Better Data is All You Need — Ari Morcos, Datology

Hey everyone, welcome to the Leading Space Podcast. This is Alessio, partner and CTO at Decibel, and I'm joined by Swyx, founder of Small AI.

Hello, hello, and we're so excited to be in the studio with Ari Morcos, CEO, co-founder of Datology. Welcome.

Thank you so much for having me.

Ari, so you first c- came across my radar. I mean, I guess Datology is, like, a relatively, I guess, exciting or well-hyped startup, at least with the fundraising and the, the high profile of the people that you hire.

I reached out to book this interview after you worked on the Arc. I don't even know how to pronounce it. Arc, Arche?

Arc, yeah.

Arc.

It's inspired by a, a real transformer that was-

Okay

... called Arc.

Yeah, the, the Arc foundation models. Uh, you guys have been doing a lot of, like, data work. How would you describe Datology today?

Yeah, so, so our mission at Datology is to take everything around the data side of, of machine learning, right? So going from you have a bunch of data sitting in storage to you're gonna feed it into a model, you know, via a data loader.

There are a ton of choices you would make in that process, ranging from how you're gonna filter the data, how you're gonna sequence the data, what synthetic data you're gonna generate, if any, um, how you're gonna batch the data, all of those things.

A- and those all have a tremendous impact on the performance of the model that you train on, on the data. One of my favorite catchphrases is, "Models are what they eat." If you show them great data, they're gonna be really high quality.

If you show them low-quality data, they're gonna be low quality. But, but this is a frontier research problem. How do you actually do this effectively? How do you do this automatically at scale, right? It has to be automatic to be able to process trillions of tokens, billions of images, things like that, and that's our mission at Datology, is to take that whole process, make it really easy so that anybody can get access to state-of-the-art data curation without needing to be an expert themselves.

Um, and in doing so, help, uh, the folks we work with to train models much faster to much better performance, um, and to also help them train much smaller models to the same or better performance, which I actually think is some of the most exciting stuff going forward.

But fundamentally, that's what we do at, do at Datology, is help people curate their data so they can train models faster, better, smaller.

So the keywords for that, data curation as a service, data efficiency, all those terms. In the pre-chat before we started recording, you mentioned that there's a cool story around how you got into data in the first place, right?

You were at GDM. You were at Meta as a research scientist. Describe how, like, that became an interest.

My PhD is actually in neuroscience. Uh, so I come much more from an empirical science sort of background. I actually spent time trying to teach mice how to count and then analyze the activity of thousands of neurons in the brain while mice did count and try to understand how did that actually happen.

How, what were the neural dynamics that, that enabled that? Um, and that's actually initially how I got into to machine learning, was as a means to analyze my neur- my neural data sets. I also started my PhD 2011.

So AlexNet c- came right after that, Charlie DQN right after that. Lots of evidence that AI was going to be very, very exciting, which, which led to me transitioning. But as a result, because I had this kind of somewhat different background of being trained as an empirical scientist rather than as a computer scientist, my real first mission when I, when I joined AI was to try to build more of a science of deep learning.

Something that I think, you know, is still true today in many cases is that deep learning is an empirical science and but most people that, um, have computer science backgrounds were trained more in the context of a, a branch of theory, um, right?

Everything was very provable. That was the initial pushback to deep learning, actually, was that you couldn't prove anything in it. Um, but deep learning is, at its core, an empirical science, right? We have to run large experiments. We understand the rules for how we design these systems, but the properties that come out of them when we actually train them on t- ton of data, um, are emergent and unexpected.

So I always really wanted to, to write these papers where they had two halves, where the first half of the paper was trying to understand why is this representation desirable or, or undesirable. Why is a model good or bad?

Um, and then understand that, and then use that understanding to then improve the model, and that was always my goal. That was kind of the perfect paper. Rather than just throwing spaghetti against the wall and seeing what stuck, we were able to really understand why something didn't work and then use that understanding to improve it.

Unfortunately, it turns out that it's not so difficult to do the first half of that, try to understand the system, but really, really difficult to actually use that understanding to improve the system. Um, a lot of times what would happen is you go to optimize for this variable.

You find, hey, here's this property of representations that makes models good. You go and you optimize for that, and then it turns out that wasn't a causal variable. That was a correlate, um, and it doesn't actually work. So I maybe wrote 30 papers where we did that first half, and maybe only three or four where we did that second half.

And, and that was always kind of frustrating and dissatisfying to me. And then around 2020, I had several papers that all kinda slapped me in the face at the same time with the same insight, which is that all that really matters is the data, and I had come into all three of these papers very much focused on inductive biases.

How do we put better inductive biases into models, either through changing the objective or through changing the architecture? Which is where most of the field was and still where you see a lot of the papers at the big conferences are about architectures and various tweaks to architectures.

But I had these multiple papers, all of which made this clear takeaway that the data is the only thing that matter. I'll give you one example. There was a paper we had called ConvIT, uh, where the idea was to take a vision transformer and initialize it as if it was a convolutional neural network.

And that way you could actually start with this inductive bias of, of convolution, but the model could choose to unlearn it if it wanted to. So the idea was it was a soft inductive bias, not a hard inductive bias.

ConvNets have a hard inductive bias. You can't not be convolutional in a ConvNet. But in this case, you, you initialize the transformer that way, and then if it wants, the model could learn not to be that. And the idea here was that this would be really helpful for models to give them this inductive bias, but then they could learn not to use it if they didn't want to.

Ju- just to follow up, you-- there's a one-to-one mapping of a ConvNet to a transformer, and you can map it directly onto the weights?

Exactly. You can map it exactly correctly, it turns out. Um, if you make it just a, you know, say you have not a, a three-by-three kernel. You can have nine heads. Each head corresponds to-

Yeah

... a different part of that kernel, and then you can initialize it so it is exactly.

So it's like a very coarse thing that can then be refined as with training.

Exactly.

Yeah.

And then it can choose to change its weight so that it, it can undo the weight tying-

Yeah, yeah

... that you impose on it this way.

Yeah.

We actually had a follow-up paper which showed you could take a trained network and actually intrain-- instantiate a trained CNN as a VIT as well. So there's a way to do this. Turns out in the small data regime, and when I say small data here, I mean, say, less than five hundred thousand data points- Um, and this was in the context of, of image, uh, self-supervised learning.

So in that small data regime, this is super helpful. And where, where this paper's actually been cited is a whole bunch of kinda niche scientific problems where there's very little data. For example, volcano prediction, where you have, like, 1,500 data points or things like that.

But the advantage of using this soft inductive bias decays as the data size increases and eventually actually becomes harmful. So if you see enough data, and the threshold at which this changes is around, like, a million data points, so it's not massive by any stretch by our current model.

So basically, once you get past a million data points, that soft inductive bias no longer helps you. Um, and it actually now i- is mildly, um, harmful. So I had this paper and a, and a couple other papers that all kind of made this, this same point, that basically, you know, when you get to enough scale, inductive biases matter not at all.

All that really matters is the learned posterior from the data distribution, and that's really what defines everything. And then, of course, the rise of the transformer really showed that actually starting with models that have fewer inductive biases built into their architecture, it, you know, is the right thing.

So we had this kind of... these... the combination of factors which ultimately, like, actually was very, very confronting for me, um, because I had spent the last six years of my career working on inductive biases. Um, and now I'm faced with, you know, several different papers, all of which show me that, "Hey, what you've been working on isn't actually really that important."

Um-

Bitter lesson built.

Bitter lesson indeed. So, you know, the bitter lesson was indeed very bitter, uh, for me. And, you know, that was really my, you know, inculcation in it, I suppose. Um, where at the end I kind of thought to myself, "Okay, clearly the bitter lesson is true here.

What should I do in this new world?" You know, and, and it, it became clear to me that there are really two options that made a ton of sense, either go work on making GPUs go brrr, and I'm not a hardware engineer, I don't know how to make GPUs go faster, or work on data.

A- and, and for a whole bunch of reasons, data has been dramatically under-invested in relative to its impact. Something I, I've said before and I'll say again is, is that data is the most under-invested in area of research relative to its impact, and I don't think it's even close.

And there, there are a whole bunch of reasons for this, um, which we can go into, some of which have to do with the culture of machine learning, some of which had to do with the incentives that have been set up.

But data has systematically generally not been considered. And even if you go and you look at the scaling laws work from Kaplan and, and Chintala and all these other things, they all assume IID data, um, which is insane.

We know that all data are not created equal, that garbage in, garbage out is, like, the oldest adage in computer science, and yet all these scaling laws assume that all data is created equal. That, that makes no sense, um, whatsoever.

That's what led me to start working on this problem. And it turns out that there's a really cool thing about data research. In addition to it being something that's impactful relative to the investment, which makes it a great research area and makes it an even better company, what I'd said be- previously was that with representations, you have this disconnect where there's the questions which are kind of scientifically interesting about understanding why a representation is good, and then the questions that are practically relevant, how do I use this to improve it?

And I think what was so frustrating to me early in my career was that those were different questions a lot of the time. Um, the questions that I wanted to ask, which were curious, um, curiosity-driven and really interesting to me as a scientist, ended up often not being the questions that were practically relevant downstream.

But it turns out with data, this is no longer true. With data, if you can understand what makes a given data point useful or what makes a given data point not informative, you can almost always use that insight to make a data set better and therefore make a model better.

So what this means is that the set of questions which are scientifically interesting and the set of questions which are practically relevant in data research are largely the same questions, and that's really rare to find in research, period.

And what this means is that we can ask the questions which, as scientists, are extremely motivating to us, but then have very high confidence that those qu- the answers to those questions are going to help us to build models that train much faster, that train to much better performance, and that can train with, with far fewer parameters.

So that's a little bit of a high level of kind of how I got into, to the data problem and, and I think the pain that I had to go through- ... to, to, to get there in the first place.

