The Swyx Mixtape

From Lex Fridman: https://lexfridman.com/demis-hassabis/
Transcripts from Karpathy's https://karpathy.ai/lexicap/0299-large.html
00:37:13.160">
link |
00:37:13.160 
So let's go to the basic building blocks of biology

00:37:16.960">link |
00:37:16.960 
that I think is another angle at which you can start

00:37:20.200">link |
00:37:20.200 
to understand the human mind, the human body,

00:37:22.280">link |
00:37:22.280 
which is quite fascinating,

00:37:23.400">link |
00:37:23.400 
which is from the basic building blocks,

00:37:26.640">link |
00:37:26.640 
start to simulate, start to model

00:37:28.960">link |
00:37:28.960 
how from those building blocks,

00:37:30.480">link |
00:37:30.480 
you can construct bigger and bigger, more complex systems,

00:37:33.080">link |
00:37:33.080 
maybe one day the entirety of the human biology.

00:37:35.820">link |
00:37:35.820 
So here's another problem that thought

00:37:39.680">link |
00:37:39.680 
to be impossible to solve, which is protein folding.

00:37:42.720">link |
00:37:42.720 
And Alpha Fold or specifically Alpha Fold 2 did just that.

00:37:48.840">link |
00:37:48.840 
It solved protein folding.

00:37:50.320">link |
00:37:50.320 
I think it's one of the biggest breakthroughs,

00:37:53.400">link |
00:37:53.400 
certainly in the history of structural biology,

00:37:55.140">link |
00:37:55.140 
but in general in science,

00:38:00.240">link |
00:38:00.240 
maybe from a high level, what is it and how does it work?

00:38:04.840">link |
00:38:04.840 
And then we can ask some fascinating questions after.

00:38:08.700">link |
00:38:08.700 
Sure.

00:38:09.980">link |
00:38:09.980 
So maybe to explain it to people not familiar

00:38:12.880">link |
00:38:12.880 
with protein folding is, you know,

00:38:14.400">link |
00:38:14.400 
first of all, explain proteins, which is, you know,

00:38:16.980">link |
00:38:16.980 
proteins are essential to all life.

00:38:18.840">link |
00:38:18.840 
Every function in your body depends on proteins.

00:38:21.520">link |
00:38:21.520 
Sometimes they're called the workhorses of biology.

00:38:23.920">link |
00:38:23.920 
And if you look into them and I've, you know,

00:38:25.340">link |
00:38:25.340 
obviously as part of Alpha Fold,

00:38:26.660">link |
00:38:26.660 
I've been researching proteins and structural biology

00:38:30.200">link |
00:38:30.200 
for the last few years, you know,

00:38:31.760">link |
00:38:31.760 
they're amazing little bio nano machines proteins.

00:38:34.760">link |
00:38:34.760 
They're incredible if you actually watch little videos

00:38:36.460">link |
00:38:36.460 
of how they work, animations of how they work.

00:38:39.000">link |
00:38:39.000 
And proteins are specified by their genetic sequence

00:38:42.600">link |
00:38:42.600 
called the amino acid sequence.

00:38:44.280">link |
00:38:44.280 
So you can think of it as their genetic makeup.

00:38:47.040">link |
00:38:47.040 
And then in the body in nature,

00:38:50.080">link |
00:38:50.080 
they fold up into a 3D structure.

00:38:53.360">link |
00:38:53.360 
So you can think of it as a string of beads

00:38:55.320">link |
00:38:55.320 
and then they fold up into a ball.

00:38:57.160">link |
00:38:57.160 
Now, the key thing is you want to know

00:38:59.100">link |
00:38:59.100 
what that 3D structure is because the structure,

00:39:02.480">link |
00:39:02.480 
the 3D structure of a protein is what helps to determine

00:39:06.120">link |
00:39:06.120 
what does it do, the function it does in your body.

