Alexandr Wang: Building Scale AI, Transforming Work With Agents & Competing With China

So before I went to MIT, I worked at Quora for a year. And so this is 2015 to 2016, or no, sorry, 2014 and 2015 was when I worked as a software engineer. And this was already at a point in the market where ML engineers, as they were called, or like machine learning engineers, made more than software engineers. So that was already like the market state at that point. I went to these summer camps that were organized by rationalists, the rationality community in San Francisco, and they were for precocious teams, but they were organized by many people who have become pivotal in the AI industry. So one of the organizers is this guy, Paul Cristiano, who used to, who's the inventor of RLHF, actually. And now he runs, or he's a research director at the USAI Safety Institute. He was at OpenAI for a long time. Greg Brockman came and gave a speech at one point. Eliezer Yudkowsky came and gave a speech at one point.

And actually, I was very, like, when I was, I don't know, it must have been 16, I was exposed to this concept that, like, potentially the most important thing to work on in my lifetime was AI and AI safety. So something I was exposed to very early on. So then when I went to MIT, I was started at MIT when I was 18. I, like, studied AI quite deeply. That was most of what I did in the sort of day job. and then um uh kind of got antsy applied to yc and then the idea was kind of like okay how initially it was like okay where can you apply um sort of like ai to things and this was um in the era of chatbots which is like crazy to think about actually um that there was like this like mini chatbot bubble boom yeah yeah 100 percent in uh in 2016 um which is uh which was I guess spurred by magic.

Right. Or some of these apps. And Facebook had a big vision around chatbots. And anyway, there's a little mini chatbot boom. So the initial thing that we wanted to work on and was was chatbots for doctors. Right. Which is like a funny idea, because do you guys know anything about doctors? Yeah. No, not at all. Like, basically, no. It was just sort of like, oh, doctors are a thing that sound expensive. And so and I think it was like I think it's like indicative of like I mean, I don't know. you guys see this all the time but i feel like most of the times young founders like first 10 ideas are like always first of all they're very mimetic so they're probably like there's a lot of like the same ideas over here there's like a dating app there's like some something for like you know social life you know there's the same ideas um and then i think that like i think young people have a very poor sense of alpha like what what are the things that they're actually like going to be uniquely positioned to do and i think you know most young people don't have a sense of So it's, you know, it's not clear. So when we were in YC, we were roommates with, with, with another YC company. And we were sort of like, we were sort of observing this, like, this, like, chatbot boom ahead of, you know, that was happening at the time.

um and but it was very clear that like um chatbots if you wanted to build them and this is funny to say in retrospect required lots of data um and required lots of like human elbow grease um to be able to get them to work effectively and so like just like kind of off the cuff at one point it was like oh like what if you just did that what if you just did the data and the the like language data and the the human data so to speak for the chatbot companies we were also very lost by the way i think you probably remember we were we were quite lost mid-batch um uh and like many yc companies i think and so then we um switched to this like concept i think the you know the initial idea was like api for um for human tasks or something along those lines and uh and one night i was just like trolling around for domains scale api.com was available and then we just bought it we launched it i think a week later on product hunt yeah i remember the the product on page is still live i was reading it last night and i remembered the tagline it was an api for human labor like that that that's my recollection of sort of like the like distilled insight that you have was like what if there is an eight what if you could call a human with like an api yeah and

that was i mean i think it was like three days for us to put up the landing page launched on product hunt i think this idea captured some amount of imagination of the like of the startup community at the time because it was sort of like this weird form of futurism where you have like humans delegated like apis delegate to humans in this in this interesting way it's like an inversion of the yes yeah yeah yeah exactly doing work for the machines and some of the other way around yeah yeah yeah it's funny because the the initial phase you know we sort of we just worked with all these engineers who reached out to us from that product hunt, which was a real grab bag of use cases. But then that was enough for us to raise money at the time and to get going.

And then a few months after that, it became clear that self-driving cars was actually the first major application that we needed to focus on. And so there were many very big decisions, I would say, in the first years of the company. One thing that was curious is at that point, there were other solutions that were already the game in town like mechanical turk from amazon was sort of the thing that people were using but you ended up capturing this whole other set of people that didn't know about it and you had a way better api and you kind of won yeah it was not clear at that point because you probably were compared a lot with mechanical turk yeah so mechanical turk was definitely the sort of like um the concept in most people's mind at the time i mean it was just it was kind of one of these things where i think a lot of people had heard about it but anyone who had used it knew it was just awful and so It's like whenever you're in a space and that's kind of the like, that's like the thing is like people mention a thing, but it sucks.