You mentioned something about the incentives in the data not being aligned. Can you unpack that? Because I think from the outside you have companies like Scale that obviously have become super successful, so people are investing a good amount of money.

But what you're basically saying, like, you know-

Still not enough

... NVIDIA's like $4 trillion and Scale is not $4 trillion. So why do you think there's that inefficiency?

Okay. So first off, we have to divide the research community from the industrial community, 'cause I think they're very different. And I think in general, data work has been far more valued, um, in industry consistently than it had been in the research community.

First and foremost part of this is that data work has just often been considered s- a second-class citizen sort of work. It's the grunt work, it's the plumbing, it's the stuff that, you know, you don't wanna work with as a, as a, you know, super hoity-toity scientist.

There are even some tweets recently going around, people saying, you know, data cleaning is, is boring, it's low-value work. Whereas I think what you'd find is that if you talk to the most talented AI researchers and you ask them, "What's the secret to your success?"

They'll largely tell you that they look at the data. Ultimately, these models are a reflection of the data that you show it. And yeah, it can be tedious, it can be challenging, but it is so critical to, to get this right.

So I think first off, there's this general perception that this is lower quality work or, or, or not quality, but, but lower prestige work, and that's been there for a long time. I think part of this had to do with the way that research incentives were set up.

The data set was viewed as the given. So if you think about research circa, say, 2018, given ImageNet, maximize performance on the val set or on the test set, right? But the, the data set ImageNet was given as something you don't change.

Even Kaggle had this framework, right? Given the data set, go and, and, and make this better. People might try things like bootstrapping or stuff like that, but generally the assumption was you're gonna improve the model through better modeling, not through improving the data set.

And part of this also was just in the supervised learning era, this made sense, right? We generally weren't compute limited. We were generally very data limited, right? Data- It was very scarce. Like, if you wanna assemble ImageNet, you have to go to MTurk and get a whole bunch of people to label the dataset.

And then there's generally some quality floor, right? Because a human has looked at every data point in this dataset. Even if there's still a lot of errors there, at least it's not gonna be as bad as just the internet scraped.

But then in, in 2019, the field underwent this pretty massive change, right? We figured out how to train without labels. And one of my, like, more controversial viewpoints, I think, is that I think the transformer is a great advance to be sure, but I think it's one of a very large set of equivalently good architectures, um, that we could have found.

And there are many, many ways we could, we could get to the same performance without the transformer. But I do not think there's any way we could get to where we are today without self-supervised learning and the ability to train on unlabeled data.

That was the, the real advance, uh, to my mind that enabled us to get these in- incredible increases in capabilities.

Which is, like, the mask objective.

It's not just masking objectives. I think mask language modeling objective is one, but even next token prediction, right?

Yeah.

But generally this notion that, hey, instead of having to get an external label from a human, we can ask the model to predict one aspect of a data point from other parts of that data, and that is really powerful.

'Cause think about it, right? That meant that we went from ImageNet, a million data points, to literally trillions of tokens, a million-fold increase-

Yeah

... in data quantity in a matter of, like, several years.

Yeah.

That's completely unheard of. And that also changed everything, 'cause now we went from data being scarce and having a high-quality floor to now all of a sudden data is absolutely massive. All of our models are basically always underfitting the data.

Um, whereas previously we would do 160 epochs on an image dataset, right? Where they would all be overfitting the data generally. So now we move to this underfitting the data regime. There's no more quality floor, and now we, we have all of these problems with redundancy, with, with low quality, with low information gain, all these various things that come with these massive unlabeled datasets.

So I think the problem also changed pretty dramatically from the 2010s to the 2020s. Um, and I think that's what makes it so exciting as a scientific question, is that this didn't really make sense to study prior to 2020.

But now this makes tremendous sense and is, I think, absolutely critical for us to solve in order for us to enable these models to continue to improve and also to enable the cost-effectiveness of these models so that they don't just stay as something that's only possible to achieve if you have hundreds and hundreds of millions of dollars.

Yeah.

Making the data better can be a massive compute multiplier. It can change the performance per dollar by orders of magnitude. Um, and, you know, in many ways that's our whole goal is how do we make that easy and effective for everyone?

Totally.

And you were at Meta from 2018 to September '23, which is both during Llama 1 and Llama 2. At what point inside of Meta maybe did some of these learnings become apparent? Like, okay, we should start to spend resources working on, on this.

You mentioned 2020, so I'm wondering if that was, like, clear to you.

I think Llama 1 was already a big breakthrough.

Yeah, Llama 1 definitely put more effort into, into data filtering, I think than, than, than many others, and definitely started to change this. But even then, I would say that actually, you know, even when I left Meta, this was still an area of, of kind of the idea of actually curating the data to figure out what's the high-quality, high-value data, I think still was fairly underappreciated.

And if you, if you talk to a lot of the folks on the data teams within the big frontier labs, um, what you'll find is that they've actually invested really heavily in crawling oftentimes. They've really worked on getting better crawlers and trying to clean up the source of the data that's coming in, which makes sense.

But ultimately, you know, I think what you really need to do is you need to take this perspective of given everything that the model has seen so far and given a potential candidate set of data, what data point is gonna teach the model the most the next time it sees a data point?

And that's a pretty different framing for how to think about this problem. And I, I think we certainly-- there's certainly been some great work done, although it's all secretive, within, I think, the bigger labs. But that's a really hard problem.

That's a frontier research problem, and, and I don't think we, we still know how to solve that. I think data curation also is a hard problem to solve, quote unquote, because it's not one where there's a single silver bullet.

There's not just do this one trick and all of a sudden things work. It's rather, here are these 50 different things that you can do, each of which provides a pretty modest gain on its own. But then if you can figure out how to make them combine, you then get a really big gain.

But you have to figure out, first off, what are all of these different things you wanna do? And then two, how do you make them play nice with each other? Because by default, they don't play nice with each other.

Yeah. I'll make a quick observation on, uh, you mentioned self-supervised learning. I, I definitely agree, like, that, uh, just getting rid of labels altogether is, is great, or forming your own labels, right? And I have a general observation that I think that extends to things that are not just learning.

So self-supervised, I don't know, optimization, self-supervised neural architecture search, self-supervised curation. If you can just automate everything, I think that's the lesson really. Just, like, just, just get the machines to do it because we are the rate limiters if we, if we must label everything.

Yeah. I think, I think that's very true, and it's actually something I think about a lot is are we actually falling prey to the bitter lesson again here by trying to have human-guided methods of, of data curation? Um, probably the best open effort on data curation is DCLM, DataComp LM.

It was led by Ludwig Schmidt, a professor at Stanford, and, you know, about 30 students across many different institutions. Really wonderful effort to kind of curate, uh, Common Crawl-style datasets.

Yeah, we've, we've actually covered DataComp and DCLM on, on the podcast.

Awesome. Great. Um, but DCLM had a really cool, uh, study at the end of the paper that, um, I don't think gets, uh, nearly enough attention as it should. So okay, so they had these 30 grad students spend, you know, two years basically trying to design what are the optimal filtering criteria for these, um, models, right?

And they built a system that's pretty good at this, um, right? So then they asked all those students, "Predict what that system is going to do." So given a data point, is the system going to say keep the data point or is it gonna say reject the data point?

These are nominally the best experts you could ever hire to do this. These are students who have just spent all of their time looking at NLP data for two years. They could not predict what the DCLM classifiers would say above chance.

So, you know, this comes up a lot of times where people often ask me, "How can you possibly do this without a human in the loop? You know, it just seems impossible. You need to have a human in to actually rate these data."

But I think that, you know, what the takeaway from, from that study is, and I think there's a number of other, uh, kind- ... piece of evidence that also suggests this is that obviously we have to be automated because humans just can't scale to billions of data points, trillions of tokens.

It's just not possible.

Yeah.

Um, but even if we could, we actually wouldn't want that. Humans are not good at this task. And to give an intuition as to why humans aren't good at this task, I think the, the easiest way to think about this is that the value of a data point is not just a function of that data point itself.

It's rather a function of how that data point relates to every other data point in the training set, right? So if I have 10,000 copies of slightly variable summaries of Hamlet, I don't need all of those. But if I were to look at any one of those individual summaries, I might say, "Hey, this is really high quality.

This is a really accurate... It, it tracks all the characters. It's well-written. It's clear." But I don't need 10,000 of those.

Yeah.

And that's just a task that a human would never be able to do because a human can't keep the whole data set in their head, obviously. So even if you could have this scale with humans, you wouldn't want to.

But so what's the right number between one and 10,000?

The unsatisfying answer is it depends. But it's also the right answer. Um, so it depends on the co- on, on how complex the concept is. So redundancy's really useful, right? And, like, removing all redundancy is a bad thing.

If I remove all redundancy, then I'd only be able to understand, say, a golden retriever in the one situation that I've ever seen it in before. I wouldn't be able to generalize, and that would be bad, right? So some redundancy is good, but I think we all have the intuitive understanding that infinite redundancy is not good.

It's bad. So where is this line for different concepts? Well, one example I like to give for this is elephants versus dogs. So elephants are pretty stereotyped. There are two kinds of elephants in the world. There are Asian elephants and African elephants.

Um, they're all gray. They all have floppy ears. They all have a trunk and some tusks. They all have, uh, you know, wrinkly skin. African elephants are bigger than Asian elephants, but largely they're all pretty similar. There's not too much variability.

So I don't need that much data or that much redundancy to understand the concept of elephants, you know, fully and completely. But dogs, on the other hand, are totally different, right? Dogs are super variable. There are hundreds of breeds, not to mention all the mixes of different dog breeds.

There are different shapes, sizes, textures, colors, all of these different things. The amount of data that I need in order to properly understand dogs is gonna be a lot higher than the amount of data I need to understand elephants.

So this comes w- to some of the challenge when you're actually trying to, to do this sort of, um, creation, at least on the filtering side, is you have to... First off, you don't get a data set where you're given, "Hey, these are a bunch of dogs.

