16:34.400)

what ideas do you have for better initialization of weights?

Lex Fridman (16:37.520)

Maybe stepping back,

Oriol Vinyals (16:38.760)

if we look at the field of machine learning,

Lex Fridman (16:41.720)

but especially deep learning, right?

Oriol Vinyals (16:44.040)

At the core of deep learning,

Lex Fridman (16:45.240)

there's this beautiful idea that is a single algorithm

Lex Fridman (16:49.240)

can solve any task, right?

Lex Fridman (16:50.920)

So it's been proven over and over

Oriol Vinyals (16:54.400)

with more increasing set of benchmarks

Lex Fridman (16:56.420)

and things that were thought impossible

Oriol Vinyals (16:58.580)

that are being cracked by this basic principle

Lex Fridman (17:01.960)

that is you take a neural network of uninitialized weights,

Lex Fridman (17:05.800)

so like a blank computational brain,

Lex Fridman (17:09.620)

then you give it, in the case of supervised learning,

Oriol Vinyals (17:12.580)

a lot ideally of examples of,

Lex Fridman (17:14.960)

hey, here is what the input looks like

Lex Fridman (17:17.120)

and the desired output should look like this.

Lex Fridman (17:19.560)

I mean, image classification is very clear example,

Oriol Vinyals (17:22.360)

images to maybe one of a thousand categories,

Lex Fridman (17:25.560)

that's what ImageNet is like,

Lex Fridman (17:26.840)

but many, many, if not all problems can be mapped this way.

Lex Fridman (17:30.720)

And then there's a generic recipe, right?

Oriol Vinyals (17:33.840)

That you can use.

Lex Fridman (17:35.240)

And this recipe with very little change,

Lex Fridman (17:38.600)

and I think that's the core of deep learning research, right?

Lex Fridman (17:41.520)

That what is the recipe that is universal?

Oriol Vinyals (17:44.420)

That for any new given task,

Lex Fridman (17:46.400)

I'll be able to use without thinking,

Oriol Vinyals (17:48.460)

without having to work very hard on the problem at stake.

Lex Fridman (17:52.600)

We have not found this recipe,

Lex Fridman (17:54.400)

but I think the field is excited to find less tweaks

Lex Fridman (18:00.160)

or tricks that people find when they work

Oriol Vinyals (18:02.640)

on important problems specific to those

Lex Fridman (18:05.280)

and more of a general algorithm, right?

Lex Fridman (18:07.540)

So at an algorithmic level,

Lex Fridman (18:09.300)

I would say we have something general already,

Oriol Vinyals (18:11.780)

which is this formula of training a very powerful model,

Lex Fridman (18:14.520)

a neural network on a lot of data.

Lex Fridman (18:17.000)

And in many cases, you need some specificity

Lex Fridman (18:21.200)

to the actual problem you're solving,

Oriol Vinyals (18:23.400)

protein folding being such an important problem

Lex Fridman (18:26.080)

has some basic recipe that is learned from before, right?

Oriol Vinyals (18:30.800)

Like transformer models, graph neural networks,

Lex Fridman (18:34.140)

ideas coming from NLP, like something called BERT,

Oriol Vinyals (18:38.600)

that is a kind of loss that you can emplace

Lex Fridman (18:41.280)

to help the knowledge distillation is another technique,

Lex Fridman (18:45.460)

right?

Lex Fridman (18:46.300)

So this is the formula.

Oriol Vinyals (18:47.120)

We still had to find some particular things

Lex Fridman (18:50.560)

that were specific to alpha fold, right?

Oriol Vinyals (18:53.600)

That's very important because protein folding

Lex Fridman (18:55.860)

is such a high value problem that as humans,

Oriol Vinyals (18:59.120)

we should solve it no matter

Lex Fridman (19:00.840)

if we need to be a bit specific.

Lex Fridman (19:02.880)

And it's possible that some of these learnings

Lex Fridman (19:04.940)

will apply then to the next iteration of this recipe

Oriol Vinyals (19:07.380)

that deep learners are about.

Lex Fridman (19:09.340)

But it is true that so far, the recipe is what's common,

Lex Fridman (19:13.200)

but the weights you generally throw away,

Lex Fridman (19:15.880)

which feels very sad.

Oriol Vinyals (19:17.800)

Although, maybe in the last,

Lex Fridman (19:20.440)

especially in the last two, three years,

Lex Fridman (19:22.280)

and when we last spoke,

Lex Fridman (19:23.560)

I mentioned this area of meta learning,

Oriol Vinyals (19:25.560)

which is the idea of learning to learn.

Lex Fridman (19:28.560)

That idea and some progress has been had starting,

Oriol Vinyals (19:32.040)

I would say, mostly from GPT3 on the language domain only,

Lex Fridman (19:36.040)

in which you could conceive a model that is trained once.

Lex Fridman (19:41.040)

And then this model is not narrow in that it only knows

Lex Fridman (19:44.720)

how to translate a pair of languages or even a set of

Oriol Vinyals (19:47.760)

or it only knows how to assign sentiment to a sentence.

Lex Fridman (19:51.520)

These actually, you could teach it by a prompting,

Oriol Vinyals (19:55.040)

it's called, and this prompting is essentially

Lex Fridman (19:56.880)

just showing it a few more examples,

Oriol Vinyals (19:59.920)

almost like you do show examples, input, output examples,

Lex Fridman (1:00:03.720)

And you show it a few examples, and now it does that.

Lex Fridman (1:00:06.880)

So it went way beyond the image net categorization of images

Lex Fridman (1:00:12.800)

that we were a bit stuck maybe before this revelation

Oriol Vinyals (1:00:17.280)

moment that happened in 2000.

Lex Fridman (1:00:19.200)

I believe it was 19, but it was after we checked.

Oriol Vinyals (1:00:21.960)

In that way, it has solved meta learning

Lex Fridman (1:00:24.400)

as was previously defined.

Oriol Vinyals (1:00:26.160)

Yes, it expanded what it meant.

Lex Fridman (1:00:27.880)

So that's what you say, what does it mean?

Lex Fridman (1:00:29.680)

So it's an evolving term.

Lex Fridman (1:00:31.400)

But here is maybe now looking forward,

Oriol Vinyals (1:00:35.320)

looking at what's happening, obviously,

Lex Fridman (1:00:38.080)

in the community with more modalities, what we can expect.

Lex Fridman (1:00:42.560)

And I would certainly hope to see the following.

Lex Fridman (1:00:45.040)

And this is a pretty drastic hope.

Lex Fridman (1:00:48.400)

But in five years, maybe we chat again.

Lex Fridman (1:00:51.200)

And we have a system, a set of weights

Oriol Vinyals (1:00:55.920)

that we can teach it to play StarCraft.

Lex Fridman (1:00:59.840)

Maybe not at the level of AlphaStar,

Lex Fridman (1:01:01.480)

but play StarCraft, a complex game,

Lex Fridman (1:01:03.720)

we teach it through interactions to prompting.

Oriol Vinyals (1:01:06.920)

You can certainly prompt a system.

Lex Fridman (1:01:08.600)

That's what Gata shows to play some simple Atari games.

Lex Fridman (1:01:11.880)

So imagine if you start talking to a system,

Lex Fridman (1:01:15.360)

teaching it a new game, showing it

Oriol Vinyals (1:01:17.280)

examples of in this particular game,

Lex Fridman (1:01:20.960)

this user did something good.

Oriol Vinyals (1:01:22.720)

Maybe the system can even play and ask you questions.

Lex Fridman (1:01:25.440)

Say, hey, I played this game.

Oriol Vinyals (1:01:27.000)

I just played this game.

Lex Fridman (1:01:28.040)

Did I do well?

Lex Fridman (1:01:29.040)

Can you teach me more?

Lex Fridman (1:01:30.440)

So five, maybe to 10 years, these capabilities,

Oriol Vinyals (1:01:34.720)

or what meta learning means, will

Lex Fridman (1:01:36.400)

be much more interactive, much more rich,

Lex Fridman (1:01:38.800)

and through domains that we were specializing.

Lex Fridman (1:01:41.640)

So you see the difference.

Oriol Vinyals (1:01:42.920)

We built AlphaStar Specialized to play StarCraft.

Lex Fridman (1:01:47.000)

The algorithms were general, but the weights were specialized.

Lex Fridman (1:01:50.400)

And what we're hoping is that we can teach a network

Lex Fridman (1:01:54.160)

to play games, to play any game, just using games as an example,

Oriol Vinyals (1:01:58.560)

through interacting with it, teaching it,

Lex Fridman (1:02:01.440)

uploading the Wikipedia page of StarCraft.

Oriol Vinyals (1:02:04.000)

This is in the horizon.

Lex Fridman (1:02:06.120)

And obviously, there are details that need to be filled

Lex Fridman (1:02:09.360)

and research needs to be done.

Lex Fridman (1:02:11.000)

But that's how I see meta learning above,

Oriol Vinyals (1:02:13.200)

which is going to be beyond prompting.

Lex Fridman (1:02:15.360)

It's going to be a bit more interactive.

Oriol Vinyals (1:02:18.080)

The system might tell us to give it feedback

Lex Fridman (1:02:20.720)

after it maybe makes mistakes or it loses a game.

Lex Fridman (1:02:24.080)

But it's nonetheless very exciting

Lex Fridman (1:02:26.240)

because if you think about this this way,

Oriol Vinyals (1:02:28.960)

the benchmarks are already there.

Lex Fridman (1:02:30.600)

We just repurposed the benchmarks.

Lex Fridman (1:02:33.120)

So in a way, I like to map the space of what

Lex Fridman (1:02:38.440)

maybe AGI means to say, OK, we went 101% performance in Go,

Oriol Vinyals (1:02:45.480)

in Chess, in StarCraft.

Lex Fridman (1:02:47.920)

The next iteration might be 20% performance

Oriol Vinyals (1:02:51.920)

across, quote unquote, all tasks.

Lex Fridman (1:02:54.720)

And even if it's not as good, it's fine.

Oriol Vinyals (1:02:57.720)

We have ways to also measure progress

Lex Fridman (1:02:59.960)

because we have those specialized agents and so on.

Lex Fridman (1:03:04.320)

So this is, to me, very exciting.

Lex Fridman (1:03:06.160)

And these next iteration models are definitely

Oriol Vinyals (1:03:10.080)

hinting at that direction of progress,

Lex Fridman (1:03:13.360)

which hopefully we can have.

Oriol Vinyals (1:03:14.720)

There are obviously some things that

Lex Fridman (1:03:16.440)

could go wrong in terms of we might not have the tools.

Oriol Vinyals (1:03:20.120)

Maybe transformers are not enough.

Lex Fridman (1:03:22.600)

There are some breakthroughs to come, which

Oriol Vinyals (1:03:24.920)

makes the field more exciting to people like me as well,

Lex Fridman (1:03:27.560)

of course.

Lex Fridman (1:03:28.600)

But that's, if you ask me, five to 10 years,

Lex Fridman (1:03:32.040)

you might see these models that start

Oriol Vinyals (1:03:33.800)

to look more like weights that are already trained.

Lex Fridman (1:03:36.880)

And then it's more about teaching or make

Oriol Vinyals (1:03:40.520)

their meta learn what you're trying

Lex Fridman (1:03:44.040)

to induce in terms of tasks and so on,

Oriol Vinyals (1:03:47.000)

well beyond the simple now tasks we're

Lex Fridman (1:03:49.920)

starting to see emerge like small arithmetic tasks

Lex Fridman (1:03:53.200)

and so on.

Lex Fridman (1:03:54.280)

So a few questions around that.

Oriol Vinyals (1:03:55.720)

This is fascinating.

Lex Fridman (1:03:57.200)

So that kind of teaching, interactive,

Lex Fridman (1:04:01.440)

so it's beyond prompting.

Lex Fridman (1:04:02.760)

So it's interacting with the neural network.

Oriol Vinyals (1:04:05.240)

That's different than the training process.

Lex Fridman (1:04:08.520)

So it's different than the optimization

Oriol Vinyals (1:04:12.440)

over differentiable functions.

Lex Fridman (1:04:15.920)

This is already trained.

Lex Fridman (1:04:17.240)

And now you're teaching, I mean, it's

Lex Fridman (1:04:21.800)

almost akin to the brain, the neurons already

Oriol Vinyals (1:04:25.560)

set with their connections.

Lex Fridman (1:04:26.960)

On top of that, you're now using that infrastructure

Oriol Vinyals (1:04:30.000)

to build up further knowledge.

Lex Fridman (1:04:33.640)

So that's a really interesting distinction that's actually

Oriol Vinyals (1:04:37.200)

not obvious from a software engineering perspective,

Lex Fridman (1:04:40.320)

that there's a line to be drawn.

Oriol Vinyals (1:04:42.560)

Because you always think for a neural network to learn,

Lex Fridman (1:04:44.880)

it has to be retrained, trained and retrained.

Lex Fridman (1:04:49.880)

And prompting is a way of teaching.

Lex Fridman (1:04:54.000)

And you'll now work a little bit of context

Oriol Vinyals (1:04:55.920)

about whatever the heck you're trying it to do.

Lex Fridman (1:04:57.960)

So you can maybe expand this prompting capability

Oriol Vinyals (1:05:00.440)

by making it interact.

Lex Fridman (1:05:03.320)

That's really, really interesting.

Oriol Vinyals (1:05:04.680)

By the way, this is not, if you look at way back

Lex Fridman (1:05:08.080)

at different ways to tackle even classification tasks.

Lex Fridman (1:05:11.840)

So this comes from longstanding literature

Lex Fridman (1:05:16.440)

in machine learning.

Lex Fridman (1:05:18.240)

What I'm suggesting could sound to some

Lex Fridman (1:05:20.800)

like a bit like nearest neighbor.

Lex Fridman (1:05:23.400)

So nearest neighbor is almost the simplest algorithm

Lex Fridman (1:05:27.120)

that does not require learning.

Lex Fridman (1:05:30.200)

So it has this interesting, you don't

Lex Fridman (1:05:32.640)

need to compute gradients.

Lex Fridman (1:05:34.320)

And what nearest neighbor does is you, quote unquote,

Lex Fridman (1:05:37.560)

have a data set or upload a data set.

Lex Fridman (1:05:39.960)

And then all you need to do is a way

Lex Fridman (1:05:42.040)

to measure distance between points.

Lex Fridman (1:05:44.720)

And then to classify a new point,

Lex Fridman (1:05:46.680)

you're just simply computing, what's

Lex Fridman (1:05:48.360)

the closest point in this massive amount of data?

Lex Fridman (1:05:51.240)

And that's my answer.

Lex Fridman (1:05:52.720)

So you can think of prompting in a way

Lex Fridman (1:05:55.440)

as you're uploading not just simple points.

Lex Fridman (1:05:58.680)

And the metric is not the distance between the images

Lex Fridman (1:06:02.480)

or something simple.

Oriol Vinyals (1:06:03.320)

It's something that you compute that's much more advanced.

Lex Fridman (1:06:06.040)

But in a way, it's very similar.

Oriol Vinyals (1:06:09.040)

You simply are uploading some knowledge

Lex Fridman (1:06:12.600)

to this pre trained system in nearest neighbor.

Oriol Vinyals (1:06:15.040)

Maybe the metric is learned or not,

Lex Fridman (1:06:17.280)

but you don't need to further train it.

Lex Fridman (1:06:19.400)

And then now you immediately get a classifier out of this.

Lex Fridman (1:06:23.680)

Now it's just an evolution of that concept,

Oriol Vinyals (1:06:25.840)

very classical concept in machine learning, which

Lex Fridman (1:06:28.080)

is just learning through what's the closest point, closest

Oriol Vinyals (1:06:32.640)

by some distance, and that's it.

Lex Fridman (1:06:34.720)

It's an evolution of that.

Lex Fridman (1:06:36.120)

And I will say how I saw meta learning when

Lex Fridman (1:06:39.400)

we worked on a few ideas in 2016 was precisely

Oriol Vinyals (1:06:44.760)

through the lens of nearest neighbor, which

Lex Fridman (1:06:47.520)

is very common in computer vision community.

Oriol Vinyals (1:06:50.160)

There's a very active area of research

Lex Fridman (1:06:52.160)

about how do you compute the distance between two images.

Lex Fridman (1:06:55.600)

But if you have a good distance metric,

Lex Fridman (1:06:57.560)

you also have a good classifier.

Oriol Vinyals (1:06:59.920)

All I'm saying is now these distances and the points

Lex Fridman (1:07:02.680)

are not just images.

Oriol Vinyals (1:07:03.800)

They're like words or sequences of words and images

Lex Fridman (1:07:08.560)

and actions that teach you something new.

Lex Fridman (1:07:10.400)

But it might be that technique wise those come back.

Lex Fridman (1:07:14.680)

And I will say that it's not necessarily true

Oriol Vinyals (1:07:18.240)

that you might not ever train the weights a bit further.

Lex Fridman (1:07:21.760)

Some aspect of meta learning, some techniques

Oriol Vinyals (1:07:24.800)

in meta learning do actually do a bit of fine tuning

Lex Fridman (1:07:28.280)

as it's called.

Oriol Vinyals (1:07:29.080)

They train the weights a little bit when they get a new task.

Lex Fridman (1:07:32.960)

So as I call the how or how we're going to achieve this,

Oriol Vinyals (1:07:37.960)

as a deep learner, I'm very skeptic.

Lex Fridman (1:07:39.840)

We're going to try a few things, whether it's

Oriol Vinyals (1:07:41.840)

a bit of training, adding a few parameters,

Lex Fridman (1:07:44.200)

thinking of these as nearest neighbor,

Oriol Vinyals (1:07:45.960)

or just simply thinking of there's a sequence of words,

Lex Fridman (1:07:49.200)

it's a prefix.

Lex Fridman (1:07:50.440)

And that's the new classifier.

Lex Fridman (1:07:53.000)

We'll see.

Oriol Vinyals (1:07:53.680)

There's the beauty of research.

Lex Fridman (1:07:55.480)

But what's important is that is a good goal in itself

Oriol Vinyals (1:08:00.160)

that I see as very worthwhile pursuing for the next stages

Lex Fridman (1:08:03.800)

of not only meta learning.

Oriol Vinyals (1:08:05.720)

I think this is basically what's exciting about machine learning

Lex Fridman (1:08:10.160)

period to me.

Oriol Vinyals (1:08:11.400)

Well, and the interactive aspect of that

Lex Fridman (1:08:13.760)

is also very interesting, the interactive version

Oriol Vinyals (1:08:16.400)

of nearest neighbor to help you pull out the classifier

Lex Fridman (1:08:22.160)

from this giant thing.

Oriol Vinyals (1:08:23.760)

OK, is this the way we can go in 5, 10 plus years

Lex Fridman (1:08:31.040)

from any task, sorry, from many tasks to any task?

Lex Fridman (1:08:38.200)

And what does that mean?

Lex Fridman (1:08:39.400)

What does it need to be actually trained on?

Lex Fridman (1:08:42.760)

Which point is the network had enough?

Lex Fridman (1:08:45.400)

So what does a network need to learn about this world

Lex Fridman (1:08:50.360)

in order to be able to perform any task?

Lex Fridman (1:08:52.440)

Is it just as simple as language, image, and action?

Lex Fridman (1:08:57.880)

Or do you need some set of representative images?

Lex Fridman (1:09:02.680)

Like if you only see land images,

Lex Fridman (1:09:05.200)

will you know anything about underwater?

Lex Fridman (1:09:06.760)

Is that some fundamentally different?

Oriol Vinyals (1:09:08.720)

I don't know.

Lex Fridman (1:09:09.800)

I mean, those are open questions, I would say.

Oriol Vinyals (1:09:12.080)

I mean, the way you put, let me maybe further your example.

Lex Fridman (1:09:15.280)

If all you see is land images but you're

Oriol Vinyals (1:09:18.920)

reading all about land and water worlds

Lex Fridman (1:09:21.560)

but in books, imagine, would that be enough?

Oriol Vinyals (1:09:25.360)

Good question.

Lex Fridman (1:09:26.400)

We don't know.

Lex Fridman (1:09:27.120)

But I guess maybe you can join us

Lex Fridman (1:09:30.440)

if you want in our quest to find this.

Oriol Vinyals (1:09:32.120)

That's precisely.

Lex Fridman (1:09:33.440)

Water world, yeah.

Oriol Vinyals (1:09:34.360)

Yes, that's precisely, I mean, the beauty of research.

Lex Fridman (1:09:37.640)

And that's the research business we're in,

Oriol Vinyals (1:09:42.680)

I guess, is to figure this out and ask the right questions

Lex Fridman (1:09:46.160)

and then iterate with the whole community,

Oriol Vinyals (1:09:49.520)

publishing findings and so on.

Lex Fridman (1:09:52.640)

But yeah, this is a question.

Oriol Vinyals (1:09:55.160)

It's not the only question, but it's certainly, as you ask,

Lex Fridman (1:09:58.640)

on my mind constantly.

Lex Fridman (1:10:00.080)

And so we'll need to wait for maybe the, let's say, five

Lex Fridman (1:10:03.920)

years, let's hope it's not 10, to see what are the answers.

Oriol Vinyals (1:10:09.400)

Some people will largely believe in unsupervised or

Lex Fridman (1:10:12.800)

self supervised learning of single modalities

Lex Fridman (1:10:15.640)

and then crossing them.

Lex Fridman (1:10:18.000)

Some people might think end to end learning is the answer.

