WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=pWXUkLP9uWM

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/pWXUkLP9uWM.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.400 --> 00:00:13.040
All right.

s3
00:00:13.040 --> 00:00:14.120
Hello everyone.

s4
00:00:14.120 --> 00:00:15.880
Really excited to be here.

s5
00:00:15.880 --> 00:00:17.520
It's a big room.

s6
00:00:17.520 --> 00:00:20.240
Very uh very cool conference so far.

s7
00:00:20.240 --> 00:00:25.120
Uh I want to talk to you today about something that's been on my mind for many many years.

s8
00:00:25.120 --> 00:00:29.360
This is actually the first time I I talk about it, sort of my version of going to Mars.

s9
00:00:29.360 --> 00:00:31.720
Um and that is the Eureka machine.

s10
00:00:31.720 --> 00:00:38.520
A machine that will eventually invent pretty much all future inventions for humanity.

s11
00:00:38.520 --> 00:00:47.240
Uh and the way we're going to get there is uh by taking a step back and thinking about what else has given us a lot of really incredible inventions,

s12
00:00:47.240 --> 00:00:55.560
uh namely evolution, and how that leads us to automating research uh and pushing the scientific frontier forward.

s13
00:00:55.560 --> 00:01:04.400
And this is a joint work with a lot of uh amazing folks uh at Recursive.com uh and even some uh folks at AIX Ventures.

s14
00:01:04.400 --> 00:01:12.760
And some of these slides are uh actually inspired by uh and taken uh partially from one of my co-founders at Recursive, Tim Rocktäschel.

s15
00:01:12.760 --> 00:01:17.120
So, uh why do I talk about evolution and why is it so important?

s16
00:01:17.120 --> 00:01:25.800
Uh I think basically evolution is this like open-ended process that has gotten us to a lot of different things that we really like.

s17
00:01:25.800 --> 00:01:30.360
Uh it started in biology, it's moving to science, technology, and eventually AI.

s18
00:01:30.360 --> 00:01:37.680
And I think it can inspire us in a lot of different ways to build better AI systems as well.

s19
00:01:37.680 --> 00:01:44.680
In fact, uh whenever we take out and there's this famous saying, "Whenever I fire a linguist, my accuracy goes up."

s20
00:01:44.680 --> 00:01:48.200
Uh I think that's true for machine translation back in the day.

s21
00:01:48.200 --> 00:02:02.120
And it may be true that we should fire all the AI engineers uh and that that are here uh and have them mostly manage an actual AI engineer that is AI and works on AI.

s22
00:02:02.120 --> 00:02:06.360
Uh and so that may be uh one of the conclusions of this talk.

s23
00:02:06.360 --> 00:02:16.160
Uh and I think most of us are going to be excited about it cuz it means that we'll all become managers of such an AI rather than having to do the nitty-gritty ourselves.

s24
00:02:16.160 --> 00:02:16.520
All right.

s25
00:02:16.520 --> 00:02:18.720
So, let's start with evolution, right?

s26
00:02:18.720 --> 00:02:22.480
The really, really big picture, 3 and 1/2 billion years or so.

s27
00:02:22.480 --> 00:02:35.480
Uh this is kind of the incredible process uh that has led from, you know, simple bacteria and plants and fish and amphibians and so on to, after many billions of years,

s28
00:02:35.480 --> 00:02:36.360
us.

s29
00:02:36.360 --> 00:02:38.080
So, I That's That's a good starting point.

s30
00:02:38.080 --> 00:02:43.680
That gives us some indication that evolutionary processes can do pretty amazing things, right?

s31
00:02:43.680 --> 00:02:49.400
But now, let's zoom in and uh go maybe down to a few million years.

s32
00:02:49.400 --> 00:03:02.239
There, we can also see how in a very first primitive ways, technological evolution uh has basically increased the world's uh sort of product uh in terms of monetary value.

s33
00:03:02.239 --> 00:03:08.400
It's a little bit harder to estimate in the beginning, but we can see these sort of sequences of exponentials.

s34
00:03:08.400 --> 00:03:10.920
And most exponentials eventually become S curves.

s35
00:03:10.920 --> 00:03:23.680
They flatten out, but humanity has done pretty well by basically developing uh many of these very basic technologies, hunting, farming, but then also thinking about science, the scientific method

s36
00:03:23.680 --> 00:03:27.640
um in the early days of the Enlightenment, and then of course the Industrial Revolution.

s37
00:03:27.640 --> 00:03:34.640
So, now we can zoom even further, uh and no worries, we're eventually going to get to Nanochat and actual auto research and and what we're doing.

