The Data Foundry for AI with Alexandr Wang from Scale

0:05Hi listeners and welcome to KnowPriors. Today I'm excited to welcome Alex Wang who started ScaleAi as a 19-year-old college dropout. Scale has since become a juggernaut in the AI industry. Modern AI is powered by three pillars, compute, data, and algorithms. While research labs are working on algorithms and AI chip companies are working on the compute pillar, Scale is the data foundry, serving almost every major LLM effort, including OpenAI, Meta, and Microsoft. This is a really special episode for me, given Alex started Scale in my house in 2016, and the company has come so far. Alex, welcome.

0:47I'm so happy to be talking to you today. Thanks for having me. Known you all for quite some time. So excited to be on the pod. Why don't we start at the beginning just for a broader audience. Talk a little bit about the founding story of Scale. Right before Scale, I was studying AI and machine learning at MIT. And this was the year when DeepMind came out with AlphaGo, where Google released TensorFlow. flow. So sort of the, maybe the beginning of the deep learning hype wave or hype cycle. And I remember I was at college, I was trying to, to use neural networks. I was trying to train, you know, image recognition, neural networks.

1:21And the thing I realized very quickly is that these models were very much so just a product of their data. And you, I sort of played this forward and thought through it. And, you know, these models or AI in general is the product of, three fundamental pillars. There's the algorithms, the compute and the computation power that goes into them, and the data. And at that time, it was clear there were companies working on the algorithms, labs like OpenAI or Google's labs or a number of AI research efforts. NVIDIA was already a very clear leader in building compute for these AI systems, but there was nobody focused on data.

2:00And it was really clear that over the long arc of this technology, data was only going to become more and more important. And so in 2016, dropped out of MIT, did YC, and really started scale to solve the data pillar of the AI ecosystem and be the organization that was going to solve all the hard problems associated with how do you actually produce and create enough data to fuel this ecosystem. And really, this was the start of scale as the data foundry for AI. It's incredible foresight because you describe it as like the beginning of the deep learning hype cycle. I don't think most people notice that a hype cycle was yet going on.

2:42And so I just distinctly remember, you know, you working through a number of early use cases, you know, building this company in my house at the time and discovering, I think far before anybody else noticed that the AV companies were spending all of their money on data. How did you think about, like, talk a little bit about how the business has evolved since then, because it's certainly not just that use case today. AI is an interesting technology because it is, at the core mathematical level, such a general purpose technology. It could be, you know, it's basically functions that can approximate nearly any function, including, like, intelligence.

3:21and so can be applied in a very wide breadth of use cases. And I think one of the challenges in building an AI over the past, you know, we've been at it for eight years now, has really been what are the applications that are gaining traction and how do you build the right infrastructure to fuel those applications? So as an infrastructure provider, you know, we provide the data foundry for all these AI applications. We, our burden is to be thinking ahead as to where are the breakthrough use cases in AI going to be and how do we basically lay down the tracks before the sort of, you know, freight train of AI comes rolling through.

3:58We, you know, when we got started in 2016, this was the very beginning of the autonomous vehicle sort of cycle. It was, I think, right when we were doing YC was when Cruise got acquired. And it was sort of the beginning of the wave of autonomous driving being one of the key tech trends. And I think that we followed the early startup advice. You have to focus early on as a company. And so we built the very first data engine that supported sensor fused data. So it supported a combination of 2D data plus 3D data. So LiDARs plus cameras that were built onto the vehicles. And then that very quickly became an industry standard across all the players, you know, working with folks like General Motors and Toyota and Stellantis and many others.

4:46In the first few years, the company were just focused on autonomous driving and a handful of other robotics use cases. But that was sort of the primetime AI use case. And then starting in about 2019, 2020, it was an interesting moment where it was actually pretty unclear where the future of AI use cases, where AI applications were going to come. And this is obviously pre-language model, pre-generative AI, and it was a period of high uncertainty. So we then started focusing on government applications. That was one of the areas where it was clear that there was high applicability, and it was one of the areas that was becoming more and more important globally.

