WEBVTT

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

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/XEd_SRVHBgU.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
00:00:12.360 --> 00:00:13.480
Amazing.

s3
00:00:13.480 --> 00:00:14.840
Um it is Oh, sorry.

s4
00:00:14.840 --> 00:00:17.160
Pardon me.

s5
00:00:17.800 --> 00:00:19.680
Um it is so lovely to be here.

s6
00:00:19.680 --> 00:00:27.360
So, I wanted to share today uh some thoughts that I have around who gets to be at the frontier of discovery.

s7
00:00:27.360 --> 00:00:32.640
So, modern computer science as a field has only existed for the last 77 years.

s8
00:00:32.640 --> 00:00:34.200
It's kind of bizarre when you think about it.

s9
00:00:34.200 --> 00:00:42.040
So, World War II, uh all the transistor technology that was developed for radio, we finally had our first versions of the computer.

s10
00:00:42.040 --> 00:00:46.880
But, when you think about it, that's only two generations of people working on these tools.

s11
00:00:46.880 --> 00:00:56.920
However, within that time, even for computer science, who and what and what topics we work on has dramatically changed.

s12
00:00:56.920 --> 00:01:06.680
And I think it's an interesting setting because actually, if you look back across science as a whole, how we do discovery has been markedly different uh at different points in time.

s13
00:01:06.680 --> 00:01:18.080
So, when we started, the whole idea of a researcher was what we call like a gentleman scientist, typically someone with adequate wealth to dabble in a discovery.

s14
00:01:18.080 --> 00:01:21.760
And these were all individuals, independent researchers.

s15
00:01:21.760 --> 00:01:32.200
Uh when we uh see like the first associations emerge with the Royal Society in the 1600s, the idea of being a scientist as a full-time job was very special.

s16
00:01:32.200 --> 00:01:35.960
And these became the predominant spaces for discovery.

s17
00:01:35.960 --> 00:01:37.600
Why am I talking about this?

s18
00:01:37.600 --> 00:01:48.840
Because to be honest, like the professionalization of science uh led to what I would call, and you know, my good friend Rosanne Liu calls, the unreasonably narrow path.

s19
00:01:48.840 --> 00:01:50.880
So, I'm an AI researcher.

s20
00:01:50.880 --> 00:01:53.160
Uh this is Yans' actual career.

s21
00:01:53.160 --> 00:01:59.480
Uh and it's interesting, if you were to be an AI researcher at the forefront, you basically had to follow this exact narrow path.

s22
00:01:59.480 --> 00:02:01.680
You had to get into the right PhD program.

s23
00:02:01.680 --> 00:02:03.880
You had to then go to the right industry lab.

s24
00:02:03.880 --> 00:02:08.840
You had to do sufficiently interesting work, and then finally you got to contribute to the frontier.

s25
00:02:08.840 --> 00:02:10.280
This was my story, too.

s26
00:02:10.280 --> 00:02:13.560
So, a lot of my work has been on efficiency at scale.

s27
00:02:13.560 --> 00:02:25.000
Uh I did my PhD and worked at DeepMind and a lot of different frontier labs, but this was in many ways um a very aggressively filtered system.

s28
00:02:25.000 --> 00:02:32.880
If you did not make it or you were not curious about the right problem at the right time, you didn't have um a place to play at a frontier lab.

s29
00:02:32.880 --> 00:02:36.160
And this is like the standard successful scientist.

s30
00:02:36.160 --> 00:02:39.000
Like you have a famous advisor, hopefully.

s31
00:02:39.000 --> 00:02:42.760
Uh you hopefully get one or two important internships.

s32
00:02:42.760 --> 00:02:49.000
And what's interesting about this is computer science was really about representing the world, and it was like all the tools that do that.

s33
00:02:49.000 --> 00:02:53.000
But the reason why most people got into computer science is the question at the end of it.

s34
00:02:53.000 --> 00:02:56.240
Like if you can represent the world, what questions can you answer?

s35
00:02:56.240 --> 00:03:06.880
So, today I'm going to talk about what I think is like one of the most profound and interesting topics, which is like why this is so important when computer science at this is changing.

s36
00:03:06.880 --> 00:03:08.280
So, I'll also speak to this.

s37
00:03:08.280 --> 00:03:12.680
It was double compounded in computer science because of the need for compute.