00:39:08.580">link |
00:39:08.580 
And also if you're interested in drugs or disease,

00:39:12.320">link |
00:39:12.320 
you need to understand that 3D structure

00:39:13.980">link |
00:39:13.980 
because if you want to target something

00:39:15.840">link |
00:39:15.840 
with a drug compound about to block something

00:39:18.640">link |
00:39:18.640 
the protein's doing, you need to understand

00:39:21.120">link |
00:39:21.120 
where it's gonna bind on the surface of the protein.

00:39:23.440">link |
00:39:23.440 
So obviously in order to do that,

00:39:24.940">link |
00:39:24.940 
you need to understand the 3D structure.

00:39:26.720">link |
00:39:26.720 
So the structure is mapped to the function.

00:39:28.640">link |
00:39:28.640 
The structure is mapped to the function

00:39:29.880">link |
00:39:29.880 
and the structure is obviously somehow specified

00:39:32.560">link |
00:39:32.560 
by the amino acid sequence.

00:39:34.840">link |
00:39:34.840 
And that's the, in essence, the protein folding problem is,

00:39:37.420">link |
00:39:37.420 
can you just from the amino acid sequence,

00:39:39.620">link |
00:39:39.620 
the one dimensional string of letters,

00:39:42.560">link |
00:39:42.560 
can you immediately computationally predict

00:39:45.600">link |
00:39:45.600 
the 3D structure?

00:39:47.120">link |
00:39:47.120 
And this has been a grand challenge in biology

00:39:50.020">link |
00:39:50.020 
for over 50 years.

00:39:51.500">link |
00:39:51.500 
So I think it was first articulated by Christian Anfinsen,

00:39:54.360">link |
00:39:54.360 
a Nobel prize winner in 1972,

00:39:57.040">link |
00:39:57.040 
as part of his Nobel prize winning lecture.

00:39:59.240">link |
00:39:59.240 
And he just speculated this should be possible

00:40:01.860">link |
00:40:01.860 
to go from the amino acid sequence to the 3D structure,

00:40:04.960">link |
00:40:04.960 
but he didn't say how.

00:40:06.060">link |
00:40:06.060 
So it's been described to me as equivalent

00:40:09.440">link |
00:40:09.440 
to Fermat's last theorem, but for biology.

00:40:12.320">link |
00:40:12.320 
You should, as somebody that very well might win

00:40:15.120">link |
00:40:15.120 
the Nobel prize in the future.

00:40:16.560">link |
00:40:16.560 
But outside of that, you should do more

00:40:19.240">link |
00:40:19.240 
of that kind of thing.

00:40:20.080">link |
00:40:20.080 
In the margin, just put random things

00:40:22.160">link |
00:40:22.160 
that will take like 200 years to solve.

00:40:24.440">link |
00:40:24.440 
Set people off for 200 years.

00:40:26.000">link |
00:40:26.000 
It should be possible.

00:40:27.720">link |
00:40:27.720 
And just don't give any details.

00:40:29.040">link |
00:40:29.040 
Exactly.

00:40:29.880">link |
00:40:29.880 
I think everyone exactly should be,

00:40:31.500">link |
00:40:31.500 
I'll have to remember that for future.

00:40:33.520">link |
00:40:33.520 
So yeah, so he set off, you know,

00:40:34.800">link |
00:40:34.800 
with this one throwaway remark, just like Fermat,

00:40:37.040">link |
00:40:37.040 
you know, he set off this whole 50 year field really

00:40:42.640">link |
00:40:42.640 
of computational biology.

00:40:44.400">link |
00:40:44.400 
And they had, you know, they got stuck.

00:40:46.240">link |
00:40:46.240 
They hadn't really got very far with doing this.

00:40:48.520">link |
00:40:48.520 
And until now, until AlphaFold came along,

00:40:52.500">link |
00:40:52.500 
this is done experimentally, right?

00:40:54.320">link |
00:40:54.320 
Very painstakingly.

00:40:55.500">link |
00:40:55.500 
So the rule of thumb is, and you have to like

00:40:57.440">link |
00:40:57.440 
crystallize the protein, which is really difficult.