These are a bunch of elephants." Instead, you just get, "Here's a bunch of data," right? So first off, you have to un- you have to, in an unsupervised way, discover what these concepts are, use something about that concept in order to make some inference about how complicated it is or how complex it is, and therefore how much data you need to understand it, figure out, "Okay, this is a really complicated concept.

I probably should keep a lot of redundancy. This is a really simple concept. I don't need that much redundancy," and then make that appropriate choice of what you want to remove. So these are-- this is, I think, where a lot of the challenge comes from, but these are the sorts of factors that you have to keep in mind when you're trying to design these systems.

How do you draw the line of a concept, though, right? Like because then it's like, well, the elephant and the dog, but what about mammals, and then what about... You know what I mean? It's like, how should people think about it?

Maybe it's that why you need Datology because- ... it's hard, it's hard to, it's hard to talk.

Yeah, no, I think that, th- I think that's, that's right to some extent. I mean, look, it's an empirical question, like, like all things are, right? Is, is that with every data set that you can choose different f- level of fine-grained.

Ultimately, it's a hyperparameter. It's a knob that you can tune, right, for, for how, um, aggressive are you gonna be with respect to creating new concepts versus keeping concepts together. Um, and it's one of these things where, you know, uh, to-- I think to your point, it, it's why we've run hundreds and hundreds of thousands of experiments to try to figure this out.

I, I think, you know, this is something where it requires just a lot of experimentation to understand how to do this. Um, and I think one of the challenges we have is not only do we have to make this so that this works on one data set, um, but we also have to build a system that can automatically adapt to any arbitrary data distribution and be able to make the appropriate inferences, you know, in zero-shot on a new data distribution.

So we kind of have these two sets of questions. First off is, like, how do we push the frontier of data curation forward? And then second of all, how do we do out-of-distribution generalization-

Yeah

... where we say, "Hey, we have this great data creation approach. How do we make sure that this generalizes to a novel data distribution?"

I don't know if this is, like, a good time, but, um, I was gonna ask for, like, a brief history of data sets. It might be too much. I don't know. Uh, I'll just list off because we've done a Data Sets 101 episode.

I think that was like one of our earliest episodes by far-

Awesome

... because, uh, we want people to, to know the data sets. And I think everyone starts with Common Crawl. I think every lab has their own web scrape. Would you, would you say that's true, or do they start from Common Crawl?

At this point, yeah, I think, I think... Uh, like I said, this is where most of the labs, I think-

Yeah

... have actually invested most of their time and effort-

Yeah, yeah

... is in building better versions of Common Crawl for themselves.

Yeah. Uh, I'll just name check some of these. If you have commentary, just, you know, just, just chime in. GitHub, the source of code, maybe Stack Overflow, even though that's cut off these days. I don't know. Do, do, do people get code from anywhere else?

Um, I mean, I think there are obviously places where you buy code-

Yeah

... uh, data, but for public code, I think those are the most common.

Yeah.

Um, I think some interesting things about those that I just personally find surprising-

Okay

... stars are not a good predictor of whether data is useful for models or not.

Not surprised.

Um, like I think that's-- Like, the most popular repos are not necessarily higher quality, at least with respect to do they improve a model's coding capa-capabilities.

You've ablated this.

I haven't done it, but the Star-Coder paper, uh, ha-has done it, and there have been a couple other papers that have all shown that. Something that I just consistently have found to be a little bit surprising. There's a lot of things that are kind of counterintuitive about data curation.

Did they, uh, this shows that I haven't read the paper, but did, did they find anything good that, that was, like, a sign of a good code base?

There wasn't anything that was super predictive.

Oh, man.

That, that... Like, honestly, in some ways, like some of them were length. Like, some of these, like, simple heuristics actually ended up being better, but nothing was super discriminative-

Yeah

... um, there, which is kind of interesting.

Okay, cool. I'm gonna keep going. Uh, arXiv is which is-

Sure

... you know, GitHub for papers. Books, Books1, Books2, and obviously Books3, controversial. I think Anthropic's, uh, getting sued over Books3.

Yeah, I think a bunch of people are getting sued.

Mm-hmm.

Meta is also getting sued over-

Sean Presser

... over Books3.

In some sense, like, ca- uh, can we just, like, look past it? I, I don't know. It's like books are transformative use. Like, uh, I don't know if you have a view on this.

Well, I think the, the law-- the recent, uh, ruling was interesting, although it was an appellate court ruling, so presumably it's gonna go to a higher court, um, afterwards. Um, but what they ruled was that it's fair use so long as you purchase the book.

So, you know, if, if you-- so you can't download Books3 and then use it because that's piracy, and that you've- ... you've stolen the books in the first place. But if you bought a copy of all of those books, then you can train it on, and then it just counts as, as fair use.

Right.

Which I think is an interesting and, uh, pret-pre-- to me, it feels pretty reasonable, um-

Mm-hmm

... line there. One fun thing about Books3 is that it, it, it also has, like, a lot of not-safe-for-work stuff, um, in, in Books3, which is kind of interesting if you actually go and look through it.

There should be a Stripe one-click checkout with, like, Books3. Just buy Books3, and then-

Download them

... get a wa-warehouse, and then get them all-

It's gonna cost a lot. I don't know

... get them all shipped there. I wonder what the cost would be. I'm sure somebody ran the numbers. I'll look it up. I don't know if you comment-- can comment on this at all, but the-- in the Meta lawsuit, I, I remember there was a, like, an email thread with, like, some of the research scientists inside of Meta talking about Books3, and Zuck was like, "Just do it."

Uh, th-this is public, right?

Yeah, that was, that was, I think, public and part of the lawsuits.

Yeah. Any reflections, comments?

All I can say is that when I was at Meta, certainly legal stuff around datasets was very challenging- ... and becoming increasingly challenging, and there are a number of situations where, you know, the only person that could approve things was Zuck, um, because of the, the scale of, of the risk, I think.

But, and, and it definitely made publishing at Meta near the end more, more challenging around just what we could do with, with any dataset because, I mean, realistically, companies like Meta and OpenAI, Anthropic are big targets for these lawsuits.

Yeah. S-so my conspiracy theory for what happened to Llama 4 is the lawyers got to it. The lawyers got to the datasets.

And they had to change what they use.

They couldn't-- yeah, they, they were just, like, hand tied behind their back when other labs were not just 'cause Me-Meta had an active lawsuit.

I think that's possible. I, I, I think probably more of it just has to do with the challenges of just continuing to scale and having that be the goal. Like, this is actually a lot of the reason why I got into data and started Datology was that the scaling laws always were terrible.

What the scaling laws paper showed was that there was a predictable relationship between-

The Kaplan one.

Yeah, the Kaplan one. That there's a predictable relationship between performance and computer data, right? That's really useful. But it was a bad, predictable relationship. Power law scaling is terrible. It means that every time you 10X your data, you get a diminishing marginal return on performance.

You know, this is why you had these prognostications, oh, you know, GPT-N is gonna cost, you know, a trillion dollars to train. It's 'cause you take that scaling curve, and you just naively extrapolate it out, and I think that's what we've seen to some extent with the failure of the, the mega models, right, with 4.5 and Llama 4 and others.

I think that there's a, a challenge of just continuing to do that naively, and you have to figure out how to break it. Um, I think there are a number of theories of, of ways how to break it, and I don't think they're mutually exclusive.

My bet is that, that data quality is a massive way to do this. And in many ways, actually, the paper that was the foundational paper for Datology, it's called Beyond Neural Scaling Laws, and was fortunate to, to get a best paper at NeurIPS.

And what that paper showed was that if you use your data correctly, you can actually bend the scaling laws themselves. And an interesting kind of technical part of this is that, you know, I mentioned what we really care about is this, how much new information do you learn from the next data point?

So technically, that's the marginal information gain per data point. Um, perplexity is another, another ver-variant of it. There's a, there's a duality-

Yeah

... um, between them. It turns out that we were able to prove in perceptrons at least, 'cause that's all, generally what, all you can ever prove- ... right? So in small scale, um, and, and this work was led by Ben Sorcher, who was a really fantastic, uh, grad student I worked with on this paper.

And what he showed was that there's a direct duality between power law scaling and the fact that you also see that the marginal information gain per data point also decays as a power law. And that's why you get power law scaling because, you know, every successive data point is teaching you less and less and less, and it follows a power law, so then you get performance decaying as a power law as well.

So if instead you can keep that so it's flat, then you bend the scaling law, and now all of a sudden you learn dramatically faster because the amount of information you're learning is not decaying with dataset size. Now, that was all in theory what you could accomplish, you know, and we proposed a couple metrics that got us one step there.

But in many ways, I would actually say that the whole point of Datology is how do we realize the potential that was shown in that paper? How do we actually make that a reality? And I think fundamentally, if we want to get scaling to, to work well fundamentally, we need to do a better job here.

Are you measuring the quality of these open datasets over time? Are the most recent open datasets better than the older ones at a good rate or, like, just marginal?

They do get better, but I think they're not relative to the headroom and potential, I would say. Right? Like, like, Nemotron is actually pretty similar in quality to DCLM. It's-- it came out about six months later. It has more unique tokens.

Um, they made a really big deal about it having more unique tokens. But on average, the quality is, is pretty straightforward. So, you know, wh-when we think about what we are able to accomplish at Datology, we usually think about along these three axes I mentioned, train faster, train better, train smaller.

So typically, basically, that's say, like, first question, train faster. Given a certain baseline dataset- How much faster can we achieve the same performance? So, you know, and how many fewer tokens? Um, so we're able to now get to the same performance as DCLM about twelve X faster.

So, you know, in fewer than ten percent of the tokens, we can match what you get from training to convergence. Um-

And when you, when you say performance, you mean like GPQA, or you mean loss or-

Yeah, so we typically take-- we take the accuracy across 15 kind of standard benchmark tasks-

Okay

... that are relevant for, you know, a given model size.

Yeah.