s38
00:03:12.680 --> 00:03:14.760
So, this resulted in jokes.

s39
00:03:14.760 --> 00:03:16.280
Um it's fun that Merve was here.

s40
00:03:16.280 --> 00:03:20.440
This is her tweet about GPU poor versus GPU rich.

s41
00:03:20.440 --> 00:03:23.959
Uh it led to barriers of entry on who can contribute frontier AI.

s42
00:03:23.959 --> 00:03:33.200
So, I put here company A, company B, company C. But to be honest, if we pulled, I think there would be significant majority votes about who those companies are.

s43
00:03:33.200 --> 00:03:38.360
But basically a handful of frontier labs have been able to build the technology we use.

s44
00:03:38.360 --> 00:03:41.959
Um it's also determined who gets to participate in breakthroughs and who doesn't.

s45
00:03:41.959 --> 00:03:54.720
So, uh this is a map of like what Stanford calls uh where statistically or significant breakthroughs have come through come from, and you can see the whole sections of the world are completely left out.

s46
00:03:54.720 --> 00:03:59.240
And so for me, uh this is a very important question worth answering.

s47
00:03:59.240 --> 00:04:00.680
Who gets to shape the frontier?

s48
00:04:00.680 --> 00:04:03.800
Who gets to answer the questions at the end of the pursuit?

s49
00:04:03.800 --> 00:04:08.240
We've seen that the shift has dramatically changed from academia to industry.

s50
00:04:08.240 --> 00:04:11.720
And it's also meant that we ship the same model to everyone.

s51
00:04:11.720 --> 00:04:18.959
Uh why that's particularly interesting is that most people intuitively understand that you shouldn't ship the same model to billions of people.

s52
00:04:18.959 --> 00:04:22.880
And they also understand that it's not a particularly good use of compute, right?

s53
00:04:22.880 --> 00:04:25.360
You're spending the same amount of compute on everything.

s54
00:04:25.360 --> 00:04:28.000
And some problems are hard and some are very easy.

s55
00:04:28.000 --> 00:04:29.200
So where does that leave us?

s56
00:04:29.200 --> 00:04:31.160
What's my talk for today?

s57
00:04:31.160 --> 00:04:34.840
Um I would like to say that we are ripe for a revolution.

s58
00:04:34.840 --> 00:04:39.800
And we are ripe for a revolution in who gets to participate at the frontier of AI.

s59
00:04:39.800 --> 00:04:42.600
Uh I'll tell you two reasons why I'm bullish on this.

s60
00:04:42.600 --> 00:04:48.960
Um I'll definitely cover one, and then I'll actually see interest and timing cuz I want to leave plenty of time for questions.

s61
00:04:48.960 --> 00:04:54.960
I think that's you know, unless you have a few questions and a bit of banter, these things can be kind of boring.

s62
00:04:54.960 --> 00:04:58.919
So we'll see, but I'll cover one definitely that I'm actively thinking about.

s63
00:04:58.919 --> 00:05:05.120
And this is what if we could allow anyone to build the same frontier intelligence um as that in labs.

s64
00:05:05.120 --> 00:05:06.320
And I've been in a few labs.

s65
00:05:06.320 --> 00:05:07.800
I've done my tour of duty.

s66
00:05:07.800 --> 00:05:15.880
And this is the core question I care about now is like how do you build intelligence that continuously adapts and that builders everywhere can have more control.

s67
00:05:15.880 --> 00:05:21.360
So instead of taking years of training to learn how to build the tools, scientists just skipped the questions.

s68
00:05:21.360 --> 00:05:24.120
So a few weeks ago we released Auto Scientist.

s69
00:05:24.120 --> 00:05:28.880
And Auto Scientist is really about how do you automate the training of models itself.

s70
00:05:28.880 --> 00:05:34.720
Um I'll share a few things that are really interesting about this is that one it's um co-optimizes the entire loop.

s71
00:05:34.720 --> 00:05:44.040
So it's like from data to alignment uh and it it chooses and self-evolves based upon the domain and the type of data.

s72
00:05:44.040 --> 00:05:48.160
Um what's interesting as well is that it actually outperforms research staff.

s73
00:05:48.160 --> 00:05:59.920
And mainly because like a lot of our research staff has experience with certain model types, and we're testing it across many different model architectures, different different size models, um as well as like dense and mixture of experts.