Oriol Vinyals (1:10:21.640)

Modularity is maybe the answer.

Lex Fridman (1:10:23.760)

So we don't know, but we're just definitely excited

Oriol Vinyals (1:10:27.040)

to find out.

Lex Fridman (1:10:27.560)

But it feels like this is the right time

Lex Fridman (1:10:29.280)

and we're at the beginning of this journey.

Lex Fridman (1:10:31.680)

We're finally ready to do these kind of general big models

Lex Fridman (1:10:36.040)

and agents.

Lex Fridman (1:10:37.960)

What do you sort of specific technical thing

Oriol Vinyals (1:10:42.480)

about Gato, Flamingo, Chinchilla, Gopher, any of these

Lex Fridman (1:10:48.040)

that is especially beautiful, that was surprising, maybe?

Lex Fridman (1:10:51.640)

Is there something that just jumps out at you?

Lex Fridman (1:10:55.200)

Of course, there's the general thing of like,

Oriol Vinyals (1:10:57.600)

you didn't think it was possible and then you

Lex Fridman (1:11:00.600)

realize it's possible in terms of the generalizability

Oriol Vinyals (1:11:03.560)

across modalities and all that kind of stuff.

Lex Fridman (1:11:05.640)

Or maybe how small of a network, relatively speaking,

Oriol Vinyals (1:11:08.920)

Gato is, all that kind of stuff.

Lex Fridman (1:11:10.440)

But is there some weird little things that were surprising?

Oriol Vinyals (1:11:15.160)

Look, I'll give you an answer that's very important

Lex Fridman (1:11:18.200)

because maybe people don't quite realize this,

Lex Fridman (1:11:22.560)

but the teams behind these efforts, the actual humans,

Lex Fridman (1:11:27.200)

that's maybe the surprising in an obviously positive way.

Lex Fridman (1:11:31.640)

So anytime you see these breakthroughs,

Lex Fridman (1:11:34.560)

I mean, it's easy to map it to a few people.

Oriol Vinyals (1:11:37.080)

There's people that are great at explaining things and so on.

Lex Fridman (1:11:39.680)

And that's very nice.

Lex Fridman (1:11:40.720)

But maybe the learnings or the method learnings

Lex Fridman (1:11:44.720)

that I get as a human about this is, sure, we can move forward.

Lex Fridman (1:11:50.480)

But the surprising bit is how important

Lex Fridman (1:11:55.640)

are all the pieces of these projects,

Lex Fridman (1:11:58.720)

how do they come together?

Lex Fridman (1:12:00.080)

So I'll give you maybe some of the ingredients of success

Oriol Vinyals (1:12:04.440)

that are common across these, but not the obvious ones

Lex Fridman (1:12:07.680)

on machine learning.

Oriol Vinyals (1:12:08.480)

I can always also give you those.

Lex Fridman (1:12:11.320)

But basically, there is engineering is critical.

Lex Fridman (1:12:17.280)

So very good engineering because ultimately we're

Lex Fridman (1:12:21.120)

collecting data sets, right?

Lex Fridman (1:12:23.720)

So the engineering of data and then

Lex Fridman (1:12:26.640)

of deploying the models at scale into some compute cluster

Oriol Vinyals (1:12:31.160)

that cannot go understated, that is a huge factor of success.

Lex Fridman (1:12:36.800)

And it's hard to believe that details matter so much.

Oriol Vinyals (1:12:41.560)

We would like to believe that it's

Lex Fridman (1:12:43.760)

true that there is more and more of a standard formula,

Oriol Vinyals (1:12:47.360)

as I was saying, like this recipe that

Lex Fridman (1:12:49.360)

works for everything.

Lex Fridman (1:12:50.520)

But then when you zoom into each of these projects,

Lex Fridman (1:12:53.680)

then you realize the devil is indeed in the details.

Lex Fridman (1:12:57.760)

And then the teams have to work together towards these goals.

Lex Fridman (1:13:03.040)

So engineering of data and obviously clusters

Lex Fridman (1:13:07.520)

and large scale is very important.

Lex Fridman (1:13:09.280)

And then one that is often not, maybe nowadays it is more clear

Lex Fridman (1:13:15.080)

is benchmark progress, right?

Lex Fridman (1:13:17.160)

So we're talking here about multiple months of tens

Oriol Vinyals (1:13:20.840)

of researchers and people that are

Lex Fridman (1:13:24.520)

trying to organize the research and so on working together.

Lex Fridman (1:13:28.080)

And you don't know that you can get there.

Lex Fridman (1:13:32.120)

I mean, this is the beauty.

Oriol Vinyals (1:13:34.520)

If you're not risking to trying to do something

Lex Fridman (1:13:37.320)

that feels impossible, you're not going to get there.

Lex Fridman (1:13:41.600)

But you need a way to measure progress.

Lex Fridman (1:13:43.920)

So the benchmarks that you build are critical.

Oriol Vinyals (1:13:47.680)

I've seen this beautifully play out in many projects.

Lex Fridman (1:13:50.520)

I mean, maybe the one I've seen it more consistently,

Oriol Vinyals (1:13:53.840)

which means we establish the metric,

Lex Fridman (1:13:56.760)

actually the community did.

Lex Fridman (1:13:58.240)

And then we leverage that massively is alpha fold.

Lex Fridman (1:14:01.520)

This is a project where the data, the metrics

Oriol Vinyals (1:14:05.160)

were all there.

Lex Fridman (1:14:06.040)

And all it took was, and it's easier said than done,

Oriol Vinyals (1:14:09.080)

an amazing team working not to try

Lex Fridman (1:14:12.840)

to find some incremental improvement

Lex Fridman (1:14:14.720)

and publish, which is one way to do research that is valid,

Lex Fridman (1:14:17.920)

but aim very high and work literally for years

Oriol Vinyals (1:14:22.440)

to iterate over that process.

Lex Fridman (1:14:24.080)

And working for years with the team,

Oriol Vinyals (1:14:25.640)

I mean, it is tricky that also happened to happen partly

Lex Fridman (1:14:30.120)

during a pandemic and so on.

Lex Fridman (1:14:32.280)

So I think my meta learning from all this

Lex Fridman (1:14:34.200)

is the teams are critical to the success.

Lex Fridman (1:14:37.960)

And then if now going to the machine learning,

Lex Fridman (1:14:40.200)

the part that's surprising is so we like architectures

Oriol Vinyals (1:14:46.880)

like neural networks.

Lex Fridman (1:14:48.680)

And I would say this was a very rapidly evolving field

Oriol Vinyals (1:14:53.040)

until the transformer came.

Lex Fridman (1:14:54.920)

So attention might indeed be all you need,

Oriol Vinyals (1:14:58.280)

which is the title, also a good title,

Lex Fridman (1:15:00.280)

although in hindsight is good.

Oriol Vinyals (1:15:02.040)

I don't think at the time I thought

Lex Fridman (1:15:03.440)

this is a great title for a paper.

Lex Fridman (1:15:05.040)

But that architecture is proving that the dream of modeling

Lex Fridman (1:15:10.960)

sequences of any bytes, there is something there that will stick.

Lex Fridman (1:15:15.320)

And I think these advance in architectures

Lex Fridman (1:15:18.280)

in how neural networks are architecture

Oriol Vinyals (1:15:21.000)

to do what they do.

Lex Fridman (1:15:23.080)

It's been hard to find one that has been so stable

Lex Fridman (1:15:26.080)

and relatively has changed very little

Lex Fridman (1:15:28.880)

since it was invented five or so years ago.

Lex Fridman (1:15:33.080)

So that is a surprising, is a surprise

Lex Fridman (1:15:35.840)

that keeps recurring into other projects.

Oriol Vinyals (1:15:38.280)

Try to, on a philosophical or technical level, introspect,

Lex Fridman (1:15:43.320)

what is the magic of attention?

Lex Fridman (1:15:45.440)

What is attention?

Lex Fridman (1:15:47.280)

That's attention in people that study cognition,

Lex Fridman (1:15:50.120)

so human attention.

Lex Fridman (1:15:52.040)

I think there's giant wars over what attention means,

Lex Fridman (1:15:55.760)

how it works in the human mind.

Lex Fridman (1:15:57.440)

So there's very simple looks at what

Oriol Vinyals (1:16:00.480)

attention is in a neural network from the days of attention

Lex Fridman (1:16:03.840)

is all you need.

Lex Fridman (1:16:04.440)

But do you think there's a general principle that's

Lex Fridman (1:16:07.520)

really powerful here?

Oriol Vinyals (1:16:08.760)

Yeah, so a distinction between transformers and LSTMs,

Lex Fridman (1:16:13.400)

which were what came before.

Lex Fridman (1:16:15.360)

And there was a transitional period

Lex Fridman (1:16:17.880)

where you could use both.

Oriol Vinyals (1:16:19.720)

In fact, when we talked about AlphaStar,

Lex Fridman (1:16:22.000)

we used transformers and LSTMs.

Lex Fridman (1:16:24.240)

So it was still the beginning of transformers.

Lex Fridman (1:16:26.400)

They were very powerful.

Lex Fridman (1:16:27.400)

But LSTMs were also very powerful sequence models.

Lex Fridman (1:16:31.480)

So the power of the transformer is

Oriol Vinyals (1:16:35.400)

that it has built in what we call

Lex Fridman (1:16:38.440)

an inductive bias of attention that makes the model.

Oriol Vinyals (1:16:43.040)

When you think of a sequence of integers,

Lex Fridman (1:16:45.720)

like we discussed this before, this is a sequence of words.

Oriol Vinyals (1:16:50.400)

When you have to do very hard tasks over these words,

Lex Fridman (1:16:54.800)

this could be we're going to translate a whole paragraph

Oriol Vinyals (1:16:57.840)

or we're going to predict the next paragraph given

Lex Fridman (1:17:00.320)

10 paragraphs before.

Oriol Vinyals (1:17:04.280)

There's some loose intuition from how we do it as a human

Lex Fridman (1:17:10.360)

that is very nicely mimicked and replicated structurally

Oriol Vinyals (1:17:15.400)

speaking in the transformer, which

Lex Fridman (1:17:16.840)

is this idea of you're looking for something.

Lex Fridman (1:17:21.160)

So you're sort of when you just read a piece of text,

Lex Fridman (1:17:25.760)

now you're thinking what comes next.

Oriol Vinyals (1:17:27.880)

You might want to relook at the text or look it from scratch.

Lex Fridman (1:17:31.800)

I mean, literally is because there's no recurrence.

Oriol Vinyals (1:17:35.040)

You're just thinking what comes next.

Lex Fridman (1:17:37.240)

And it's almost hypothesis driven.

Lex Fridman (1:17:40.040)

So if I'm thinking the next word that I write is cat or dog,

Lex Fridman (1:17:46.600)

the way the transformer works almost philosophically

Oriol Vinyals (1:17:49.880)

is it has these two hypotheses.

Lex Fridman (1:17:52.840)

Is it going to be cat or is it going to be dog?

Lex Fridman (1:17:55.640)

And then it says, OK, if it's cat,

Lex Fridman (1:17:58.360)

I'm going to look for certain words.

Oriol Vinyals (1:17:59.920)

Not necessarily cat, although cat is an obvious word

Lex Fridman (1:18:01.920)

you would look in the past to see

Oriol Vinyals (1:18:03.520)

whether it makes more sense to output cat or dog.

Lex Fridman (1:18:05.960)

And then it does some very deep computation

Oriol Vinyals (1:18:09.480)

over the words and beyond.

Lex Fridman (1:18:11.480)

So it combines the words, but it has the query

Oriol Vinyals (1:18:16.200)

as we call it that is cat.

Lex Fridman (1:18:18.440)

And then similarly for dog.

Lex Fridman (1:18:20.680)

And so it's a very computational way to think about, look,

Lex Fridman (1:18:24.760)

if I'm thinking deeply about text,

Oriol Vinyals (1:18:27.000)

I need to go back to look at all of the text, attend over it.

Lex Fridman (1:18:30.600)

But it's not just attention.

Lex Fridman (1:18:32.200)

What is guiding the attention?

Lex Fridman (1:18:34.000)

And that was the key insight from an earlier paper

Lex Fridman (1:18:36.680)

is not how far away is it?

Lex Fridman (1:18:39.120)

I mean, how far away is it is important?

Lex Fridman (1:18:40.840)

What did I just write about?

Lex Fridman (1:18:42.720)

That's critical.

Lex Fridman (1:18:44.120)

But what you wrote about 10 pages ago

Lex Fridman (1:18:46.760)

might also be critical.

Lex Fridman (1:18:48.480)

So you're looking not positionally, but content wise.

Lex Fridman (1:18:53.160)

And transformers have this beautiful way

Oriol Vinyals (1:18:56.120)

to query for certain content and pull it out

Lex Fridman (1:18:59.480)

in a compressed way.

Lex Fridman (1:19:00.440)

So then you can make a more informed decision.

Lex Fridman (1:19:02.960)

I mean, that's one way to explain transformers.

Lex Fridman (1:19:05.920)

But I think it's a very powerful inductive bias.

Lex Fridman (1:19:10.000)

There might be some details that might change over time,

Lex Fridman (1:19:12.480)

but I think that is what makes transformers so much more

Lex Fridman (1:19:17.360)

powerful than the recurrent networks that

Oriol Vinyals (1:19:20.080)

were more recency bias based, which obviously works

Lex Fridman (1:19:23.600)

in some tasks, but it has major flaws.

Oriol Vinyals (1:19:26.720)

Transformer itself has flaws.

Lex Fridman (1:19:29.240)

And I think the main one, the main challenge

Oriol Vinyals (1:19:31.680)

is these prompts that we just were talking about,

Lex Fridman (1:19:35.760)

they can be 1,000 words long.

Lex Fridman (1:19:38.040)

But if I'm teaching you StarGraph,

Lex Fridman (1:19:40.440)

I'll have to show you videos.

Oriol Vinyals (1:19:41.880)

I'll have to point you to whole Wikipedia articles

Lex Fridman (1:19:44.600)

about the game.

Oriol Vinyals (1:19:46.120)

We'll have to interact probably as you play.

Lex Fridman (1:19:48.040)

You'll ask me questions.

Oriol Vinyals (1:19:49.480)

The context required for us to achieve

Lex Fridman (1:19:52.320)

me being a good teacher to you on the game

Oriol Vinyals (1:19:54.720)

as you would want to do it with a model, I think

Lex Fridman (1:19:58.920)

goes well beyond the current capabilities.

Lex Fridman (1:20:01.720)

So the question is, how do we benchmark this?

Lex Fridman (1:20:03.920)

And then how do we change the structure of the architectures?

Oriol Vinyals (1:20:07.320)

I think there's ideas on both sides,

Lex Fridman (1:20:08.840)

but we'll have to see empirically, obviously,

Lex Fridman (1:20:11.800)

what ends up working.

Lex Fridman (1:20:13.320)

And as you talked about, some of the ideas

Oriol Vinyals (1:20:15.280)

could be keeping the constraint of that length in place,

Lex Fridman (1:20:19.440)

but then forming hierarchical representations

Oriol Vinyals (1:20:23.000)

to where you can start being much clever in how

Lex Fridman (1:20:26.600)

you use those 1,000 tokens.

Oriol Vinyals (1:20:28.800)

Indeed.

Lex Fridman (1:20:30.920)

Yeah, that's really interesting.

Lex Fridman (1:20:32.240)

But it also is possible that this attentional mechanism

Lex Fridman (1:20:34.840)

where you basically, you don't have a recency bias,

Lex Fridman (1:20:37.560)

but you look more generally, you make it learnable.

Lex Fridman (1:20:42.000)

The mechanism in which way you look back into the past,

Oriol Vinyals (1:20:45.240)

you make that learnable.

Lex Fridman (1:20:46.760)

It's also possible we're at the very beginning of that

Oriol Vinyals (1:20:50.160)

because that, you might become smarter and smarter

Lex Fridman (1:20:54.400)

in the way you query the past.

Lex Fridman (1:20:58.400)

So recent past and distant past and maybe very, very distant

Lex Fridman (1:21:01.800)

past.

Lex Fridman (1:21:02.320)

So almost like the attention mechanism

Lex Fridman (1:21:04.960)

will have to improve and evolve as good as the tokenization

Oriol Vinyals (1:21:11.280)

mechanism so you can represent long term memory somehow.

Lex Fridman (1:21:14.960)

Yes.

Lex Fridman (1:21:16.080)

And I mean, hierarchies are very,

Lex Fridman (1:21:18.240)

I mean, it's a very nice word that sounds appealing.

Oriol Vinyals (1:21:22.160)

There's lots of work adding hierarchy to the memories.

Lex Fridman (1:21:25.920)

In practice, it does seem like we keep coming back

Oriol Vinyals (1:21:29.480)

to the main formula or main architecture.

Lex Fridman (1:21:33.880)

That sometimes tells us something.

Oriol Vinyals (1:21:35.320)

There is such a sentence that a friend of mine told me,

Lex Fridman (1:21:38.560)

like, whether it wants to work or not.

Lex Fridman (1:21:41.000)

So Transformer was clearly an idea that wanted to work.

Lex Fridman (1:21:44.920)

And then I think there's some principles

Oriol Vinyals (1:21:47.520)

we believe will be needed.

Lex Fridman (1:21:49.080)

But finding the exact details, details matter so much.

Oriol Vinyals (1:21:52.880)

That's going to be tricky.

Lex Fridman (1:21:54.200)

I love the idea that there's like you as a human being,

Oriol Vinyals (1:21:59.440)

you want some ideas to work.

Lex Fridman (1:22:01.280)

And then there's the model that wants some ideas

Oriol Vinyals (1:22:03.800)

to work and you get to have a conversation

Lex Fridman (1:22:05.960)

to see which more likely the model will win in the end.

Oriol Vinyals (1:22:10.520)

Because it's the one, you don't have to do any work.

Lex Fridman (1:22:12.800)

The model is the one that has to do the work.

Lex Fridman (1:22:14.360)

So you should listen to the model.

Lex Fridman (1:22:15.840)

And I really love this idea that you

Oriol Vinyals (1:22:17.440)

talked about the humans in this picture.

Lex Fridman (1:22:19.160)

If I could just briefly ask, one is you're

Oriol Vinyals (1:22:21.840)

saying the benchmarks about the modular humans working on this,

Lex Fridman (1:22:28.960)

the benchmarks providing a sturdy ground of a wish

Oriol Vinyals (1:22:32.160)

to do these things that seem impossible.

Lex Fridman (1:22:34.680)

They give you, in the darkest of times,

Oriol Vinyals (1:22:37.880)

give you hope because little signs of improvement.

Lex Fridman (1:22:41.520)

Yes.

Oriol Vinyals (1:22:42.000)

Like somehow you're not lost if you have metrics

Lex Fridman (1:22:46.560)

to measure your improvement.

Lex Fridman (1:22:48.680)

And then there's other aspect.

Lex Fridman (1:22:50.800)

You said elsewhere and here today, like titles matter.

Oriol Vinyals (1:22:56.560)

I wonder how much humans matter in the evolution

Lex Fridman (1:23:01.280)

of all of this, meaning individual humans.

Oriol Vinyals (1:23:06.760)

Something about their interactions,

Lex Fridman (1:23:08.160)

something about their ideas, how much they change

Oriol Vinyals (1:23:11.200)

the direction of all of this.

Lex Fridman (1:23:12.920)

Like if you change the humans in this picture,

Oriol Vinyals (1:23:15.440)

is it that the model is sitting there

Lex Fridman (1:23:18.160)

and it wants some idea to work?

Oriol Vinyals (1:23:22.480)

Or is it the humans, or maybe the model

Lex Fridman (1:23:25.000)

is providing you 20 ideas that could work.

Lex Fridman (1:23:27.000)

And depending on the humans you pick,

Lex Fridman (1:23:29.080)

they're going to be able to hear some of those ideas.

Oriol Vinyals (1:23:33.160)

Because you're now directing all of deep learning and deep mind,

Lex Fridman (1:23:35.920)

you get to interact with a lot of projects,

Oriol Vinyals (1:23:37.720)

a lot of brilliant researchers.

Lex Fridman (1:23:40.600)

How much variability is created by the humans in all of this?

Oriol Vinyals (1:23:44.080)

Yeah, I mean, I do believe humans matter a lot,

Lex Fridman (1:23:47.320)

at the very least at the time scale of years

Oriol Vinyals (1:23:53.360)

on when things are happening and what's the sequencing of it.

Lex Fridman (1:23:56.880)

So you get to interact with people that, I mean,

Oriol Vinyals (1:24:00.800)

you mentioned this.

Lex Fridman (1:24:02.200)

Some people really want some idea to work

Lex Fridman (1:24:05.080)

and they'll persist.

Lex Fridman (1:24:07.040)

And then some other people might be more practical,

Oriol Vinyals (1:24:09.400)

like I don't care what idea works.

Lex Fridman (1:24:12.840)

I care about cracking protein folding.

Lex Fridman (1:24:16.800)

And at least these two kind of seem opposite sides.

Lex Fridman (1:24:21.240)

We need both.

Lex Fridman (1:24:22.400)

And we've clearly had both historically,

Lex Fridman (1:24:25.680)

and that made certain things happen earlier or later.

Lex Fridman (1:24:28.960)

So definitely humans involved in all of this endeavor

Lex Fridman (1:24:33.400)

have had, I would say, years of change or of ordering

Lex Fridman (1:24:38.640)

how things have happened, which breakthroughs came before,

Lex Fridman (1:24:41.840)

which other breakthroughs, and so on.