5:27So we built the very first data engines to support government data. This would support mostly geospatial and satellite and other overhead imagery. This ended up fueling the first AI program of record for the USDOD and was sort of the start of our government business. And that technology ended up being critical years later in the Ukraine conflict. And then also around that time was when we started working on generative AI. So we partnered with OpenAI at that time to do the very first experiments on RLHF on top of GPT-2. This was like the primordial days of RLHF. And the models back then were really rudimentary.

6:11Like they didn't, they truly, it did not seem like anything to us. But we were just like, you know, OpenAI, they're a bunch of smart people, we should work with them, we should partner with them. And so we partnered with a team that originally invented RLHF. And then we basically continued innovating with them from 2019 onwards. But we didn't think that much about the underlying technological trend. You know, they integrate this, all of this technology into GPT-3. There was a paper InstructGPT, which is kind of the precursor to ChatGPT that we worked with them on. And then ultimately, you know, in 2022, Dolly2 and ChatGPG rolled around and we ended up focusing a lot of our effort as a company into how do we fuel the data for Genitive AI?

6:56How do we be the data foundry for Genitive AI? And today, you know, fast forward to today, our data foundry fuels basically every major large language model in the industry, work with OpenAI, Meta, Microsoft, many of the other players, partner with them very closely in fueling their AI development. And in that timeframe, the ambitions of AI have just, you know, totally exploded. I mean, we've gone from, you know, GPT-3, I think, was a landmark model, but it was, you know, there was a modesty to GPT-3 at the time. And now, you know, we're looking at building, you know, agents and very complex reasoning capabilities, multimodality, multilinguality.

7:36I mean, the infrastructure that we have to build to support all the directions that developers want to take this technology has been really staggering and quite incredible. Yeah, you've basically surfed multiple waves of AI. And one of the big shifts that's happening right now is there's other types of parties that are starting to engage with this technology. So you're obviously now working with a lot of the technology giants, with government, with automotive companies. It seems like there's emergence now of enterprise customers and a platform for that. There's emergence of sovereign AI. How are you engaging with these other massive use cases that are coming now on the generative AI side?

8:10It's quite an exciting time because I think for the first time in maybe the entire history of AI, AI truly feels like a general purpose technology, which can be applied in a very large number of business use cases. I contrast this to the autonomous vehicle era where it really felt like we were building a very specific use case that happened to be very, very valuable. Now its general purpose can be encompassed across the broad span. And as we think about what are the infrastructure requirements to support this broad industry, and what is the broad arc of the technology, it's really one where we think, how do we empower data abundance, right?

8:48There's this question that comes up a lot, you know, are we going to run out of tokens? And what happens when we do? And I think that that's a choice. I think we as an industry can either choose data abundance or data scarcity. and we view our role and our job in the ecosystem to build data abundance, the key to the scaling of these large language models and these language models in general is the ability to scale data. And I think that one of the fundamental bottlenecks to what's in the way of us getting from GPT-4 to GPT-10 is data abundance. Are we going to have the data to actually get there?

9:29And our goal is, you know, how do we ensure that we have enough tokens to do that? And we've sort of, as a community, we have easy data, which is all the data on the internet. And we've kind of exhausted all the easy data. And now it's about, you know, forward data production that has high supervisory signal that is basically very valuable. And we think about this as, you know, frontier data production. And the kinds of data that are really relevant and valuable to the models today, the quality requirements have just increased dramatically. It's not anymore the case that these models can learn that much more from various comments on Reddit or whatnot.

10:08They need truly frontier data. And what does this look like? This is reasoning chain of thoughts from the world's experts or from mathematicians or physicists or biologists or chemists or lawyers or doctors. This is agent workflow data of agents in enterprise use cases or in consumer use cases or even coding agents and other agents like that. This is multilingual data, so data that encompasses the full span of, you know, the many, many languages that are spoken in the world. This includes all the multimodal data, to your point, like, you know, how do we integrate video data, audio data, you know, start including more of the esoteric data types that exist within enterprises and exist within a lot of industrial use cases into these models.