So your MMLUs, your arcs, your races, you know, et cetera.

The q- the problem with those is like, are you training to the test, right?

Yeah.

Like, are you... You know, I'm sure you know this.

And, and that's something that we're super careful about, um, because it's really easy to overfit to these benchmarks, of course, and then end up with models that are really brittle. And I think this is something that we've seen, especially with synthetic data, and synthetic data is a big part of what we do at Datology.

We found that it can drive, like, pretty dramatic gains if you do it correctly. Um, there are lots of ways to do synthetic data incorrectly. We've seen a number of models, right, that are trained on h- on a lot of synthetic data and end up doing really well on benchmarks, but then kind of don't pass vibe checks, uh, and people don't really use.

So we do a lot to try to prevent this. Um, first and foremost, we keep a held out set of tests- sets that we only look at very occasionally, um, and we also don't evaluate on a whole bunch of other evals that we then have, you know, models that end up getting evaled on later to try to really ensure this.

But yeah, this is fundamentally how we measure. We look at a b-- an average of benchmarks, just trying to kind of think what's, what's fair and reasonable-

Yeah

... um, with respect to what we can do. So, you know, that's like the first thing we t-typically look at. Then we look at train better, of course. Under the same compute budget, how much better can you do, uh, with a given dataset?

Um, we're able to beat kind of the best open datasets by anywhere from four to five points, depending on the specific dataset and, um-

Eval

... and eval. Some of the evals actually are much bigger than four to five points. Four to five points on average, um, and those are absolute, absolute points. We generally find that in order to get that same performance from training longer on the baseline datasets, you'd have to train on those baseline datasets, you know, at least five to ten times longer to try to match that performance because every successive point of accuracy of course gets harder and harder, um, to achieve.

Um, and then finally, train smaller. Basically say, "Okay, given-- holding performance constant, what's the smallest parameter count model that we can get, uh, to outperform?" We can already get models that have fewer than half the parameters and also train faster and also outperform the larger models trained on the uncurated or alternatively curated datasets by for-- by a, a large margin.

So, you know, th-this is a, a big roundabout way of, of, of getting to this answer of, you know, have the open datasets, I think, kept up with, with the, this improvement. You know, with a fairly small team, we're now a team of about 30, you know, most of the results that I've discussed, like, were achieved with a team of under 20, um, 'cause we've grown quite a bit in the last couple months.

And with not that much compute, uh, by, uh, kinda common standards. You know, more than academics, but certainly nowhere close to the, the frontier labs. We've been able to achieve, I, I think, pretty dramatic results. I think the reason for this is because there's so much headroom here.

You know, we've already been able to get 10X gains. I think there's at least another 100X behind this that is still to be done. There's so much stuff that we're just not even doing right now that I know makes sense to do, uh, let alone all the things that we are doing that I know we can be doing better, that we're still very suboptimal with respect to how we're doing this.

Like, I know that the way we do our synthetic data right now could be much better, that the way we do our filtering could be much better, the way we do our model-based filtering, our embedding-based filtering, all these different aspects could be much stronger.

So I think there's just so much headroom here. I think the challenge is that, um, there's not a huge incentive to do this in the open dataset community. I mean, the, the labs which have the biggest incentives obviously have strong incentives not to share anything with respect to that.

Um, so you're left to kind of, you know, the Allen Institute, things like DCLM, Hugging Face, et cetera, to make progress there. But I do think that this is su- this is a hard enough problem that it really demands a, a whole company that is really focused on this.

I think you s- what you see in all the frontier labs is that they have data teams. Um, and if you talk to the folks that work on those data teams, what you'll kinda systematically hear is that typically they're, they're under-resourced relative to, to the gains that they're delivering, um, that they're always having to fight for attention.

And this is just like a fundamental thing that I, I saw at Meta, I saw at DeepMind, and, and I've heard at all these other places. It was a big part of why I decided to start Datology instead of doing this within Meta.

You know, I had the opportunity to, to start a data team there, and that was to try to centralize this. But fundamentally, I think that this is such an important problem that it, it's a problem that needs to be the end itself, not just the means to the end, um, which I think is what you see in many of these big groups.

You need to have a large team of really talented people who are really passionate about looking at the data, um, and there aren't that many people who are that passionate about it, to just focus on how do we build the best possible datasets for model training.

I think it's hard to do this as a data team. I, I think there's a real benefit of being a data company, and that's a lot of why I started Datology.

How do you think the almost economics of like the open source datasets world evolve? Because you basically have these like open source datasets that are like good, but maybe they're not quite as good to make production data, um, systems, and then you have companies like yourselves that are sitting on top of it.

Do you think at some point there's gonna be some sort of rupture between like, "Hey, why are you just taking my open source dataset and making it better in private for people without contributing back?" And do you guys have plans to then open source other sets?

I, I think there's like kind of this open question of are these things actually useful in the open, or should you just do it in, in private?

Yeah, it's a great question, and one that we've thought a lot about. I mean, so first off, one thing to note is, right, is that, that in, in-- while we do work with folks who are, who are just training on open models, in general, we really built our product and designed it to be able to work with companies that are training on a combination of open source and proprietary data.

And that proprietary data could just be data they've been collecting as a matter of business for the last decade, or that could be data that they've sourced from a data annotator or, or, you know, another data provider. And some folks we work with have all three, right?

They're gonna use open data, they're gonna use data that they've acquired, and then they're gonna use data that they-- that's part of their business, um, to begin with. So, you know, and, and, and that's like, I f- I think a, a lot of where our, our focus goes, although of course, we are excited about working with lots of folks who are training on more open datasets.

So I published for, you know, a decade, uh, more than that even. Like, you know, this was, was very near and dear to my heart, and it's something that we've thought a lot about at Datology. I think one of the challenges of building a startup today, especially a startup for which science is a critical component, um, which is, as I mentioned, is one of the things that really attracted me, um, to starting Datology, is this tension, right?

Fundamentally, we have to, we have to build a business. In order to do that, we have to have a moat. And you can think about kind of three places, I think, where a moat could come from. You know, one is from science knowhow.

One is from engineering infrastructure and the challenge of, of just implementing, uh, this yourself. Um, and then finally, there's a brand moat that you can eventually reach. We're very far from a brand moat, uh, at this point in our, in our journey.

Eventually, I would love to have a brand moat where whenever anyone thinks data and AI, they think Datology and know that's where I should go first. Um, I hope that we get to that point. But in the meantime, uh, you know, we have to rely on the other two moats, on, on the science know-how and the engineering infrastructure.

I think on the open data side, what we've seen is that the engineering infrastructure definitely can be a moat, but unfortunately, I think that science know-how moat is actually pretty important. And a lot of the evidence that we've seen so far has suggested that that is something that's meaningful.

As an example, you know, uh, many of the customers we talk to, the... one of the first things they'll ask is, "Hey, compare to the best open source data set." Um, right. So if, if, if we were giving away everything we needed to in order to build that best open source data set, some folks would just, would just go there.

So I think that's been where our challenge has been. Now, what we've tried to do, and I think we've done a good job of, and I'm generally happy with the balance we've struck, is try to, in the blog posts that we put out, give a lot of intuition as to kind of w- what we're doing and how it works without necessarily getting to that point of, of, of reproducibility.

You know, that's, I think, much more open than you see most of the, the big labs be.

Yeah.

If you look at, like, the data section of, like, the Gemini tech report, it basically says, like, data quality was the single most important thing for making great models.

One paragraph.

We used algorithms- ... and heuristics. It's like, "Great." You know, so like e- like I think some people were even pointing out, um, you know, like recently the, the... there's been a lot more attention on rephrasing as a method for using synthetic data.

Was it the Apple paper?

Um, the Apple paper, the Kimi paper has mentioned this, uh, a bunch of others. And, you know, some po- some folks recently pointed out that, like, "Hey, in our blog post from November, we were talking a lot about that.

That's something that we do a, a lot of." Pritish Manning, the guy who first came up with rephrasing is, was one of our first employees. So, you know, we've improved on that pretty dramatically and, and, and taking it to new places.

But that's something that, you know, we very... like, I think that there would've been an incentive to just, like, not even talk about that at all.

Sorry, just on that, do you feel like this is, like, a great example of you were talking about it in the data, and then the Kimi paper comes out with a model, and then people are like, "Oh, the rephrasing is important," but you're like, "Hey, I was telling you that before- "...

but I just didn't have a model to show you that it was important." Do you think that's still even in sci- in, in open science, like, a limiter for people that, like, if you don't have a model, people don't care?

Same with DeepSeek. A lot of the things in the paper were, like, kind of known, but then once you have them applied, people care.

I think that's certainly something that happens, and I think, uh, speaks to the same sort of cultural incentives, um, that we, we talked about earlier, where I think that, you know, people tend to think about this very much in, you know, ultimately the, the, it being a means to an end.

And I understand why that is, of course. And ultimately, like, you know, when we sell better data, like, ultimately we're selling a better model, um, at the end of it. We're, you know, more cost-effective model. But I think that the fact that people don't care about it as much, unless if they...

it's really s- you're smacked in the face with it, I think is both a tragedy and an opportunity. And, you know, I would love it if it weren't that case, but, but, but given that it is, you know, that's, I think, the opportunity we see at Datology to, to really make an impact here.

This might be a little bit of a tangent, but you, you mentioned synthetic data, you mentioned rephrasing, so I figured now's a good time to go into it. Um, you know, I, I f- I figured that most of the work on Datology is filtering, but I see synthetic data as something slightly different.

It, it is in a general domain of improved data quality, but it's different than filtering.

Yeah.

Am I right to equate synthetic data with rephrasing, or is there, are there other parts to synthetic data in your mind?

Yes, I think there are different t- parts of synthetic data. There are two parts. But let me first actually just comment on the, on the filtering-

Yeah

... um, versus thing. So I, I used to actually use the word data, um, filtering or data pruning. Um, and a- actually that paper I mentioned that was at NeurIPS, that one actually has data pruning, uh, in the title, and that's how you beat scaling laws through data pruning.