Lex Fridman (1:24:43.280)

So certainly that does happen.

Lex Fridman (1:24:45.800)

And so one other, maybe one other axis of distinction

Oriol Vinyals (1:24:50.600)

is what I called, and this is most commonly used

Lex Fridman (1:24:53.840)

in reinforcement learning, is the exploration exploitation

Oriol Vinyals (1:24:56.920)

trade off as well.

Lex Fridman (1:24:57.800)

It's not exactly what I meant, although quite related.

Lex Fridman (1:25:00.960)

So when you start trying to help others,

Lex Fridman (1:25:07.000)

like you become a bit more of a mentor

Oriol Vinyals (1:25:11.440)

to a large group of people, be it a project or the deep

Lex Fridman (1:25:14.600)

learning team or something, or even in the community

Oriol Vinyals (1:25:17.440)

when you interact with people in conferences and so on,

Lex Fridman (1:25:20.760)

you're identifying quickly some things that are explorative

Oriol Vinyals (1:25:26.040)

or exploitative.

Lex Fridman (1:25:27.080)

And it's tempting to try to guide people, obviously.

Oriol Vinyals (1:25:30.720)

I mean, that's what makes our experience.

Lex Fridman (1:25:33.160)

We bring it, and we try to shape things sometimes wrongly.

Lex Fridman (1:25:36.760)

And there's many times that I've been wrong in the past.

Lex Fridman (1:25:39.560)

That's great.

Lex Fridman (1:25:40.800)

But it would be wrong to dismiss any sort of the research

Lex Fridman (1:25:47.800)

styles that I'm observing.

Lex Fridman (1:25:49.880)

And I often get asked, well, you're in industry, right?

Lex Fridman (1:25:52.760)

So we do have access to large compute scale and so on.

Lex Fridman (1:25:55.640)

So there are certain kinds of research

Lex Fridman (1:25:57.360)

I almost feel like we need to do responsibly and so on.

Lex Fridman (1:26:01.640)

But it is, Carlos, we have the particle accelerator here,

Lex Fridman (1:26:05.160)

so to speak, in physics.

Lex Fridman (1:26:06.280)

So we need to use it.

Lex Fridman (1:26:07.480)

We need to answer the questions that we

Oriol Vinyals (1:26:09.240)

should be answering right now for the scientific progress.

Lex Fridman (1:26:12.320)

But then at the same time, I look at many advances,

Oriol Vinyals (1:26:15.200)

including attention, which was discovered in Montreal

Lex Fridman (1:26:19.280)

initially because of lack of compute, right?

Lex Fridman (1:26:22.440)

So we were working on sequence to sequence

Lex Fridman (1:26:24.920)

with my friends over at Google Brain at the time.

Lex Fridman (1:26:27.840)

And we were using, I think, eight GPUs,

Lex Fridman (1:26:30.400)

which was somehow a lot at the time.

Lex Fridman (1:26:32.360)

And then I think Montreal was a bit more limited in the scale.

Lex Fridman (1:26:36.080)

But then they discovered this content based attention

Oriol Vinyals (1:26:38.800)

concept that then has obviously triggered things

Lex Fridman (1:26:42.240)

like Transformer.

Oriol Vinyals (1:26:43.320)

Not everything obviously starts Transformer.

Lex Fridman (1:26:46.280)

There's always a history that is important to recognize

Oriol Vinyals (1:26:49.920)

because then you can make sure that then those who might feel

Lex Fridman (1:26:53.680)

now, well, we don't have so much compute,

Oriol Vinyals (1:26:56.320)

you need to then help them optimize

Lex Fridman (1:27:00.320)

that kind of research that might actually

Oriol Vinyals (1:27:02.320)

produce amazing change.

Lex Fridman (1:27:04.240)

Perhaps it's not as short term as some of these advancements

Oriol Vinyals (1:27:07.960)

or perhaps it's a different time scale.

Lex Fridman (1:27:09.720)

But the people and the diversity of the field

Oriol Vinyals (1:27:13.040)

is quite critical that we maintain it.

Lex Fridman (1:27:15.680)

And at times, especially mixed a bit with hype or other things,

Oriol Vinyals (1:27:19.800)

it's a bit tricky to be observing maybe

Lex Fridman (1:27:23.600)

too much of the same thinking across the board.

Lex Fridman (1:27:27.760)

But the humans definitely are critical.

Lex Fridman (1:27:30.520)

And I can think of quite a few personal examples

Oriol Vinyals (1:27:33.960)

where also someone told me something

Lex Fridman (1:27:36.640)

that had a huge effect onto some idea.

Lex Fridman (1:27:40.320)

And then that's why I'm saying at least in terms of years,

Lex Fridman (1:27:43.360)

probably some things do happen.

Oriol Vinyals (1:27:44.920)

Yeah, it's fascinating.

Lex Fridman (1:27:46.040)

And it's also fascinating how constraints somehow

Oriol Vinyals (1:27:48.240)

are essential for innovation.

Lex Fridman (1:27:51.040)

And the other thing you mentioned about engineering,

Oriol Vinyals (1:27:53.440)

I have a sneaking suspicion.

Lex Fridman (1:27:54.960)

Maybe I over, my love is with engineering.

Lex Fridman (1:28:00.040)

So I have a sneaky suspicion that all the genius,

Lex Fridman (1:28:04.480)

a large percentage of the genius is

Oriol Vinyals (1:28:06.600)

in the tiny details of engineering.

Lex Fridman (1:28:09.320)

So I think we like to think our genius,

Oriol Vinyals (1:28:14.000)

the genius is in the big ideas.

Lex Fridman (1:28:17.600)

I have a sneaking suspicion that because I've

Oriol Vinyals (1:28:20.600)

seen the genius of details, of engineering details,

Lex Fridman (1:28:24.440)

make the night and day difference.

Lex Fridman (1:28:28.840)

And I wonder if those kind of have a ripple effect over time.

Lex Fridman (1:28:32.960)

So that too, so that's sort of taking the engineering

Oriol Vinyals (1:28:36.360)

perspective that sometimes that quiet innovation

Lex Fridman (1:28:39.400)

at the level of an individual engineer

Oriol Vinyals (1:28:41.800)

or maybe at the small scale of a few engineers

Lex Fridman (1:28:44.680)

can make all the difference.

Oriol Vinyals (1:28:46.840)

Because we're working on computers that

Lex Fridman (1:28:50.200)

are scaled across large groups, that one engineering decision

Oriol Vinyals (1:28:55.080)

can lead to ripple effects.

Lex Fridman (1:28:57.320)

It's interesting to think about.

Oriol Vinyals (1:28:59.000)

Yeah, I mean, engineering, there's

Lex Fridman (1:29:01.160)

also kind of a historical, it might be a bit random.

Oriol Vinyals (1:29:06.360)

Because if you think of the history of how especially

Lex Fridman (1:29:10.240)

deep learning and neural networks took off,

Oriol Vinyals (1:29:12.360)

feels like a bit random because GPUs happened

Lex Fridman (1:29:16.600)

to be there at the right time for a different purpose, which

Oriol Vinyals (1:29:19.120)

was to play video games.

Lex Fridman (1:29:20.640)

So even the engineering that goes into the hardware

Lex Fridman (1:29:24.920)

and it might have a time, the time frame

Lex Fridman (1:29:27.160)

might be very different.

Oriol Vinyals (1:29:28.160)

I mean, the GPUs were evolved throughout many years

Lex Fridman (1:29:31.640)

where we didn't even were looking at that.

Lex Fridman (1:29:33.920)

So even at that level, that revolution, so to speak,

Lex Fridman (1:29:38.680)

the ripples are like, we'll see when they stop.

Lex Fridman (1:29:42.200)

But in terms of thinking of why is this happening,

Lex Fridman (1:29:46.960)

I think that when I try to categorize it

Oriol Vinyals (1:29:49.760)

in sort of things that might not be so obvious,

Lex Fridman (1:29:52.720)

I mean, clearly there's a hardware revolution.

Oriol Vinyals (1:29:54.920)

We are surfing thanks to that.

Lex Fridman (1:29:58.360)

Data centers as well.

Oriol Vinyals (1:29:59.720)

I mean, data centers are like, I mean, at Google,

Lex Fridman (1:30:02.680)

for instance, obviously they're serving Google.

Lex Fridman (1:30:04.840)

But there's also now thanks to that

Lex Fridman (1:30:06.920)

and to have built such amazing data centers,

Oriol Vinyals (1:30:09.680)

we can train these models.

Lex Fridman (1:30:11.720)

Software is an important one.

Oriol Vinyals (1:30:13.400)

I think if I look at the state of how

Lex Fridman (1:30:16.640)

I had to implement things to implement my ideas,

Lex Fridman (1:30:20.040)

how I discarded ideas because they were too hard

Lex Fridman (1:30:22.080)

to implement.

Oriol Vinyals (1:30:23.120)

Yeah, clearly the times have changed.

Lex Fridman (1:30:25.280)

And thankfully, we are in a much better software position

Oriol Vinyals (1:30:28.440)

as well.

Lex Fridman (1:30:29.400)

And then, I mean, obviously there's

Oriol Vinyals (1:30:31.680)

research that happens at scale and more people

Lex Fridman (1:30:34.360)

enter the field.

Oriol Vinyals (1:30:35.120)

That's great to see.

Lex Fridman (1:30:35.920)

But it's almost enabled by these other things.

Lex Fridman (1:30:38.200)

And last but not least is also data, right?

Lex Fridman (1:30:40.560)

Curating data sets, labeling data sets,

Oriol Vinyals (1:30:43.120)

these benchmarks we think about.

Lex Fridman (1:30:44.920)

Maybe we'll want to have all the benchmarks in one system.

Lex Fridman (1:30:48.880)

But it's still very valuable that someone

Lex Fridman (1:30:51.240)

put the thought and the time and the vision

Oriol Vinyals (1:30:53.600)

to build certain benchmarks.

Lex Fridman (1:30:54.880)

We've seen progress thanks to.

Lex Fridman (1:30:56.640)

But we're going to repurpose the benchmarks.

Lex Fridman (1:30:59.280)

That's the beauty of Atari is like we solved it in a way.

Lex Fridman (1:31:04.160)

But we use it in Gato.

Lex Fridman (1:31:06.000)

It was critical.

Lex Fridman (1:31:06.840)

And I'm sure there's still a lot more

Lex Fridman (1:31:09.080)

to do thanks to that amazing benchmark

Oriol Vinyals (1:31:10.960)

that someone took the time to put,

Lex Fridman (1:31:13.160)

even though at the time maybe, oh, you

Oriol Vinyals (1:31:15.560)

have to think what's the next iteration of architectures.

Lex Fridman (1:31:19.480)

That's what maybe the field recognizes.

Lex Fridman (1:31:21.440)

But that's another thing we need to balance

Lex Fridman (1:31:24.040)

in terms of humans behind.

Oriol Vinyals (1:31:25.760)

We need to recognize all these aspects

Lex Fridman (1:31:27.960)

because they're all critical.

Lex Fridman (1:31:29.440)

And we tend to think of the genius, the scientist,

Lex Fridman (1:31:33.600)

and so on.

Lex Fridman (1:31:34.080)

But I'm glad I know you have a strong engineering background.

Lex Fridman (1:31:38.000)

But also, I'm a lover of data.

Lex Fridman (1:31:40.040)

And the pushback on the engineering comment

Lex Fridman (1:31:43.200)

ultimately could be the creators of benchmarks

Oriol Vinyals (1:31:46.120)

who have the most impact.

Lex Fridman (1:31:47.400)

Andrej Karpathy, who you mentioned,

Oriol Vinyals (1:31:49.160)

has recently been talking a lot of trash about ImageNet, which

Lex Fridman (1:31:52.240)

he has the right to do because of how critical he is about

Oriol Vinyals (1:31:54.960)

ImageNet, how essential he is to the development

Lex Fridman (1:31:57.760)

and the success of deep learning around ImageNet.

Lex Fridman (1:32:01.480)

And he's saying that that's actually

Lex Fridman (1:32:02.960)

that benchmark is holding back the field.

Oriol Vinyals (1:32:05.520)

Because I mean, especially in his context on Tesla Autopilot,

Lex Fridman (1:32:09.000)

that's looking at real world behavior of a system.

Oriol Vinyals (1:32:14.280)

There's something fundamentally missing

Lex Fridman (1:32:16.280)

about ImageNet that doesn't capture

Oriol Vinyals (1:32:17.920)

the real worldness of things.

Lex Fridman (1:32:20.400)

That we need to have data sets, benchmarks that

Oriol Vinyals (1:32:23.560)

have the unpredictability, the edge cases, whatever

Lex Fridman (1:32:27.600)

the heck it is that makes the real world so

Oriol Vinyals (1:32:30.760)

difficult to operate in.

Lex Fridman (1:32:32.280)

We need to have benchmarks of that.

Lex Fridman (1:32:34.640)

But just to think about the impact of ImageNet

Lex Fridman (1:32:37.760)

as a benchmark, and that really puts a lot of emphasis

Oriol Vinyals (1:32:42.120)

on the importance of a benchmark,

Lex Fridman (1:32:43.720)

both sort of internally a deep mind and as a community.

Lex Fridman (1:32:46.640)

So one is coming in from within, like,

Lex Fridman (1:32:50.120)

how do I create a benchmark for me to mark and make progress?

Lex Fridman (1:32:55.280)

And how do I make benchmark for the community

Lex Fridman (1:32:58.120)

to mark and push progress?

Oriol Vinyals (1:33:02.520)

You have this amazing paper you coauthored,

Lex Fridman (1:33:05.840)

a survey paper called Emergent Abilities

Oriol Vinyals (1:33:08.600)

of Large Language Models.

Lex Fridman (1:33:10.480)

It has, again, the philosophy here

Oriol Vinyals (1:33:12.520)

that I'd love to ask you about.

Lex Fridman (1:33:14.520)

What's the intuition about the phenomena of emergence

Lex Fridman (1:33:17.320)

in neural networks transformed as language models?

Lex Fridman (1:33:20.600)

Is there a magic threshold beyond which

Lex Fridman (1:33:24.160)

we start to see certain performance?

Lex Fridman (1:33:27.080)

And is that different from task to task?

Lex Fridman (1:33:29.880)

Is that us humans just being poetic and romantic?

Lex Fridman (1:33:32.640)

Or is there literally some level at which we start

Lex Fridman (1:33:36.160)

to see breakthrough performance?

Lex Fridman (1:33:38.120)

Yeah, I mean, this is a property that we start seeing in systems

Oriol Vinyals (1:33:43.520)

that actually tend to be so in machine learning,

Lex Fridman (1:33:48.160)

traditionally, again, going to benchmarks.

Oriol Vinyals (1:33:51.680)

I mean, if you have some input, output, right,

Lex Fridman (1:33:54.840)

like that is just a single input and a single output,

Oriol Vinyals (1:33:58.200)

you generally, when you train these systems,

Lex Fridman (1:34:01.200)

you see reasonably smooth curves when

Oriol Vinyals (1:34:04.760)

you analyze how much the data set size affects

Lex Fridman (1:34:10.040)

the performance, or how the model size affects

Oriol Vinyals (1:34:12.280)

the performance, or how long you train the system for affects

Lex Fridman (1:34:18.200)

the performance, right?

Lex Fridman (1:34:19.280)

So if we think of ImageNet, the training curves

Lex Fridman (1:34:23.080)

look fairly smooth and predictable in a way.

Lex Fridman (1:34:28.080)

And I would say that's probably because it's

Lex Fridman (1:34:31.520)

kind of a one hop reasoning task, right?

Oriol Vinyals (1:34:36.520)

It's like, here is an input, and you

Lex Fridman (1:34:39.160)

think for a few milliseconds or 100 milliseconds, 300

Oriol Vinyals (1:34:42.760)

as a human, and then you tell me,

Lex Fridman (1:34:44.560)

yeah, there's an alpaca in this image.

Lex Fridman (1:34:47.840)

So in language, we are seeing benchmarks that require more

Lex Fridman (1:34:55.560)

pondering and more thought in a way, right?

Oriol Vinyals (1:34:58.200)

This is just kind of you need to look for some subtleties.

Lex Fridman (1:35:02.440)

It involves inputs that you might think of,

Oriol Vinyals (1:35:05.840)

even if the input is a sentence describing

Lex Fridman (1:35:08.360)

a mathematical problem, there is a bit more processing

Oriol Vinyals (1:35:13.080)

required as a human and more introspection.

Lex Fridman (1:35:15.640)

So I think how these benchmarks work

Oriol Vinyals (1:35:20.440)

means that there is actually a threshold.

Lex Fridman (1:35:24.720)

Just going back to how transformers

Oriol Vinyals (1:35:26.480)

work in this way of querying for the right questions

Lex Fridman (1:35:29.520)

to get the right answers, that might

Oriol Vinyals (1:35:31.760)

mean that performance becomes random

Lex Fridman (1:35:35.400)

until the right question is asked

Oriol Vinyals (1:35:37.720)

by the querying system of a transformer or of a language

Lex Fridman (1:35:40.920)

model like a transformer.

Lex Fridman (1:35:42.760)

And then only then you might start

Lex Fridman (1:35:46.240)

seeing performance going from random to nonrandom.

Lex Fridman (1:35:50.000)

And this is more empirical.

Lex Fridman (1:35:53.080)

There's no formalism or theory behind this yet,

Oriol Vinyals (1:35:56.320)

although it might be quite important.

Lex Fridman (1:35:57.760)

But we are seeing these phase transitions

Oriol Vinyals (1:36:00.320)

of random performance until some,

Lex Fridman (1:36:03.200)

let's say, scale of a model.

Lex Fridman (1:36:04.880)

And then it goes beyond that.

Lex Fridman (1:36:06.680)

And it might be that you need to fit

Oriol Vinyals (1:36:10.440)

a few low order bits of thought before you can make progress

Lex Fridman (1:36:16.040)

on the whole task.

Lex Fridman (1:36:17.200)

And if you could measure, actually,

Lex Fridman (1:36:19.720)

those breakdown of the task, maybe you

Oriol Vinyals (1:36:22.240)

would see more smooth, like, yeah,

Lex Fridman (1:36:25.320)

once you get these and these and these and these and these,

Oriol Vinyals (1:36:27.760)

then you start making progress in the task.

Lex Fridman (1:36:30.240)

But it's somehow a bit annoying because then it

Oriol Vinyals (1:36:35.240)

means that certain questions we might ask about architectures

Lex Fridman (1:36:40.240)

possibly can only be done at a certain scale.

Lex Fridman (1:36:42.960)

And one thing that, conversely, I've

Lex Fridman (1:36:46.320)

seen great progress on in the last couple of years

Oriol Vinyals (1:36:49.200)

is this notion of science of deep learning and science

Lex Fridman (1:36:53.120)

of scale in particular.

Lex Fridman (1:36:55.000)

So on the negative is that there are

Lex Fridman (1:36:57.520)

some benchmarks for which progress might

Oriol Vinyals (1:37:01.000)

need to be measured at minimum at a certain scale

Lex Fridman (1:37:04.000)

until you see then what details of the model

Oriol Vinyals (1:37:07.040)

matter to make that performance better.

Lex Fridman (1:37:09.960)

So that's a bit of a con.

Lex Fridman (1:37:11.880)

But what we've also seen is that you can empirically

Lex Fridman (1:37:17.960)

analyze behavior of models at scales that are smaller.

Lex Fridman (1:37:22.880)

So let's say, to put an example, we

Lex Fridman (1:37:25.920)

had this Chinchilla paper that revised the so called scaling

Oriol Vinyals (1:37:30.080)

laws of models.

Lex Fridman (1:37:31.320)

And that whole study is done at a reasonably small scale,

Oriol Vinyals (1:37:35.000)

that may be hundreds of millions up to 1 billion parameters.

Lex Fridman (1:37:38.600)

And then the cool thing is that you create some loss,

Oriol Vinyals (1:37:41.880)

some loss that some trends, you extract trends from data

Lex Fridman (1:37:45.840)

that you see, OK, it looks like the amount of data required

Oriol Vinyals (1:37:49.400)

to train now a 10x larger model would be this.

Lex Fridman (1:37:52.080)

And these laws so far, these extrapolations

Oriol Vinyals (1:37:55.200)

have helped us save compute and just get to a better place

Lex Fridman (1:37:59.880)

in terms of the science of how should we

Oriol Vinyals (1:38:02.520)

run these models at scale, how much data, how much depth,

Lex Fridman (1:38:05.640)

and all sorts of questions we start

Oriol Vinyals (1:38:07.360)

asking extrapolating from a small scale.

Lex Fridman (1:38:10.560)

But then these emergence is sadly that not everything

Oriol Vinyals (1:38:13.720)

can be extrapolated from scale depending on the benchmark.

Lex Fridman (1:38:16.920)

And maybe the harder benchmarks are not

Lex Fridman (1:38:19.840)

so good for extracting these laws.

Lex Fridman (1:38:21.920)

But we have a variety of benchmarks at least.

Lex Fridman (1:38:24.160)

So I wonder to which degree the threshold, the phase shift

Lex Fridman (1:38:29.240)

scale is a function of the benchmark.