10:56There's this very large mandate, I think, for our industry to actually figure out what is the means of production by which we're actually going to be able to generate and produce more tokens to fuel the future of this industry. And I think there's a few sources or there's a few answers to this. So the first is we need the best and brightest minds in the world to be contributing data. I think it's one of the things I think is actually quite interesting about this technology is, you know, very smart humans. And so PhDs or doctors or lawyers or experts in all these various fields actually can have an extremely high impact into the future of this technology by producing data that ultimately feeds into the algorithms.

11:40If you think about it, it's actually their work is one of the ways that they can have a very scaled society-level impact. You know, there's an argument that you can make that producing high-quality data for AI systems is near infinite impact because, you know, even if you improve the model just a little bit, if you were to integrate that over all of the future invocations of that model, that's like a ridiculous amount of impact. So I think that's something that's quite exciting. It's kind of interesting because Google's original mission was to organize the world's information and make it universally accessible and useful.

12:13And they would go and they would scan in books, right, from library archives, and they were trying to find different ways to collect all the world's information. And effectively, that's what you folks are doing or helping others do. You're effectively saying, where is all the expert knowledge and how do we translate that into data that can then be used by machines so that people can ultimately use that information? And that's super exciting. It's exciting to the contributors who are in our network as well, because they think that there's obviously a monetary component and they're excited to do this work.

12:44But there's a very meaningful motivation, which is how do I leverage my expert knowledge and expert insight and use that to fuel this entire AI movement, which I think is is like a deep, you know, that's kind of like the deepest scientific motivation, which is how do I use my knowledge and capability and intelligence to fuel humanity and progress and knowledge going into the future? I think the somewhat undervalued thing where it's going to age me, but like there was a decade or so where like the biggest thing happening in technology was digitization of different processes. And I think there's actually some belief that like, oh, that's happened, right?

13:23Like, you know, interactions are digital and like information is captured in relational database systems on, you know, customers and employees or whatever. But one of the big discoveries as a investor in this field over the last five years has been like the data is not actually captured for almost any use case you might imagine for AI, right? Because Because I have multiple companies, and I'm sure Alad does too, and you and your personal investing where, you know, the first six months of the company is a question of where are we going to get this data? You go to many of the incumbent software and services vendors.

13:57And despite having done this task, you know, for years, they have not actually captured the information you'd want to teach a model. And like that, you know, that knowledge capture era, I think is happening and scale is a really important part. to make a Dune 2 analogy, I mean, I think it really is, you know, data production is very similar to spice production. It is the, it will be the lifeblood of all the future of these AI systems. And, you know, so, so I think best and brightest people is one key source. Proprietary data is definitely a very important source as well. You know, crazy stat, but JP Morgan's proprietary data set is 150 petabytes of data.

14:35GPT-4 is trained on less than one petabyte of data. So there's clearly so much data that exists within enterprises and governments that is proprietary data that can be used for training incredibly powerful AI systems. And then I think there's this key question of what's the future of synthetic data and how synthetic data needs to emerge. And our perspective is that the critical thing is what we call hybrid human AI synthetic data. So how can you build hybrid human AI systems such that AI are doing a lot of the heavy lifting, but human experts and people, you know, the basically best and brightest, the smartest people, the sort of best at reasoning can contribute all of their insight and capability to ensure that you produce data that's of extremely high quality, of high fidelity to ultimately fuel the future of these models.