When I started Datology, I, I really changed the language to be data curation-

Yeah

... um, over, uh, data pruning or data filtering, and that's because curation is a lot more than just filtering. Filtering and saying, "Hey, this is a bad data point, we wanna get rid of it," i- is absolutely an important part of what we do, but it's also about rebalancing data sets, upweighting and upsampling certain data distributionally and downsampling others.

Which that might not mean filtering, it might just be changing the, the weighting with, with which you take it. The order in which you present data can be really impactful. Curricula, and we now have seen this with discrete curricula, you know, for multi-phase training and things like that.

That's not filtering. You know, the way you batch the data, uh, can be an important factor. Um, synthetic data can be an important factor. The way you mix sources, all of these sorts of things beyond just filtering. So filtering is a very important part of what we do, and it will always be something that we care a lot about, but it's, it's much more th- than that.

Okay, so now to the question about synthetic data. I think at, at a high level, there are two approaches to synthetic data, and we have focused more on one of them, the rephrasing one, than the other, although I think there is opportunity in the other one.

So the first approach is create new data where the knowledge that's in that data is largely coming from the model that's generating that synthetic data.

Oh, that's distillation then.

It's a version of distillation, and I think that this version of synthetic data could be construed as distillation in disguise.

Yeah.

Um, and I, I think it is a very clear version of this. And when you think about, like, the criticisms of synthetic data around model collapse and stuff like that, I think they largely apply to this version of you have a net new data creation that's coming out of these models.

So that's like top one.

I'll slip, I'll slip one in there. There's also model steganography, where you can sort of hide preferences in a model and distill it down.

Absolutely, and now we've seen, like, the recent, like, owl stuff around that.

If people search anthropic owls, you'll see it.

Yeah, exactly. The other way-

Yeah

... is this rephrasing, rewriting-

Okay

... um, approach. Um, so this is the information that's in the data that's actually coming from the data that you're conditioning the rephrasing on in the first place.

Yeah.

And all the model's doing is it's reformatting the data-

Cleaning

... or presenting it in a new, a new way that maybe is easier for a model to learn.

Yeah, cleaning, right?

Um, it's cleaning it in some way.

Okay.

It could be cleaning it, it could be making it, you know, the information more accessible. It could be putting that information in a format that is more representative of what the model's gonna be faced with downstream. So I do think that, like, one of the things that definitely happens with synthetic data is we are bringing more post-training like data, um, into pre-training.

Yeah, sounds like SFT.

Um, and in, in general, like, one of my beliefs is that most of what we do in post-training is better done in pre- and mid-training and earlier on in training in general. Um-

It's just the scale, you know, you don't have that scale until now.

It's just that, yeah, ex- exactly. I think, I think, I think if you assume this paradigm where, you know, pre-training is incredibly expensive and something that you can only do very, very rarely, and then post-training is cheap, then it makes sense.

But as soon as you break that assumption, and I think DeepSeek showed and, uh, that already you can get a frontier model for a marginal cost of a couple million dollars.

Yeah.

Um, that's gone down since then because we've gotten better at it and compute has come down in price, um, since then. Like, I believe that getting to a frontier model should cost a million dollars or less for most organizations, at, at least, at least in a specialized domain, right?

And when you think about what enterprises need, that's generally what they need. They don't need a model that can do everything. They need a model that can do a constrained set of tasks to very high accuracy for as low an inference cost as possible.

And I think that, uh, that will be, uh, you know- Under a million dollars very, very s- soon, and th- and that changes a lot of these dynamics. But going back to the synthetic data, uh, question of these two different types.

So I think there's one towards this net new creation. I think that's where you have a lot of risk, that's where you get the model collapse concerns, where, you know, I train a model, uh, I train a generative model on a given data distribution, it overfits the modes and it underfits the tails.

So then if I have it generate a bunch of data, it's gonna be more mode and, and less tail, and then I do that a bunch of times, and eventually I get a spike. I get a delta function.

Yeah, only mode.

Um, only mode, exactly. Um, like that makes sense, um, why that happens. Um, I will note that if you filter the data after each point, that's now information injection, and that can break all of this, um, and I think can prevent model collapse.

Which a little bit is what RL is.

Which is a little bit what RL is. I think you can absolutely view it that way. And I think actually a lot of the work that has suggested that, um, you know, RL is really just eliciting the, the capabilities of pre-trained models, like random rewards or a single example, and then it's just changing the distrib- it's like aligning to the distribution the model has in the first place-

Yeah

... are, I think, very in line with that way of thinking about it.

You're, you're distilling from a perfect model, which is the environment-

Mm-hmm

... or the, the verifier, whatever, and then you're distilling that into the, the thing, so-

Yep

... like it's, it's amazing. It's beautiful.

Um, but, but the cool thing about rewriting is that because the model that's doing the rephrasing just needs to know how to rephrase, it doesn't need to know anything about the content itself, it doesn't need to understand it, it means you can use a pretty weak model to do the rephrasing and have it generalize and generate data that can teach a model that's much better than the model that's doing that rephrasing.

So I think with this distillation in disguise, I, I'm generally quite skeptical-

Yeah

... that you can get a model that will be better than the teacher that's generating the synthetic data when you do this sort of net new data creation. Um, it's possible you could, through some sort of heavy rejection sampling on the big model, because you're effectively inserting new information when you say which of the synthetic outputs is good or bad, right?

There's some new supervision coming in there. Um, but I'm generally skeptical of that. Whereas we've seen this, we actually have a blog post coming out in the next week or two about kind of our synthetic data generation, which we call Beyond Web.

Um-

Wow.

And we'll have some cool scientific experiments in there too, to our point of trying to figure out this balance where we can share some of the science, but also do so in a way that, you know, i- is, is sustainable for our business.

And one of the things we show there actually is that by doing this, you can actually go do-- get a model to do much, much better than if you had trained on all of the data, all raw tokens in the first place.

So that by doing this rephrasing effectively, you actually can break this data wall and now get models that are better than either of the models that generated the data. With rephrasing, I think this is super possible because most of the information is coming from the data.

It's not coming from the model itself.

Couple follow-ups on that, just things I've always wondered. Uh, are textbooks all you need?

No. They are not all you need. I think textbooks are great, and I think there's a lot of really, uh, great content and, and high-quality data points like that. Um, but obviously textbooks are also a very narrow data distribution.

And the sing-- if I can o-- if there's only one thing that you should take away from this entire interview- ... about what is good for, for data quality, it's diversity.

Mm-hmm.

Um, like in many ways, right, there, there was this w-- like I, I used to do all this work on out-of-distribution generalization, and we had all of these like, you know, very careful studies where we would say, "Okay, let's, you know, make this corner of the data distribution, then we leave this held out where it's never seen this combination of things, and let's see if it can generalize."

And then like, you know, LLMs and, and the modern way of training models came along and said, "Hey, what if nothing was out of distribution?" "What if we just made it so that we train on-

All the things

... everything, and everything's now in distribution?"

And by the way, you know, that, that is in line with AGI, right? So you might as well.

And that's basically what we've done, and it's worked. Um, it's worked shockingly well, like, like way beyond anyone I, I think, or most people would have expected.

Yeah.

Um, I certainly was shocked by it. I made a strong bet that there was no way you could get compositionality just from scaling.

Yeah.

And well, you can.

Yeah.

Turns out, um, it does work when you get big enough.

Uh, what I was really referencing was, uh, th- this is the Microsoft Fee papers, right?

Yeah.

One, two, three, four. Uh, a lot of them do the rephrasing or rewrite- rewriting in textbook format. And I feel like that's a little bit of cargo culting of like, "Oh, just 'cause you like, write like Wikipedia or write like textbooks, the models learn better."

That's not prove- I don't know, that's not automatically proven to be the case so...

I think that's also partly part of the reason why you see a big difference between the benchmark scores of those models and their real-world use.

Oh.

They went to too narrow a distribution, and I think this is the problem with synthetic data fundamentally, is that you're always gonna have some bias here. I think you can do a lot to make it more diverse-

Yeah

... and we have put a lot of effort into finding ways to do that. Um, for example, we rephrase into many, many-

Mm

... many different styles and formats. Um, that's really important to, to get stuff that's good. But I think this is the risk, right? That you, you go on way too narrow a distribution, and models all are, are always going to be fairly peaky with their output distribution, and then that actually results in reducing diversity.

That said, I will say that there is a takeaway of that are textbooks all you need that I think is correct, which is repeating higher quality tokens is almost always better than seeing net new lower quality tokens. So, like epoching over higher quality data, almost always better than getting the same amount of new data of an unknown quality or of average quality.

Average in this case being like what you just get from an internet dump or something like that, or even a reasonably filtered internet dump.

Yeah.

It's always better.

The modification I made or n- the study I would want to commission out of that is like instead of having another epoch on, on high-quality data, if you found high-quality data, good, go and paraphrase it, and then, and then train on that.

Maybe that'll get additional gains. I, I don't think I've seen any papers that have been to that effect.

Um, the Kimi paper actually had an experiment to that effect, where they tried adding, um, multiple epochs, and they looked at how many rephrasings they did-

Yeah

... uh, of each of them and had some results there that were interesting-

Okay

... to that, to that effect.

Amazing. And then the other question was more on curriculum. Uh, curriculum learning had a bad rep for a while. How come it's back? What's changed?

Yeah, so a bunch of things, and this was really interesting 'cause w- when I was going out and, um, you know, initially deciding whether to start Datology and raising and, like, talking to various, you know, initial recruits and stuff, it was, like, mid '23.

Um, and at the time I was saying, you know, curricula are gonna be a really- ... important aspect. Um, and a lot of people were basically just like, "No, curricula don't work." Like, "We tried this a bunch of times, and curricula don't work."