Lex Fridman (1:38:32.440)

So some of the science of scale might

Lex Fridman (1:38:35.120)

be engineering benchmarks where that threshold is low,

Oriol Vinyals (1:38:40.400)

sort of taking a main benchmark and reducing it somehow

Lex Fridman (1:38:46.160)

where the essential difficulty is left

Lex Fridman (1:38:48.480)

but the scale of which the emergence happens

Lex Fridman (1:38:51.880)

is lower just for the science aspect of it

Oriol Vinyals (1:38:54.320)

versus the actual real world aspect.

Lex Fridman (1:38:56.960)

Yeah, so luckily we have quite a few benchmarks, some of which

Oriol Vinyals (1:38:59.920)

are simpler or maybe they're more like I think people might

Lex Fridman (1:39:02.640)

call these systems one versus systems two style.

Lex Fridman (1:39:05.920)

So I think what we're not seeing luckily

Lex Fridman (1:39:09.920)

is that extrapolations from maybe slightly more smooth

Oriol Vinyals (1:39:14.080)

or simpler benchmarks are translating to the harder ones.

Lex Fridman (1:39:18.560)

But that is not to say that this extrapolation will

Oriol Vinyals (1:39:21.480)

hit its limits.

Lex Fridman (1:39:22.560)

And when it does, then how much we scale or how we scale

Oriol Vinyals (1:39:27.560)

will sadly be a bit suboptimal until we find better laws.

Lex Fridman (1:39:31.760)

And these laws, again, are very empirical laws.

Oriol Vinyals (1:39:33.840)

They're not like physical laws of models,

Lex Fridman (1:39:35.960)

although I wish there would be better theory about these

Oriol Vinyals (1:39:39.520)

things as well.

Lex Fridman (1:39:40.240)

But so far, I would say empirical theory,

Oriol Vinyals (1:39:43.040)

as I call it, is way ahead than actual theory

Lex Fridman (1:39:46.000)

of machine learning.

Oriol Vinyals (1:39:47.800)

Let me ask you almost for fun.

Lex Fridman (1:39:50.560)

So this is not, Oriol, as a deep mind person or anything

Oriol Vinyals (1:39:55.840)

to do with deep mind or Google, just as a human being,

Lex Fridman (1:39:59.080)

looking at these news of a Google engineer who claimed

Oriol Vinyals (1:40:04.320)

that, I guess, the lambda language model was sentient.

Lex Fridman (1:40:11.120)

And you still need to look into the details of this.

Lex Fridman (1:40:14.080)

But making an official report and the claim

Lex Fridman (1:40:19.440)

that he believes there's evidence that this system has

Oriol Vinyals (1:40:23.880)

achieved sentience.

Lex Fridman (1:40:25.160)

And I think this is a really interesting case

Oriol Vinyals (1:40:29.480)

on a human level, on a psychological level,

Lex Fridman (1:40:31.720)

on a technical machine learning level of how language models

Oriol Vinyals (1:40:37.240)

transform our world, and also just philosophical level

Lex Fridman (1:40:39.840)

of the role of AI systems in a human world.

Lex Fridman (1:40:44.200)

So what do you find interesting?

Lex Fridman (1:40:48.080)

What's your take on all of this as a machine learning

Lex Fridman (1:40:51.080)

engineer and a researcher and also as a human being?

Lex Fridman (1:40:54.240)

Yeah, I mean, a few reactions.

Oriol Vinyals (1:40:57.440)

Quite a few, actually.

Lex Fridman (1:40:58.680)

Have you ever briefly thought, is this thing sentient?

Oriol Vinyals (1:41:02.560)

Right, so never, absolutely never.

Lex Fridman (1:41:04.800)

Like even with Alpha Star?

Oriol Vinyals (1:41:06.240)

Wait a minute.

Lex Fridman (1:41:08.080)

Sadly, though, I think, yeah, sadly, I have not.

Oriol Vinyals (1:41:11.840)

Yeah, I think the current, any of the current models,

Lex Fridman (1:41:15.280)

although very useful and very good,

Oriol Vinyals (1:41:18.880)

yeah, I think we're quite far from that.

Lex Fridman (1:41:22.320)

And there's kind of a converse side story.

Lex Fridman (1:41:25.320)

So one of my passions is about science in general.

Lex Fridman (1:41:30.320)

And I think I feel I'm a bit of a failed scientist.

Oriol Vinyals (1:41:34.440)

That's why I came to machine learning,

Lex Fridman (1:41:36.520)

because you always feel, and you start seeing this,

Oriol Vinyals (1:41:40.080)

that machine learning is maybe the science that

Lex Fridman (1:41:43.320)

can help other sciences, as we've seen.

Oriol Vinyals (1:41:46.400)

It's such a powerful tool.

Lex Fridman (1:41:48.640)

So thanks to that angle, that, OK, I love science.

Oriol Vinyals (1:41:52.480)

I love, I mean, I love astronomy.

Lex Fridman (1:41:53.880)

I love biology.

Lex Fridman (1:41:54.880)

But I'm not an expert.

Lex Fridman (1:41:56.000)

And I decided, well, the thing I can do better

Oriol Vinyals (1:41:58.600)

at is computers.

Lex Fridman (1:41:59.960)

But having, especially with when I was a bit more involved

Oriol Vinyals (1:42:04.720)

in AlphaFold, learning a bit about proteins

Lex Fridman (1:42:07.400)

and about biology and about life,

Oriol Vinyals (1:42:11.440)

the complexity, it feels like it really is.

Lex Fridman (1:42:14.840)

I mean, if you start looking at the things that are going on

Oriol Vinyals (1:42:19.200)

at the atomic level, and also, I mean, there's obviously the,

Lex Fridman (1:42:26.360)

we are maybe inclined to try to think of neural networks

Oriol Vinyals (1:42:29.280)

as like the brain.

Lex Fridman (1:42:30.400)

But the complexities and the amount of magic

Oriol Vinyals (1:42:33.760)

that it feels when, I mean, I'm not an expert,

Lex Fridman (1:42:37.080)

so it naturally feels more magic.

Lex Fridman (1:42:38.560)

But looking at biological systems,

Lex Fridman (1:42:40.800)

as opposed to these computational brains,

Oriol Vinyals (1:42:46.640)

just makes me like, wow, there's such a level of complexity

Lex Fridman (1:42:50.320)

difference still, like orders of magnitude complexity that,

Oriol Vinyals (1:42:54.840)

sure, these weights, I mean, we train them

Lex Fridman (1:42:56.640)

and they do nice things.

Lex Fridman (1:42:58.040)

But they're not at the level of biological entities, brains,

Lex Fridman (1:43:04.320)

cells.

Oriol Vinyals (1:43:06.000)

It just feels like it's just not possible to achieve

Lex Fridman (1:43:09.680)

the same level of complexity behavior.

Lex Fridman (1:43:12.400)

And my belief, when I talk to other beings,

Lex Fridman (1:43:16.240)

is certainly shaped by this amazement of biology

Oriol Vinyals (1:43:20.360)

that, maybe because I know too much,

Lex Fridman (1:43:22.360)

I don't have about machine learning,

Lex Fridman (1:43:23.800)

but I certainly feel it's very far fetched and far

Lex Fridman (1:43:28.120)

in the future to be calling or to be thinking,

Oriol Vinyals (1:43:31.720)

well, this mathematical function that is differentiable

Lex Fridman (1:43:35.640)

is, in fact, sentient and so on.

Oriol Vinyals (1:43:39.200)

There's something on that point that is very interesting.

Lex Fridman (1:43:42.000)

So you know enough about machines and enough

Oriol Vinyals (1:43:46.120)

about biology to know that there's

Lex Fridman (1:43:47.760)

many orders of magnitude of difference and complexity.

Lex Fridman (1:43:51.880)

But you know how machine learning works.

Lex Fridman (1:43:56.080)

So the interesting question for human beings

Oriol Vinyals (1:43:58.160)

that are interacting with a system that don't know

Lex Fridman (1:44:00.080)

about the underlying complexity.

Lex Fridman (1:44:02.280)

And I've seen people, probably including myself,

Lex Fridman (1:44:05.240)

that have fallen in love with things that are quite simple.

Lex Fridman (1:44:08.400)

And so maybe the complexity is one part of the picture,

Lex Fridman (1:44:11.520)

but maybe that's not a necessary condition for sentience,

Oriol Vinyals (1:44:18.840)

for perception or emulation of sentience.

Lex Fridman (1:44:24.760)

Right.

Lex Fridman (1:44:25.280)

So I mean, I guess the other side of this

Lex Fridman (1:44:27.560)

is that's how I feel personally.

Lex Fridman (1:44:29.560)

I mean, you asked me about the person, right?

Lex Fridman (1:44:32.360)

Now, it's very interesting to see how other humans feel

Lex Fridman (1:44:35.560)

about things, right?

Lex Fridman (1:44:37.080)

We are, again, I'm not as amazed about things

Oriol Vinyals (1:44:41.640)

that I feel this is not as magical as this other thing

Lex Fridman (1:44:44.560)

because of maybe how I got to learn about it

Lex Fridman (1:44:48.040)

and how I see the curve a bit more smooth

Lex Fridman (1:44:50.480)

because I've just seen the progress of language models

Oriol Vinyals (1:44:54.000)

since Shannon in the 50s.

Lex Fridman (1:44:56.000)

And actually looking at that time scale,

Lex Fridman (1:44:58.920)

we're not that fast progress, right?

Lex Fridman (1:45:00.880)

I mean, what we were thinking at the time almost 100 years ago

Oriol Vinyals (1:45:06.040)

is not that dissimilar to what we're doing now.

Lex Fridman (1:45:08.880)

But at the same time, yeah, obviously others,

Oriol Vinyals (1:45:11.440)

my experience, the personal experience,

Lex Fridman (1:45:14.440)

I think no one should tell others how they should feel.

Lex Fridman (1:45:20.680)

I mean, the feelings are very personal, right?

Lex Fridman (1:45:22.920)

So how others might feel about the models and so on.

Oriol Vinyals (1:45:26.080)

That's one part of the story that

Lex Fridman (1:45:27.840)

is important to understand for me personally as a researcher.

Lex Fridman (1:45:31.960)

And then when I maybe disagree or I

Lex Fridman (1:45:35.200)

don't understand or see that, yeah, maybe this is not

Oriol Vinyals (1:45:38.200)

something I think right now is reasonable,

Lex Fridman (1:45:39.920)

knowing all that I know, one of the other things

Lex Fridman (1:45:42.840)

and perhaps partly why it's great to be talking to you

Lex Fridman (1:45:46.480)

and reaching out to the world about machine learning

Oriol Vinyals (1:45:49.200)

is, hey, let's demystify a bit the magic

Lex Fridman (1:45:53.440)

and try to see a bit more of the math

Lex Fridman (1:45:56.200)

and the fact that literally to create these models,

Lex Fridman (1:45:59.800)

if we had the right software, it would be 10 lines of code

Lex Fridman (1:46:03.520)

and then just a dump of the internet.

Lex Fridman (1:46:06.760)

Versus then the complexity of the creation of humans

Lex Fridman (1:46:11.600)

from their inception, right?

Lex Fridman (1:46:13.520)

And also the complexity of evolution of the whole universe

Oriol Vinyals (1:46:17.600)

to where we are that feels orders of magnitude

Lex Fridman (1:46:21.040)

more complex and fascinating to me.

Lex Fridman (1:46:23.400)

So I think, yeah, maybe part of the only thing

Lex Fridman (1:46:26.520)

I'm thinking about trying to tell you is, yeah, I think

Oriol Vinyals (1:46:30.240)

explaining a bit of the magic.

Lex Fridman (1:46:32.560)

There is a bit of magic.

Oriol Vinyals (1:46:33.600)

It's good to be in love, obviously,

Lex Fridman (1:46:35.240)

with what you do at work.

Lex Fridman (1:46:36.920)

And I'm certainly fascinated and surprised quite often as well.

Lex Fridman (1:46:41.320)

But I think, hopefully, as experts in biology,

Oriol Vinyals (1:46:45.040)

hopefully will tell me this is not as magic.

Lex Fridman (1:46:47.080)

And I'm happy to learn that through interactions

Oriol Vinyals (1:46:50.840)

with the larger community, we can also

Lex Fridman (1:46:54.000)

have a certain level of education

Oriol Vinyals (1:46:56.000)

that in practice also will matter because, I mean,

Lex Fridman (1:46:58.920)

one question is how you feel about this.

Lex Fridman (1:47:00.800)

But then the other very important is

Lex Fridman (1:47:03.000)

you starting to interact with these in products and so on.

Oriol Vinyals (1:47:06.960)

It's good to understand a bit what's going on,

Lex Fridman (1:47:09.240)

what's not going on, what's safe, what's not safe,

Lex Fridman (1:47:12.280)

and so on, right?

Lex Fridman (1:47:13.000)

Otherwise, the technology will not

Oriol Vinyals (1:47:15.280)

be used properly for good, which is obviously

Lex Fridman (1:47:18.120)

the goal of all of us, I hope.

Lex Fridman (1:47:20.480)

So let me then ask the next question.

Lex Fridman (1:47:22.920)

Do you think in order to solve intelligence

Oriol Vinyals (1:47:25.760)

or to replace the leg spot that does interviews

Lex Fridman (1:47:29.480)

as we started this conversation with,

Lex Fridman (1:47:31.480)

do you think the system needs to be sentient?

Lex Fridman (1:47:34.880)

Do you think it needs to achieve something like consciousness?

Lex Fridman (1:47:38.720)

And do you think about what consciousness

Lex Fridman (1:47:41.120)

is in the human mind that could be instructive for creating AI

Lex Fridman (1:47:45.360)

systems?

Lex Fridman (1:47:46.720)

Yeah.

Oriol Vinyals (1:47:47.760)

Honestly, I think probably not to the degree of intelligence

Lex Fridman (1:47:53.480)

that there's this brain that can learn,

Oriol Vinyals (1:47:58.760)

can be extremely useful, can challenge you, can teach you.

Lex Fridman (1:48:02.960)

Conversely, you can teach it to do things.

Oriol Vinyals (1:48:05.600)

I'm not sure it's necessary, personally speaking.

Lex Fridman (1:48:09.080)

But if consciousness or any other biological or evolutionary

Oriol Vinyals (1:48:15.680)

lesson can be repurposed to then influence

Lex Fridman (1:48:20.880)

our next set of algorithms, that is a great way

Lex Fridman (1:48:24.360)

to actually make progress, right?

Lex Fridman (1:48:25.680)

And the same way I try to explain transformers a bit

Lex Fridman (1:48:28.240)

how it feels we operate when we look at text specifically,

Lex Fridman (1:48:33.360)

these insights are very important, right?

Lex Fridman (1:48:36.000)

So there's a distinction between details of how the brain might

Lex Fridman (1:48:41.240)

be doing computation.

Oriol Vinyals (1:48:43.200)

I think my understanding is, sure, there's neurons

Lex Fridman (1:48:46.560)

and there's some resemblance to neural networks,

Lex Fridman (1:48:48.520)

but we don't quite understand enough of the brain in detail,

Lex Fridman (1:48:52.200)

right, to be able to replicate it.

Lex Fridman (1:48:55.240)

But then if you zoom out a bit, our thought process,

Lex Fridman (1:49:01.320)

how memory works, maybe even how evolution got us here,

Oriol Vinyals (1:49:05.560)

what's exploration, exploitation,

Lex Fridman (1:49:07.280)

like how these things happen, I think

Oriol Vinyals (1:49:09.080)

these clearly can inform algorithmic level research.

Lex Fridman (1:49:12.960)

And I've seen some examples of this

Oriol Vinyals (1:49:17.040)

being quite useful to then guide the research,

Lex Fridman (1:49:19.720)

even it might be for the wrong reasons, right?

Lex Fridman (1:49:21.640)

So I think biology and what we know about ourselves

Lex Fridman (1:49:26.080)

can help a whole lot to build, essentially,

Lex Fridman (1:49:30.000)

what we call AGI, this general, the real ghetto, right?

Lex Fridman (1:49:34.480)

The last step of the chain, hopefully.

Lex Fridman (1:49:36.480)

But consciousness in particular, I don't myself

Lex Fridman (1:49:40.760)

at least think too hard about how to add that to the system.

Lex Fridman (1:49:44.760)

But maybe my understanding is also very personal

Lex Fridman (1:49:47.840)

about what it means, right?

Oriol Vinyals (1:49:48.840)

I think even that in itself is a long debate

Lex Fridman (1:49:51.760)

that I know people have often.

Lex Fridman (1:49:55.240)

And maybe I should learn more about this.

Lex Fridman (1:49:57.720)

Yeah, and I personally, I notice the magic often

Oriol Vinyals (1:50:01.680)

on a personal level, especially with physical systems

Lex Fridman (1:50:04.960)

like robots.

Oriol Vinyals (1:50:06.120)

I have a lot of legged robots now in Austin

Lex Fridman (1:50:10.440)

that I play with.

Lex Fridman (1:50:11.680)

And even when you program them, when

Lex Fridman (1:50:13.480)

they do things you didn't expect,

Oriol Vinyals (1:50:15.560)

there's an immediate anthropomorphization.

Lex Fridman (1:50:18.560)

And you notice the magic, and you

Oriol Vinyals (1:50:19.960)

start to think about things like sentience

Lex Fridman (1:50:22.600)

that has to do more with effective communication

Lex Fridman (1:50:26.000)

and less with any of these kind of dramatic things.

Lex Fridman (1:50:30.160)

It seems like a useful part of communication.

Oriol Vinyals (1:50:32.840)

Having the perception of consciousness

Lex Fridman (1:50:36.560)

seems like useful for us humans.

Oriol Vinyals (1:50:38.800)

We treat each other more seriously.

Lex Fridman (1:50:40.840)

We are able to do a nearest neighbor shoving of that entity

Oriol Vinyals (1:50:46.000)

into your memory correctly, all that kind of stuff.

Lex Fridman (1:50:48.640)

It seems useful, at least to fake it,

Oriol Vinyals (1:50:50.800)

even if you never make it.

Lex Fridman (1:50:52.440)

So maybe, like, yeah, mirroring the question.

Lex Fridman (1:50:55.560)

And since you talked to a few people,

Lex Fridman (1:50:57.440)

then you do think that we'll need

Oriol Vinyals (1:50:59.880)

to figure something out in order to achieve intelligence

Lex Fridman (1:51:04.560)

in a grander sense of the word.

Oriol Vinyals (1:51:06.520)

Yeah, I personally believe yes, but I don't even

Lex Fridman (1:51:09.360)

think it'll be like a separate island we'll have to travel to.

Oriol Vinyals (1:51:14.160)

I think it will emerge quite naturally.

Lex Fridman (1:51:16.400)

OK, that's easier for us then.

Oriol Vinyals (1:51:19.040)

Thank you.

Lex Fridman (1:51:20.080)

But the reason I think it's important to think about

Oriol Vinyals (1:51:22.760)

is you will start, I believe, like with this Google

Lex Fridman (1:51:25.800)

engineer, you will start seeing this a lot more, especially

Oriol Vinyals (1:51:29.320)

when you have AI systems that are actually interacting

Lex Fridman (1:51:31.600)

with human beings that don't have an engineering background.

Lex Fridman (1:51:35.120)

And we have to prepare for that.

Lex Fridman (1:51:38.520)

Because I do believe there will be a civil rights

Oriol Vinyals (1:51:41.160)

movement for robots, as silly as it is to say.

Lex Fridman (1:51:44.520)

There's going to be a large number of people

Oriol Vinyals (1:51:46.720)

that realize there's these intelligent entities with whom

Lex Fridman (1:51:49.760)

I have a deep relationship, and I don't want to lose them.

Oriol Vinyals (1:51:53.160)

They've come to be a part of my life, and they mean a lot.

Lex Fridman (1:51:55.920)

They have a name.

Oriol Vinyals (1:51:57.120)

They have a story.

Lex Fridman (1:51:58.040)

They have a memory.

Lex Fridman (1:51:59.120)

And we start to ask questions about ourselves.

Lex Fridman (1:52:01.240)

Well, this thing sure seems like it's capable of suffering,

Oriol Vinyals (1:52:07.520)

because it tells all these stories of suffering.

Lex Fridman (1:52:09.800)

It doesn't want to die and all those kinds of things.

Lex Fridman (1:52:11.960)

And we have to start to ask ourselves questions.

Lex Fridman (1:52:14.400)

What is the difference between a human being and this thing?

Lex Fridman (1:52:16.960)

And so when you engineer, I believe

Lex Fridman (1:52:20.120)

from an engineering perspective, from a deep mind or anybody

Oriol Vinyals (1:52:23.400)

that builds systems, there might be laws in the future

Lex Fridman (1:52:26.440)

where you're not allowed to engineer systems

Oriol Vinyals (1:52:29.120)

with displays of sentience, unless they're explicitly

Lex Fridman (1:52:35.120)

designed to be that, unless it's a pet.

Lex Fridman (1:52:37.320)

So if you have a system that's just doing customer support,

Lex Fridman (1:52:41.160)

you're legally not allowed to display sentience.

Oriol Vinyals (1:52:44.160)

We'll start to ask ourselves that question.

Lex Fridman (1:52:47.200)

And then so that's going to be part of the software

Oriol Vinyals (1:52:49.920)

engineering process.

Lex Fridman (1:52:52.080)

Which features do we have?

Lex Fridman (1:52:53.320)

And one of them is communications of the sentience.

Lex Fridman (1:52:56.440)

But it's important to start thinking about that stuff,

Oriol Vinyals (1:52:58.680)

especially how much it captivates public attention.