00:40:59.820">link |
00:40:59.820 
Some proteins can't be crystallized like membrane proteins.

00:41:03.060">link |
00:41:03.060 
And then you have to use very expensive electron microscopes

00:41:05.940">link |
00:41:05.940 
or X ray crystallography machines.

00:41:08.200">link |
00:41:08.200 
Really painstaking work to get the 3D structure

00:41:10.680">link |
00:41:10.680 
and visualize the 3D structure.

00:41:12.400">link |
00:41:12.400 
So the rule of thumb in experimental biology

00:41:14.840">link |
00:41:14.840 
is that it takes one PhD student,

00:41:16.860">link |
00:41:16.860 
their entire PhD to do one protein.

00:41:20.320">link |
00:41:20.320 
And with AlphaFold 2, we were able to predict

00:41:23.440">link |
00:41:23.440 
the 3D structure in a matter of seconds.

00:41:26.400">link |
00:41:26.400 
And so we were, you know, over Christmas,

00:41:28.700">link |
00:41:28.700 
we did the whole human proteome

00:41:30.240">link |
00:41:30.240 
or every protein in the human body or 20,000 proteins.

00:41:33.280">link |
00:41:33.280 
So the human proteomes like the equivalent

00:41:34.760">link |
00:41:34.760 
of the human genome, but on protein space.

00:41:37.560">link |
00:41:37.560 
And sort of revolutionized really

00:41:40.240">link |
00:41:40.240 
what a structural biologist can do.

00:41:43.300">link |
00:41:43.300 
Because now they don't have to worry

00:41:45.720">link |
00:41:45.720 
about these painstaking experimental,

00:41:47.960">link |
00:41:47.960 
should they put all of that effort in or not?

00:41:49.560">link |
00:41:49.560 
They can almost just look up the structure

00:41:51.120">link |
00:41:51.120 
of their proteins like a Google search.

00:41:53.280">link |
00:41:53.280 
And so there's a data set on which it's trained

00:41:56.880">link |
00:41:56.880 
and how to map this amino acid sequence.

00:41:58.800">link |
00:41:58.800 
First of all, it's incredible that a protein,

00:42:00.760">link |
00:42:00.760 
this little chemical computer is able to do

00:42:02.480">link |
00:42:02.480 
that computation itself in some kind of distributed way

00:42:05.720">link |
00:42:05.720 
and do it very quickly.

00:42:07.800">link |
00:42:07.800 
That's a weird thing.

00:42:08.840">link |
00:42:08.840 
And they evolve that way because, you know,

00:42:10.480">link |
00:42:10.480 
in the beginning, I mean, that's a great invention,

00:42:13.200">link |
00:42:13.200 
just the protein itself.

00:42:14.760">link |
00:42:14.760 
And then there's, I think, probably a history

00:42:18.240">link |
00:42:18.240 
of like they evolved to have many of these proteins

00:42:22.740">link |
00:42:22.740 
and those proteins figure out how to be computers themselves

00:42:26.600">link |
00:42:26.600 
in such a way that you can create structures

00:42:28.560">link |
00:42:28.560 
that can interact in complexes with each other

00:42:30.540">link |
00:42:30.540 
in order to form high level functions.

00:42:32.660">link |
00:42:32.660 
I mean, it's a weird system that they figured it out.

00:42:35.520">link |
00:42:35.520 
Well, for sure.

00:42:36.360">link |
00:42:36.360 
I mean, you know, maybe we should talk

00:42:37.640">link |
00:42:37.640 
about the origins of life too,

00:42:39.000">link |
00:42:39.000 
but proteins themselves, I think are magical

00:42:41.180">link |
00:42:41.180 
and incredible, as I said, little bio nano machines.