s38
00:03:34.640 --> 00:03:37.120
Uh it's a very, very quick zoom.

s39
00:03:37.120 --> 00:03:41.600
Um and now we can zoom down to last few thousands of years.

s40
00:03:41.600 --> 00:03:47.200
And what we're seeing there is that with more technology, we were able to sustain more people, right?

s41
00:03:47.200 --> 00:03:54.280
So, when we're working on pushing that frontier forward, uh we're very certain that that will lead to more human flourishing, right?

s42
00:03:54.280 --> 00:04:04.480
And especially in the last few uh, hundred years, we're seeing this incredible explosion in the population of people because of technology.

s43
00:04:04.480 --> 00:04:07.160
And the evolution uh, that it brings.

s44
00:04:07.160 --> 00:04:12.200
And in many cases, that evolutionary process is run by us, so it's sort of conscious.

s45
00:04:12.200 --> 00:04:20.200
Uh, but there are sort of interesting uh, inspirations that we can take from that as we're thinking about the evolution of AI in the next cycles.

s46
00:04:20.200 --> 00:04:27.320
Uh, in fact, and I might not agree with everything with Marc Andreessen, but uh, he is very smart and we agree on a lot of things.

s47
00:04:27.320 --> 00:04:38.080
Uh, and so I think he wrote this really great uh, techno-optimist manifesto in which he, I think, correctly points out that the only perpetual source of growth for the entire economy,

s48
00:04:38.080 --> 00:04:46.720
a lot of people worry about AI taking jobs and things like that, but the truth is it will very, very likely increase uh, the economy massively and that will benefit

s49
00:04:46.720 --> 00:04:48.480
uh, benefit a lot of us.

s50
00:04:48.480 --> 00:04:51.320
And so the perpetual source of growth is technology.

s51
00:04:51.320 --> 00:05:00.800
Uh, in fact, we can go even further and say that there's no material problem, and again, it's not sort of psychological problems and things like that, but no material problems

s52
00:05:00.800 --> 00:05:03.680
uh, that cannot be solved with even more technology.

s53
00:05:03.680 --> 00:05:04.440
Right?

s54
00:05:04.440 --> 00:05:14.160
For problem of starvation, we invented the green revolution, darkness, light, uh, cold, indoor heating, heat, air conditioning, and the list goes on.

s55
00:05:14.160 --> 00:05:23.280
So, I think we can kind of realize that this evolutionary process has been going on for a very long time and continues to make a huge amount of progress.

s56
00:05:23.280 --> 00:05:31.840
In fact, the progress is so fast that there can within one lifetime be a major, major shift.

s57
00:05:31.840 --> 00:05:32.120
Right?

s58
00:05:32.120 --> 00:05:42.240
If you're born in 1900, uh, then 3 years, when you're 3 years old, the first human ever was able to, thanks to the Wright brothers, kind of have sustained

s59
00:05:42.240 --> 00:05:43.919
motored flight.

s60
00:05:43.919 --> 00:05:51.240
And then about 60-ish years later, in 1969, humans flew all the way to the moon, right?

s61
00:05:51.240 --> 00:06:04.040
So, that within one lifetime, humanity went from like no one can fly for a very long time other than sort of gliding down a hill or something, no one can really fly to we all fly to the moon, right?

s62
00:06:04.040 --> 00:06:14.520
And so, for us, I think, what that means is we're probably, and I sometimes say this, we're like too late to explore Earth, we're too early to explore the stars,

s63
00:06:14.520 --> 00:06:24.400
but we're right on time to build an AI that could actually do what flying did for some in one lifetime due to intelligence.

s64
00:06:24.400 --> 00:06:32.840
We can build and move from AI being worse at everything that we do to possibly being better at any specific task that we do.

s65
00:06:32.840 --> 00:06:33.040
Right?

s66
00:06:33.040 --> 00:06:40.720
And that that will probably be our our 60-year time frame, and because everything moves faster, it might only be 30 years or so.

s67
00:06:40.720 --> 00:06:45.880
So, then, uh there's an interesting connection between technology and science and theory, right?

s68
00:06:45.880 --> 00:06:51.440
Like sometimes the application comes first, and then we develop the theory later, and then improve at the technology.

s69
00:06:51.440 --> 00:06:56.440
Sometimes the theory comes first, and from that we can build new kinds of technologies.

s70
00:06:56.440 --> 00:07:08.320
And so, it's very helpful to think a little bit about the philosophy of science, and no better uh to be inspired there than Karl Popper, wrote that just like in other types of evolution,

s71
00:07:08.320 --> 00:07:13.880
when we choose a theory, we also choose one that is best uh in competition with other theories.