s74
00:05:59.920 --> 00:06:01.880
And that search space is a lot broader.

s75
00:06:01.880 --> 00:06:08.840
And so, exploiting it using like how do you how do you self-improve from experience and scale is very effective.

s76
00:06:08.840 --> 00:06:10.720
Um I think this is very interesting.

s77
00:06:10.720 --> 00:06:13.640
This only worked when we co-optimized the data.

s78
00:06:13.640 --> 00:06:23.720
Uh so, there's a lot of auto research projects right now, which basically treat data as like um the agent is it decides whether to create data or not or what to do.

s79
00:06:23.720 --> 00:06:30.680
Frankly, we did not get the returns for like how much you can squeeze out of performance until you control for data quality.

s80
00:06:30.680 --> 00:06:35.960
So, we actually co-optimized based on all the adaptation we did with the data exactly what we would do with the model.

s81
00:06:35.960 --> 00:06:37.600
And that was super interesting.

s82
00:06:37.600 --> 00:06:40.800
It speaks to like the need to control the entire flow.

s83
00:06:40.800 --> 00:06:54.880
Um What's fun about this is like really what auto scientists do and it's a combines all the knowledge it gained from the adaptive data component with also the knowledge of the domain and also the ability to self-improve for a domain and to learn from other components of that domain.

s84
00:06:54.880 --> 00:06:57.560
Um What What is a cheeky fact and this is quite fun.

s85
00:06:57.560 --> 00:07:01.440
You'll notice all these percentages for win rates are like 60 plus.

s86
00:07:01.440 --> 00:07:05.800
Um that's because we put the budget stopping it stopping it above 60.

s87
00:07:05.800 --> 00:07:10.440
So, like once it was above 60, the our our genetic flow could exit.

s88
00:07:10.440 --> 00:07:16.120
But we since like removed that barrier and like you can see it just go up over time, which is super fascinating.

s89
00:07:16.120 --> 00:07:20.640
Um and then I think what's interesting is like it changes a lot of the hyperparameters.

s90
00:07:20.640 --> 00:07:24.200
Typically, that humans are much more wary about changing all at once.

s91
00:07:24.200 --> 00:07:27.520
And so, you get massive exploitation of the search space.

s92
00:07:27.520 --> 00:07:35.200
I see this is crucial for like how do you reduce the amount of compute you use for customization because you train with much more predictability.

s93
00:07:35.200 --> 00:07:42.240
But also, how do you um leverage like your domain knowledge to really unlock how you build frontier AI.

s94
00:07:42.240 --> 00:07:43.240
Um and this was fun.

s95
00:07:43.240 --> 00:07:47.040
We we did uh we announced a beta like 4 weeks ago.

s96
00:07:47.040 --> 00:07:50.840
The excitement is most acute for like medical and sciences.

s97
00:07:50.840 --> 00:07:53.960
And that's largely I think and legal and and code as well.

s98
00:07:53.960 --> 00:07:58.360
But these are like domains where typically current models fall short.

s99
00:07:58.360 --> 00:08:04.960
And also domains where um in many ways like the degree of last mile customization is really acute.

s100
00:08:04.960 --> 00:08:07.160
Um and this is the core point.

s101
00:08:07.160 --> 00:08:08.560
I you know, I think this is fun.

s102
00:08:08.560 --> 00:08:15.520
I think I might even have time to cover The other point I wanted to make about why now is very important for like changing who shapes.

s103
00:08:15.520 --> 00:08:21.280
But this really the main factor that this does is it increases your innovation cycle.

s104
00:08:21.280 --> 00:08:26.680
Like and it also increases the likelihood that when you train and spend compute you'll succeed.

s105
00:08:26.680 --> 00:08:28.840
And those combined factors are super interesting.

s106
00:08:28.840 --> 00:08:31.000
Like one thing that we're doing next is extending that.

s107
00:08:31.000 --> 00:08:34.599
So even your test time compute should be adaptive based on your task.

s108
00:08:34.599 --> 00:08:39.200
Um So this kind of brings me back to where I started.

s109
00:08:39.200 --> 00:08:47.280
And like kind of the grumpy statement I said, which is like you know, we have this rarefied super narrow compounding issue of barriers to entry.

s110
00:08:47.280 --> 00:08:52.520
One is like that you need to do this um very narrow funnel of who gets to build frontier AI.