That's usually like a pretty good sign. And so that was that was enough to give us like early confidence. But then I think the thing that like really I would say that the the thing that was as actually fundamental to the success of the company was actually focusing on this like on this like seemingly very narrow problem of self-driving cars. I think that, you know, I remember very early on when it was maybe like six months after we were out of YC. Basically, there was another YC company, Cruise, that had reached out to us on our website. And sort of like in the blink of an eye, they became our largest customer.

And they found you just from your launch? Yeah, just, yeah, I think maybe even Google. Like, it's not even totally obvious, but just, yeah, vaguely from our launch and vaguely. It was actually an XYC founder that was working at Cruise that reached out to us. So maybe some YC mumbo-jumbo. We're a caretzy, man. Who knows? The world works in mysterious ways. So they grew very, very large. So then early on, we made this decision. And I remember we went to our lead investor at the time, and we had this conversation. It was like, hey, actually, we think we should probably just focus on this self-driving thing.

Alexandr WangIt was actually a very interesting conversation because the reaction was like, oh, that's just obviously way too small a market.

um and like you know you're never going to build like a gigantic business that way um and we're like we think it's probably a much bigger market than than you think it is because there's like you know all these self-driving companies are getting crazy amounts of funding and the automotive companies are doing huge programs in self-driving and it clearly is the future like it feels like something that that um that should exist and so we're like if we focus on it we think we can build like build the business much more quickly and it's funny looking back because both things are true It is both true that it enabled us to build the business to be to get to scale pretty very quickly.

And it's also true that that was not a big enough market to sustain a gigantic business. The story of scale in many ways is like this progression of like, how do you continue? You know, AI is this incredibly dynamic space. Lots of things are constantly changing. And a lot of, I think, what we pride ourselves on at the company is how we've been able to continue building on and contributing to this very fast-moving industry. When did you become much more aware of the scaling laws? Because one of the interesting facts that sort of emerged is that you're a little bit the Jensen Huang of data. I think that in self-driving, scaling laws were not really a thing.

um because and the fundamental the biggest reason actually was that like one of the biggest problems in self-driving is that your whole algorithm needs to run on the car and so you're very constrained by the amount of compute you have access to and is available to you so like a lot of the engineers and a lot of the companies working on self-driving never really thought about scaling laws they were just all thinking about like okay how do you keep grinding these algorithms to be better and better better that are like small enough to fit onto these um onto these cars but then we started working with open ai in 2019 this was like gpt2 era um and i would say like gpt1 gpt was sort of like this curiosity gpt2 um i remember open ai like they would have a booth at these like large ai conferences and they would like you know their demo would be to allow researchers to like talk to gpt2 and it was like mildly like it was it wasn't like particularly impressive but it was like kind of cool it's like kind of this thing and then um i think by gpt3 uh it was sort of this like that's when the scaling laws clearly um you know felt very real and that was i mean i think gpt3 was 2020 um so it was actually like long before before the world caught on to what was

happening yeah did did you You know, as early as 2020, did you have a strong inkling that this was really going to be, like, the next big chapter of scale? Or not until ChatGPT took off? Was that clear? Like, 3.5 or was it 4? I think that, like, in 2020, I think it was clear that scaling laws were going to be a big thing. But it was still not totally obvious. I remember this, like, interaction. You know, I got early access to GPT-3. And then it was, like, in the playground. and then i i was like playing with it with a friend of mine and i told the friend of mine oh you can like talk to this model and during the conversation um uh my friend got like visibly frustrated and angry at the ai but in a way that wasn't just like oh this is a dumb like toy it was like it was in a way that was like somewhat personal and that's when i was i realized like whoa this is like somehow qualitatively different from anything that existed before it feel like was passing the turing test at that point kind of it was like semblances yeah semblances it was like sort of like the the the glimpses of it potentially passing the turing test right but i think the thing that really um caused the recognition of i would say generative ai which

is still even the term in some ways it was really dolly i think that that um that convinced um that convinced everyone but i think i think my my personal um journey was like gbd3 like was like highly interesting and then and so it was like one of many bets at the company and then in 2022 over the course of dolly and and then um and then later chat gbt and you know um gbd4 etc and we worked with open ai on instruct gbt which is kind of a precursor to chat gbt it became very obvious that that was like the at the far moment for the for the company and for frankly the world that's when we saw it as well with the big shifting companies because it was that 3.5 moment release end of 2022 and we started seeing a bunch of companies and smart people changing directions and pivoting their companies in 2023 and that was that moment this dynamic that you referenced which is kind of the you know scales the NVIDIA for data kind of thing um I think that became quite obvious.