Lex Fridman (1:53:01.640)

Yeah, absolutely.

Oriol Vinyals (1:53:03.120)

It's definitely a topic that is important.

Lex Fridman (1:53:06.360)

We think about.

Lex Fridman (1:53:07.880)

And I think in a way, I always see not every movie

Lex Fridman (1:53:12.560)

is equally on point with certain things.

Lex Fridman (1:53:16.080)

But certainly science fiction in this sense

Lex Fridman (1:53:19.000)

at least has prepared society to start

Oriol Vinyals (1:53:22.120)

thinking about certain topics that even if it's

Lex Fridman (1:53:25.360)

too early to talk about, as long as we are reasonable,

Oriol Vinyals (1:53:29.400)

it's certainly going to prepare us for both the research

Lex Fridman (1:53:33.840)

to come and how to.

Oriol Vinyals (1:53:34.920)

I mean, there's many important challenges and topics

Lex Fridman (1:53:38.080)

that come with building an intelligent system, many of

Oriol Vinyals (1:53:43.200)

which you just mentioned.

Lex Fridman (1:53:44.640)

So I think we're never going to be fully ready

Oriol Vinyals (1:53:49.880)

unless we talk about these.

Lex Fridman (1:53:51.360)

And we start also, as I said, just expanding the people

Oriol Vinyals (1:53:58.840)

we talk to not include only our own researchers and so on.

Lex Fridman (1:54:03.240)

And in fact, places like DeepMind but elsewhere,

Oriol Vinyals (1:54:06.480)

there's more interdisciplinary groups forming up

Lex Fridman (1:54:10.320)

to start asking and really working

Oriol Vinyals (1:54:12.880)

with us on these questions.

Lex Fridman (1:54:14.880)

Because obviously, this is not initially

Lex Fridman (1:54:17.400)

what your passion is when you do your PhD,

Lex Fridman (1:54:19.360)

but certainly it is coming.

Lex Fridman (1:54:21.440)

So it's fascinating.

Lex Fridman (1:54:23.120)

It's the thing that brings me to one of my passions

Oriol Vinyals (1:54:27.160)

that is learning.

Lex Fridman (1:54:28.000)

So in this sense, this is a new area

Oriol Vinyals (1:54:31.680)

that, as a learning system myself,

Lex Fridman (1:54:35.120)

I want to keep exploring.

Lex Fridman (1:54:36.640)

And I think it's great to see parts of the debate.

Lex Fridman (1:54:41.000)

And even I've seen a level of maturity

Oriol Vinyals (1:54:43.720)

in the conferences that deal with AI.

Lex Fridman (1:54:46.400)

If you look five years ago to now,

Oriol Vinyals (1:54:49.840)

just the amount of workshops and so on has changed so much.

Lex Fridman (1:54:53.040)

It's impressive to see how much topics of safety, ethics,

Lex Fridman (1:54:58.520)

and so on come to the surface, which is great.

Lex Fridman (1:55:01.720)

And if it were too early, clearly it's fine.

Oriol Vinyals (1:55:03.800)

I mean, it's a big field, and there's

Lex Fridman (1:55:05.920)

lots of people with lots of interests

Oriol Vinyals (1:55:09.040)

that will do progress or make progress.

Lex Fridman (1:55:11.880)

And obviously, I don't believe we're too late.

Lex Fridman (1:55:14.160)

So in that sense, I think it's great

Lex Fridman (1:55:16.440)

that we're doing this already.

Oriol Vinyals (1:55:18.160)

It better be too early than too late

Lex Fridman (1:55:20.200)

when it comes to super intelligent AI systems.

Oriol Vinyals (1:55:22.720)

Let me ask, speaking of sentient AIs,

Lex Fridman (1:55:25.480)

you gave props to your friend Ilyas Etzgever

Oriol Vinyals (1:55:28.680)

for being elected the fellow of the Royal Society.

Lex Fridman (1:55:31.960)

So just as a shout out to a fellow researcher

Lex Fridman (1:55:34.680)

and a friend, what's the secret to the genius of Ilyas

Lex Fridman (1:55:38.240)

Etzgever?

Lex Fridman (1:55:39.400)

And also, do you believe that his tweets,

Lex Fridman (1:55:42.640)

as you've hypothesized and Andrej Karpathy did as well,

Lex Fridman (1:55:46.000)

are generated by a language model?

Lex Fridman (1:55:48.680)

Yeah.

Lex Fridman (1:55:49.360)

So I strongly believe Ilya is going to visit in a few weeks,

Lex Fridman (1:55:54.240)

actually.

Lex Fridman (1:55:54.720)

So I'll ask him in person.

Lex Fridman (1:55:58.000)

Will he tell you the truth?

Oriol Vinyals (1:55:59.160)

Yes, of course, hopefully.

Lex Fridman (1:56:00.720)

I mean, ultimately, we all have shared paths,

Lex Fridman (1:56:04.040)

and there's friendships that go beyond, obviously,

Lex Fridman (1:56:08.280)

institutions and so on.

Lex Fridman (1:56:09.960)

So I hope he tells me the truth.

Lex Fridman (1:56:11.680)

Well, maybe the AI system is holding him hostage somehow.

Oriol Vinyals (1:56:14.400)

Maybe he has some videos that he doesn't want to release.

Lex Fridman (1:56:16.920)

So maybe it has taken control over him.

Lex Fridman (1:56:19.720)

So he can't tell the truth.

Lex Fridman (1:56:20.960)

Well, if I see him in person, then I think he will know.

Lex Fridman (1:56:23.920)

But I think Ilya's personality, just knowing him for a while,

Lex Fridman (1:56:33.920)

everyone in Twitter, I guess, gets a different persona.

Lex Fridman (1:56:36.640)

And I think Ilya's one does not surprise me.

Lex Fridman (1:56:40.920)

So I think knowing Ilya from before social media

Lex Fridman (1:56:43.600)

and before AI was so prevalent, I

Lex Fridman (1:56:46.000)

recognize a lot of his character.

Lex Fridman (1:56:47.560)

So that's something for me that I

Lex Fridman (1:56:49.200)

feel good about a friend that hasn't changed

Oriol Vinyals (1:56:52.520)

or is still true to himself.

Lex Fridman (1:56:55.960)

Obviously, there is, though, a fact

Oriol Vinyals (1:56:58.960)

that your field becomes more popular,

Lex Fridman (1:57:02.080)

and he is obviously one of the main figures in the field,

Oriol Vinyals (1:57:05.440)

having done a lot of advancement.

Lex Fridman (1:57:07.040)

So I think that the tricky bit here

Oriol Vinyals (1:57:09.080)

is how to balance your true self with the responsibility

Lex Fridman (1:57:12.200)

that your words carry.

Lex Fridman (1:57:13.560)

So in this sense, I appreciate the style, and I understand it.

Lex Fridman (1:57:19.360)

But it created debates on some of his tweets

Oriol Vinyals (1:57:24.160)

that maybe it's good we have them early anyways.

Lex Fridman (1:57:27.920)

But yeah, then the reactions are usually polarizing.

Oriol Vinyals (1:57:31.040)

I think we're just seeing the reality of social media

Lex Fridman (1:57:34.160)

be there as well, reflected on that particular topic

Oriol Vinyals (1:57:38.120)

or set of topics he's tweeting about.

Lex Fridman (1:57:40.200)

Yeah, I mean, it's funny that he used to speak to this tension.

Oriol Vinyals (1:57:42.960)

He was one of the early seminal figures

Lex Fridman (1:57:46.160)

in the field of deep learning, so there's

Oriol Vinyals (1:57:47.800)

a responsibility with that.

Lex Fridman (1:57:48.960)

But he's also, from having interacted with him quite a bit,

Oriol Vinyals (1:57:53.200)

he's just a brilliant thinker about ideas, which, as are you.

Lex Fridman (1:58:01.280)

And there's a tension between becoming

Oriol Vinyals (1:58:03.120)

the manager versus the actual thinking

Lex Fridman (1:58:06.960)

through very novel ideas, the scientist versus the manager.

Lex Fridman (1:58:13.640)

And he's one of the great scientists of our time.

Lex Fridman (1:58:17.680)

So this was quite interesting.

Lex Fridman (1:58:18.960)

And also, people tell me quite silly,

Lex Fridman (1:58:20.840)

which I haven't quite detected yet.

Lex Fridman (1:58:23.200)

But in private, we'll have to see about that.

Lex Fridman (1:58:26.000)

Yeah, yeah.

Oriol Vinyals (1:58:27.480)

I mean, just on the point of, I mean,

Lex Fridman (1:58:30.000)

Ilya has been an inspiration.

Oriol Vinyals (1:58:33.360)

I mean, quite a few colleagues, I can think,

Lex Fridman (1:58:35.480)

shaped the person you are.

Oriol Vinyals (1:58:38.080)

Like, Ilya certainly gets probably the top spot,

Lex Fridman (1:58:42.320)

if not close to the top.

Lex Fridman (1:58:43.800)

And if we go back to the question about people in the field,

Lex Fridman (1:58:47.960)

like how their role would have changed the field or not,

Oriol Vinyals (1:58:51.680)

I think Ilya's case is interesting

Lex Fridman (1:58:54.000)

because he really has a deep belief in the scaling up

Oriol Vinyals (1:58:58.760)

of neural networks.

Lex Fridman (1:58:59.640)

There was a talk that is still famous to this day

Oriol Vinyals (1:59:03.680)

from the Sequence to Sequence paper, where he was just

Lex Fridman (1:59:07.720)

claiming, just give me supervised data

Lex Fridman (1:59:10.560)

and a large neural network, and then you'll

Lex Fridman (1:59:12.800)

solve basically all the problems.

Oriol Vinyals (1:59:16.240)

That vision was already there many years ago.

Lex Fridman (1:59:19.800)

So it's good to see someone who is, in this case,

Oriol Vinyals (1:59:22.880)

very deeply into this style of research

Lex Fridman (1:59:27.160)

and clearly has had a tremendous track record of successes

Lex Fridman (1:59:32.800)

and so on.

Lex Fridman (1:59:34.160)

The funny bit about that talk is that we rehearsed the talk

Oriol Vinyals (1:59:37.520)

in a hotel room before, and the original version of that talk

Lex Fridman (1:59:42.040)

would have been even more controversial.

Lex Fridman (1:59:44.000)

So maybe I'm the only person that

Lex Fridman (1:59:46.760)

has seen the unfiltered version of the talk.

Lex Fridman (1:59:49.520)

And maybe when the time comes, maybe we

Lex Fridman (1:59:52.160)

should revisit some of the skip slides

Oriol Vinyals (1:59:55.120)

from the talk from Ilya.

Lex Fridman (1:59:57.560)

But I really think the deep belief

Oriol Vinyals (20:03.040)

algorithmically speaking to the process

Lex Fridman (20:04.900)

of creating this model.

Lex Fridman (20:06.320)

But now you're doing it through language,

Lex Fridman (20:07.840)

which is very natural way for us to learn from one another.

Oriol Vinyals (20:11.080)

I tell you, hey, you should do this new task.

Lex Fridman (20:13.180)

I'll tell you a bit more.

Oriol Vinyals (20:14.600)

Maybe you ask me some questions

Lex Fridman (20:16.080)

and now you know the task, right?

Oriol Vinyals (20:17.840)

You didn't need to retrain it from scratch.

Lex Fridman (20:20.320)

And we've seen these magical moments almost

Oriol Vinyals (20:24.080)

in this way to do few shot promptings through language

Lex Fridman (20:26.960)

on language only domain.

Lex Fridman (20:28.560)

And then in the last two years,

Lex Fridman (20:30.960)

we've seen these expanded to beyond language,

Oriol Vinyals (20:34.640)

adding vision, adding actions and games,

Lex Fridman (20:38.040)

lots of progress to be had.

Lex Fridman (20:39.480)

But this is maybe, if you ask me like about

Lex Fridman (20:42.160)

how are we gonna crack this problem?

Oriol Vinyals (20:43.720)

This is perhaps one way in which you have a single model.

Lex Fridman (20:48.760)

The problem of this model is it's hard to grow

Oriol Vinyals (20:52.160)

in weights or capacity,

Lex Fridman (20:54.320)

but the model is certainly so powerful

Lex Fridman (20:56.400)

that you can teach it some tasks, right?

Lex Fridman (20:58.960)

In this way that I teach you,

Oriol Vinyals (21:00.600)

I could teach you a new task now,

Lex Fridman (21:02.000)

if we were all at a text based task

Oriol Vinyals (21:05.120)

or a classification of vision style task.

Lex Fridman (21:08.440)

But it still feels like more breakthroughs should be had,

Lex Fridman (21:12.860)

but it's a great beginning, right?

Lex Fridman (21:14.040)

We have a good baseline.

Oriol Vinyals (21:15.440)

We have an idea that this maybe is the way we want

Lex Fridman (21:18.160)

to benchmark progress towards AGI.

Lex Fridman (21:20.800)

And I think in my view, that's critical

Lex Fridman (21:22.880)

to always have a way to benchmark the community

Oriol Vinyals (21:25.760)

sort of converging to these overall,

Lex Fridman (21:27.840)

which is good to see.

Lex Fridman (21:29.240)

And then this is actually what excites me

Lex Fridman (21:33.520)

in terms of also next steps for deep learning

Oriol Vinyals (21:36.640)

is how to make these models more powerful,

Lex Fridman (21:39.120)

how do you train them, how to grow them

Oriol Vinyals (21:41.760)

if they must grow, should they change their weights

Lex Fridman (21:44.520)

as you teach it task or not?

Oriol Vinyals (21:46.120)

There's some interesting questions, many to be answered.

Lex Fridman (21:48.560)

Yeah, you've opened the door

Oriol Vinyals (21:49.760)

about to a bunch of questions I want to ask,

Lex Fridman (21:52.320)

but let's first return to your tweet

Lex Fridman (21:55.720)

and read it like a Shakespeare.

Lex Fridman (21:57.160)

You wrote, God is not the end, it's the beginning.

Lex Fridman (22:01.280)

And then you wrote meow and then an emoji of a cat.

Lex Fridman (22:06.200)

So first two questions.

Lex Fridman (22:07.740)

First, can you explain the meow and the cat emoji?

Lex Fridman (22:10.080)

And second, can you explain what Godot is and how it works?

Oriol Vinyals (22:13.680)

Right, indeed.

Lex Fridman (22:14.640)

I mean, thanks for reminding me

Oriol Vinyals (22:16.520)

that we're all exposing on Twitter and.

Lex Fridman (22:19.920)

Permanently there.

Oriol Vinyals (22:20.960)

Yes, permanently there.

Lex Fridman (22:21.920)

One of the greatest AI researchers of all time,

Oriol Vinyals (22:25.120)

meow and cat emoji.

Lex Fridman (22:27.200)

Yes. There you go.

Oriol Vinyals (22:28.280)

Right, so.

Lex Fridman (22:29.120)

Can you imagine like touring, tweeting, meow and cat,

Oriol Vinyals (22:32.720)

probably he would, probably would.

Lex Fridman (22:34.360)

Probably.

Lex Fridman (22:35.200)

So yeah, the tweet is important actually.

Lex Fridman (22:38.020)

You know, I put thought on the tweets, I hope people.

Lex Fridman (22:40.720)

Which part do you think?

Lex Fridman (22:41.720)

Okay, so there's three sentences.

Oriol Vinyals (22:44.840)

Godot is not the end, Godot is the beginning,

Lex Fridman (22:48.640)

meow, cat emoji.

Lex Fridman (22:50.120)

Okay, which is the important part?

Lex Fridman (22:51.720)

The meow, no, no.

Oriol Vinyals (22:53.120)

Definitely that it is the beginning.

Lex Fridman (22:56.080)

I mean, I probably was just explaining a bit

Oriol Vinyals (23:00.340)

where the field is going, but let me tell you about Godot.

Lex Fridman (23:03.760)

So first the name Godot comes from maybe a sequence

Oriol Vinyals (23:08.120)

of releases that DeepMind had that named,

Lex Fridman (23:11.820)

like used animal names to name some of their models

Oriol Vinyals (23:15.100)

that are based on this idea of large sequence models.

Lex Fridman (23:19.120)

Initially they're only language,

Lex Fridman (23:20.620)

but we are expanding to other modalities.

Lex Fridman (23:23.180)

So we had, you know, we had Gopher, Chinchilla,

Oriol Vinyals (23:28.800)

these were language only.

Lex Fridman (23:29.960)

And then more recently we released Flamingo,

Oriol Vinyals (23:32.720)

which adds vision to the equation.

Lex Fridman (23:35.460)

And then Godot, which adds vision

Lex Fridman (23:38.160)

and then also actions in the mix, right?

Lex Fridman (23:41.660)

As we discuss actually actions,

Oriol Vinyals (23:44.520)

especially discrete actions like up, down, left, right.

Lex Fridman (23:47.600)

I just told you the actions, but they're words.

Lex Fridman (23:49.520)

So you can kind of see how actions naturally map

Lex Fridman (23:52.800)

to sequence modeling of words,

Oriol Vinyals (23:54.560)

which these models are very powerful.

Lex Fridman (23:57.100)

So Godot was named after, I believe,

Lex Fridman (24:01.720)

I can only from memory, right?

Lex Fridman (24:03.640)

These, you know, these things always happen

Oriol Vinyals (24:06.080)

with an amazing team of researchers behind.

Lex Fridman (24:08.520)

So before the release, we had a discussion

Lex Fridman (24:12.200)

about which animal would we pick, right?

Lex Fridman (24:14.240)

And I think because of the word general agent, right?

Lex Fridman (24:18.380)

And this is a property quite unique to Godot.

Lex Fridman (24:21.920)

We kind of were playing with the GA words

Lex Fridman (24:24.760)

and then, you know, Godot.

Lex Fridman (24:26.040)

Rhymes with cat.

Oriol Vinyals (24:26.960)

Yes.

Lex Fridman (24:28.080)

And Godot is obviously a Spanish version of cat.

Oriol Vinyals (24:30.280)

I had nothing to do with it, although I'm from Spain.

Lex Fridman (24:32.280)

Oh, how do you, wait, sorry.

Lex Fridman (24:33.320)

How do you say cat in Spanish?

Lex Fridman (24:34.700)

Gato.

Oriol Vinyals (24:35.540)

Oh, gato, okay.

Lex Fridman (24:36.360)

Now it all makes sense.

Oriol Vinyals (24:37.200)

Okay, okay, I see, I see, I see.

Lex Fridman (24:38.160)

Now it all makes sense.

Oriol Vinyals (24:39.120)

Okay, so.

Lex Fridman (24:39.960)

How do you say meow in Spanish?

Oriol Vinyals (24:40.840)

No, that's probably the same.

Lex Fridman (24:41.960)

I think you say it the same way,

Lex Fridman (24:44.440)

but you write it as M, I, A, U.

Lex Fridman (24:48.120)

Okay, it's universal.

Oriol Vinyals (24:49.240)

Yes.

Lex Fridman (24:50.080)

All right, so then how does the thing work?

Lex Fridman (24:51.680)

So you said general is, so you said language, vision.

Lex Fridman (24:57.520)

And action. Action.

Lex Fridman (24:59.240)

How does this, can you explain

Lex Fridman (25:01.840)

what kind of neural networks are involved?

Lex Fridman (25:04.240)

What does the training look like?

Lex Fridman (25:06.360)

And maybe what do you,

Lex Fridman (25:09.380)

are some beautiful ideas within the system?

Lex Fridman (25:11.840)

Yeah, so maybe the basics of Gato

Oriol Vinyals (25:16.060)

are not that dissimilar from many, many work that come.

Lex Fridman (25:19.920)

So here is where the sort of the recipe,

Oriol Vinyals (25:22.880)

I mean, hasn't changed too much.

Lex Fridman (25:24.200)

There is a transformer model

Oriol Vinyals (25:25.600)

that's the kind of recurrent neural network

Lex Fridman (25:28.640)

that essentially takes a sequence of modalities,

Oriol Vinyals (25:33.320)

observations that could be words,

Lex Fridman (25:36.360)

could be vision or could be actions.

Lex Fridman (25:38.800)

And then its own objective that you train it to do

Lex Fridman (25:42.120)

when you train it is to predict what the next anything is.

Lex Fridman (25:46.360)

And anything means what's the next action.

Lex Fridman (25:48.760)

If this sequence that I'm showing you to train

Oriol Vinyals (25:51.220)

is a sequence of actions and observations,

Lex Fridman (25:53.500)

then you're predicting what's the next action

Lex Fridman (25:55.600)

and the next observation, right?

Lex Fridman (25:57.100)

So you think of these really as a sequence of bites, right?

Lex Fridman (26:00.880)

So take any sequence of words,

Lex Fridman (26:04.220)

a sequence of interleaved words and images,

Oriol Vinyals (26:07.000)

a sequence of maybe observations that are images

Lex Fridman (26:11.280)

and moves in Atari up, down, left, right.

Lex Fridman (26:14.280)

And these you just think of them as bites

Lex Fridman (26:17.640)

and you're modeling what's the next bite gonna be like.

Lex Fridman (26:20.580)

And you might interpret that as an action

Lex Fridman (26:23.440)

and then play it in a game,

Oriol Vinyals (26:25.880)

or you could interpret it as a word

Lex Fridman (26:27.720)

and then write it down

Oriol Vinyals (26:29.120)

if you're chatting with the system and so on.