Curricula are one of these ideas that I think always, like, had to work in the sense that it just made too much sense. There are a number of these things where it's like it might be hard to figure out how to make it work well, but, like, it always had to work.

There was actually a really cool paper from Stanford that had a nice way of conceptualizing this, which is, uh, imagine a graph where each of the nodes are a different concept, um, or, you know, idea that you want the model to understand, and then the edges are basically the dependency between those concepts, right?

So if, if concept A helps you learn concept B, there would be an, an edge from concept A to concept B, right? So now this is the graph. Imagine this graph of, you know, all concepts in the world and all the different edges between them, right?

Huge graph. If that graph is empty, then it would mean that nothing is helpful for learning anything else. Um, right? And then curricula would not make any sense. You should just randomly order things. If that graph was complete, so that the edges b- there was an, an edge of e- equivalent weight between every pair of nodes, then similarly it would mean that everything is equally useful for learning everything else, and curricula don't work, and you shouldn't use them.

Any other graph- ... besides those two graphs, curricula makes sense. I think it's o- pretty obvious that neither of those is a graph of the actual world that we live in. Clearly the world does have dependencies, some very, very obvious, like the fact that, you know, it'd be hard for me to do division and multiplication if I don't understand addition and subtraction, and, you know, some much more vague.

But I have always believed, like, that this has to work, and, and the challenge has largely been that if you're fully saturating your data, then there's really no advantage of curricula. Unless if you, unless if you wouldn't be able to learn it otherwise, generally I think the idea behind curricula is that it makes you much more efficient.

But in the supervised learning world, we were fully saturating these data sets, so, you know, maybe a curricula would get you there faster, but that wasn't the bottleneck or the limiting factor. So there wasn't a clear incentive to actually go and do these hard experiments to try to figure out how to make a good curriculum because, like, who cares if I can get you to ImageNet performance in 80 epochs instead of 160 epochs?

Like, that's nice, but, like, it's, it's not a big deal in the first place. But now we're in this totally different world where now all of a sudden all of our models are underfitting the data. This is super important, and getting a curriculum right could literally make the difference between, you know, spending 10 times as much on a model training, you know, hundreds of millions of dollars potentially.

And now all of a sudden curricula make a ton of sense. So I think that's why the problem didn't really make sense to really put a lot of effort into, um, previously. Um, and you know, now we've seen pretty clearly with discrete curricula that this makes a big impact.

And, like, largely what we talk about when we say mid-training is really just, like, a later phase of, of your discrete curriculum, um, I think is another way of thinking about it, right? You could even think of post-training as part of a curriculum.

In fact, one of the things that I'm really excited about is, you know, we've mostly focused on pre and mid-training at Datology so far. One of the kind of most consistent asks from every one of our customers has been, "Can you do more on post-training?

Um, can you also help us curate the post-training data?" So we're starting to invest pretty heavily there. Um, and one of the things I'm really excited about is actually viewing this whole thing from pre-training to mid-training to post-training holistically as a single process.

Um, and then asking questions like, "How do we optimize our pre-training data to make post-training more effective?" or things like that. You know, these are, I think, really exciting questions, and something that you don't see happen even at the big labs because they have entirely separate teams, right?

There's a pre-training team. There's a mid-training team. There's a post-training team. And, like, the, uh, mid-training team is a customer of the pre-training team, and the post-training team is, like, a customer of the mid-training and pre-training team. But it's quite hard to actually have signals propagate through all these.

Um, so, so I think this is a really exciting o- area.

I'll push you a bit on this.

Yeah.

Um, you know, I think a, a p- popular view is post-training is elicitation of-

Mm-hmm

... uh, capabilities that you already trained in pre- in pre-training. So what dependencies can you have that feed back into the, into pre-training?

So, so I'm inclined to, to agree with that view, and, and I think that that view would lead very strongly to the fact that you should be trying to optimize your pre-training data to make post-training processes more effective.

Wow.

So you should try to figure out, how do I optimize my pre-training data so that the slope of the test time compute curve or so that the slope of the RL curve is as steep as you possibly can be?

Um, or alternatively, how do I optimize my pre-training data so that the slope of the jailbreaking curve is as shallow as possible, right? Like, fundamentally, I think alignment in post-training doesn't really make sense as a, as a long-term solution.

If you can easily align a model through post-training, you can easily misalign a model through post-training. If it's easy to put it in, it's easy to take it out. If it's really hard to put it in, it's really hard to take it out.

That's just, like, a truism of models, right?

Mm-hmm.

So if you do alignment during pre-training, you'll actually end up with models that are, I think, largely impossible to misalign without putting a massive amount of data into them. Um, I think there are a lot of benefits to that.

Um, and I think we've also seen evidence for this, like looking at the difference between Llama and Qwen with respect to their ability to be post-trained, right? It's much easier to RL Qwen-

Qwen, yeah

... than it is to do Llama. Likely that has to do with the fact that Qwen put a lot of synthetic reasoning traces, uh, into their training data.

Even with wrong examples.

Yeah, but even with wrong examples- ... that's where, that's, it still is showing there, um, which is wild, right? Um, but I think that pretty clearly shows that it's the base model that's doing it. It's not the rewards you're giving.

If, if you give random rewards and the model s- still learns, it's probably not the reward signal- ... that's doing it.

That's cool.

I'm just curious on the customer, uh, usage. How many people are doing post-training? Obviously nobody today because you don't have it. But when people come to you, are people looking mostly to do post-training on open models, on OpenAI models, or

What do they ask for?

Yeah. So we usually work with folks who are either, um, training their own models from scratch, um, or doing continued pre-training on, uh, an open model with a bunch of domain-specific data that they have that's unique to, you know, their use cases and, and their business.

We typically focus on folks that are, uh, you know, tr- doing training with significant cost. So typically that means, you know, at, at least a couple tens of billions of tokens, oftentimes more. So kind of the standard small-scale post-training, fine-tuning case, um, we, we don't focus as much on.

That said, I think this has been a question that a lot of people have asked us consistently. Like, "Hey, who's actually training their own models? Like, why don't I just rely on this, rely on the open models?" And I think there are a number of reasons why we see people, uh, do this.

So first off, I think sovereign AI has been a, a pretty big place where we've seen a lot of demand.

It's a big focus for us.

Lots of countries, they wanna have models that they own, that are unique to, to, to their language, their culture, and, you know, that requires them to have really good data curation-

Yeah

... of course, in order to do this effectively.

Just to double-click, countries owning models isn't actually a thing that I know about. Like, the-- You know, I f- I'm from Singapore, we have the SEAL model, but it's not, like, owned by a country, and I can't name any other country that owns a model.

Yeah, I think that's actually correct. Like-

Okay

... it's, it's largely-- What you see right now is these public-private partnerships-

Yeah

... where governments are making pretty large grants.

TII UAE is, like, the closest.

Yeah, I think you have those. I think you also have these places, right, where the funding is, is the country, and it becomes-

I see

... a little unclear where it comes from.

Okay.

But yeah, I think usually what you see is that countries are doing big grants, um, to private companies, um, or public-private partnerships-

Like Mistral

... to go and build-

Yeah

... you know, that sort of, sort of thing. Um, so that's a big thing. I think we, I think we've seen a lot of, um, you know, uh, larger enterprises, um, that have a lot of their own data that wanna do this.

And when you think about this, ultimately what we see is that, okay, across those three value props, train faster, train better, train smaller, like, which matters and when? Um, like, train faster, in principle, that's the easiest one to compute.

You know, I say, "Okay, this model would have cost you ten million dollars to train. I get it to you for a million dollars or for eight hundred thousand or whatever," right? Great, I saved you a ton of money.

In practice, though, nobody wants to train a ten million dollar model for a million dollars. They wanna train-

But they already have the model.

They already have that. They wanna train a hundred million dollar model for ten million dollars. Um, you know, they, they want, they want train better. So train faster, um, usually doesn't matter so much from the perspective of, "Hey, this model's now a lot cheaper."

It does matter a lot more from the perspective of you can iterate much faster, right? 'Cause when you think of the, the workflow of most ML engineers, you t- you start a training, you go, and you, you sit on your hands until the training finishes.

You know, you find something else to do, but largely you're waiting, and your iteration is bounded by how long that takes. Um, if you can take something from taking ten days for a model to finish training to being overnight, now your existing team is way more productive and can do far more iterations and stuff like that.

So that's where we usually see that matter the most. Most people care the most about train better, right? I can get a better model for the same compute, and, and we can absolutely deliver that through data. Data is effectively a compute multiplier, right?

Because all models are underfitting their datasets, if you can make your model more data efficient, you effectively make your compute more valuable. Um, because if you think about compute as I inject a certain number of dollars, and I get a certain performance back, if I use better data, then I will get more performance back per dollar invested, and now my compute is more valuable.

So that's where train better, I think, it tends to be the, the most meaningful thing. Um, but interestingly, for the most companies that are most advanced on their AI transformation journey, train smaller is the one that I think actually means the most.

Um, because when you think about the total cost of ownership of these models, it's gonna be very, very heavily weighted towards inference. It's all inference. And, you know, you think about a company that's spending, say, fifty mil a year on inference, which in the scheme of things is, is not very much, right?

If you deploy a model that's twice as big as it needs to be, that's gonna cost you twenty-five mil in year one. The cost to train a model that has fewer than half the parameters but is just as good or even better at your particular use cases is, say, two or three million dollars.

That's a no-brainer if you can do it easily, right? If it's really hard, then you're never gonna do that. But if you can do it easily, and you can get it right on the first try, that's a no-brainer.

Um, and then as, uh, you know, and then fif- fifty mil a year is, like, not gonna be very much, right? We know that all of these products have, you know, a tiny, tiny fraction of what their eventual user bases will be, right?

We're still very much in the first inning here. You know, e- everyone that listens to this podcast is using AI nonstop, but the rest of the world is not yet. So the inference costs are gonna skyrocket with these models, and if you use a general purpose model that then you constrain to say, "Hey, this model knows about everything, but now only do this one thing," that model's gonna have a ton of parameters that do not need to be there that are gonna massively increase the cost of serving that model.