s111
00:08:52.520 --> 00:08:57.040
And the other is that um typically compute and cost really dominate.

s112
00:08:57.040 --> 00:08:59.440
Um we want to change that.

s113
00:08:59.440 --> 00:09:04.520
Like we decided okay, we're going to cover languages from day one, 242 languages.

s114
00:09:04.520 --> 00:09:07.640
And also like a big interest for us is actually non-verifiable tasks.

s115
00:09:07.640 --> 00:09:14.440
I think this is super interesting because um this is really the bulk of like everyday tasks that people do.

s116
00:09:14.440 --> 00:09:21.520
And um it's really where the meat of like what is interesting for progress is going to be over the next year.

s117
00:09:21.520 --> 00:09:23.960
Um and this leads me into our mandate.

s118
00:09:23.960 --> 00:09:31.320
We care deeply about how do you accelerate learning in a way that models should be able to um learn from their environment.

s119
00:09:31.320 --> 00:09:36.400
So right now we've moved from a era of like the the model is monolithic.

s120
00:09:36.400 --> 00:09:42.600
You know, when I was at um different parts of my research career, basically a whole team would be around building a model.

s121
00:09:42.600 --> 00:09:46.960
You give it to someone else to serve, and like you have someone else do the front end.

s122
00:09:46.960 --> 00:09:50.600
And actually now, like the most important intelligence is a model who interacts.

s123
00:09:50.600 --> 00:09:56.680
And so, this idea of how efficiently are you going to interact, how will you continuously learn from the environment, is pretty core.

s124
00:09:56.680 --> 00:09:58.440
And um I think about it a lot.

s125
00:09:58.440 --> 00:10:00.600
So, um let's see.

s126
00:10:00.600 --> 00:10:02.120
I think I do have time, right?

s127
00:10:02.120 --> 00:10:03.480
How are we doing for time?

s128
00:10:03.480 --> 00:10:04.160
Oh, I do.

s129
00:10:04.160 --> 00:10:05.040
I have plenty.

s130
00:10:05.040 --> 00:10:05.680
This is lovely.

s131
00:10:05.680 --> 00:10:13.240
So, we'll have time for questions, and I'll share a little bit about what I think um the next uh component is.

s132
00:10:13.240 --> 00:10:22.160
I think core to this, so if we just did auto scientist, but it still took enormous compute to do frontier AI trainings, I think we'd be in a bit of a pickle, right?

s133
00:10:22.160 --> 00:10:28.680
Like I'd be saying, "Oh great, you can use this agent, but don't worry, just bring your uh 10,000 GPUs with you."

s134
00:10:28.680 --> 00:10:40.160
But, I think there's another trend, which I think makes this very important timing, and rooms like this probably much more optimistic than like have been in a few years ago about who can build frontier AI.

s135
00:10:40.160 --> 00:10:45.680
Um and one of that is like the rules of like where you get rate of return for computer are totally changing.

s136
00:10:45.680 --> 00:10:51.320
So, um I wrote a very paper about this called So Death of Scaling.

s137
00:10:51.320 --> 00:10:59.240
Um but empirically, we do now know that um pre-training size in particular is not your most lucrative axis of scale.

s138
00:10:59.240 --> 00:11:00.360
And what does this mean?

s139
00:11:00.360 --> 00:11:11.600
Like if pre-training scale isn't going to dominate performance, it actually really greatly changes who can create the best recipes for innovation, because pre-training compute typically has to be co-located.

s140
00:11:11.600 --> 00:11:17.400
Um it has to be uh in many ways large volume to accommodate for redundancy.

s141
00:11:17.400 --> 00:11:23.760
Inference compute and other places where you actually apply compute, typically you can have much more like distributed.

s142
00:11:23.760 --> 00:11:27.160
It's also much more higher return given the amount of flops.

s143
00:11:27.160 --> 00:11:35.480
And so, it's interesting when we talk about what is the state of pre-training compute, we know it's not giving the same returns largely because our architecture is saturated.

s144
00:11:35.480 --> 00:11:39.800
So, um we see much smaller models outperforming much larger ones.

s145
00:11:39.800 --> 00:11:49.800
Uh this is like the Open LLM leaderboard, and this is like the daily submission of like the best small model under 13B versus all the larger models.

s146
00:11:49.800 --> 00:11:53.600
Um and you can see over time that ratio totally flips.