s72
00:07:13.880 --> 00:07:19.440
Of course, you need if you wanted LLMs to do that, they need to find them, you need web search, for instance.

s73
00:07:19.440 --> 00:07:28.600
Um but, uh in the theory that best holds its own, uh it's one that, just like evolution, has a certain natural selection process, right?

s74
00:07:28.600 --> 00:07:36.560
It proves itself, uh and there is also a sort of survival of the fittest uh going on in scientific theories.

s75
00:07:36.560 --> 00:07:49.000
And uh in fact, uh a lot of science, according to Popper, is basically us proposing a new theory hypothesis or explanation or description and then subjecting it to rigorous empirical testing.

s76
00:07:49.000 --> 00:07:57.960
That is the essentially evolution evolutionary pressure of scientific theories.

s77
00:07:58.520 --> 00:08:13.080
And basically that was a very short run through so the history of open-ended evolution which hopefully makes us all realize that more science lead to more technology which will lead to more growth which lead to more human flourishing.

s78
00:08:13.080 --> 00:08:23.000
And so that then begs the question does it make sense for us to try to just scale up and spend a lot of our resources as humanity to scale up scientific discovery

s79
00:08:23.000 --> 00:08:26.080
in order to lead to this flourishing.

s80
00:08:26.080 --> 00:08:39.200
When when you double click into that you kind of realize which Stanisław Lem already realized a long time ago that the exponential growth of science will actually be at some point halted by the lack of people working on it, right?

s81
00:08:39.200 --> 00:08:48.520
There's so many niche subfields now in all the different areas of science that is very hard to get a million people to work on that particular thing.

s82
00:08:48.520 --> 00:08:58.160
And so as a result of this incredible widening of the scope he says the number of people focusing on any single section of it has decreased.

s83
00:08:58.160 --> 00:09:09.760
And that then leads us to really thinking about how could we automate this and automate scientific discovery and that then leads us to what I call the Eureka machine.

s84
00:09:09.760 --> 00:09:19.480
This is basically our attempt at trying to build a machine that automates the process of scientific discoveries.

s85
00:09:19.480 --> 00:09:33.360
And in fact I like in a couple months I'll have a book coming out on on this exact idea and so I'll just give you a super high-level highlight of how such a Eureka machine could be built for basically everything from physics, chemistry, biology, neuroscience,

s86
00:09:33.360 --> 00:09:36.880
medicine, economics, astrophysics and so on.

s87
00:09:36.880 --> 00:09:41.840
And there are essentially four pillars that are all extremely important to this machine.

s88
00:09:41.840 --> 00:09:50.080
One is, of course, you have to understand what knowledge is already out there, what things humanity has already invented.

s89
00:09:50.080 --> 00:09:57.480
You have to get all the scientific measurement data into, as in the second pillar, this machine.

s90
00:09:57.480 --> 00:10:05.080
Then, for things that you cannot yet measure, we don't yet know, you should try to then build simulations.

s91
00:10:05.080 --> 00:10:10.640
Anything you can simulate, you can verify, and you can then solve with AI.

s92
00:10:10.640 --> 00:10:22.720
And if all else fails, or at the very end of these processes, you still need to have some kind of physical industrial like lab that actually can run real experiments in the real world.

s93
00:10:22.720 --> 00:10:35.880
And on top of all of this, you'll have a basically an agent swarm that will deal with all of these different sources of knowledge and data and experimentations and and rewards.

s94
00:10:35.880 --> 00:10:48.839
And in terms of, you know, the foundational model of knowledge, of course, we also, you know, basically is is a good example of how every single technology we've built so far, especially

s95
00:10:48.839 --> 00:11:00.440
in AI, but also before that, the internet, browsers, GPUs, and so on, we can rethink, and there are a lot of startups possible in rethinking every single one of the layers of technology

s96
00:11:00.440 --> 00:11:03.400
as infrastructure for superintelligence.

s97
00:11:03.400 --> 00:11:09.680
And at you.com, for instance, we work on web search for LLMs, right, and agents and so on.

s98
00:11:09.680 --> 00:11:12.280
And that actually is quite different, right?

s99
00:11:12.280 --> 00:11:18.440
Agents can read thousands of very long snippets, rather than just 10 blue links with like a very short snippet.

s100
00:11:18.440 --> 00:11:31.480
And so, you can rethink each of these different layers of technology that we've built for people, and rebuild them for AI in order to use them as tools to then build

s101
00:11:31.480 --> 00:11:33.360
a superintelligence.

s102
00:11:33.360 --> 00:11:37.280
Now, that is essentially uh the sort of why.