I would say GPT-4 really was the moment where it was like, wow, scaling laws are very real. The need for data will basically grow to consume all available information and knowledge that humans have. And so it was like, wow, this is like this astronomically large opportunity. Yeah, four seemed like the first time it was something that you could get to not hallucinate basically ever. You could actually have a zero hallucination experience in limited domains, which is we're still sort of in that regime, even at this point. You know, the classic view is that if it's hallucinating, you're not giving it the correct data in the prompt or context, or you're trying to do too much in one step.

Yeah, I mean, I think I think like the reasoning paradigm is is has a lot of legs. And it's actually been interesting, this last era of the of model improvement, because the gains are not really coming from pre-training, which is so so we're like moving on to a new scaling curve of reasoning and reinforcement learning. But it's like shockingly effective. And I think that, you know, it's the analogies between like AI and Moore's law are pretty clear, which is like, you know, you'll get on different like technical curves, but like you zoom way out, it'll just feel like there's like smooth improvement of models.

One of the things that has been popping up with some of the really big, well-known rappers is they're getting access to full parameter fine tunes of the base models, especially the frontier-based closed source models. Is that a big part of your business or something that people are sort of coming to you for? Or just, like, these verticalized full parameter fine-tuned, like, data sets? Yeah, I think this is going to be a, like, blueprint for the future, right? So right now, I mean, like, the total number of large-scale parameter fine-tuned or reinforcement fine-tuned models is, like, still pretty small.

But if you kind of think about it, like, one version of the future is that every firm's core IP is actually their specialized model or their own fine-tuned model.

Alexandr WangAnd just in the same way that like, you know, today, you would generally think that the IP of most tech companies is their code base. In the future, you would generally think that their specialized IP might be the model that powers all of their internal workflows.

and what are the special things they can add on top? Well, they can add on data and environments that are somehow specific, very, very specific to the day-to-day problems or information or challenges or business problems that they see on a day-to-day level. And that's the kind of like really gritty real world information that nobody else will have because nobody else is like doing the same, the exact same business motion as them. There's a lot of weird tension in that though. So I remember friends of ours from one of the top model companies came by and they were like, hey, do you think YC and YC companies would give us their evals so we could train against it?

And we were like, no, dude, what are you talking about? Why would they do that? Because that's like their moat. And then I guess now that based on this conversation, it's actually, I mean, evals are pretty important as a part of RL cycles. And then even the evals are not really the valuable part. But the valuable part is actually the, like, properly fine-tuned model for your data set and your set of, you know, sort of problems. Yeah, it's like these Lego blocks, right? If you have the data and you have the environments and then you have the, you have, you know, a base model, you, like, you know, can stack those on top of each other, get a fine-tuned model.

And obviously the evals are important. This is some of the tension. And this is basically, you know, in a nutshell, the sort of, like, does AGI become a Borg that just sort of, like, swallows the whole economy in, like, you know, as one firm? or do you still have a specialized economy? My belief, generally speaking, is that you still do have a specialized economy. Like these models are platforms, but the like alpha in the modern world will be determined by, you know, to what degree you're able to sort of like encapsulate your business problems into data sets or environments that are then conducive towards building like differentiated models or differentiated AI capabilities.