s147
00:11:53.600 --> 00:12:08.480
Um and also there's kind of the grumpy assessment that most recent models that have severely um played with just increasing model size haven't provided the same stepwise change um as their predecessors.

s148
00:12:08.480 --> 00:12:16.280
And a lot of that is because where the most returns for performance are now are on a broader action space.

s149
00:12:16.280 --> 00:12:24.520
Um and this is really what I was getting at when we move from an algorithm to we are expanding optimization space in new places.

s150
00:12:24.520 --> 00:12:36.880
And what's the fun about that is that these are new places where um the barriers to entry are much more nimble, and where recipe and algorithm and research matters again.

s151
00:12:36.880 --> 00:12:40.560
Um and things like how do you automate that discovery?

s152
00:12:40.560 --> 00:12:45.000
Um and so this is what I'll state, and I think then we should open up for questions.

s153
00:12:45.000 --> 00:12:48.880
Um and I would encourage good grumpy questions or fun positions.

s154
00:12:48.880 --> 00:12:50.040
Let's make the use of the time.

s155
00:12:50.040 --> 00:12:53.520
I know I was told earlier that almost no talks have time for questions.

s156
00:12:53.520 --> 00:12:59.600
I find that so disappointing, so um so we'll need some brave people to start the conversation.

s157
00:12:59.600 --> 00:13:09.160
Um but I will say this means we're better off, and I would say it's a very good time to be like working on intelligence because instead of just a handful of people getting to

s158
00:13:09.160 --> 00:13:14.400
getting to create it, it's much more now about the question you want to answer at the end of the day.

s159
00:13:14.400 --> 00:13:23.480
The reason why people did a computer science PhD was to learn the tools to get to the question, and now you can just get to the question, which is super meaningful.

s160
00:13:23.480 --> 00:13:25.080
Um okay, let me open up.

s161
00:13:25.080 --> 00:13:26.280
Where should we start?

s162
00:13:26.280 --> 00:13:28.280
We um have an abundance.

s163
00:13:28.280 --> 00:13:29.840
I hear there's no microphone.

s164
00:13:29.840 --> 00:13:36.320
So, if you want to ask a question, you want to make a statement, I will indulge a statement if it's interesting.

s165
00:13:36.640 --> 00:13:38.200
And yeah, go for it.

s166
00:13:38.200 --> 00:13:40.480
Just raise your hand and I'll I'll repeat it afterwards.

s167
00:13:40.480 --> 00:13:44.960
Let me just get to the end of this in case people want to reach me afterwards.

s168
00:13:44.960 --> 00:13:45.880
Nice.

s169
00:13:45.880 --> 00:13:46.480
Yes, go ahead.

s170
00:13:46.480 --> 00:13:47.960
Gentleman in the fourth row.

s171
00:13:47.960 --> 00:13:49.240
Go for it.

s172
00:13:49.240 --> 00:13:55.720
You mentioned uh you know uh AI from here will be more democratized.

s173
00:13:55.720 --> 00:13:58.200
Can you point to like how?

s174
00:13:58.200 --> 00:14:00.120
So, I think how it's twofold.

s175
00:14:00.120 --> 00:14:03.880
One is there's very few people who know how to train frontier models.

s176
00:14:03.880 --> 00:14:08.160
I would say realistically probably less than 5,000 in the world at scale.

s177
00:14:08.160 --> 00:14:11.560
I think that type of knowledge, that's a very exploitable search space.

s178
00:14:11.560 --> 00:14:15.800
And actually as humans, all those configurations, we're not particularly good at.

s179
00:14:15.800 --> 00:14:20.400
It's kind of like secret knowledge we pass as if, you know, we're apprentices.

s180
00:14:20.400 --> 00:14:21.360
So, that's one.

s181
00:14:21.360 --> 00:14:27.280
Like once you automate a lot of that knowledge, you just accelerate innovation cycles.

s182
00:14:27.280 --> 00:14:30.480
Which means that you can explore and do more questions.

s183
00:14:30.480 --> 00:14:38.360
Typically, what people I think often miss is that the cost of asking something informs what is asked.

s184
00:14:38.360 --> 00:14:43.720
And if you make it cheaper to ask something, you change like the volume of things that are asked, which is super interesting.

s185
00:14:43.720 --> 00:14:48.400
The other reason though, I do think it's very much a facet of like the changing nature of compute.