s103
00:11:37.280 --> 00:11:43.280
Like like we want to build superintelligence in order to automate science.

s104
00:11:43.280 --> 00:11:53.800
Uh and to me, that will be the next big step function change uh in in humanity uh and technology as we know it.

s105
00:11:53.800 --> 00:11:56.480
Now, how do we actually build it?

s106
00:11:56.480 --> 00:12:00.400
Uh I think the best way to build it is to have it built itself, right?

s107
00:12:00.400 --> 00:12:05.280
We moved as a field and especially natural language processing for instance, which I've worked on for many years.

s108
00:12:05.280 --> 00:12:08.320
We moved from not having linguists.

s109
00:12:08.320 --> 00:12:20.400
This feels like ancient, you know, BC uh history, uh but before ChatGPT, um we we moved from having linguists tell us a bunch of things about language and then training statistical models on top of that.

s110
00:12:20.400 --> 00:12:32.440
And when we allowed neural networks to actually automate learning those features with word vectors and uh other neural network architectures and back-to-back uh and end-to-end learning and backpropagation, we basically

s111
00:12:32.440 --> 00:12:35.760
uh were able to get much bigger improvements.

s112
00:12:35.760 --> 00:12:38.160
Uh then we did a bunch of architecture engineering.

s113
00:12:38.160 --> 00:12:45.280
Now a bunch of people at least are working on a unified architecture, uh but even that unified architecture has a lot of manual processes.

s114
00:12:45.280 --> 00:12:54.760
And so, it's clear over and over again in AI that when we take out a manual process and we replace it with a learned system, improvements will follow.

s115
00:12:54.760 --> 00:13:02.400
Uh and so, that's why I think uh we should try to build a supermachine by having uh an RSI that builds itself.

s116
00:13:02.400 --> 00:13:13.400
And the beauty is that only now um AI can actually do this because AI is code and AI can code now.

s117
00:13:13.400 --> 00:13:21.160
This this ability to really code in longer and longer time horizons has really only happened in the last like 6 to 8 months.

s118
00:13:21.160 --> 00:13:31.120
And that now enables such an RSI to work on itself, to develop almost a certain sense of self-awareness of its own shortcomings and then fix those shortcomings.

s119
00:13:31.120 --> 00:13:41.040
Uh and then once we have that machine that has gotten really, really good at doing research in AI itself, we can then use it to do AI research for a lot of other things

s120
00:13:41.040 --> 00:13:44.240
uh in in other scientific fields.

s121
00:13:44.240 --> 00:13:46.880
And so at a high level it's quite easy, right?

s122
00:13:46.880 --> 00:13:52.000
We have three steps: ideation, implementation, and validation of ideas.

s123
00:13:52.000 --> 00:13:56.880
That's true for basically almost every scientific field.

s124
00:13:56.880 --> 00:14:10.320
And so uh to end maybe on some very specific examples, uh we have built this first kind of version of such a Eureka machine uh and we wanted to just show that it works on some small

s125
00:14:10.320 --> 00:14:13.680
uh samples that a lot of people know uh and are aware of.

s126
00:14:13.680 --> 00:14:25.120
And so we basically started uh with three things that show you and give you a very first glimpse of an and sort of simple proof points uh of what such a machinery can do.

s127
00:14:25.120 --> 00:14:31.040
And that was basically better training, faster training, and and better kernels uh for for Nvidia GPUs.

s128
00:14:31.040 --> 00:14:35.839
Um the first one, NanoChat, um I'm sure many of you have heard of it.

s129
00:14:35.839 --> 00:14:45.520
A lot of people think that's already recursive self-improvement and it's kind of a weak form in the sense that usually when you do auto research, it's it's not recursive self-improvement,

s130
00:14:45.520 --> 00:14:45.760
right?

s131
00:14:45.760 --> 00:14:58.080
True recursive self-improvement is when you have an AI that has a sense of self-awareness of its own shortcomings, full access over everything uh in its arsenal from pre-training to RL training

s132
00:14:58.080 --> 00:15:04.560
and harnesses and everything, and then actually updates that entire system in the next version of itself.

s133
00:15:04.560 --> 00:15:14.440
Now you can also take such a system and just ask it to improve some other process, some other AI, like a small NanoChat run where you can train something in 5 minutes.

s134
00:15:14.440 --> 00:15:18.280
And that is really exciting and it's an important milestone, but it's not actual RSI.

s135
00:15:18.280 --> 00:15:31.760
So here we basically showed three examples of such an auto research system and what it can do and after a very very short time it essentially was able to outperform

s136
00:15:31.760 --> 00:15:36.960
many different teams and teams that also use other AI research.