Yeah, that's why asking for evals was so crazy to me because it's like, okay, you get the evals, the base model is way better. And then not, you know, now all your competitors have exactly the same thing that used to be your advantage. I think we will undergo a process in AI where we learn what the bright lines are, right? I mean, I think that like, it's like very obvious and intuitive to tech companies that they should not give away their code base and they should not give away their database. Like they should not give away their data, they should not give away their code base. The analogs of that in a highly AI-fueled economy, I think, will identify over time, but are the evals, your data, your environments, et cetera. I think you have a very techno-optimistic view of what the future is going to be with how jobs are going to be shaped. Can you talk more about that? Because I think you hinted at it before. It's going to be more specialized. It's not that all these jobs are gonna go away right first off it's undeniably true that we're we're uh at the beginning of an era of like a new a new way of working like like you know this there's this term that people have used a long time which is like the future of work well uh we're like entering the future of work or

certainly the next era and so work fundamentally will change but i do think um humans own the future and we we are we're like We have a lot of agency, actually, and a lot of a lot of choice in how this sort of like reformatting of of work or how the reformatting of sort of like workflows ends up playing out. You know, I think you kind of see this play out in in coding right now. And I think coding in some ways is is really the sort of like case study for other fields and other, you know, other areas of work where sort of the initial phase is this sort of like assistant style thing where, you know, you're kind of doing your work and then the models are kind of like assisting you a little bit here and there and then you go to a you know the sort of like cursor agent mode kind of thing where you're you're like synchronously asking the the models to like carry out these workflows and you're sort of like you're you're managing like one agent kind of or you're sort of like you're kind of like pair programming with a single agent and then and then now with like codecs or other systems like it's it's very clear the paradigm is like oh you have this like you have this like swarm of agents that you're going to deploy on like all these various tasks and you're just

going to like sort of like you know deploy like um give all these tasks and you'll have this sort of like um this this cohort of agents that are sort of like you know doing this work that you you think is appropriate and that last job um uh has a has a semantic meaning in the in the current workforce it's a manager you know you're basically managing this sort of like this set of agents to do um actual work and so and and i think that like agi or you know agi or doomers or whatnot like they take this view that like oh even this job of like managing the agents will just be done by the agents so like humans will be taken out of the of the process entirely but our belief my personal belief is that you know this is um management is very complicated um management is also like more about like what's the vision that you have and what's the sort of like what's the like end result you're aiming towards and those will be fundamentally i think like you know we have a human demand and human desire driven economy so those will be driven by humans and so i think the terminal state of the economy is just is large-scale humans manage agents in a nutshell i have a funny story where um founder friend of mine is trying to promote uh one of his you know

junior employees but they're really really smart and they're working on the agent infrastructure and then he was like hey do you want to like you know I'm looking for someone who could step into management you've never managed people before do you you know if we hired some people under you like how would you feel about that and this you know mid 20 something really smart you know sort of he's just like he's an engineer and he's like why would I do that like just give me like more compute like you know the model like look at what just happened to the model literally like last month and you know i didn't have to do anything it just started doing things that it couldn't do a month ago why would i want to manage people like just give me like i will just manage more agents for you and it's fine okay so what are the unique things that that um that humans will do over time i mean i think this like this like element of vision um is very important this element of like kind of like debugging or sort of like um fixing when things go wrong like most of a manager's job speaking as a manager is is just like putting out fires dealing with problems dealing with like like issues that come up like i think intuitively

you know the idealistic manager job seems like this very cushy job because you're like oh yeah all the other people do all the work and i'm just sort of like i just vaguely supervise and then the reality is obviously like highly chaotic i think people often jump to this like you know extreme reality where it's like oh yeah these like you know you're just going to manage the agents and you're going to sort of like live this like you know kind of victorian life where all your problems are solved but but no i think it's still going to be pretty complicated like getting agents to like coordinate well with one another and like coordinating the workflows and and then debugging the issues that come up like these are still complicated issues and You know, having seen what happened in self-driving, which was more or less that, like, you know, it's easy to get to 90%, very, very hard to get to 99%.

I think that, like, something similar will happen with large-scale agent deployments and that, like, you know, final 10% of accuracy will be, like, you know, will require a lot of work. Yeah, even for self-driving cars right now, there's the remote assist for all these super edge case. So there's still a human at the end managing the car.

Alexandr WangYeah. And the ratio, by the way, I mean, the companies don't publish them, but I think the ratio is something like five cars to one teleoperator or maybe even less than maybe three cars per teleoperator.

So the ratio is like, you know, much lower than people think. I think that like humans are much more involved even in self-driving cars than I think most people appreciate. I mean, which if you put it in that perspective, I think it's still very optimistic. It's just the output of getting rides instead of doing, in today's world, if you're an Uber driver, you just do one car. In this world, you can do five cars, right? Well, you have to believe for this, like, for an optimistic version of the future where, you know, unemployment is still low, et cetera. You just have to believe that humans are, like, almost insatiable in their desire and their demand.