s186
00:14:48.400 --> 00:14:54.240
So, agentic compute, post-training compute matters a significant amount for performance.

s187
00:14:54.240 --> 00:15:01.945
That does not require the same type of um I dare I say hoarding of GPUs.

s188
00:15:01.945 --> 00:15:01.960
[laughter]

s189
00:15:01.960 --> 00:15:05.640
But like I think it it's very different compute purchasing dynamics.

s190
00:15:05.640 --> 00:15:09.440
And again, it means that the person with the best idea has a higher chance of winning.

s191
00:15:09.440 --> 00:15:10.320
Which is fun.

s192
00:15:10.320 --> 00:15:10.760
Nice.

s193
00:15:10.760 --> 00:15:11.360
What else?

s194
00:15:11.360 --> 00:15:11.960
Who wants to go?

s195
00:15:11.960 --> 00:15:15.800
I see Yeah, we can go up here and then I saw a hand back there.

s196
00:15:15.800 --> 00:15:16.839
Okay, yes, I do see you.

s197
00:15:16.839 --> 00:15:20.720
The glare is high, but you go first and then we'll come up here.

s198
00:15:20.720 --> 00:15:23.680
So, Frontier Labs care a lot about safety.

s199
00:15:23.680 --> 00:15:26.640
So, one of the challenges in safety

s200
00:15:32.480 --> 00:15:37.760
Yeah, so the question I'll just repeat it so cuz I think there's probably people who want to know who are in the room.

s201
00:15:37.760 --> 00:15:48.880
So the question was one of the I guess the counterpoints on some Frontier Labs about not enabling Frontier AI but outside is that is a safety question.

s202
00:15:48.880 --> 00:15:56.160
So I think it would be I definitely am not one of those people who who says that open source doesn't carry any risk.

s203
00:15:56.160 --> 00:16:01.400
So when you make a tool more readily available, there's a profile of risk associated with it.

s204
00:16:01.400 --> 00:16:06.320
Um Auto Scientists is to be fair like it's about enabling people to customize their models.

s205
00:16:06.320 --> 00:16:09.960
You can think of that as like a slightly different question from whether those are open source.

s206
00:16:09.960 --> 00:16:11.960
It's giving people way more control.

s207
00:16:11.960 --> 00:16:18.960
Whether that's local or private or within that company, it's about like how do they own their own intelligence?

s208
00:16:18.960 --> 00:16:25.080
Um what do I think broadly about the impact of open source on safety?

s209
00:16:25.080 --> 00:16:41.400
You can do so much the dynamic has often conflated like that real risk of like wider access with slight sense that that it it it restrains who can actually participate and I think that's a delicate balance.

s210
00:16:41.400 --> 00:16:46.640
And I think you have to acknowledge risk by also navigating that and acknowledging that it limits who can participate.

s211
00:16:46.640 --> 00:16:48.200
Yeah, so nuanced answer.

s212
00:16:48.200 --> 00:16:56.760
So I I guess I should be more bombastic on that one but I I guess I have been in this discussion a few times and I find the binary views on other sides

s213
00:16:56.760 --> 00:16:57.760
I feel like they miss a lot.

s214
00:16:57.760 --> 00:17:00.720
But anyways, okay, go ahead.

s215
00:17:08.040 --> 00:17:17.160
Oh, I think for for automating and speeding up learning, one of the core questions is how do you balance what you store in the parametric space and the non-parametric

s216
00:17:17.160 --> 00:17:17.880
space?

s217
00:17:17.880 --> 00:17:28.199
And actually, one of the most interesting things, I mentioned that this only work because we co-optimized data and model, um it will only work to do like an AutoScientist

s218
00:17:28.199 --> 00:17:31.200
for harnesses if you also co-optimize it with a model.

s219
00:17:31.200 --> 00:17:39.200
And so, it's interesting, it's actually a long horizon problem, and that's super fascinating to think about where you're optimizing the choices for each and co-training, which is cool.

s220
00:17:39.200 --> 00:17:39.640
Nice.

s221
00:17:39.640 --> 00:17:43.240
Uh I think we have time for maybe two more and then we can I'll pass on to the next speaker.

s222
00:17:43.240 --> 00:17:44.360
Nice, go ahead.

s223
00:17:44.360 --> 00:17:45.720
I have a question.