15:26I want to pull this thread a little bit because something you and I were talking about both in the context of data collection and evals is like, what do you do when the models are actually quite good, right? Better than humans on many measured dimensions. And so, like, can you talk about that from both the data and perhaps, you know, we should talk about evaluation as well? I mean, I think philosophically, the question is not, is a model better than a human unassisted from a model? The question is, is a human plus a model together going to be able to produce better output than a model alone? And I think that'll be the case for a very, very, very long time, that humans are still, you know, human intelligence is complementary to machine intelligence that we're building.

16:06And they're going to be able to combine to build, you know, to do things that are strictly better than what the models are going to be able to do on their own. I have this optimism. Laud and I had a debate at one point that was challenging for me philosophically about whether or not Centaur play or like machine and human intelligence were complementary. My simple case for this is when we look at the machine intelligence, like the models that are produced, you know, we always, you know, you see things that are really weird. You know, there's like the Rot 13 versus Rot 8 thing, for example, where the models know how to do Rot 13, they don't know how to do Rot 8.

16:40There's the reversal curse. You know, there's all these artifacts that indicate somehow that it is not like human intelligence or not like biological intelligence. And I think that's the bold case for humanity, which is that, you know, there are certain qualities and attributes of human intelligence, which are somehow distinct from the very separate and very different process by which we're training these algorithms. And so then I think, you know, what does this look like in practice? It's, you know, if a model produces an answer or response, how can a human critique that response to improve it?

17:12How can a human expert, you know, highlight where there's factuality errors or where there's reasoning errors to improve the quality of it? How can the human aid in guiding the model over like a long period of time to produce reasoning chains that are very correct and deep and are able to drive the capability of these models forward? And so I think there's a lot that goes into, this is what we spend all of our time thinking about. What is the human expert plus model teaming that's going to help us keep pushing the boundary of what the models are capable of doing? How long do you think human expertise continues to play a role in that.

17:46So if I look at certain models, MedPalm 2 would be a good example where Google released a model where they showed that the model output was better than the average physician. You could still get better output from a cardiologist, but if you just asked a GP a cardiology question, the model would do better as ranked by physician experts. So it showed that already for certain types of capabilities, the model provided better insights or output than and people who are trained to do some aspects of that. How far do you think that goes in terms of, or when do you think human expertise no longer is additive to these models?

18:19Is that never? Is it three years from now? I'm sort of curious of the timeframe. I think it's never because I think that, you know, the key quality of human intelligence or biological intelligence is this ability to reason and optimize over very long time horizons. So, and this is biological, right? Because our goals as biological entities is to optimize over, you know, our lifetimes, optimize for reproduction, et cetera. So we have the ability as human intelligences to produce long-term goals, continue optimizing, adjusting, and reasoning over very long time horizons. You know, current models don't have this capability because the models are trained on these like little nuggets of human intelligence.

19:03So they're very good at almost like a shot glass full of human intelligence, but they're very bad at continuing that intelligence over a long time period or a long time horizon. And so this fundamental quality of biological intelligence, I think, is something that will only be taught to the model over time through a direct transfer via data to fuel these models. You don't think there's like an architectural breakthrough in planning that solves it? I think there will be architectural breakthroughs that improve performance dramatically. But I think if you think about it inherently, like these models are not trained to optimize over long time horizons in any way.

19:45And we don't have the environments to be able to get them to optimize for these like, you know, amorphous goals over long time horizons. So I think this is a somewhat fundamental limitation. Before we talk about some of the cool releases you guys have coming out and what's next for scale, maybe we can zoom out and just congratulate you on the fundraise that you guys just did. A billion dollars at almost 14 billion in valuation with really interesting investors like AMD, Cisco, Meta. I want to hear a little bit about the strategics. Our mission is to serve the entire AI ecosystem and the broader AI industry.

20:23We're an infrastructure provider. that's our role is to be as much as possible supporting the entire industry to flourish as much as possible. And we thought an important part of that was how can we be an important part of the ecosystem and build as much ecosystem around this data foundry, which is going to fuel the future of the industry as much as possible, which is one of the reasons why we wanted to bring along, A, other infrastructure providers like Intel and AMD and folks who are also laying the groundwork for the future of the technology, but also key players in the industry like Meta. Folks like Cisco as well, our view is that ultimately there's the stack that we think about.