And that, like, you know, prices will go down, things will become, you know, the economy will become more efficient, and we'll just, like, want more. And I think this has been a pretty reliable trend for like the history of humanity is that like, you know, we have somewhat insatiable demand. And so I have like conviction that like, you know, the economy can kind of get as efficient as it needs or as it like can get like hyper, hyper efficient. And then human demand will just like continue to sort of like fill the bucket. In the 20th century, when you said computer, maybe early 20th century, people didn't think of a computer as it is today.

They thought of a human being that would sit in front of a punch card tabulator. And that was what a computer was doing. It was a job title. Literally, that was a real person's job. And then, of course, now today, it's like, where are all the computers? Well, they're actually real computers now. I don't know. That was the Apollo mission. It was a bunch of people just crunching numbers with the trajectories of the Apollo. And that was it. Because the computer that went on the rocket actually was a microcontroller with, I think, only, like, single-digit hertz. It was, like, very tiny amount of computations.

It was just humans doing it. Totally. And even this, like, I mean, I think the concept of being a programmer is, like, highly esoteric. um in the sense that like oh you're like writing the instructions for these like machines to just like you know just continue to do repetitively and in some ways it's like the leverage boost that all humans will get is like similar to the leverage boost that like programmers have had historically for a long time i think a lot like a lot of people in silicon valley say it's like the the closest thing to alchemy in our world pre-ai let's say is programming because you sort of like you can do something that uh creates like like an infinite there's these infinite replicas of whatever you build and they can sort of like run an infinite number of times and um and i think the entire human workforce will soon see that kind that large of a leverage boost which is extremely exciting because i think that like programmers are are um have like benefited over the past few decades from this like unique perch where they they have like you know one 10x or 100x engineer can like can build something like absolutely incredible and like very very valuable and like very um uh shockingly productive and all of a sudden i think like like humans in all trades

i think will like gain this like level of leverage so i'm curious to return to a point that you made earlier about like how scale has kept reinventing itself if you had to like describe the arc of scale like what's what's what's the story and what were the turning points. Our initial business was all around producing data, generating data for various AI applications. And primarily self-driving car companies, right? For the early years, it was really, like you were saying, you were really focused on that. Yeah. For the first three years, fully focused on that. One of the properties of focusing on that business, of building that business is over time, we had this obligation to really get ahead of most of the waves of ai if that makes sense because you know for ai to be successful in any vertical area it needed data and so like our demand for our our products would precede a lot of times the actual sort of like evolution of ai into those industries so you know as an example we started working with open ai on language models in 2019 um we started working with the dod on government AI applications, um, and defense AI applications in 2020. It was like long before I think the, you know, recent sort of like drone fueled, um, you know, AI, uh, AI craze in the,

in the department of defense, we started working with enterprises long before, um, there was sort of like this, uh, you know, the recent sort of like larger waves around, uh, enterprise AI implementation. So, um, almost, uh, uh, sort of systemically or, or intrinsically we've had to, basically build ahead of the waves of ai i think it's actually quite similar to NVIDIA you know whenever like jensen gives his annual presentations about you know NVIDIA and its future and its outlook like he always is so ahead of the trends um and that's because he has to get there on the trend before the trend can even happen that's i think been one um one way in which our businesses continue to adapt because ai is like this you know it's this this like it's the fastest moving industry, I think, ever in the history of the world. And so, you know, that each turn, each evolution has been, has moved incredibly quickly. The other thing that happened late 2021, early 2022, we started working on applications. And so we started building out AI-based applications and now much more so agentic workflows and agentic applications for enterprises and government customers. And this was an interesting evolution of our business, because historically, like our

core business is highly operational. You know, we build this like data foundry, we have all these processes to produce data. It's a very operational process that involves like lots of humans and human experts to be able to produce data with quality control systems in place. That highly operational business and the success of that business is what created the momentum for us to, you know sort of dream about building an applications business when we went into it uh i had studied other businesses that had basically successfully um added on very different businesses and what are sort of like the unique traits or why do some of those work and one of them that is probably the most interesting um i think is like the most singular in modern business history is Amazon building AWS.