Lex Fridman (26:32.480)

So Gato basically can be thought as inputs,

Oriol Vinyals (26:36.600)

images, text, video, actions.

Lex Fridman (26:41.480)

It also actually inputs some sort of proprioception sensors

Oriol Vinyals (26:45.800)

from robotics because robotics is one of the tasks

Lex Fridman (26:48.280)

that it's been trained to do.

Lex Fridman (26:49.860)

And then at the output, similarly,

Lex Fridman (26:51.920)

it outputs words, actions.

Oriol Vinyals (26:53.720)

It does not output images, that's just by design,

Lex Fridman (26:57.440)

we decided not to go that way for now.

Oriol Vinyals (27:00.880)

That's also in part why it's the beginning

Lex Fridman (27:02.760)

because there's more to do clearly.

Lex Fridman (27:04.920)

But that's kind of what the Gato is,

Lex Fridman (27:06.440)

is this brain that essentially you give it any sequence

Oriol Vinyals (27:09.200)

of these observations and modalities

Lex Fridman (27:11.940)

and it outputs the next step.

Lex Fridman (27:13.760)

And then off you go, you feed the next step into

Lex Fridman (27:17.380)

and predict the next one and so on.

Oriol Vinyals (27:20.060)

Now, it is more than a language model

Lex Fridman (27:24.160)

because even though you can chat with Gato,

Oriol Vinyals (27:26.780)

like you can chat with Chinchilla or Flamingo,

Lex Fridman (27:30.520)

it also is an agent, right?

Lex Fridman (27:33.200)

So that's why we call it A of Gato,

Lex Fridman (27:37.200)

like the letter A and also it's general.

Oriol Vinyals (27:41.340)

It's not an agent that's been trained to be good

Lex Fridman (27:43.960)

at only StarCraft or only Atari or only Go.

Oriol Vinyals (27:47.860)

It's been trained on a vast variety of datasets.

Lex Fridman (27:51.640)

What makes it an agent, if I may interrupt,

Lex Fridman (27:53.840)

the fact that it can generate actions?

Lex Fridman (27:56.000)

Yes, so when we call it, I mean, it's a good question, right?

Lex Fridman (28:00.080)

When do we call a model?

Lex Fridman (28:02.760)

I mean, everything is a model,

Lex Fridman (28:03.840)

but what is an agent in my view is indeed the capacity

Lex Fridman (28:07.360)

to take actions in an environment that you then send to it

Lex Fridman (28:11.680)

and then the environment might return

Lex Fridman (28:13.480)

with a new observation

Lex Fridman (28:15.040)

and then you generate the next action.

Lex Fridman (28:17.560)

This actually, this reminds me of the question

Lex Fridman (28:20.440)

from the side of biology, what is life?

Lex Fridman (28:23.000)

Which is actually a very difficult question as well.

Lex Fridman (28:25.380)

What is living, what is living when you think about life

Lex Fridman (28:29.200)

here on this planet Earth?

Lex Fridman (28:31.000)

And a question interesting to me about aliens,

Lex Fridman (28:33.420)

what is life when we visit another planet?

Lex Fridman (28:35.720)

Would we be able to recognize it?

Lex Fridman (28:37.200)

And this feels like, it sounds perhaps silly,

Lex Fridman (28:40.220)

but I don't think it is.

Lex Fridman (28:41.360)

At which point is the neural network a being versus a tool?

Lex Fridman (28:48.260)

And it feels like action, ability to modify its environment

Lex Fridman (28:52.400)

is that fundamental leap.

Oriol Vinyals (28:54.560)

Yeah, I think it certainly feels like action

Lex Fridman (28:57.440)

is a necessary condition to be more alive,

Lex Fridman (29:01.960)

but probably not sufficient either.

Lex Fridman (29:04.400)

So sadly I...

Oriol Vinyals (29:05.240)

It's a soul consciousness thing, whatever.

Lex Fridman (29:06.880)

Yeah, yeah, we can get back to that later.

Lex Fridman (29:09.080)

But anyways, going back to the meow and the gato, right?

Lex Fridman (29:12.320)

So one of the leaps forward and what took the team a lot

Oriol Vinyals (29:17.640)

of effort and time was, as you were asking,

Lex Fridman (29:21.280)

how has gato been trained?

Lex Fridman (29:23.080)

So I told you gato is this transformer neural network,

Lex Fridman (29:26.080)

models actions, sequences of actions, words, et cetera.

Lex Fridman (29:30.840)

And then the way we train it is by essentially pulling

Lex Fridman (29:35.520)

data sets of observations, right?

Lex Fridman (29:39.400)

So it's a massive imitation learning algorithm

Lex Fridman (29:42.600)

that it imitates obviously to what

Oriol Vinyals (29:45.320)

is the next word that comes next from the usual data

Lex Fridman (29:48.520)

sets we use before, right?

Lex Fridman (29:50.160)

So these are these web scale style data sets of people

Lex Fridman (29:54.600)

writing on webs or chatting or whatnot, right?

Lex Fridman (29:58.520)

So that's an obvious source that we use on all language work.

Lex Fridman (2:00:01.040)

into some certain style of research

Oriol Vinyals (2:00:03.240)

pays out, is good to be practical sometimes.

Lex Fridman (2:00:06.400)

And I actually think Ilya and myself are practical,

Lex Fridman (2:00:09.400)

but it's also good.

Lex Fridman (2:00:10.440)

There's some sort of long term belief and trajectory.

Oriol Vinyals (2:00:14.840)

Obviously, there's a bit of lack involved,

Lex Fridman (2:00:16.720)

but it might be that that's the right path.

Oriol Vinyals (2:00:18.840)

Then you clearly are ahead and hugely influential to the field

Lex Fridman (2:00:22.320)

as he has been.

Lex Fridman (2:00:23.560)

Do you agree with that intuition that maybe

Lex Fridman (2:00:26.440)

was written about by Rich Sutton in The Bitter Lesson,

Oriol Vinyals (2:00:33.600)

that the biggest lesson that can be read from 70 years of AI

Lex Fridman (2:00:36.480)

research is that general methods that leverage computation

Lex Fridman (2:00:40.080)

are ultimately the most effective?

Lex Fridman (2:00:42.800)

Do you think that intuition is ultimately correct?

Oriol Vinyals (2:00:48.560)

General methods that leverage computation,

Lex Fridman (2:00:52.240)

allowing the scaling of computation

Oriol Vinyals (2:00:54.360)

to do a lot of the work.

Lex Fridman (2:00:56.240)

And so the basic task of us humans

Oriol Vinyals (2:00:59.640)

is to design methods that are more

Lex Fridman (2:01:01.440)

and more general versus more and more specific to the tasks

Oriol Vinyals (2:01:05.960)

at hand.

Lex Fridman (2:01:07.040)

I certainly think this essentially mimics

Oriol Vinyals (2:01:10.320)

a bit of the deep learning research,

Lex Fridman (2:01:14.680)

almost like philosophy, that on the one hand,

Oriol Vinyals (2:01:18.840)

we want to be data agnostic.

Lex Fridman (2:01:20.480)

We don't want to preprocess data sets.

Oriol Vinyals (2:01:22.160)

We want to see the bytes, the true data as it is,

Lex Fridman (2:01:25.560)

and then learn everything on top.

Lex Fridman (2:01:27.440)

So very much agree with that.

Lex Fridman (2:01:30.120)

And I think scaling up feels, at the very least, again,

Oriol Vinyals (2:01:33.360)

necessary for building incredible complex systems.

Lex Fridman (2:01:38.960)

It's possibly not sufficient, barring that we

Oriol Vinyals (2:01:42.880)

need a couple of breakthroughs.

Lex Fridman (2:01:45.080)

I think Reed Sutton mentioned search

Oriol Vinyals (2:01:47.960)

being part of the equation of scale and search.

Lex Fridman (2:01:52.200)

I think search, I've seen it, that's

Oriol Vinyals (2:01:55.720)

been more mixed in my experience.

Lex Fridman (2:01:57.400)

So from that lesson in particular,

Oriol Vinyals (2:01:59.320)

search is a bit more tricky because it

Lex Fridman (2:02:02.480)

is very appealing to search in domains like Go,

Oriol Vinyals (2:02:05.320)

where you have a clear reward function that you can then

Lex Fridman (2:02:08.080)

discard some search traces.

Lex Fridman (2:02:10.560)

But then in some other tasks, it's

Lex Fridman (2:02:13.160)

not very clear how you would do that,

Oriol Vinyals (2:02:15.160)

although recently one of our recent works, which actually

Lex Fridman (2:02:19.320)

was mostly mimicking or a continuation,

Lex Fridman (2:02:22.120)

and even the team and the people involved were pretty much very

Lex Fridman (2:02:25.840)

intersecting with AlphaStar, was AlphaCode,

Oriol Vinyals (2:02:28.400)

in which we actually saw the bitter lesson how

Lex Fridman (2:02:31.440)

scale of the models and then a massive amount of search

Oriol Vinyals (2:02:34.240)

yielded this kind of very interesting result

Lex Fridman (2:02:36.760)

of being able to have human level code competition.

Lex Fridman (2:02:41.280)

So I've seen examples of it being

Lex Fridman (2:02:43.640)

literally mapped to search and scale.

Oriol Vinyals (2:02:46.320)

I'm not so convinced about the search bit,

Lex Fridman (2:02:48.120)

but certainly I'm convinced scale will be needed.

Lex Fridman (2:02:51.000)

So we need general methods.

Lex Fridman (2:02:52.600)

We need to test them, and maybe we

Oriol Vinyals (2:02:54.080)

need to make sure that we can scale them given the hardware

Lex Fridman (2:02:57.080)

that we have in practice.

Lex Fridman (2:02:59.080)

But then maybe we should also shape how the hardware looks

Lex Fridman (2:03:01.920)

like based on which methods might be needed to scale.

Lex Fridman (2:03:05.640)

And that's an interesting contrast of these GPU comments

Lex Fridman (2:03:11.600)

that is we got it for free almost because games

Oriol Vinyals (2:03:14.280)

were using these.

Lex Fridman (2:03:15.080)

But maybe now if sparsity is required,

Oriol Vinyals (2:03:19.440)

we don't have the hardware.

Lex Fridman (2:03:20.560)

Although in theory, many people are

Oriol Vinyals (2:03:22.840)

building different kinds of hardware these days.

Lex Fridman (2:03:24.800)

But there's a bit of this notion of hardware lottery

Oriol Vinyals (2:03:27.760)

for scale that might actually have an impact at least

Lex Fridman (2:03:31.560)

on the scale of years on how fast we will make progress

Oriol Vinyals (2:03:35.240)

to maybe a version of neural nets

Lex Fridman (2:03:37.680)

or whatever comes next that might enable

Oriol Vinyals (2:03:41.920)

truly intelligent agents.

Lex Fridman (2:03:44.360)

Do you think in your lifetime we will build an AGI system that

Oriol Vinyals (2:03:50.520)

would undeniably be a thing that achieves human level

Lex Fridman (2:03:55.640)

intelligence and goes far beyond?

Oriol Vinyals (2:03:58.480)

I definitely think it's possible that it will go far beyond.

Lex Fridman (2:04:03.720)

But I'm definitely convinced that it will

Oriol Vinyals (2:04:05.520)

be human level intelligence.

Lex Fridman (2:04:08.480)

And I'm hypothesizing about the beyond

Oriol Vinyals (2:04:11.000)

because the beyond bit is a bit tricky to define,

Lex Fridman (2:04:16.520)

especially when we look at the current formula of starting

Oriol Vinyals (2:04:21.280)

from this imitation learning standpoint.

Lex Fridman (2:04:23.760)

So we can certainly imitate humans at language and beyond.

Lex Fridman (2:04:30.760)

So getting at human level through imitation

Lex Fridman (2:04:33.440)

feels very possible.

Oriol Vinyals (2:04:34.920)

Going beyond will require reinforcement learning

Lex Fridman (2:04:39.120)

and other things.

Lex Fridman (2:04:39.880)

And I think in some areas that certainly already has paid out.

Lex Fridman (2:04:43.600)

I mean, Go being an example that's

Oriol Vinyals (2:04:46.000)

my favorite so far in terms of going

Lex Fridman (2:04:48.240)

beyond human capabilities.

Lex Fridman (2:04:50.440)

But in general, I'm not sure we can define reward functions

Lex Fridman (2:04:55.600)

that from a seed of imitating human level

Oriol Vinyals (2:04:59.360)

intelligence that is general and then going beyond.

Lex Fridman (2:05:02.920)

That bit is not so clear in my lifetime.

Lex Fridman (2:05:05.280)

But certainly, human level, yes.

Lex Fridman (2:05:08.240)

And I mean, that in itself is already quite powerful,

Oriol Vinyals (2:05:11.000)

I think.

Lex Fridman (2:05:11.520)

So going beyond, I think it's obviously not.

Oriol Vinyals (2:05:14.560)

We're not going to not try that if then we

Lex Fridman (2:05:17.680)

get to superhuman scientists and discovery

Lex Fridman (2:05:20.760)

and advancing the world.

Lex Fridman (2:05:22.160)

But at least human level in general

Oriol Vinyals (2:05:25.600)

is also very, very powerful.

Lex Fridman (2:05:27.560)

Well, especially if human level or slightly beyond

Oriol Vinyals (2:05:31.560)

is integrated deeply with human society

Lex Fridman (2:05:33.760)

and there's billions of agents like that,

Lex Fridman (2:05:36.520)

do you think there's a singularity moment beyond which

Lex Fridman (2:05:39.960)

our world will be just very deeply transformed

Lex Fridman (2:05:44.200)

by these kinds of systems?

Lex Fridman (2:05:45.640)

Because now you're talking about intelligence systems

Oriol Vinyals (2:05:47.840)

that are just, I mean, this is no longer just going

Lex Fridman (2:05:53.040)

from horse and buggy to the car.

Oriol Vinyals (2:05:56.440)

It feels like a very different kind of shift

Lex Fridman (2:05:59.760)

in what it means to be a living entity on Earth.

Lex Fridman (2:06:03.280)

Are you afraid?

Lex Fridman (2:06:04.240)

Are you excited of this world?

Oriol Vinyals (2:06:06.280)

I'm afraid if there's a lot more.

Lex Fridman (2:06:09.360)

So I think maybe we'll need to think about if we truly

Oriol Vinyals (2:06:13.680)

get there just thinking of limited resources

Lex Fridman (2:06:18.400)

like humanity clearly hit some limits

Lex Fridman (2:06:21.480)

and then there's some balance, hopefully,

Lex Fridman (2:06:23.440)

that biologically the planet is imposing.

Lex Fridman (2:06:26.320)

And we should actually try to get better at this.

Lex Fridman (2:06:28.600)

As we know, there's quite a few issues

Oriol Vinyals (2:06:31.600)

with having too many people coexisting

Lex Fridman (2:06:35.840)

in a resource limited way.

Lex Fridman (2:06:37.720)

So for digital entities, it's an interesting question.

Lex Fridman (2:06:40.360)

I think such a limit maybe should exist.

Lex Fridman (2:06:43.520)

But maybe it's going to be imposed by energy availability

Lex Fridman (2:06:47.680)

because this also consumes energy.

Oriol Vinyals (2:06:49.760)

In fact, most systems are more inefficient

Lex Fridman (2:06:53.560)

than we are in terms of energy required.

Lex Fridman (2:06:56.720)

But definitely, I think as a society,

Lex Fridman (2:06:59.480)

we'll need to just work together to find

Lex Fridman (2:07:03.520)

what would be reasonable in terms of growth

Lex Fridman (2:07:06.400)

or how we coexist if that is to happen.

Oriol Vinyals (2:07:11.400)

I am very excited about, obviously,

Lex Fridman (2:07:14.640)

the aspects of automation that make people

Oriol Vinyals (2:07:17.720)

that obviously don't have access to certain resources

Lex Fridman (2:07:20.120)

or knowledge, for them to have that access.

Oriol Vinyals (2:07:23.920)

I think those are the applications in a way

Lex Fridman (2:07:26.280)

that I'm most excited to see and to personally work towards.

Oriol Vinyals (2:07:30.960)

Yeah, there's going to be significant improvements

Lex Fridman (2:07:32.640)

in productivity and the quality of life

Oriol Vinyals (2:07:34.320)

across the whole population, which is very interesting.

Lex Fridman (2:07:36.960)

But I'm looking even far beyond

Oriol Vinyals (2:07:39.200)

us becoming a multiplanetary species.

Lex Fridman (2:07:42.680)

And just as a quick bet, last question.

Lex Fridman (2:07:45.360)

Do you think as humans become multiplanetary species,

Lex Fridman (2:07:49.200)

go outside our solar system, all that kind of stuff,

Lex Fridman (2:07:52.480)

do you think there will be more humans

Lex Fridman (2:07:54.440)

or more robots in that future world?

Lex Fridman (2:07:57.200)

So will humans be the quirky, intelligent being of the past

Lex Fridman (2:08:04.480)

or is there something deeply fundamental

Oriol Vinyals (2:08:07.000)

to human intelligence that's truly special,

Lex Fridman (2:08:09.560)

where we will be part of those other planets,

Lex Fridman (2:08:12.120)

not just AI systems?

Lex Fridman (2:08:13.920)

I think we're all excited to build AGI

Oriol Vinyals (2:08:18.640)

to empower or make us more powerful as human species.

Lex Fridman (2:08:25.080)

Not to say there might be some hybridization.

Oriol Vinyals (2:08:27.560)

I mean, this is obviously speculation,

Lex Fridman (2:08:29.680)

but there are companies also trying to,

Oriol Vinyals (2:08:32.480)

the same way medicine is making us better.

Lex Fridman (2:08:35.640)

Maybe there are other things that are yet to happen on that.

Lex Fridman (2:08:39.080)

But if the ratio is not at most one to one,

Lex Fridman (2:08:43.320)

I would not be happy.

Lex Fridman (2:08:44.520)

So I would hope that we are part of the equation,

Lex Fridman (2:08:49.200)

but maybe there's maybe a one to one ratio feels

Oriol Vinyals (2:08:53.280)

like possible, constructive and so on,

Lex Fridman (2:08:56.200)

but it would not be good to have a misbalance,

Oriol Vinyals (2:08:59.600)

at least from my core beliefs and the why I'm doing

Lex Fridman (2:09:03.280)

what I'm doing when I go to work and I research

Lex Fridman (2:09:05.760)

what I research.

Lex Fridman (2:09:07.120)

Well, this is how I know you're human

Lex Fridman (2:09:09.520)

and this is how you've passed the Turing test.

Lex Fridman (2:09:12.720)

And you are one of the special humans, Oriel.

Oriol Vinyals (2:09:15.000)

It's a huge honor that you would talk with me

Lex Fridman (2:09:17.120)

and I hope we get the chance to speak again,

Oriol Vinyals (2:09:19.920)

maybe once before the singularity, once after

Lex Fridman (2:09:23.040)

and see how our view of the world changes.

Oriol Vinyals (2:09:25.440)

Thank you again for talking today.

Lex Fridman (2:09:26.600)

Thank you for the amazing work you do.

Oriol Vinyals (2:09:28.200)

You're a shining example of a research

Lex Fridman (2:09:31.320)

and a human being in this community.

Oriol Vinyals (2:09:32.960)

Thanks a lot.

Lex Fridman (2:09:33.800)

Like yeah, looking forward to before the singularity

Oriol Vinyals (2:09:36.240)

certainly and maybe after.

Lex Fridman (2:09:39.920)

Thanks for listening to this conversation

Oriol Vinyals (2:09:41.480)

with Oriel Venialis.

Lex Fridman (2:09:43.120)

To support this podcast, please check out our sponsors

Oriol Vinyals (2:09:45.520)

in the description.

Lex Fridman (2:09:46.960)

And now let me leave you with some words from Alan Turing.

Oriol Vinyals (2:09:51.160)

Those who can imagine anything can create the impossible.

Lex Fridman (2:09:55.080)

Thank you for listening and hope to see you next time.

Lex Fridman (30:02.000)

But then we also took a lot of agents

Lex Fridman (30:05.640)

that we have at DeepMind.

Oriol Vinyals (30:06.720)

I mean, as you know, DeepMind, we're quite interested

Lex Fridman (30:10.920)

in learning reinforcement learning and learning agents

Oriol Vinyals (30:14.960)

that play in different environments.

Lex Fridman (30:17.040)

So we kind of created a data set of these trajectories,

Oriol Vinyals (30:20.760)

as we call them, or agent experiences.

Lex Fridman (30:23.120)

So in a way, there are other agents

Oriol Vinyals (30:25.240)

we train for a single mind purpose to, let's say,

Lex Fridman (30:29.560)

control a 3D game environment and navigate a maze.

Lex Fridman (30:33.320)

So we had all the experience that

Lex Fridman (30:35.320)

was created through one agent interacting

Oriol Vinyals (30:38.160)

with that environment.

Lex Fridman (30:39.480)

And we added this to the data set, right?

Lex Fridman (30:41.800)

And as I said, we just see all the data,

Lex Fridman (30:44.400)

all these sequences of words or sequences

Oriol Vinyals (30:46.440)

of this agent interacting with that environment or agents

Lex Fridman (30:51.120)

playing Atari and so on.

Oriol Vinyals (30:52.200)

We see it as the same kind of data.

Lex Fridman (30:54.920)

And so we mix these data sets together.

Lex Fridman (30:57.440)

And we train Gato.

Lex Fridman (31:00.120)

That's the G part, right?