00:42:45.760">link |
00:42:45.760 
And actually Leventhal, who was another scientist,

00:42:50.280">link |
00:42:50.280 
a contemporary of Amphinson, he coined this Leventhal,

00:42:55.120">link |
00:42:55.120 
what became known as Leventhal's paradox,

00:42:56.820">link |
00:42:56.820 
which is exactly what you're saying.

00:42:58.320">link |
00:42:58.320 
He calculated roughly an average protein,

00:43:01.580">link |
00:43:01.580 
which is maybe 2000 amino acids base as long,

00:43:05.080">link |
00:43:05.080 
is can fold in maybe 10 to the power 300

00:43:09.960">link |
00:43:09.960 
different confirmations.

00:43:11.480">link |
00:43:11.480 
So there's 10 to the power 300 different ways

00:43:13.320">link |
00:43:13.320 
that protein could fold up.

00:43:14.800">link |
00:43:14.800 
And yet somehow in nature, physics solves this,

00:43:18.160">link |
00:43:18.160 
solves this in a matter of milliseconds.

00:43:20.520">link |
00:43:20.520 
So proteins fold up in your body in, you know,

00:43:23.080">link |
00:43:23.080 
sometimes in fractions of a second.

00:43:25.600">link |
00:43:25.600 
So physics is somehow solving that search problem.

00:43:29.080">link |
00:43:29.080 
And just to be clear, in many of these cases,

00:43:31.200">link |
00:43:31.200 
maybe you can correct me if I'm wrong,

00:43:33.040">link |
00:43:33.040 
there's often a unique way for that sequence to form itself.

00:43:37.680">link |
00:43:37.680 
So among that huge number of possibilities,

00:43:41.240">link |
00:43:41.240 
it figures out a way how to stably,

00:43:45.320">link |
00:43:45.320 
in some cases there might be a misfunction, so on,

00:43:47.800">link |
00:43:47.800 
which leads to a lot of the disorders and stuff like that.

00:43:50.040">link |
00:43:50.040 
But most of the time it's a unique mapping

00:43:52.720">link |
00:43:52.720 
and that unique mapping is not obvious.

00:43:54.820">link |
00:43:54.820 
No, exactly.

00:43:55.660">link |
00:43:55.660 
Which is what the problem is.

00:43:57.120">link |
00:43:57.120 
Exactly, so there's a unique mapping usually in a healthy,

00:44:00.720">link |
00:44:00.720 
if it's healthy, and as you say in disease,

00:44:04.040">link |
00:44:04.040 
so for example, Alzheimer's,

00:44:05.400">link |
00:44:05.400 
one conjecture is that it's because of misfolded protein,

00:44:09.000">link |
00:44:09.000 
a protein that folds in the wrong way, amyloid beta protein.

00:44:12.040">link |
00:44:12.040 
So, and then because it folds in the wrong way,

00:44:14.560">link |
00:44:14.560 
it gets tangled up, right, in your neurons.

00:44:17.600">link |
00:44:17.600 
So it's super important to understand

00:44:20.560">link |
00:44:20.560 
both healthy functioning and also disease

00:44:23.600">link |
00:44:23.600 
is to understand, you know, what these things are doing

00:44:26.480">link |
00:44:26.480 
and how they're structuring.

00:44:27.600">link |
00:44:27.600 
Of course, the next step is sometimes proteins change shape

00:44:30.540">link |
00:44:30.540 
when they interact with something.

00:44:32.160">link |
00:44:32.160 
So they're not just static necessarily in biology.

00:44:37.200">link |
00:44:37.200 
Maybe you can give some interesting,

00:44:39.780">link |
00:44:39.780 
so beautiful things to you about these early days

00:44:43.260">link |
00:44:43.260 
of AlphaFold, of solving this problem,

00:44:46.160">link |
00:44:46.160 
because unlike games, this is real physical systems

00:44:51.280">link |
00:44:51.280 
that are less amenable to self play type of mechanisms.

00:44:55.640">link |
00:44:55.640 
Sure.