You know, if in 2000, you had written a short story that said that like, you know, this large online retailer would build this like large scale cloud computing rent-a-server business, like it would seem like nonsensical. I remember when they launched AWS in 2006, Amazon stock went down because all the analysts thought it was such a terrible idea. It had never been done before. It doesn't seem related at all to their core business. It's this weird thing. But the wisdom of that was, I think, twofold. I think first, and from talking to people who are there at the genesis moment of this business, one thing, probably the most important thing, was that they had conviction that the underlying business model of AWS would basically be this infinitely large and growing market.

that market would literally grow forever. There would be this exponential of the amount of compute that needed to be built up in the world. And if you did that, there was sufficient cost advantages from economies of scale. I think startups, you kind of have to switch modes at a certain point. Where early on, you're trying to go for very, very narrow markets. Almost the narrowest markets you can. And then you're just trying to gain momentum and then sort of like slowly grow out from those hyper-narrow markets. And then at some point,

Alexandr Wangif you like have ambitions to be a $100 billion company or more, then you have to sort of like switch gears and say, where are the infinite markets?

And how do you build towards those infinite markets? And so this was sort of like the moment where we realized that. And the simple realization was that every business and every organization was just going to have to reformat their entire businesses with AI-driven technology. and now obviously like agent-driven technology. And that would just be like over time that would swallow the entire economy. And so it was like another one of these like, okay, that's an infinite business to build out AI applications and AI deployments for large enterprises and governments. I think a lot of people don't realize that you guys are in the middle of this transformation.

They still think of scale as the data labeling company. But like, if you fast forward 10 years, do you think most of scale will actually be the agent business? Yeah, it's growing much faster at this point. I think it's an infinite market. So the crappy thing about most markets is that they have like a pretty shallow S-curve. But then, you know, you look at hyperscalers or like, you know, these like mega cap tech companies and they just have like these like ridiculously large markets. So you really want to get into these like infinite markets. So our strategy so far has been to focus on building use cases for, you know, focus on a small number of customers and be quite selective. So we work with, you know, the number one pharma company in the world, the number one telco in the world, the number one bank, the number one healthcare provider. And we work a lot with the U.S. government, you know, the Department of Defense and other government agencies. And the whole thing is, like, how do we take a very focused approach towards building stuff that resemble, you know, real differentiated AI capabilities? And all of this, I think, sounds somewhat trite, But we have this multi-hundred million dollar business in building all these applications. By my count, I think it's one of the largest AI application businesses in the industry. That's certainly what our investors tell us. And it's fueled by our differentiation in the data business because our belief fundamentally is that kind of what we talked about before.

The end state for every enterprise or every organization is some form of specialization imbued to them by their own data. Our day jobs historically have been producing highly differentiated data for, you know, these like large scale model builders in the world. And then we can apply that wisdom and that capability and those operational capabilities towards enterprises and their unique problem sets and give them specialized applications. Honestly, it kind of sounds like Palantir. At the most zoomed out level, you sort of squint. In that you're a technology provider. We're like a technology provider to some of the largest organizations in the world with a focus on data.

And I think the key difference is Palantir has built a real focus around these data ontologies and really solving this messy data integration problem for enterprises. And then our whole viewpoint is, like, what is the, like, most strategic data that will enable differentiation for your AI strategy? And how do we, like, generate or harness that data from within your enterprise towards developing that? I guess you will end up being pretty big competitors in another five, ten. years but for now like it's basically so green field i mean i think it's an infinitely large market yeah so you might not ever meet actually which is interesting yeah yeah i think in practice now we actually like frankly we work we're more partnered with palantir than competitive with them yeah um and uh well that's because the problems that these giant organizations are actually so massive and intractable that they throw up their hands it's like they have no shot at ever hiring people who could possibly solve the problem uh but a company like scale or a company like Palantir can actually hire kind of the same kind of people who would apply to YC, actually. It's kind of like this. Yeah, I don't know. The through line in my head right now is

realizing like, you know, there's plenty of capital. And then the limiting agent is actually really great, technical, smart people who are optimistic and actually work really hard. There's like not enough of those people. That's true for the world. And by the way, i think one of the cool things about um agents as we were talking about before is that like all of a sudden those people get near infinite leverage so um i think we're gonna i think that bottleneck gets exploded now hopefully um due to due to ai again i i think you know just like how in cloud aws is the largest by far but there's so many other cloud providers that actually are all at like like it's not a winner-take-all kind of business per se and it doesn't have to be Yeah, exactly. And I think that it's just too big of a market to even be close to winner takes all. There's no single organization that can have the operational breadth to be able to swallow the whole market.