s137
00:15:36.960 --> 00:15:39.480
So, let's double click into some of these.

s138
00:15:39.480 --> 00:15:42.040
Nano chat is really exciting example.

s139
00:15:42.040 --> 00:15:57.280
Basically, you train a very small chat model in less than 5 minutes and you basically want to have it get to the best possible bits per byte number.

s140
00:15:57.280 --> 00:16:10.000
And so, the whole community had worked on this for quite some time and got to 0.93 and after training this for a little more than a day or two,

s141
00:16:10.000 --> 00:16:13.120
we basically got it down to 0.91.

s142
00:16:13.120 --> 00:16:14.800
Which is pretty exciting.

s143
00:16:14.800 --> 00:16:25.839
Now, it wouldn't be that exciting if all it did was just find a couple of hyper parameters and tune them carefully, but it actually did find truly interesting novel

s144
00:16:25.839 --> 00:16:37.880
ideas like hash bi-grams and tri-gram embeddings and tables for those and mixing that into various value paths of the intention through variety of learned gates.

s145
00:16:37.880 --> 00:16:44.320
So, it actually started to doing more and more interesting things rather than just kind of tuning hyper parameters.

s146
00:16:44.320 --> 00:16:47.320
Another one nano GPT speed run.

s147
00:16:47.320 --> 00:16:49.120
Obviously, speed is very important.

s148
00:16:49.120 --> 00:17:00.040
So, here we're able to work on this again, apply the system and after very short amount of time it got better than people working often together with the AI

s149
00:17:00.040 --> 00:17:10.079
for over a year on on this very on this benchmark and made the whole thing another two seconds over two seconds faster at 70 seconds.

s150
00:17:10.079 --> 00:17:16.040
And again, discovering very interesting ideas in the process.

s151
00:17:16.319 --> 00:17:18.120
And then the third one is CUDA kernels.

s152
00:17:18.120 --> 00:17:25.319
Of course, we all care about not burning through our GPU budgets too quickly and trying to be very efficient.

s153
00:17:25.319 --> 00:17:35.880
I think in general it's actually kind of shocking how inefficient a lot of mixture of expert models still are running very large clusters that cost billions of dollars and then only have like 30% or so

s154
00:17:35.880 --> 00:17:36.920
utilization.

s155
00:17:36.920 --> 00:17:47.560
So a lot of work that's ongoing in the world to improve that and different fields or different groups of people are various different stages of that.

s156
00:17:47.560 --> 00:17:59.520
But long story short, lots of different CUDA kernels are used during training and testing and here we basically again took that system and after a couple days it discovered

s157
00:17:59.520 --> 00:18:13.320
better kernels than the leader boards best on the Nvidia benchmark website by again quite quite a sizable margin across all the different categories of those kernels.

s158
00:18:13.320 --> 00:18:25.680
And while we are pretty good at AI and we actually in the team didn't have any particular CUDA kernel experts who just spent their entire careers writing good kernels.

s159
00:18:25.680 --> 00:18:33.760
But still, you know, we do just enough to make sure and work together with Nvidia to make sure that there no reward hacks here and and other issues.

s160
00:18:33.760 --> 00:18:44.280
But actually found that eventually these all checked out and were indeed pretty much all the different kernels found the best solutions there.

s161
00:18:44.280 --> 00:18:57.640
And so with that I hope I could convince you that indeed RSI could be that next big S-curve an exponential that gets layered on top of previous exponentials and

s162
00:18:57.640 --> 00:19:09.360
that should help us with not just AI but eventually science and then all of technology and then allowing many more people to flourish on our planet.

s163
00:19:09.360 --> 00:19:16.840
And so maybe I'll end on this note here, which is a lot of people wonder how much longer I can go, right?

s164
00:19:16.840 --> 00:19:24.320
Every exponential eventually flattens out and it's actually quite hard to know like when we even talk about exponential growth in the eye, what does that even mean?

s165
00:19:24.320 --> 00:19:32.800
There are many different, I call them spaces of intelligence and we won't have time to go into all of all of these, but as soon as you actually try to define

s166
00:19:32.800 --> 00:19:46.560
multiple different dimensions of each of these 10 spaces that make up this complex sort of volumetric uh thing that is intelligence, you'll realize that there's still so much more to go.

s167
00:19:46.560 --> 00:19:59.520
Like on the upper bounds of intelligence, we're still astronomically far away from reaching those and across pretty much every single one of these dimensions and the spaces that they make up.

s168
00:19:59.520 --> 00:20:05.680
So, if any of that is interesting and you want to help us build that, um we'd love to hear from you.

s169
00:20:05.680 --> 00:20:07.880
Thank you.