Oriol Vinyals (31:01.600)

It's general because it really has mixed.

Lex Fridman (31:05.240)

It doesn't have different brains for each modality

Oriol Vinyals (31:07.560)

or each narrow task.

Lex Fridman (31:09.040)

It has a single brain.

Oriol Vinyals (31:10.480)

It's not that big of a brain compared

Lex Fridman (31:12.120)

to most of the neural networks we see these days.

Oriol Vinyals (31:14.760)

It has 1 billion parameters.

Lex Fridman (31:18.200)

Some models we're seeing get in the trillions these days.

Lex Fridman (31:21.120)

And certainly, 100 billion feels like a size

Lex Fridman (31:25.080)

that is very common from when you train these jobs.

Lex Fridman (31:29.000)

So the actual agent is relatively small.

Lex Fridman (31:32.640)

But it's been trained on a very challenging, diverse data set,

Oriol Vinyals (31:36.280)

not only containing all of the internet

Lex Fridman (31:38.000)

but containing all these agent experience playing

Oriol Vinyals (31:40.680)

very different, distinct environments.

Lex Fridman (31:43.240)

So this brings us to the part of the tweet of this

Oriol Vinyals (31:46.720)

is not the end, it's the beginning.

Lex Fridman (31:48.880)

It feels very cool to see Gato, in principle,

Oriol Vinyals (31:53.120)

is able to control any sort of environments, especially

Lex Fridman (31:57.360)

the ones that it's been trained to do, these 3D games, Atari

Oriol Vinyals (32:00.360)

games, all sorts of robotics tasks, and so on.

Lex Fridman (32:04.600)

But obviously, it's not as proficient

Oriol Vinyals (32:07.760)

as the teachers it learned from on these environments.

Lex Fridman (32:10.520)

Not obvious.

Oriol Vinyals (32:11.680)

It's not obvious that it wouldn't be more proficient.

Lex Fridman (32:15.040)

It's just the current beginning part

Oriol Vinyals (32:17.960)

is that the performance is such that it's not as good

Lex Fridman (32:21.760)

as if it's specialized to that task.

Oriol Vinyals (32:23.360)

Right.

Lex Fridman (32:23.840)

So it's not as good, although I would argue size matters here.

Lex Fridman (32:28.040)

So the fact that I would argue always size always matters.

Lex Fridman (32:31.360)

That's a different conversation.

Lex Fridman (32:33.320)

But for neural networks, certainly size does matter.

Lex Fridman (32:36.200)

So it's the beginning because it's relatively small.

Lex Fridman (32:39.600)

So obviously, scaling this idea up

Lex Fridman (32:42.560)

might make the connections that exist between text

Oriol Vinyals (32:48.240)

on the internet and playing Atari and so on more

Lex Fridman (32:51.560)

synergistic with one another.

Lex Fridman (32:53.320)

And you might gain.

Lex Fridman (32:54.240)

And that moment, we didn't quite see.

Lex Fridman (32:56.320)

But obviously, that's why it's the beginning.

Lex Fridman (32:58.600)

That synergy might emerge with scale.

Oriol Vinyals (33:00.920)

Right, might emerge with scale.

Lex Fridman (33:02.160)

And also, I believe there's some new research or ways

Oriol Vinyals (33:05.160)

in which you prepare the data that you

Lex Fridman (33:08.480)

might need to make it more clear to the model

Oriol Vinyals (33:11.560)

that you're not only playing Atari,

Lex Fridman (33:14.120)

and you start from a screen.

Lex Fridman (33:16.240)

And here is up and a screen and down.

Lex Fridman (33:18.400)

Maybe you can think of playing Atari

Oriol Vinyals (33:20.640)

as there's some sort of context that is needed for the agent

Lex Fridman (33:23.880)

before it starts seeing, oh, this is an Atari screen.

Oriol Vinyals (33:26.920)

I'm going to start playing.

Lex Fridman (33:28.640)

You might require, for instance, to be told in words,

Oriol Vinyals (33:33.360)

hey, in this sequence that I'm showing,

Lex Fridman (33:36.840)

you're going to be playing an Atari game.

Lex Fridman (33:39.080)

So text might actually be a good driver to enhance the data.

Lex Fridman (33:44.400)

So then these connections might be made more easily.

Lex Fridman (33:46.960)

So that's an idea that we start seeing in language.

Lex Fridman (33:51.200)

But obviously, beyond, this is going to be effective.

Oriol Vinyals (33:55.080)

It's not like I don't show you a screen,

Lex Fridman (33:57.400)

and you, from scratch, you're supposed to learn a game.

Oriol Vinyals (34:01.000)

There is a lot of context we might set.

Lex Fridman (34:03.360)

So there might be some work needed as well

Oriol Vinyals (34:05.840)

to set that context.

Lex Fridman (34:07.720)

But anyways, there's a lot of work.

Lex Fridman (34:10.640)

So that context puts all the different modalities

Lex Fridman (34:13.520)

on the same level ground to provide the context best.

Lex Fridman (34:16.640)

So maybe on that point, so there's

Lex Fridman (34:19.880)

this task, which may not seem trivial, of tokenizing the data,

Oriol Vinyals (34:25.480)

of converting the data into pieces,

Lex Fridman (34:28.480)

into basic atomic elements that then could cross modalities

Oriol Vinyals (34:34.400)

somehow.

Lex Fridman (34:35.240)

So what's tokenization?

Lex Fridman (34:37.840)

How do you tokenize text?

Lex Fridman (34:39.640)

How do you tokenize images?

Lex Fridman (34:42.160)

How do you tokenize games and actions and robotics tasks?

Lex Fridman (34:47.080)

Yeah, that's a great question.

Lex Fridman (34:48.320)

So tokenization is the entry point

Lex Fridman (34:52.840)

to actually make all the data look like a sequence,

Oriol Vinyals (34:55.600)

because tokens then are just these little puzzle pieces.

Lex Fridman (34:59.480)

We break down anything into these puzzle pieces,

Lex Fridman (35:01.760)

and then we just model, what's this puzzle look like when

Lex Fridman (35:05.560)

you make it lay down in a line, so to speak, in a sequence?

Lex Fridman (35:09.600)

So in Gato, the text, there's a lot of work.

Lex Fridman (35:15.440)

You tokenize text usually by looking

Lex Fridman (35:17.400)

at commonly used substrings, right?

Lex Fridman (35:20.040)

So there's ING in English is a very common substring,

Lex Fridman (35:23.680)

so that becomes a token.

Lex Fridman (35:25.480)

There's quite a well studied problem on tokenizing text.

Lex Fridman (35:29.080)

And Gato just used the standard techniques

Lex Fridman (35:31.560)

that have been developed from many years,

Oriol Vinyals (35:34.280)

even starting from ngram models in the 1950s and so on.

Lex Fridman (35:38.000)

Just for context, how many tokens,

Lex Fridman (35:40.440)

what order, magnitude, number of tokens

Lex Fridman (35:42.640)

is required for a word, usually?

Lex Fridman (35:45.120)

What are we talking about?

Lex Fridman (35:46.200)

Yeah, for a word in English, I mean,

Oriol Vinyals (35:48.880)

every language is very different.

Lex Fridman (35:51.120)

The current level or granularity of tokenization

Oriol Vinyals (35:53.920)

generally means it's maybe two to five.

Lex Fridman (35:57.840)

I mean, I don't know the statistics exactly,

Lex Fridman (36:00.200)

but to give you an idea, we don't tokenize

Lex Fridman (36:03.000)

at the level of letters.

Oriol Vinyals (36:04.160)

Then it would probably be, I don't

Lex Fridman (36:05.720)

know what the average length of a word is in English,

Lex Fridman (36:08.080)

but that would be the minimum set of tokens you could use.

Lex Fridman (36:11.400)

It was bigger than letters, smaller than words.

Oriol Vinyals (36:13.200)

Yes, yes.

Lex Fridman (36:13.880)

And you could think of very, very common words like the.

Oriol Vinyals (36:16.840)

I mean, that would be a single token,

Lex Fridman (36:18.760)

but very quickly you're talking two, three, four tokens or so.

Lex Fridman (36:22.360)

Have you ever tried to tokenize emojis?

Lex Fridman (36:24.680)

Emojis are actually just sequences of letters, so.

Oriol Vinyals (36:30.080)

Maybe to you, but to me they mean so much more.

Lex Fridman (36:33.000)

Yeah, you can render the emoji, but you

Oriol Vinyals (36:35.080)

might if you actually just.

Lex Fridman (36:36.840)

Yeah, this is a philosophical question.

Lex Fridman (36:39.360)

Is emojis an image or a text?

Lex Fridman (36:43.320)

The way we do these things is they're actually

Oriol Vinyals (36:46.640)

mapped to small sequences of characters.

Lex Fridman (36:49.520)

So you can actually play with these models

Lex Fridman (36:52.600)

and input emojis, it will output emojis back,

Lex Fridman (36:55.760)

which is actually quite a fun exercise.

Oriol Vinyals (36:57.960)

You probably can find other tweets about these out there.

Lex Fridman (37:02.240)

But yeah, so anyways, text.

Oriol Vinyals (37:04.440)

It's very clear how this is done.

Lex Fridman (37:06.720)

And then in Gato, what we did for images

Oriol Vinyals (37:10.560)

is we map images to essentially we compressed images,

Lex Fridman (37:14.880)

so to speak, into something that looks more like less

Oriol Vinyals (37:19.120)

like every pixel with every intensity.

Lex Fridman (37:21.320)

That would mean we have a very long sequence, right?

Oriol Vinyals (37:23.840)

Like if we were talking about 100 by 100 pixel images,

Lex Fridman (37:27.320)

that would make the sequences far too long.

Lex Fridman (37:30.000)

So what was done there is you just

Lex Fridman (37:32.520)

use a technique that essentially compresses an image

Oriol Vinyals (37:35.760)

into maybe 16 by 16 patches of pixels,

Lex Fridman (37:40.160)

and then that is mapped, again, tokenized.

Oriol Vinyals (37:42.720)

You just essentially quantize this space

Lex Fridman (37:45.360)

into a special word that actually

Oriol Vinyals (37:48.720)

maps to these little sequence of pixels.

Lex Fridman (37:51.720)

And then you put the pixels together in some raster order,

Lex Fridman (37:55.120)

and then that's how you get out or in the image

Lex Fridman (37:59.360)

that you're processing.

Lex Fridman (38:00.720)

But there's no semantic aspect to that,

Lex Fridman (38:04.040)

so you're doing some kind of,

Oriol Vinyals (38:05.840)

you don't need to understand anything about the image

Lex Fridman (38:07.760)

in order to tokenize it currently.

Oriol Vinyals (38:09.640)

No, you're only using this notion of compression.

Lex Fridman (38:12.600)

So you're trying to find common,

Oriol Vinyals (38:15.080)

it's like JPG or all these algorithms.

Lex Fridman (38:17.640)

It's actually very similar at the tokenization level.

Oriol Vinyals (38:20.520)

All we're doing is finding common patterns

Lex Fridman (38:23.320)

and then making sure in a lossy way we compress these images

Oriol Vinyals (38:27.200)

given the statistics of the images

Lex Fridman (38:29.480)

that are contained in all the data we deal with.

Oriol Vinyals (38:31.840)

Although you could probably argue that JPEG

Lex Fridman (38:34.200)

does have some understanding of images.

Oriol Vinyals (38:38.720)

Because visual information, maybe color,

Lex Fridman (38:44.000)

compressing crudely based on color

Oriol Vinyals (38:46.920)

does capture something important about an image

Lex Fridman (38:51.160)

that's about its meaning, not just about some statistics.

Oriol Vinyals (38:54.640)

Yeah, I mean, JP, as I said,

Lex Fridman (38:56.640)

the algorithms look actually very similar

Oriol Vinyals (38:58.640)

to they use the cosine transform in JPG.

Lex Fridman (39:04.120)

The approach we usually do in machine learning

Oriol Vinyals (39:07.120)

when we deal with images and we do this quantization step

Lex Fridman (39:10.120)

is a bit more data driven.

Lex Fridman (39:11.400)

So rather than have some sort of Fourier basis

Lex Fridman (39:14.120)

for how frequencies appear in the natural world,

Oriol Vinyals (39:18.880)

we actually just use the statistics of the images

Lex Fridman (39:23.840)

and then quantize them based on the statistics,

Lex Fridman (39:27.000)

much like you do in words, right?

Lex Fridman (39:28.280)

So common substrings are allocated a token

Lex Fridman (39:32.400)

and images is very similar.

Lex Fridman (39:34.400)

But there's no connection.

Oriol Vinyals (39:36.960)

The token space, if you think of,

Lex Fridman (39:39.240)

oh, like the tokens are an integer

Lex Fridman (39:41.080)

and in the end of the day.

Lex Fridman (39:42.440)

So now like we work on, maybe we have about,

Oriol Vinyals (39:46.200)

let's say, I don't know the exact numbers,

Lex Fridman (39:48.000)

but let's say 10,000 tokens for text, right?

Oriol Vinyals (39:51.160)

Certainly more than characters

Lex Fridman (39:52.840)

because we have groups of characters and so on.

Lex Fridman (39:55.320)

So from one to 10,000, those are representing

Lex Fridman (39:58.280)

all the language and the words we'll see.

Lex Fridman (40:01.000)

And then images occupy the next set of integers.

Lex Fridman (40:04.160)

So they're completely independent, right?

Lex Fridman (40:05.800)

So from 10,001 to 20,000,

Lex Fridman (40:08.920)

those are the tokens that represent

Oriol Vinyals (40:10.640)

these other modality images.

Lex Fridman (40:12.760)

And that is an interesting aspect that makes it orthogonal.

Lex Fridman (40:18.640)

So what connects these concepts is the data, right?

Lex Fridman (40:21.600)

Once you have a data set,

Oriol Vinyals (40:23.760)

for instance, that captions images that tells you,

Lex Fridman (40:26.880)

oh, this is someone playing a frisbee on a green field.

Oriol Vinyals (40:30.480)

Now the model will need to predict the tokens

Lex Fridman (40:34.560)

from the text green field to then the pixels.

Lex Fridman (40:37.800)

And that will start making the connections

Lex Fridman (40:39.760)

between the tokens.

Lex Fridman (40:40.600)

So these connections happen as the algorithm learns.

Lex Fridman (40:43.640)

And then the last, if we think of these integers,

Oriol Vinyals (40:45.840)

the first few are words, the next few are images.

Lex Fridman (40:48.760)

In Gato, we also allocated the highest order of integers

Lex Fridman (40:55.240)

to actions, right?

Lex Fridman (40:56.240)

Which we discretize and actions are very diverse, right?

Oriol Vinyals (40:59.920)

In Atari, there's, I don't know if 17 discrete actions.

Lex Fridman (41:04.120)

In robotics, actions might be torques

Lex Fridman (41:06.960)

and forces that we apply.

Lex Fridman (41:08.240)

So we just use kind of similar ideas

Oriol Vinyals (41:11.200)

to compress these actions into tokens.

Lex Fridman (41:14.320)

And then we just, that's how we map now

Oriol Vinyals (41:18.000)

all the space to these sequence of integers.

Lex Fridman (41:20.800)

But they occupy different space

Lex Fridman (41:22.480)

and what connects them is then the learning algorithm.

Lex Fridman (41:24.840)

That's where the magic happens.

Lex Fridman (41:26.320)

So the modalities are orthogonal

Lex Fridman (41:28.840)

to each other in token space.

Lex Fridman (41:30.760)

So in the input, everything you add, you add extra tokens.

Lex Fridman (41:35.760)

And then you're shoving all of that into one place.

Oriol Vinyals (41:40.440)

Yes, the transformer.

Lex Fridman (41:41.640)

And that transformer, that transformer tries

Oriol Vinyals (41:46.400)

to look at this gigantic token space

Lex Fridman (41:49.360)

and tries to form some kind of representation,

Oriol Vinyals (41:52.240)

some kind of unique wisdom

Lex Fridman (41:56.760)

about all of these different modalities.

Lex Fridman (41:59.240)

How's that possible?

Lex Fridman (42:02.120)

If you were to sort of like put your psychoanalysis hat on

Lex Fridman (42:06.520)

and try to psychoanalyze this neural network,

Lex Fridman (42:09.400)

is it schizophrenic?

Oriol Vinyals (42:11.760)

Does it try to, given this very few weights,

Lex Fridman (42:17.160)

represent multiple disjoint things

Lex Fridman (42:19.560)

and somehow have them not interfere with each other?

Lex Fridman (42:22.800)

Or is it somehow building on the joint strength,

Lex Fridman (42:27.960)

on whatever is common to all the different modalities?

Lex Fridman (42:31.800)

Like what, if you were to ask a question,

Lex Fridman (42:34.520)

is it schizophrenic or is it of one mind?

Lex Fridman (42:38.720)

I mean, it is one mind and it's actually

Oriol Vinyals (42:42.640)

the simplest algorithm, which that's kind of in a way

Lex Fridman (42:46.800)

how it feels like the field hasn't changed

Oriol Vinyals (42:49.840)

since back propagation and gradient descent

Lex Fridman (42:52.600)

was purpose for learning neural networks.

Lex Fridman (42:55.760)

So there is obviously details on the architecture.

Lex Fridman (42:58.720)

This has evolved.

Oriol Vinyals (42:59.640)

The current iteration is still the transformer,

Lex Fridman (43:03.080)

which is a powerful sequence modeling architecture.

Lex Fridman (43:07.440)

But then the goal of this, you know,

Lex Fridman (43:11.000)

setting these weights to predict the data

Oriol Vinyals (43:13.840)

is essentially the same as basically I could describe.

Lex Fridman (43:17.240)

I mean, we described a few years ago,

Lex Fridman (43:18.680)

Alpha star language modeling and so on, right?

Lex Fridman (43:21.600)

We take, let's say an Atari game,

Oriol Vinyals (43:24.600)

we map it to a string of numbers

Lex Fridman (43:27.640)

that will all be probably image space

Lex Fridman (43:30.360)

and action space interleaved.

Lex Fridman (43:32.440)

And all we're gonna do is say, okay, given the numbers,

Oriol Vinyals (43:37.280)

you know, 10,001, 10,004, 10,005,

Lex Fridman (43:40.400)

the next number that comes is 20,006,

Oriol Vinyals (43:43.280)

which is in the action space.

Lex Fridman (43:45.400)

And you're just optimizing these weights

Oriol Vinyals (43:48.880)

via very simple gradients.

Lex Fridman (43:51.720)

Like, you know, mathematical is almost

Oriol Vinyals (43:53.520)

the most boring algorithm you could imagine.

Lex Fridman (43:55.880)

We settle the weights so that

Oriol Vinyals (43:57.800)

given this particular instance,

Lex Fridman (44:00.200)

these weights are set to maximize the probability

Oriol Vinyals (44:04.080)

of having seen this particular sequence of integers

Lex Fridman (44:07.280)

for this particular game.

Lex Fridman (44:09.120)

And then the algorithm does this

Lex Fridman (44:11.640)

for many, many, many iterations,

Lex Fridman (44:14.800)

looking at different modalities, different games, right?

Lex Fridman (44:17.920)

That's the mixture of the dataset we discussed.

Lex Fridman (44:20.480)

So in a way, it's a very simple algorithm

Lex Fridman (44:24.040)

and the weights, right, they're all shared, right?

Lex Fridman (44:27.560)

So in terms of, is it focusing on one modality or not?

Lex Fridman (44:30.920)

The intermediate weights that are converting

Oriol Vinyals (44:33.240)

from these input of integers

Lex Fridman (44:35.160)

to the target integer you're predicting next,

Oriol Vinyals (44:37.720)

those weights certainly are common.

Lex Fridman (44:40.320)

And then the way that tokenization happens,

Oriol Vinyals (44:43.400)

there is a special place in the neural network,

Lex Fridman (44:45.840)

which is we map this integer, like number 10,001,

Oriol Vinyals (44:49.800)

to a vector of real numbers.

Lex Fridman (44:51.920)

Like real numbers, we can optimize them

Lex Fridman (44:54.760)

with gradient descent, right?

Lex Fridman (44:56.120)

The functions we learn

Oriol Vinyals (44:57.080)

are actually surprisingly differentiable.

Lex Fridman (44:59.720)

That's why we compute gradients.

Lex Fridman (45:01.720)

So this step is the only one

Lex Fridman (45:03.920)

that this orthogonality you mentioned applies.

Lex Fridman (45:06.560)

So mapping a certain token for text or image or actions,

Lex Fridman (45:12.520)

each of these tokens gets its own little vector

Oriol Vinyals (45:15.040)

of real numbers that represents this.

Lex Fridman (45:17.200)

If you look at the field back many years ago,

Oriol Vinyals (45:19.560)

people were talking about word vectors or word embeddings.

Lex Fridman (45:23.480)

These are the same.

Oriol Vinyals (45:24.320)

We have word vectors or embeddings.

Lex Fridman (45:26.040)

We have image vector or embeddings

Lex Fridman (45:28.880)

and action vector of embeddings.

Lex Fridman (45:30.880)

And the beauty here is that as you train this model,

Oriol Vinyals (45:33.920)

if you visualize these little vectors,

Lex Fridman (45:36.640)

it might be that they start aligning

Oriol Vinyals (45:38.480)

even though they're independent parameters.