Talking about operations, you clearly are living in the future, which is super cool. I'm sure you're running scale with all these agents and tools already to make it very efficient. Could you share some of the things that you're doing internally as a company and agents you're adopting so you can do more with less people? We saw this early because when the model developers were starting to develop agents and starting to develop using reinforcement learning, like actual reasoning models where the models could actually really do end-to-end workflows, we were responsible for producing a lot of the data sets that enabled the agents to get there.

And then we saw just how effective that training process is. I think that like the efficacy of reinforcement learning for for agent deployments is like is pretty insane.

Alexandr WangSo then once we realized that, we realized, like, okay, if you can actually, like, you know, turn existing human-driven workflows into environments and data for reinforcement learning, then you have this ability to convert these, like, human workflows, especially ones where you're, like, okay with some level of faultiness and okay with a certain level of reliability.

you can convert those into agentic workflows. So there's all sorts of agent workflows that happen in our hiring processes and happen in our quality control processes and happen to sort of just automate away certain data analyses and data processes, as well as various sales reporting. It's sort of embedded at every major org of the company. And the whole thing is just like mindset. that like, can you identify these like very repetitive human workflows and basically like undergo this process where you convert that into data sets that enable you to build automation tools? What do these data sets actually look like? I mean, for browser uses, like, is it an environment? And then, you know, here's a video of a human being going through this process of like filling out this form and decide like, yes, no on this drop down or something. I mean, you know, what's a concrete example just for the audience? One of the processes that we go through is like you'll take a sort of like full packet of a from a candidate and you'll like want to distill that into like you know a brief of some sort that sort of like gives all the salient details about that candidate for like decision by a sort of like broader committee um and these kinds of cases you know broadly speaking

like deep research plus plus kind of things are like the lowest hanging fruit it's just sort of like can you take these processes that like more or less look like you know you have to like click around a bunch of places and pull a bunch of pieces of information and then blend them together and then produce some analysis on top of that. Like that process, that fundamental like information-driven sort of like analysis process is the easiest thing to drive via agentic workloads. And the kinds of data you need are just like, you know, we call them kind of environments, but usually it's just like, what is the task?

What is the full sort of like data set that's necessary to conduct that task? And what is like the rubric for how you conduct that effectively? Do you need RL and fine-tuning when, like, prompt engineering and meta-prompting seems so good? I think that, yeah. I mean, I think prompting – I mean, as the models get better, prompting will get better. But, like, prompting gets you to a certain level, and then reinforcement learning gets you beyond that level. And actually, this is a good point. I think that, like, probably most of the time in our business, it's mostly prompting that just, like, works really well.

I mean, that's the weird thing is, like, oh, shoot, you don't have to crack open the models. and then frankly like the next models are going to be so good and then the evals are mainly about picking which model or you know at what point do you switch to the next one i do think startups need basically like a strategy for how they like will um walk up the complexity curve so to speak like you need to like your you know whatever product or business you build like needs to like really um benefit from like the ability to like race up this complexity curve which is the broad broader curve of capability of the models.

I mean, you actually created this leaderboard that has a lot of these super hard tasks that are trying to go into this next curve of reasoning. Can you tell us about it? One of the things that we built in partnership with the Center for Eye Safety is Humanity's Last Exam. It was a funny name. I think, unfortunately, there will be yet another exam beyond it. But, you know, the idea was how, like, let's effectively work with, you know, the smartest scientists in the field. And, you know, we worked with many very brilliant professors, but also very many, like, individual researchers who were, like, quite brilliant.

And we just collated and aggregated this data set of what the smartest researchers in the world would say the hardest scientific problems they've worked on recently are. They solved them or they sort of like came to the right, you know, they were able to solve the problems, but they're sort of like the hardest problems that they're aware of and know of. I was curious how you came up with these problems. So each of the professors contributed new problems. So these are not these are problems that have never appeared in any textbook or any exam ever. They just like came out of their brains and they like typed up like a new problem, like from scratch.