00:44:56.460">link |
00:44:56.460 
The size of the data set is smaller

00:44:58.440">link |
00:44:58.440 
than you might otherwise like,

00:44:59.760">link |
00:44:59.760 
so you have to be very clever about certain things.

00:45:01.800">link |
00:45:01.800 
Is there something you could speak to

00:45:04.800">link |
00:45:04.800 
what was very hard to solve

00:45:06.680">link |
00:45:06.680 
and what are some beautiful aspects about the solution?

00:45:09.920">link |
00:45:09.920 
Yeah, I would say AlphaFold is the most complex

00:45:12.800">link |
00:45:12.800 
and also probably most meaningful system

00:45:14.600">link |
00:45:14.600 
we've built so far.

00:45:15.860">link |
00:45:15.860 
So it's been an amazing time actually in the last,

00:45:18.400">link |
00:45:18.400 
you know, two, three years to see that come through

00:45:20.520">link |
00:45:20.520 
because as we talked about earlier, you know,

00:45:23.200">link |
00:45:23.200 
games is what we started on

00:45:25.480">link |
00:45:25.480 
building things like AlphaGo and AlphaZero,

00:45:27.900">link |
00:45:27.900 
but really the ultimate goal was to,

00:45:30.400">link |
00:45:30.400 
not just to crack games,

00:45:31.520">link |
00:45:31.520 
it was just to build,

00:45:33.120">link |
00:45:33.120 
use them to bootstrap general learning systems

00:45:35.320">link |
00:45:35.320 
we could then apply to real world challenges.

00:45:37.440">link |
00:45:37.440 
Specifically, my passion is scientific challenges

00:45:40.640">link |
00:45:40.640 
like protein folding.

00:45:41.920">link |
00:45:41.920 
And then AlphaFold of course

00:45:43.280">link |
00:45:43.280 
is our first big proof point of that.

00:45:45.360">link |
00:45:45.360 
And so, you know, in terms of the data

00:45:49.040">link |
00:45:49.040 
and the amount of innovations that had to go into it,

00:45:50.920">link |
00:45:50.920 
we, you know, it was like

00:45:52.280">link |
00:45:52.280 
more than 30 different component algorithms

00:45:54.480">link |
00:45:54.480 
needed to be put together to crack the protein folding.

00:45:57.960">link |
00:45:57.960 
I think some of the big innovations were that

00:46:00.800">link |
00:46:00.800 
kind of building in some hard coded constraints

00:46:04.220">link |
00:46:04.220 
around physics and evolutionary biology

00:46:07.760">link |
00:46:07.760 
to constrain sort of things like the bond angles

00:46:11.640">link |
00:46:11.640 
in the protein and things like that,

00:46:15.400">link |
00:46:15.400 
a lot, but not to impact the learning system.

00:46:18.040">link |
00:46:18.040 
So still allowing the system to be able to learn

00:46:21.000">link |
00:46:21.000 
the physics itself from the examples that we had.

00:46:25.540">link |
00:46:25.540 
And the examples, as you say,

00:46:26.640">link |
00:46:26.640 
there are only about 150,000 proteins,

00:46:28.840">link |
00:46:28.840 
even after 40 years of experimental biology,

00:46:31.240">link |
00:46:31.240 
only around 150,000 proteins have been,

00:46:33.880">link |
00:46:33.880 
the structures have been found out about.

00:46:35.920">link |
00:46:35.920 
So that was our training set,

00:46:37.120">link |
00:46:37.120 
which is much less than normally we would like to use,

00:46:41.120">link |
00:46:41.120 
but using various tricks, things like self distillation.

00:46:43.840">link |
00:46:43.840 
So actually using AlphaFold predictions,

00:46:48.280">link |
00:46:48.280 
some of the best predictions

00:46:49.480">link |
00:46:49.480 
that it thought was highly confident in,

00:46:51.000">link |
00:46:51.000 
we put them back into the training set, right?

00:46:53.320">link |
00:46:53.320 
To make the training set bigger,

00:46:55.440">link |
00:46:55.440 
that was critical to AlphaFold working.