Lex Fridman (45:41.400)

There could be anything,

Lex Fridman (45:42.840)

but then it might be that you take the word gato or cat,

Lex Fridman (45:47.480)

which maybe is common enough

Oriol Vinyals (45:48.520)

that it actually has its own token.

Lex Fridman (45:50.200)

And then you take pixels that have a cat

Lex Fridman (45:52.400)

and you might start seeing

Lex Fridman (45:53.960)

that these vectors look like they align, right?

Lex Fridman (45:57.400)

So by learning from this vast amount of data,

Lex Fridman (46:00.640)

the model is realizing the potential connections

Oriol Vinyals (46:03.920)

between these modalities.

Lex Fridman (46:05.640)

Now, I will say there will be another way,

Oriol Vinyals (46:07.840)

at least in part, to not have these different vectors

Lex Fridman (46:13.160)

for each different modality.

Oriol Vinyals (46:15.520)

For instance, when I tell you about actions

Lex Fridman (46:18.360)

in certain space, I'm defining actions by words, right?

Lex Fridman (46:22.800)

So you could imagine a world in which I'm not learning

Lex Fridman (46:26.520)

that the action app in Atari is its own number.

Oriol Vinyals (46:31.240)

The action app in Atari maybe is literally the word

Lex Fridman (46:34.400)

or the sentence app in Atari, right?

Lex Fridman (46:37.320)

And that would mean we now leverage

Lex Fridman (46:39.400)

much more from the language.

Oriol Vinyals (46:41.040)

This is not what we did here,

Lex Fridman (46:42.520)

but certainly it might make these connections

Oriol Vinyals (46:45.680)

much easier to learn and also to teach the model

Lex Fridman (46:49.080)

to correct its own actions and so on, right?

Lex Fridman (46:51.280)

So all these to say that gato is indeed the beginning,

Lex Fridman (46:55.840)

that it is a radical idea to do this this way,

Lex Fridman (46:59.400)

but there's probably a lot more to be done

Lex Fridman (47:02.320)

and the results to be more impressive,

Oriol Vinyals (47:04.440)

not only through scale, but also through some new research

Lex Fridman (47:07.920)

that will come hopefully in the years to come.

Lex Fridman (47:10.480)

So just to elaborate quickly,

Lex Fridman (47:12.280)

you mean one possible next step

Oriol Vinyals (47:16.680)

or one of the paths that you might take next

Lex Fridman (47:20.200)

is doing the tokenization fundamentally

Oriol Vinyals (47:25.200)

as a kind of linguistic communication.

Lex Fridman (47:28.240)

So like you convert even images into language.

Lex Fridman (47:31.320)

So doing something like a crude semantic segmentation,

Lex Fridman (47:35.520)

trying to just assign a bunch of words to an image

Oriol Vinyals (47:38.360)

that like have almost like a dumb entity

Lex Fridman (47:42.280)

explaining as much as it can about the image.

Lex Fridman (47:45.320)

And so you convert that into words

Lex Fridman (47:46.920)

and then you convert games into words

Lex Fridman (47:49.280)

and then you provide the context in words and all of it.

Lex Fridman (47:52.120)

And eventually getting to a point

Oriol Vinyals (47:56.320)

where everybody agrees with Noam Chomsky

Lex Fridman (47:58.080)

that language is actually at the core of everything.

Oriol Vinyals (48:00.920)

That's it's the base layer of intelligence

Lex Fridman (48:04.240)

and consciousness and all that kind of stuff, okay.

Oriol Vinyals (48:07.520)

You mentioned early on like size, it's hard to grow.

Lex Fridman (48:11.240)

What did you mean by that?

Oriol Vinyals (48:12.800)

Because we're talking about scale might change.

Lex Fridman (48:17.000)

There might be, and we'll talk about this too,

Oriol Vinyals (48:18.960)

like there's a emergent, there's certain things

Lex Fridman (48:23.880)

about these neural networks that are emergent.

Lex Fridman (48:25.640)

So certain like performance we can see only with scale

Lex Fridman (48:28.960)

and there's some kind of threshold of scale.

Lex Fridman (48:30.960)

So why is it hard to grow something like this Meow network?

Lex Fridman (48:36.640)

So the Meow network, it's not hard to grow

Oriol Vinyals (48:41.120)

if you retrain it.

Lex Fridman (48:42.600)

What's hard is, well, we have now 1 billion parameters.

Oriol Vinyals (48:46.840)

We train them for a while.

Lex Fridman (48:48.120)

We spend some amount of work towards building these weights

Oriol Vinyals (48:53.120)

that are an amazing initial brain

Lex Fridman (48:55.840)

for doing these kinds of tasks we care about.

Lex Fridman (48:58.800)

Could we reuse the weights and expand to a larger brain?

Lex Fridman (49:03.880)

And that is extraordinarily hard,

Lex Fridman (49:06.680)

but also exciting from a research perspective

Lex Fridman (49:10.040)

and a practical perspective point of view, right?

Lex Fridman (49:12.520)

So there's this notion of modularity in software engineering

Lex Fridman (49:17.520)

and we starting to see some examples

Lex Fridman (49:20.360)

and work that leverages modularity.

Lex Fridman (49:23.160)

In fact, if we go back one step from Gato

Oriol Vinyals (49:26.200)

to a work that I would say train much larger,

Lex Fridman (49:29.560)

much more capable network called Flamingo.

Oriol Vinyals (49:32.400)

Flamingo did not deal with actions,

Lex Fridman (49:34.160)

but it definitely dealt with images in an interesting way,

Oriol Vinyals (49:38.280)

kind of akin to what Gato did,

Lex Fridman (49:40.120)

but slightly different technique for tokenizing,

Lex Fridman (49:42.840)

but we don't need to go into that detail.

Lex Fridman (49:45.280)

But what Flamingo also did, which Gato didn't do,

Lex Fridman (49:49.240)

and that just happens because these projects,

Lex Fridman (49:51.560)

they're different, it's a bit of like the exploratory nature

Oriol Vinyals (49:55.760)

of research, which is great.

Lex Fridman (49:57.120)

The research behind these projects is also modular.

Oriol Vinyals (50:00.480)

Yes, exactly.

Lex Fridman (50:01.720)

And it has to be, right?

Oriol Vinyals (50:02.640)

We need to have creativity

Lex Fridman (50:05.480)

and sometimes you need to protect pockets of people,

Oriol Vinyals (50:09.120)

researchers and so on.

Lex Fridman (50:10.200)

By we, you mean humans.

Oriol Vinyals (50:11.760)

Yes.

Lex Fridman (50:12.720)

And also in particular researchers

Lex Fridman (50:14.480)

and maybe even further DeepMind or other such labs.

Lex Fridman (50:18.720)

And then the neural networks themselves.

Lex Fridman (50:20.880)

So it's modularity all the way down.

Lex Fridman (50:23.480)

All the way down.

Lex Fridman (50:24.320)

So the way that we did modularity very beautifully

Lex Fridman (50:27.400)

in Flamingo is we took Chinchilla,

Oriol Vinyals (50:30.000)

which is a language only model, not an agent,

Lex Fridman (50:33.440)

if we think of actions being necessary for agency.

Lex Fridman (50:36.600)

So we took Chinchilla, we took the weights of Chinchilla

Lex Fridman (50:40.840)

and then we froze them.

Oriol Vinyals (50:42.640)

We said, these don't change.

Lex Fridman (50:44.680)

We train them to be very good at predicting the next word.

Oriol Vinyals (50:47.400)

It's a very good language model, state of the art

Lex Fridman (50:50.120)

at the time you release it, et cetera, et cetera.

Lex Fridman (50:52.800)

We're going to add a capability to see, right?

Lex Fridman (50:55.360)

We are going to add the ability to see

Oriol Vinyals (50:56.800)

to this language model.

Lex Fridman (50:58.200)

So we're going to attach small pieces of neural networks

Oriol Vinyals (51:01.800)

at the right places in the model.

Lex Fridman (51:03.760)

It's almost like I'm injecting the network

Oriol Vinyals (51:07.760)

with some weights and some substructures

Lex Fridman (51:10.640)

in a good way, right?

Lex Fridman (51:12.760)

So you need the research to say, what is effective?

Lex Fridman (51:15.160)

How do you add this capability

Oriol Vinyals (51:16.600)

without destroying others, et cetera.

Lex Fridman (51:18.720)

So we created a small sub network initialized,

Oriol Vinyals (51:24.280)

not from random, but actually from self supervised learning

Lex Fridman (51:28.680)

that a model that understands vision in general.

Lex Fridman (51:32.720)

And then we took data sets that connect the two modalities,

Lex Fridman (51:37.160)

vision and language.

Lex Fridman (51:38.680)

And then we froze the main part,

Lex Fridman (51:41.120)

the largest portion of the network, which was Chinchilla,

Oriol Vinyals (51:43.640)

that is 70 billion parameters.

Lex Fridman (51:45.880)

And then we added a few more parameters on top,

Oriol Vinyals (51:49.160)

trained from scratch, and then some others

Lex Fridman (51:51.360)

that were pre trained with the capacity to see,

Oriol Vinyals (51:55.200)

like it was not tokenization

Lex Fridman (51:57.320)

in the way I described for Gato, but it's a similar idea.

Lex Fridman (52:01.360)

And then we trained the whole system.

Lex Fridman (52:03.560)

Parts of it were frozen, parts of it were new.

Lex Fridman (52:06.520)

And all of a sudden, we developed Flamingo,

Lex Fridman (52:09.640)

which is an amazing model that is essentially,

Oriol Vinyals (52:12.520)

I mean, describing it is a chatbot

Lex Fridman (52:14.960)

where you can also upload images

Lex Fridman (52:16.920)

and start conversing about images.

Lex Fridman (52:19.880)

But it's also kind of a dialogue style chatbot.

Lex Fridman (52:23.680)

So the input is images and text and the output is text.

Lex Fridman (52:26.600)

Exactly.

Lex Fridman (52:28.480)

How many parameters, you said 70 billion for Chinchilla?

Lex Fridman (52:31.760)

Yeah, Chinchilla is 70 billion.

Lex Fridman (52:33.200)

And then the ones we add on top,

Lex Fridman (52:34.600)

which kind of almost is almost like a way

Oriol Vinyals (52:38.000)

to overwrite its little activations

Lex Fridman (52:40.920)

so that when it sees vision,

Oriol Vinyals (52:42.400)

it does kind of a correct computation of what it's seeing,

Lex Fridman (52:45.280)

mapping it back towards, so to speak.

Lex Fridman (52:47.920)

That adds an extra 10 billion parameters, right?

Lex Fridman (52:50.800)

So it's total 80 billion, the largest one we released.

Lex Fridman (52:53.920)

And then you train it on a few datasets

Lex Fridman (52:57.320)

that contain vision and language.

Lex Fridman (52:59.280)

And once you interact with the model,

Lex Fridman (53:01.120)

you start seeing that you can upload an image

Lex Fridman (53:04.160)

and start sort of having a dialogue about the image,

Lex Fridman (53:07.960)

which is actually not something,

Oriol Vinyals (53:09.480)

it's very similar and akin to what we saw in language only.

Lex Fridman (53:12.520)

These prompting abilities that it has,

Lex Fridman (53:15.240)

you can teach it a new vision task, right?

Lex Fridman (53:17.720)

It does things beyond the capabilities

Oriol Vinyals (53:20.440)

that in theory the datasets provided in themselves,

Lex Fridman (53:24.480)

but because it leverages a lot of the language knowledge

Oriol Vinyals (53:27.080)

acquired from Chinchilla,

Lex Fridman (53:28.880)

it actually has this few shot learning ability

Lex Fridman (53:31.760)

and these emerging abilities

Lex Fridman (53:33.080)

that we didn't even measure

Oriol Vinyals (53:34.640)

once we were developing the model,

Lex Fridman (53:36.400)

but once developed, then as you play with the interface,

Oriol Vinyals (53:40.080)

you can start seeing, wow, okay, yeah, it's cool.

Lex Fridman (53:42.320)

We can upload, I think one of the tweets

Oriol Vinyals (53:45.000)

talking about Twitter was this image from Obama

Lex Fridman (53:47.840)

that is placing a weight

Lex Fridman (53:49.840)

and someone is kind of weighting themselves

Lex Fridman (53:52.400)

and it's kind of a joke style image.

Lex Fridman (53:54.880)

And it's notable because I think Andrew Carpati

Lex Fridman (53:57.840)

a few years ago said,

Oriol Vinyals (53:59.400)

no computer vision system can understand

Lex Fridman (54:02.320)

the subtlety of this joke in this image,

Oriol Vinyals (54:04.720)

all the things that go on.

Lex Fridman (54:06.360)

And so what we try to do, and it's very anecdotally,

Oriol Vinyals (54:09.600)

I mean, this is not a proof that we solved this issue,

Lex Fridman (54:12.120)

but it just shows that you can upload now this image

Lex Fridman (54:15.720)

and start conversing with the model,

Lex Fridman (54:17.560)

trying to make out if it gets that there's a joke

Oriol Vinyals (54:21.360)

because the person weighting themselves

Lex Fridman (54:23.520)

doesn't see that someone behind

Oriol Vinyals (54:25.040)

is making the weight higher and so on and so forth.

Lex Fridman (54:27.840)

So it's a fascinating capability

Lex Fridman (54:30.760)

and it comes from this key idea of modularity

Lex Fridman (54:33.240)

where we took a frozen brain

Lex Fridman (54:34.800)

and we just added a new capability.

Lex Fridman (54:37.760)

So the question is, should we,

Lex Fridman (54:40.600)

so in a way you can see even from DeepMind,

Lex Fridman (54:42.720)

we have Flamingo that this moderate approach

Lex Fridman (54:46.280)

and thus could leverage the scale a bit more reasonably

Lex Fridman (54:49.040)

because we didn't need to retrain a system from scratch.

Lex Fridman (54:52.200)

And on the other hand, we had Gato,

Lex Fridman (54:54.080)

which used the same data sets,

Lex Fridman (54:55.800)

but then he trained it from scratch, right?

Lex Fridman (54:57.400)

And so I guess big question for the community

Oriol Vinyals (55:00.480)

is should we train from scratch

Lex Fridman (55:02.760)

or should we embrace modularity?

Lex Fridman (55:04.640)

And this lies, like this goes back to modularity

Lex Fridman (55:08.640)

as a way to grow, but reuse seems like natural

Lex Fridman (55:12.040)

and it was very effective, certainly.

Lex Fridman (55:14.920)

The next question is, if you go the way of modularity,

Oriol Vinyals (55:18.960)

is there a systematic way of freezing weights

Lex Fridman (55:22.680)

and joining different modalities across,

Oriol Vinyals (55:27.040)

you know, not just two or three or four networks,

Lex Fridman (55:29.200)

but hundreds of networks from all different kinds of places,

Oriol Vinyals (55:32.280)

maybe open source network that looks at weather patterns

Lex Fridman (55:36.280)

and you shove that in somehow

Lex Fridman (55:37.880)

and then you have networks that, I don't know,

Lex Fridman (55:40.360)

do all kinds of stuff, play StarCraft

Lex Fridman (55:42.000)

and play all the other video games

Lex Fridman (55:43.960)

and you can keep adding them in without significant effort,

Oriol Vinyals (55:49.480)

like maybe the effort scales linearly or something like that

Lex Fridman (55:53.160)

as opposed to like the more network you add,

Oriol Vinyals (55:54.880)

the more you have to worry about the instabilities created.

Lex Fridman (55:57.840)

Yeah, so that vision is beautiful.

Oriol Vinyals (55:59.840)

I think there's still the question

Lex Fridman (56:03.400)

about within single modalities, like Chinchilla was reused,

Lex Fridman (56:06.720)

but now if we train a next iteration of language models,

Lex Fridman (56:10.120)

are we gonna use Chinchilla or not?

Lex Fridman (56:11.720)

Yeah, how do you swap out Chinchilla?

Lex Fridman (56:13.040)

Right, so there's still big questions,

Lex Fridman (56:15.840)

but that idea is actually really akin to software engineering,

Lex Fridman (56:19.280)

which we're not reimplementing libraries from scratch,

Oriol Vinyals (56:22.280)

we're reusing and then building ever more amazing things,

Lex Fridman (56:25.280)

including neural networks with software that we're reusing.

Lex Fridman (56:28.880)

So I think this idea of modularity, I like it,

Lex Fridman (56:32.120)

I think it's here to stay

Lex Fridman (56:33.800)

and that's also why I mentioned

Lex Fridman (56:35.840)

it's just the beginning, not the end.

Oriol Vinyals (56:38.160)

You've mentioned meta learning,

Lex Fridman (56:39.360)

so given this promise of Gato,

Oriol Vinyals (56:42.760)

can we try to redefine this term

Lex Fridman (56:45.960)

that's almost akin to consciousness

Oriol Vinyals (56:47.560)

because it means different things to different people

Lex Fridman (56:50.120)

throughout the history of artificial intelligence,

Lex Fridman (56:52.360)

but what do you think meta learning is

Lex Fridman (56:56.600)

and looks like now in the five years, 10 years,

Lex Fridman (57:00.040)

will it look like the system like Gato, but scaled?

Lex Fridman (57:03.160)

What's your sense of, what does meta learning look like?

Lex Fridman (57:07.000)

Do you think with all the wisdom we've learned so far?

Lex Fridman (57:10.480)

Yeah, great question.

Oriol Vinyals (57:11.520)

Maybe it's good to give another data point

Lex Fridman (57:14.480)

looking backwards rather than forward.

Lex Fridman (57:16.160)

So when we talk in 2019,

Lex Fridman (57:22.880)

meta learning meant something that has changed

Oriol Vinyals (57:26.480)

mostly through the revolution of GPT3 and beyond.

Lex Fridman (57:31.120)

So what meta learning meant at the time

Oriol Vinyals (57:35.000)

was driven by what benchmarks people care about

Lex Fridman (57:37.640)

in meta learning.

Lex Fridman (57:38.800)

And the benchmarks were about a capability

Lex Fridman (57:42.560)

to learn about object identities.

Lex Fridman (57:44.960)

So it was very much overfitted to vision

Lex Fridman (57:48.480)

and object classification.

Lex Fridman (57:50.360)

And the part that was meta about that was that,

Lex Fridman (57:52.880)

oh, we're not just learning a thousand categories

Oriol Vinyals (57:55.320)

that ImageNet tells us to learn.

Lex Fridman (57:57.040)

We're going to learn object categories

Oriol Vinyals (57:59.200)

that can be defined when we interact with the model.

Lex Fridman (58:03.280)

So it's interesting to see the evolution, right?

Oriol Vinyals (58:06.640)

The way this started was we have a special language

Lex Fridman (58:10.720)

that was a data set, a small data set

Oriol Vinyals (58:13.200)

that we prompted the model with saying,

Lex Fridman (58:15.920)

hey, here is a new classification task.

Oriol Vinyals (58:18.960)

I'll give you one image and the name,

Lex Fridman (58:21.720)

which was an integer at the time of the image

Lex Fridman (58:24.320)

and a different image and so on.

Lex Fridman (58:25.920)

So you have a small prompt in the form of a data set,

Oriol Vinyals (58:30.000)

a machine learning data set.

Lex Fridman (58:31.600)

And then you got then a system that could then predict

Oriol Vinyals (58:35.480)

or classify these objects that you just

Lex Fridman (58:37.440)

defined kind of on the fly.

Lex Fridman (58:40.280)

So fast forward, it was revealed that language models

Lex Fridman (58:46.480)

are few shot learners.

Oriol Vinyals (58:47.440)

That's the title of the paper.

Lex Fridman (58:49.120)

So very good title.

Oriol Vinyals (58:50.080)

Sometimes titles are really good.

Lex Fridman (58:51.480)

So this one is really, really good.

Oriol Vinyals (58:53.520)

Because that's the point of GPT3 that showed that, look, sure,

Lex Fridman (58:58.800)

we can focus on object classification

Lex Fridman (59:00.960)

and what meta learning means within the space of learning

Lex Fridman (59:04.160)

object categories.

Oriol Vinyals (59:05.400)

This goes beyond, or before rather,

Lex Fridman (59:07.440)

to also Omniglot, before ImageNet and so on.

Lex Fridman (59:10.120)

So there's a few benchmarks.

Lex Fridman (59:11.560)

To now, all of a sudden, we're a bit unlocked from benchmarks.

Lex Fridman (59:15.240)

And through language, we can define tasks.

Lex Fridman (59:17.960)

So we're literally telling the model

Oriol Vinyals (59:20.280)

some logical task or a little thing that we wanted to do.

Lex Fridman (59:23.920)

We prompt it much like we did before,

Lex Fridman (59:26.000)

but now we prompt it through natural language.

Lex Fridman (59:28.440)

And then not perfectly, I mean, these models have failure modes

Lex Fridman (59:32.280)

and that's fine, but these models then

Lex Fridman (59:35.560)

are now doing a new task.

Lex Fridman (59:37.240)

And so they meta learn this new capability.

Lex Fridman (59:40.520)

Now, that's where we are now.

Oriol Vinyals (59:43.480)

Flamingo expanded this to visual and language,

Lex Fridman (59:47.320)

but it basically has the same abilities.

Oriol Vinyals (59:49.440)

You can teach it, for instance, an emergent property

Lex Fridman (59:52.720)

was that you can take pictures of numbers

Lex Fridman (59:55.320)

and then do arithmetic with the numbers just by teaching it,

Lex Fridman (59:59.040)

oh, when I show you 3 plus 6, I want you to output 9.