21:04There's the infrastructure, there's the technology, and there's the application. And our goal as much as possible is how do we leverage this data capability, this data foundry to empower every layer of that stack as much as possible and build a broader industry viewpoint around what's needed for the future of data. I mean, I think that this is an exciting moment for us. I mean, we see our role, you know, going back to the framing of what's holding us back from GPT-10, what's in the way from GPT-4 to GPT-10. We want to be investing into actually enabling that pretty incredible technology journey.

21:42And, you know, there's tens of billions, maybe hundreds of billions of dollars investment going into the compute side of this equation. And one of the reasons why we thought it was important to raise the money and continue investing is, you know, there's real investment that's going to have to be made into the data production to actually get us there. With great power comes great responsibility. If, you know, if these AI systems are what we think they are in terms of societal impact, like trust in those systems is a crucial question. Like, how do you guys think about this as part of your work at scale?

22:14A lot of what we think about is how do we utilize, how does the data foundry enhance the entire AI lifecycle, right? And that lifecycle goes from, you know, A, ensuring there's data abundance, as well as data quality going into the systems, but also being able to measure the AI systems, which builds confidence in AI, and also enables further development and further adoption of the technology. And this is the fundamental loop that I think every AI company goes through. You know, they get a bunch of data, or they generate a bunch of data, they train their models, they evaluate those systems, and they sort of, you know, go again in the loop.

22:48And so evaluation and measurement of the AI systems is a critical component of the life cycle, but also a critical component, I think, of society being able to build trust in these systems. You know, how are governments going to know that these AI systems are safe and secure and fit for, you know, broader adoption within their countries? How are enterprises going to know that when they deploy an AI agent or an AI system, that it's actually going to be good for the consumers and that it's not going to create greater risk for them? How are labs going to be able to consistently measure what are the intelligences of the AI systems that we build?

23:24And how do they make sure they continue to develop responsibly as a result? Can you give our listeners a little bit of intuition for what makes evals hard? One of the hard things that, you know, because we're building systems that we're trying to approximate and build human intelligence, grading one of these AI systems is not something that's very easy to do automatically. And it's sort of like, you know, you have to kind of build IQ tests for these models, which in and of itself is a very fraught philosophical question. It's like, how do you measure the intelligence of a system? And there's very practical problems as well.

23:59So most of the benchmarks that we as a community look at for... The academic benchmarks. Yeah, the academic benchmarks that are what the industry used to measure the performance of these algorithms are fraught with issues. Many of the models are overfit on these benchmarks. They're sort of in the training data sets of these models. And so... You guys just did some interesting research here. Yes. Published them. Yep. So one of the things we did is we published DSM-1K, which was a held out eval. So we basically produce a new evaluation of the math capabilities of models that there's no way would ever exist in the training data set to really see how much of the, how were the performance of the models, what were the reported performance of the model capability versus the actual capability.

24:41And what you notice is some of the models perform really well, but some of them perform much worse than the reported performance. And so this whole question of how we decide we're actually going to measure these models is a really tough one. And our answer is we have to leverage the same human experts and kind of the best and brightest minds to do expert evaluations on top of these models to understand, you know, where are they powerful, where are they weak, and what are the sort of risks associated with these models? So, you know, one of the things that we're very passionate about is there needs to be sort of public visibility and transparency into the performance of these models.

25:20So there need to be leaderboards, there need to be evaluations that are public that demonstrate in a very rigorous scientific way what the performance of these models are. And then we need to build the platforms and capabilities for governments, enterprises, labs to be able to do constant evaluation on top of these models to ensure that we're always developing the technology in a safe way and we're always deploying it in a safe way. So this is something that we think is, you know, just in the same way that our role as an infrastructure provider is to support the data needs for the entire ecosystem.