00:46:58.400">link |
00:46:58.400 
So there was actually a huge number

00:47:00.160">link |
00:47:00.160 
of different innovations like that,

00:47:02.720">link |
00:47:02.720 
that were required to ultimately crack the problem.

00:47:06.080">link |
00:47:06.080 
AlphaFold one, what it produced was a distrogram.

00:47:09.720">link |
00:47:09.720 
So a kind of a matrix of the pairwise distances

00:47:13.600">link |
00:47:13.600 
between all of the molecules in the protein.

00:47:17.880">link |
00:47:17.880 
And then there had to be a separate optimization process

00:47:20.440">link |
00:47:20.440 
to create the 3D structure.

00:47:23.640">link |
00:47:23.640 
And what we did for AlphaFold two

00:47:25.120">link |
00:47:25.120 
is make it truly end to end.

00:47:26.920">link |
00:47:26.920 
So we went straight from the amino acid sequence of bases

00:47:31.720">link |
00:47:31.720 
to the 3D structure directly

00:47:33.860">link |
00:47:33.860 
without going through this intermediate step.

00:47:36.080">link |
00:47:36.080 
And in machine learning, what we've always found is

00:47:38.600">link |
00:47:38.600 
that the more end to end you can make it,

00:47:40.920">link |
00:47:40.920 
the better the system.

00:47:42.160">link |
00:47:42.160 
And it's probably because in the end,

00:47:46.160">link |
00:47:46.160 
the system's better at learning what the constraints are

00:47:48.560">link |
00:47:48.560 
than we are as the human designers of specifying it.

00:47:51.920">link |
00:47:51.920 
So anytime you can let it flow end to end

00:47:54.040">link |
00:47:54.040 
and actually just generate what it is

00:47:55.400">link |
00:47:55.400 
you're really looking for, in this case, the 3D structure,

00:47:58.440">link |
00:47:58.440 
you're better off than having this intermediate step,

00:48:00.560">link |
00:48:00.560 
which you then have to handcraft the next step for.

00:48:03.360">link |
00:48:03.360 
So it's better to let the gradients and the learning

00:48:06.160">link |
00:48:06.160 
flow all the way through the system from the end point,

00:48:09.000">link |
00:48:09.000 
the end output you want to the inputs.

00:48:10.880">link |
00:48:10.880 
So that's a good way to start on a new problem.

00:48:13.040">link |
00:48:13.040 
Handcraft a bunch of stuff,

00:48:14.360">link |
00:48:14.360 
add a bunch of manual constraints

00:48:16.640">link |
00:48:16.640 
with a small end to end learning piece

00:48:18.640">link |
00:48:18.640 
or a small learning piece and grow that learning piece

00:48:21.560">link |
00:48:21.560 
until it consumes the whole thing.

00:48:22.840">link |
00:48:22.840 
That's right.

00:48:23.680">link |
00:48:23.680 
And so you can also see,

00:48:25.320">link |
00:48:25.320 
this is a bit of a method we've developed

00:48:26.960">link |
00:48:26.960 
over doing many sort of successful alpha,

00:48:29.640">link |
00:48:29.640 
we call them alpha X projects, right?

00:48:32.200">link |
00:48:32.200 
And the easiest way to see that is the evolution

00:48:34.600">link |
00:48:34.600 
of alpha go to alpha zero.

00:48:36.720">link |
00:48:36.720 
So alpha go was a learning system,

00:48:39.640">link |
00:48:39.640 
but it was specifically trained to only play go, right?

00:48:42.280">link |
00:48:42.280 
So, and what we wanted to do with first version of alpha go

00:48:45.360">link |
00:48:45.360 
is just get to world champion performance

00:48:47.520">link |
00:48:47.520 
no matter how we did it, right?

00:48:49.200">link |
00:48:49.200 
And then of course, alpha go zero,

00:48:51.400">link |
00:48:51.400 
we remove the need to use human games as a starting point,

00:48:55.280">link |
00:48:55.280 
right?