So I think that, you know, when you think about the use case of an enterprise where they need a model that's an inch wide and a mile deep, it can do a, a small handful of things, but it can do that really, really effectively to five nines of reliability, and it can do it for as low a cost as possible.

The economics make it so that it, it, it really makes a lot of sense, um, to do this yourself if you can do it easily. Um, and the way we think about it is that there are kinda two big barriers.

First, you have to get training right, and then you gotta get data right. And on the training side, um, I think three years ago, this was super hard, right? But Mosaic was the first one to really recognize that there was a huge opportunity in making this easy, and now this has largely been commoditized, um, by things like SageMaker and Together and, and lots of different folks that help you on the training side.

But on the data side, the barrier is just as high as ever. Um, and in many ways, that's our mission at Datology, is how do we bring that barrier down so that anyone who wants to train a model can do so with the best quality data on their first try.

They don't have to go and spend forty years in the desert. They don't have to get it wrong a hundred times first, which is what will happen if you don't have this experience. But instead, on the first shot, they get a really great model.

Yeah. Uh, just a follow-up question on train smaller. Uh, yeah, I, I, I fully agree, and I think that this is, uh, something a lot of people are investing in. Um, you are f- primarily doing work on the, on the data side, data pruning, which maybe is a bad word now, data curation, whatever.

I think a lot of people, uh, you know, w- Jon Den Frankel was on the podcast very early on, but a lot of people were betting on pruning the, the, the model itself.

Mm-hmm.

Like, you have a working model at size, and you just lop off anything above, like, a certain epsilon. Is that confirmed to just be dead?

So, so it's funny, Jonathan actually interned with me when I was at Meta-

Yeah

... um, and we worked on this stuff together. You know, he had the lottery ticket hypothesis, which is a really beautiful paper.

Which he now completely disowns.

Which he largely disowns. Um, you know, I, I had, I, I, I had this whole idea when Jonathan and I worked together that we wanted to create a lottery ticket initialization. It would just be, uh, an initialization you'd sample from, um, for initializing the weights that would then be one of these, like, perfect winning ticket initializations.

The last initialization.

Um, but we actually found out that the problem was that the lottery ticket was actually data dependent, and that was where the fundamental problem came, that as soon as you change the data distribution a little bit, like, the winning tickets ch- changed in a, in a really big way.

Um, I don't think pruning is dead. Uh, parameter pruning still absolutely has a, has a place, but I think certainly we found it challenging to really realize the potential of it. I think one of the big trick, uh, tricks with s- with pruning, parameter pruning, just to be clear, was that unstructured pruning, when you would, you know, prune weights randomly.

So you view all the weights as, as a smorgasbord, um, and just prune them randomly. That worked really well, and you could remove massive quantities of the weights with unstructured pruning. The problem is that unstructured pruning doesn't really give you a clear compute advantage because you need to have a sparse matrix now to, to reflect this, and there's a pretty huge overhead of sparse matrix multiplies.

Um, GPUs are not very good at sparse matrix multiplies. Like, there's some support for them now, but they-

There's some hardware optimizations for that, yeah.

And, and there's some hardware, and people tr- talked about, like, building ASICs that would be really good at unstructured pruning, but I don't think I've seen one that works super well. I think if someone did make something that worked really well for, for kind of m- models that were pruned in an unstructured way, that could be effective.

Um, structured pruning, in which case you just, like, remove a unit, you just remove a neuron, that, uh, is really easy to make as a, a, a-

Faster, yeah

... as a faster in a GPU, but that just doesn't work, uh, nearly as, as well. So, you know, I, I think there's still potential here. I don't think it's the panacea that, that I and I think many others had hoped.

That said, I think one thing that's cool about using better data to train smaller models is that it's complementary with any other approaches for optimizing inference. So, you know, I think pruning and quantization obviously still have a lot of, uh, role to play in, in helping inference go faster, and that would stack o- on top of anything that we're doing, um, which I think is kind of cool.

Yeah. One also I think, uh, kind of a grand challenge, golden question that would be very valuable for you, and just in general, is, is this idea of, like, what is the smallest possible model for a given capability?

Do you have any insights on that? I, I did a podcast with, um, Jack Morris, who's out of, uh, Cornell, and, uh, you know, I think, like, there's, there's, like, some information limit, and he-- I think he, he had some answer like, you know, it's, it's, like, eight bits per parameter or something like that.

I, I forget what the, the conclusion was.

Yeah, I'm not sure what... I, what, if I, if I'd put out a specific number- ... but I, but I would definitely say far, far smaller than what our current models are trained to be.

Right.

Like, we are nowhere close to, to this. And, and, and, and, like, I am generally of the belief that most of the models that the vast majority of people will be using in, say, three years, will be single-digit B or smaller.

I think we've seen this very clearly. Like, you look at just, like, the Llama series, you know, and if, if you want to exclude Llama Four, do so. Um, but you know, Llama One through Three, you can see pretty clearly that, you know, the seven B variant from one gener- uh, from the n- from, you know, N plus One generation is, like, pretty close to the 70B gen- v- variant from the prior generation, you know?

If it's not quite there, but there's still a very clear trend here. We're seeing this with the Qwen models, right? You look at some of these small Qwen models, and they're just incredibly performant relative to what state-of-the-art was, you know, a year ago.

I, I think it's pretty clear that these models are way too big. I personally would bet against kind of the, the next frontier being trillion-parameter models, and rather that, you know, we're gonna really optimize the inference cost of it.

I think also test-time compute as a paradigm really pushes you towards smaller models, right? Because if your cost of, of solving a problem is cost of inference times number of thinking steps, and you have to do a lot of thinking steps- ...

well, well, now this is, like, a really... like, minimizing the cost of inference is, is really important. And I think that, you know, anything we can do to make it so that you can just make that inference model that is doing the f- the one step of thinking a lot faster enables test-time compute to be a lot more effective.

Yeah. I think there's, there's another version of this, which is the sort of Andrej Karpathy cognitive core concept of a model that doesn't know anything but can use tools a lot-

Yeah

... to, to fi- to find, figure out. Again, another information theoretical limit that would be very helpful to figure out is what is the minimal viable model for that stuff? Like, uh, zero on GPQA, hundred on BrowseConf.

I, I, I really like that idea, and I think it's very possible to do that. 'Cause, like, knowledge s-storing takes a lot of capacity.

Yeah.

It takes a lot of parameters.

You don't need it.

And you don't need it. And you know, we can just look... like, there are, there-- like, actually, one of my first papers, um, that I ever wrote was actually about showing that when you train models on randomized labels, 'cause this was something that was kind of a common test to do.

You show-- that was a-- the one way you could prove that a model was memorizing would be that you randomize all the labels, and now there's no actual-

Recall

... true association.

Yeah.

It would have to memorize it. And, like, models could do this really well. There was, like, an ICLR best paper from twenty seventeen that showed this, that people were really surprised that, that models could memorize all of ImageNet.

Now this seems crazy, because of course, models can memorize the whole internet. Um, but at the time, that was, like, crazy. Wait, they could just memorize a million labels? Like, that's wild. And what we found there, actually, was that if you went and you just deleted units with a model that memorized, it would be really damaging to the model that memorized.

But a model that actually learned a generalizing solution, you could delete a lot of units, um, and it would be pretty robust, um, to that. So it's actually a very clear demonstration of exactly this concept, that the more you memorize, the more capacity you're using.

Drop out regula- regularization.

There's a lot of dualities to drop out.

Yeah.

And, and, and, like, I think there's an argument to be made that drop out, like, pr- you know, helps to prevent memorization-

Yeah

... and it helps to learn more generalizable solutions, and that's part of why it worked well. But yeah, I think it's very possible to do this. And, like, I think we're wasting a ton of capacity in these models on knowledge that is just totally unnecessary, uh, for, for them to have.

Before we wrap, just because we started with the RC models and then we never talked about them. Um, the, uh, I think the most interesting thing to me was they started with 23 trillion tokens of data, and then you helped them get down to 6.6 trillion.

Any learnings from that? And this is a 4.5B model, which is par with Gemma 4B and a little worse than Qwen 3, but roughly the same. Any learnings there, experiences, things that auto model should adopt?

So yeah, so we started for that one, we started with a combination of DCLM, Nemotron, and FineWeb. We basically just concatenate them all together. It's about 25 trillion tokens to, to combine for all those to produce seven trillion out of that.

You know, I mean, I think what, what was exciting to us about that was, in general, you know, seeing the speed at which the model learned. So, you know, it was beating Gemma pretty consistently before the one trillion mark, which was, which was pretty cool, uh, to see, and I think really highlighted in many ways, you know, how higher quality data can get you much better performance much more quickly.

General insights, I think, uh, or, or takeaways from that, I mean, I think it was exciting for us as kind of one of our, our first real, like RC is the first customer that we're talking about and, and being public about, you know, since starting the company, so obviously, that was an exciting moment.

But I think really generally, it, it, it's a good showcase about the fact that combining all of these different techniques can give you a really big gain. You know, I think that's one of the things we've been saying, but it's nice to have a real demonstration about that.

You know, this is not something where it was synthetic data taking us here, or it was filtering taking us here. It was really about thinking about how do we actually combine all of these techniques? And one of the things we've consistently found actually is that when you take these different techniques and you try to make them work together, they don't, generally.

Um, you can make them work together, but it's quite hard to do so. So I think what was quite exciting for us there was showing that that's possible. And then combined with that, I think people, um, first off tend to think that you can't stack curation.

I think the fact that we, we started with some of the best curated open data sets and were able to make them dramatically better is a, a pretty good insight to the fact that there's still a ton of headroom left here.

Um, like we didn't need to go to Common Crawl to get those tokens. We are of course, doing work on that. Uh, and we think there's a lot we can do to improve there. But just starting from that, and, and we actually now are making bigger data sets from that.