s224
00:17:45.720 --> 00:17:53.320
We talked a little bit about this actually working on the post-training side of the models is actually cheaper than doing pre-training with public data.

s225
00:17:53.320 --> 00:18:04.240
But like still, especially talking like really large models, like reinforcement learning, even fine-tuning is still pretty like GPU-intensive, so like you talked a little bit about like smaller models.

s226
00:18:04.240 --> 00:18:13.520
But I still think that most like frontier smaller models still rely on like the bigger models to distill knowledge like downwards, not necessarily to be trained.

s227
00:18:13.520 --> 00:18:24.680
So, how do you see this like like for the future, how to work like with these smaller models, how they can be start working without it being trained without depending on these larger models?

s228
00:18:24.680 --> 00:18:26.360
Yeah, actually that's an excellent point.

s229
00:18:26.360 --> 00:18:28.080
I think the question amounts to two points.

s230
00:18:28.080 --> 00:18:31.120
One, are larger models necessary for distillation benefits?

s231
00:18:31.120 --> 00:18:35.560
And then second, so frontier models are still pretty large.

s232
00:18:35.560 --> 00:18:40.000
Um so, I think for the second one, frontier models are still pretty large, yes.

s233
00:18:40.000 --> 00:18:49.280
I don't think I'm arguing that you um I I My argument is slightly different, and my argument is in no frontier AI lab is going to four exercise their model again for pre-training.

s234
00:18:49.280 --> 00:18:52.760
So, it's almost like we know we're at an opposite for this architecture.

s235
00:18:52.760 --> 00:19:05.280
If someone comes out with a new architecture, that's totally different, you can the architecture determines your ceiling, and I'm saying we are probably at the ceiling of size, which means that that's fine because it means

s236
00:19:05.280 --> 00:19:07.520
okay, it's what you innovate within that.

s237
00:19:07.520 --> 00:19:09.760
Um so, so size does matter.

s238
00:19:09.760 --> 00:19:13.120
I I that's that's a very good point to bring up.

s239
00:19:13.120 --> 00:19:20.720
Meaning I'm not advocating everyone use a 0.8B, but I am saying that we now have a more equal feeling for playing field at the the top.

s240
00:19:20.720 --> 00:19:25.560
Second point is interesting, distillation like the impact certainly.

s241
00:19:25.560 --> 00:19:28.760
So data quality in general means you use capacity a lot more.

s242
00:19:28.760 --> 00:19:39.040
So you what you will see in pre-training is instead of the size people are just moving post-training further back, which is very fascinating and a bigger lever.

s243
00:19:39.040 --> 00:19:41.600
So I agree distillation is helpful.

s244
00:19:41.600 --> 00:19:53.040
It's just that again we've hit the ceiling and so it's almost like no one is going to supersize their model or if they do it's not clear it's beneficial

s245
00:19:53.040 --> 00:20:01.960
except for small size of the distribution, which is very much the long tail and that's kind of interesting like where that trade-off is where worth that much pre-training compute.

s246
00:20:01.960 --> 00:20:02.960
So very good question.

s247
00:20:02.960 --> 00:20:05.000
One more and then I think we are done.

s248
00:20:05.000 --> 00:20:07.320
Yes, go ahead.

s249
00:20:10.080 --> 00:20:11.120
Yeah, it's actually in beta.

s250
00:20:11.120 --> 00:20:13.320
So you can I shared here.

s251
00:20:13.320 --> 00:20:14.320
You can try it in beta.

s252
00:20:14.320 --> 00:20:17.240
So we actually are offering the GPUs for free.

s253
00:20:17.240 --> 00:20:18.969
Okay, oh that's a nice question.

s254
00:20:18.969 --> 00:20:20.080
[laughter]

s255
00:20:20.080 --> 00:20:22.480
I promise I don't know this gentleman.

s256
00:20:22.480 --> 00:20:27.600
But yes, I think actually we're trying to remove the compute hurdle and I think it's quite cool to see.

s257
00:20:27.600 --> 00:20:30.280
So feel free to take a look at the beta.

s258
00:20:30.280 --> 00:20:30.880
Nice, lovely.

s259
00:20:30.880 --> 00:20:31.720
Thank you so much.

s260
00:20:31.720 --> 00:20:32.320
Really nice.

s261
00:20:32.320 --> 00:20:34.560
Thank you.