00:48:56.120">link |
00:48:56.120 
So it could just play against itself

00:48:57.960">link |
00:48:57.960 
from random starting point from the beginning.

00:49:00.280">link |
00:49:00.280 
So that removed the need for human knowledge about go.

00:49:03.720">link |
00:49:03.720 
And then finally alpha zero then generalized it

00:49:05.960">link |
00:49:05.960 
so that any things we had in there, the system,

00:49:08.920">link |
00:49:08.920 
including things like symmetry of the go board were removed.

00:49:12.240">link |
00:49:12.240 
So the alpha zero could play from scratch

00:49:14.600">link |
00:49:14.600 
any two player game and then mu zero,

00:49:16.440">link |
00:49:16.440 
which is the final, our latest version

00:49:18.360">link |
00:49:18.360 
of that set of things was then extending it

00:49:20.680">link |
00:49:20.680 
so that you didn't even have to give it

00:49:22.120">link |
00:49:22.120 
the rules of the game.

00:49:23.200">link |
00:49:23.200 
It would learn that for itself.

00:49:24.880">link |
00:49:24.880 
So it could also deal with computer games

00:49:26.600">link |
00:49:26.600 
as well as board games.

00:49:27.760">link |
00:49:27.760 
So that line of alpha go, alpha go zero, alpha zero,

00:49:30.400">link |
00:49:30.400 
mu zero, that's the full trajectory

00:49:33.480">link |
00:49:33.480 
of what you can take from imitation learning

00:49:37.200">link |
00:49:37.200 
to full self supervised learning.

00:49:40.440">link |
00:49:40.440 
Yeah, exactly.

00:49:41.640">link |
00:49:41.640 
And learning the entire structure

00:49:44.720">link |
00:49:44.720 
of the environment you're put in from scratch, right?

00:49:47.640">link |
00:49:47.640 
And bootstrapping it through self play yourself.

00:49:51.840">link |
00:49:51.840 
But the thing is it would have been impossible, I think,

00:49:53.720">link |
00:49:53.720 
or very hard for us to build alpha zero

00:49:55.960">link |
00:49:55.960 
or mu zero first out of the box.

00:49:58.600">link |
00:49:58.600 
Even psychologically, because you have to believe

00:50:01.400">link |
00:50:01.400 
in yourself for a very long time.

00:50:03.040">link |
00:50:03.040 
You're constantly dealing with doubt

00:50:04.640">link |
00:50:04.640 
because a lot of people say that it's impossible.

00:50:06.680">link |
00:50:06.680 
Exactly, so it's hard enough just to do go.

00:50:08.640">link |
00:50:08.640 
As you were saying, everyone thought that was impossible

00:50:10.920">link |
00:50:10.920 
or at least a decade away from when we did it

00:50:14.160">link |
00:50:14.160 
back in 2015, 2016.

00:50:17.320">link |
00:50:17.320 
And so yes, it would have been psychologically

00:50:20.960">link |
00:50:20.960 
probably very difficult as well as the fact

00:50:22.960">link |
00:50:22.960 
that of course we learn a lot by building alpha go first.

00:50:26.400">link |
00:50:26.400 
Right, so I think this is why I call AI

00:50:28.520">link |
00:50:28.520 
an engineering science.

00:50:29.880">link |
00:50:29.880 
It's one of the most fascinating science disciplines,

00:50:32.280">link |
00:50:32.280 
but it's also an engineering science in the sense

00:50:34.200">link |
00:50:34.200 
that unlike natural sciences, the phenomenon you're studying

00:50:38.200">link |
00:50:38.200 
doesn't exist out in nature.

00:50:39.440">link |
00:50:39.440 
You have to build it first.

00:50:40.880">link |
00:50:40.880 
So you have to build the artifact first,

00:50:42.480">link |
00:50:42.480 
and then you can study and pull it apart and how it works.