I think we can get up to 15 trillion just starting from that corpus and still have pretty identical quality to that, which is pretty neat. So I think showing that you can get there. And then the other-- and, and then it really stacks.

Like one of the other things we consistently find is that if we apply our curation on top of, um, say DCLM, and then we apply it on top of FineWeb, the gap between FineWeb and DCLM is maintained, um, in the gap between kind of Datology curated DCLM and Datology curated FineWeb.

Um, they both get a lot better, but Datology D-DCLM is still better than Datology FineWeb. So, you know, there really is a lot that we can do here, and I think that would be the biggest thing that I would just say.

There's so much still left to do here. We're just scratching the surface. Um, we're pretty excited about what these results showed. Um, we already have better data sets than what, uh, RC trained on, uh, 'cause that model was largely trained in, in, in May, and pretty excited about all the next, uh, trainings that, that we'll have that go even bigger.

I have a couple more lightning fun questions.

Sure.

What data does everyone want based on your customer conversation? What data does everyone want but is really hard to get?

I mean, I think expert data is the, the pretty obvious thing that, that, that-

Just domain experts

... domain expertise. That said, I would also note that like most people don't know what data they actually should be getting. Um-

They just show up with whatever they have and-

Yeah, and I think something we've actually found shockingly frequently is we talk to folks who, you know, have been planning for a really expensive training run, you know, millions and millions of dollars training run. They've been thinking about the architecture they're gonna use.

They've been thinking about all this stuff. And then, you know, they reach out to us and they're like, "Hey, like, we realize we need a good data set, and we're planning to kick off training in two weeks. Like, can you help us?"

Um, and a lot of it's like, "Hey, you probably should be thinking about your data set, um, before all the other things." Um, if, if anything, that's actually the, the most important thing. So I think, but I would say the most surprising thing is maybe how often people don't even have a conception of what good data is.

And oftentimes, I think when people-- what they think is good data often isn't, um, which, you know, goes to the DCLM point, I, I think that we mentioned in the past. It's, it's, it's very counterintuitive and really hard for humans to identify this is high quality, this is low quality.

Uh, this is a little bit of a recruiting question. Uh, what data efficiency question, if somebody had an answer, they should join Datology immediately?

The first thing I would just say is like, if you are one of these people that keeps on finding yourself just like staring at the data, you keep on going into the data set, if you can tell me what your, you know, favorite and, and least favorite C4 example is- ...

like, you, you belong at Datology. You could c-- you should come join us, um, and, and join a bunch of other nerds that, that love doing that exact same thing. I think in many ways, that's kind of the, the single biggest predictor of whether someone, um, is gonna be really happy at Datology, is like how much do you just look at the data, um, in, in your own work?

'Cause like I think you'd be surprised by how many really talented researchers don't do it very often, that they really just view it as a given. I, I think it's been p-pretty surprising, um, across the board. Um, that said, there are so many questions that I am-- uh, on, on the science side that I'm, I'm just super excited about.

Uh, you know, I mentioned the interactions between pre- and post-training. That's definitely one, um, that we're really excited about. One of the things that we really care a lot about is making it so that our product and curation automatically adapts to novel data distributions, right?

If you have this where it has to be fully automated, and, you know, we didn't talk about this too much, but one of our challenges often is if we're working with an enterprise that has a lot of proprietary data, they obviously don't wanna give that to us.

Um, so we bring our curation to their data, but this means that it has to adapt automatically. You know, we have pretty limited access into going and looking at that data. So that's actually a really hairy and interesting out-of-distribution generalization problem.

Um, but it's also really important because there's no golden curation. A curation is only optimal with respect to a given set of downstream use cases or tasks, right? So we need to be able to define based off of, you know, if the model needs to be able to do XYZ, how should we use that information to adjust the curation that we do to make sure that we're giving the data that's most relevant for solving tasks S- X- XYZ?

And that needs to happen automatically. So we have a number of ways that we can do that for a number of our techniques, but that's a very broad and general question that we want to apply to every part of our pipeline so that the way we do synthetic data differs based off of the downstream use cases.

The way we're doing this, the way we're doing every different part, filtering, et cetera, is gonna change based off of that. So that's another question that we're just, uh, really excited about. And fundamentally, you know, anything about really trying to answer this question about, you know, how do you value data with respect to a target?

You know, when, when I think of, of Datology and our core competency, I think every company needs to have an unfair advantage, um, or some core competency that they do better than anyone else, you know? And, and for us at Datology, you know, I want us to be, and, and I think we already are, the best in the world at valuing data with respect to a downstream use case.

In many ways, I think that's kind of the NP-complete problem of AI. Um, if you can do that, you can kind of do anything, and that's the thing that we're really focused on. And of course, curation is, like, the very obvious direct application of that core competency.

But when we think about kind of the vision for the company in the long term, it's about saying and taking what are all, all the other ways we can operationalize that same core skill set. And I think there are tons of really interesting ways, things you can do there.

But that's the fundamental question that we really wanna answer, and then there are tons of different, you know, entry points to that question. But if, if that's a question that excites you, if you have, you know, been working on data somewhere else and you have felt this pain- ...

of, of being, you know, uh, a second-class citizen or having the data team be kind of dismissed and you wanna be in a place where literally the only poi- reason that the company exists is because data is all we care about, I mean, the, the name of the company, Datology, the science of data, that's why we're here, um, then you should absolutely talk to us.

Awesome. Um, and just to wrap on some gossip, uh, let's talk about Meta- ... and super intelligence. So and just in the notes, you know, what-- when you talk about science moat and whatnot, you raised a lot of money from very prominent people.

So you have, you know, uh, Yann LeCun is one of your investors, Jeffrey Hinton, Jeff Dean. Um, so, uh, when Ari says-

Kudos.

When Ari says that they have a science moat, believe him. Um, so maybe since you have Yann as an investor, this is more of a touchy question, but what, what do you make of the whole- ... Meta super intelligence team and, you know, Yann was also on LinkedIn, and he was like, "Hey, you know, I'm actually working on, you know, at FAIR, we're focused on the next generation of AI, not on this current generation, so my role is the same."

But then maybe people might say, you know, "Then why didn't you do the current generation 10 years ago?" What do you make of the whole, of the whole change and whether or not you think this is an interesting direction for Meta, especially given the large platform and user base that they have?

Well, first, with respect to Yann specifically, I mean, Yann's an incredibly talented scientist, of course. But I think that, you know, his preference has always been to do science rather than to run an organization. So he, he-- I think he ran FAIR, um, like organizationally for a year or two right at the very beginning.

But pretty quickly, uh, he handed that off to other people, and, like, when I was there, it was Joelle Pineau and, and Antoine Bordes, uh, b- and then Joelle for most of it that, that really were running FAIR.

And, and she was, uh, an incredible leader. I, I, I, I really respect her deeply and, and couldn't have asked for a better kind of advocate for science, uh, within FAIR. Um, but I think-

When, when she left, uh, people were saying, like, "This is the end of FAIR."

I hope that's not true. Um, but, but I also had that concern. But I think Yann always really wanted to just actually do the science himself and, and, you know, he's generally for much of-- most of the time I was at FAIR, he kind of operated with his own group of a couple kind of postdocs and visiting scientists, and then he'd have a couple students through NYU, and he would kind of do his own research there.

So I don't think he was ever, um, you know, or at least not since the beginning, in a role where he was defining AI strategy for Meta. I don't think that's the role he wanted, um, at any point.

You know, I think he really wanted to be doing that research, and I think-- so I don't think that his role probably is changing very significantly in the sense that he wasn't, uh, doing that previously, and I don't think it was what he wanted, um, to do.

I mean, I think one thing that's pretty cool about it, obviously, is it showcases the importance of data, um, that Meta is willing to spend, uh, quite this much on, uh, you know, scale, uh- ... kind of, uh, not k- like acquisition, not acquisition, uh, that we're seeing today.

Alex Wang is not going to underrate data, let's put it that way.

Yes, he's definitely not going to underrate, uh, the importance of data. You know, and I do think that this is an area where, you know, the stuff we've done, uh, i- is quite different than, than I think what, what we've seen from, from the data annotators, which have been more focused on collecting the data versus actually optimizing and curing it, uh, curating it.

I think there's quite a bit you can do on top of those, um, things. So I think it definitely draws some attention to that. I will also just say generally when Zuck makes a very big bet, it's not proven wise, uh, to bet against him.

Just historically, that's been the case. And like most, most of the, the big bets I think have panned out. I think the one that's still really up in the air is the metaverse. But I would actually argue that I think that's gonna end up paying off in the long run.

I think the Ray-Ban glasses, pretty darn cool, and a lot of the foundations of, of what was in Reality Labs will go into those. Also, FAIR was part of Reality Labs, actually, for, like, a year and a half, uh-

Oh

... after one reorg. Like initially, FAIR wasn't, and then it got reorg-ed into Reality Labs. So I think when I left, actually, FAIR was officially part of Reality Labs-

Wow

... um, if I recall correctly. Um, there was at least a, a one-and-a-half to two-year period where, where that was the case. Um, so some of the AI investment actually that, like, laid the foundations came out of Real- out, out of that metaverse-

Yeah

... um, investment in the first place. You know, that said, I think, you know, we talk about data as being a compute multiplier all the time. Uh, talent, I think obviously is a compute m- multiplier. And given the amounts that they're spending on compute, I think you can make a good argument as to why spending in a crazy amount on talent, uh, is also worth it.

So I'm excited to see what they do. Um, I, I hope that they, uh, put a lot of focus on data.

And become customers.

Yes.

Awesome. Well, thank you so much for, uh, chatting and coming by, uh, and insisting on in person because you are actually very charismatic in person, so I'm glad you did this.

Well, thank you very much. Thanks for having me.

Yeah.

And, uh, a joy to, to get to chat, uh, in real life.

Awesome. Cool.