Transformers: The best idea in AI | Andrej Karpathy and Lex Fridman

looking back what is the most beautiful

or surprising idea in deep learning or

AI in general that you've come across

you've seen this field explode

and grow in interesting ways just what

what cool ideas like like we made you

sit back and go hmm small big or small

well the one that I've been thinking

about recently the most probably is the

the Transformer architecture

um so basically uh neural networks have

a lot of architectures that were trendy

have come and gone for different sensory

modalities like for Vision Audio text

you would process them with different

looking neural nuts and recently we've

seen these convergence towards one

architecture the Transformer and you can

feed it video or you can feed it you

know images or speech or text and it

just gobbles it up and it's kind of like

a bit of a general purpose uh computer

that is also trainable and very

efficient to run on our Hardware

and so this paper came out in 2016 I

want to say

um attention is all you need attention

is all you need you criticize the paper

title in retrospect that it wasn't

um it didn't foresee the bigness of the

impact yeah that it was going to have

yeah I'm not sure if the authors were

aware of the impact that that paper

would go on to have probably they

weren't but I think they were aware of

some of the motivations and design

decisions behind the Transformer and

they chose not to I think uh expand on

it in that way in a paper and so I think

they had an idea that there was more

um than just the surface of just like oh

we're just doing translation and here's

a better architecture you're not just

doing translation this is like a really

cool differentiable optimizable

efficient computer that you've proposed

and maybe they didn't have all of that

foresight but I think is really

interesting isn't it funny sorry to

interrupt that title is memeable that

they went for such a profound idea they

went with the I don't think anyone used

that kind of title before right

protection is all you need yeah it's

like a meme or something basically it's

not funny that one like uh maybe if it

was a more serious title it wouldn't

have the impact honestly I yeah there is

an element of me that honestly agrees

with you and prefers it this way yes

if it was two grand it would over

promise and then under deliver

potentially so you want to just uh meme

your way to greatness

that should be a t-shirt so you you

tweeted the Transformers the Magnificent

neural network architecture because it

is a general purpose differentiable

computer it is simultaneously expressive

in the forward pass optimizable via back

propagation gradient descent and

efficient High parallelism compute graph

can you discuss some of those details

expressive optimizable efficient

yeah from memory or or in general

whatever comes to your heart you want to

have a general purpose computer that you

can train on arbitrary problems uh like

say the task of next word prediction or

detecting if there's a cat in the image

or something like that and you want to

train this computer so you want to set

its its weights and I think there's a

number of design criteria that sort of

overlap in the Transformer

simultaneously that made it very

successful and I think the authors were

kind of uh deliberately trying to make

this really uh powerful architecture and

um so in a basically it's very powerful

in the forward pass because it's able to

express

um very uh General computation as a sort

of something that looks like message

passing you have nodes and they all

store vectors and these nodes get to

basically look at each other and it's

each other's vectors and they get to

communicate and basically notes get to

broadcast hey I'm looking for certain

things and then other nodes get to

broadcast hey these are the things I

have those are the keys and the values

so it's not just the tension yeah

exactly Transformer is much more than

just the attention component it's got

many pieces architectural that went into

it the residual connection of the way

it's arranged there's a multi-layer

perceptron in there the way it's stacked

and so on

um but basically there's a message

passing scheme where nodes get to look

at each other decide what's interesting

and then update each other and uh so I

think the um when you get to the details

of it I think it's a very expressive

function uh so it can express lots of

different types of algorithms and

forward paths not only that but the way

it's designed with the residual

connections layer normalizations the

soft Max attention and everything it's

also optimizable this is a really big

deal because there's lots of computers

that are powerful that you can't

optimize or they're not easy to optimize

using the techniques that we have which

is back propagation and gradient and

send these are first order methods very

simple optimizers really and so um you

also need it to be optimizable

um and then lastly you want it to run

efficiently in the hardware our Hardware

is a massive throughput machine like

gpus they prefer lots of parallelism so

you don't want to do lots of sequential

operations you want to do a lot of

operations serially and the Transformer

is designed with that in mind as well

and so it's designed for our hardware

and it's designed to both be very

expressive in a forward pass but also

very optimizable in the backward pass

and you said that uh the residual

connections support a kind of ability to

learn short algorithms fast them first

and then gradually extend them longer

during training yeah what's what's the

idea of learning short algorithms right

think of it as a so basically a

Transformer is a series of uh blocks

right and these blocks have attention

and a little multi-layer perceptron and

so you you go off into a block and you

come back to this residual pathway and

then you go off and you come back and

then you have a number of layers

arranged sequentially and so the way to

look at it I think is because of the

residual pathway in the backward path

the gradients uh sort of flow along it

uninterrupted because addition

distributes the gradient equally to all

of its branches so the gradient from the

supervision at the top uh just floats

directly to the first layer and the all

these residual connections are arranged

so that in the beginning during

initialization they contribute nothing

to the residual pathway

um so what it kind of looks like is

imagine the Transformer is kind of like

a uh python function like a death and um

you get to do various kinds of like

lines of code so you have a hundred

layers deep uh Transformer typically

they would be much shorter say 20. so

you have 20 lines of code then you can

do something in them and so think of

during the optimization basically what

it looks like is first you optimize the

first line of code and then the second

line of code can kick in and the third

line of code can and I kind of feel like

because of the residual pathway and the

Dynamics of the optimization uh you can

sort of learn a very short algorithm

that gets the approximate tensor but

then the other layers can sort of kick

in and start to create a contribution

and at the end of it you're you're

optimizing over an algorithm that is uh

20 lines of code

except these lines of code are very

complex because it's an entire block of

a transformer you can do a lot in there

what's really interesting is that this

Transformer architecture actually has

been a remarkably resilient basically

the Transformer that came out in 2016 is

the Transformer you would use today

except you reshuffle some of the layer

norms the player normalizations have

been reshuffled to a pre-norm

formulation and so it's been remarkably

stable but there's a lot of bells and

whistles that people have attached to it

and try to uh improve it I do think that

basically it's a it's a big step in

simultaneously optimizing for lots of

properties of a desirable neural network

architecture and I think people have

been trying to change it but it's proven

remarkably resilient but I do think that

there should be even better

architectures potentially but it's uh

your you admire the resilience here yeah

there's something profound about this

architecture that at least so maybe we

can everything could be turned into a

uh into a problem that Transformers can

solve currently definitely looks like

the Transformers taking over Ai and you

can feed basically arbitrary problems

into it and it's a general

differentiable computer and it's

extremely powerful and uh this

convergence in AI has been really

interesting to watch uh for me

personally what else do you think could

be discovered here about Transformers

like a surprising thing or or is it a

stable

um

we're in a stable place is there

something interesting we might discover

about Transformers like aha moments

maybe has to do with memory uh maybe

knowledge representation that kind of

stuff

definitely the Zeitgeist today is just

pushing like basically right now there's

that guys is do not touch the

Transformer touch everything else yes so

people are scaling up the data sets

making them much much bigger they're

working on the evaluation making the

evaluation much much bigger and uh

um they're basically keeping the

architecture unchanged and that's how

we've um that's the last five years of

progress in AI kind of