Am I saying this right? Yeah, and the general guidance was like, you know, what has come up recently in your research that you think is a particularly hard problem, right? The problems are stupidly hard, incidentally. They're like insane. I don't know if you guys have looked at these problems. They're totally crazy. Yeah, it's totally crazy. And by the way- They cannot be searched on the internet. It's like you need to have a lot of expertise and actually think about them for quite a long time. Yeah, they require a lot of reasoning. I'm recently like right now. So we have a time limit where the models can only think for I think it's 15 minutes or 30 minutes.

And one of the most recent requests from one of the labs is like, can you please increase that time limit to like a day so that the model has like up to a day to think about the to think about the problems. But, yeah, no, they're they're deviously hard problems. Unless you have expertise in the specific problem, you probably don't have a chance of getting it right. But even this evaluation, like I think when we first launched it...

Alexandr WangYou know, and this was earlier this year, the best models were scoring like 7%, 8% on it. Now the best models score north of 20%. It's moved really, really quickly.

And I think, you know, I think... Do you think we're going to get a benchmark saturation for this one as well? I think eventually, yeah, it'll be saturated. And then we have to move on to new evaluations. I mean, I think the like, the saving grace for the naming was that it is the last exam. The new evals will be sort of like real-world tasks, real-world activities, which are sort of like fundamentally fuzzier and more complicated. Have you solved any of the problems yourself, Alex? I know you were a competitive math person for a long time. Yeah, yeah. I mean, the math problems require a lot of – they're like very deep in the fields.

I think I managed to get a handful, but like most of them are like hopeless. Yeah. I looked at the ones that the models can solve. So that was that was one of the evals. And we've produced a number of other evaluations. But but yeah, I think that like the the in the AI industry really, I think, continues to suffer from a lack of very hard evals and very hard tests that show really like the frontier of of model capabilities. and these evals when you get when you build an eval that sort of like becomes popular in the industry it has this like deeper effect which is that that's all of a sudden the like north star in the yardstick that in that researchers are trying to um optimize for and so it's it's actually this like very gratifying activity you know we built humanities last exam um you know most of the like all the model providers um you know will always report their their their their results there's like tons of researchers who are really motivated by by doing a good job i mean it's it's And the models are going to get, you know, deviously good at like, you know, frontier research problems.

I guess Sam's starting to talk about, you know, that sort of stage four innovators of AGI is coming. And, you know, that's the prognostication for the next year. Do you think that's, you know, correct? The next 12 to 24 months is like really the moment that literally new scientific breakthrough is coming from the operation of reasoning and these models. I mean, I think it's super plausible. Well, you know, in fields like biology, and this is probably one of the ones that comes up the most, but there's like there's probably intuitions that the models have about biology that humans don't even have because it's just, you know, they have this like different form of intelligence.

Right. And so you'd expect there to be some areas in some fields where the models have some fundamental deep advantage versus humans. And so I think it's like very realistic to expect in those fields. Biology, I think, is sort of like the clearest one for me. It kind of already happened for chemistry. Last year, the Nobel Prize went to the Google team, Demes and John Jumper with AlphaFold. Yeah, exactly. That was like a huge jump. Before that, there was like this competition where they were trying to get more printing fold structures that were going to get solved. And it was like abysmal and AlphaFold destroyed it.

It's a strange time to be a scientist, but an exciting time for science. there's this um uh short story it talks about this future where like you know there's uh effectively AIs that are like that are conducting all the frontier of R&D research and um and scientists you know what scientists do is they just sort of like look at the discoveries that the AIs make and sort of like try to understand them yeah I mean I think that like very exciting time to to see the how the frontier of human knowledge expands and then I mean I think that be great because in areas like um in biology will fuel breakthroughs in medicine and healthcare and and all these other and all these other things and then the majority of the economy will will chug along you know giving humans what they want china open sourcing or deep seek open sourcing their models is like another very interesting question like how does that play out and um and there's this awkward sort of thing that you know the best open source models in the world now come out of China.

It's sort of this, like, awkward reality to contend with. What do you think we can do to just make sure that it's the American models that are ahead? Or, you know, is that written in the stars? Or, you know, something tells me that's not.

Alexandr WangThe simplest explanation for me about why the Chinese models are so good is espionage.

I think that there's a lot of secrets in how these frontier models are trained. And when I say secrets, it sounds more interesting than they are, but there's just a lot of tacit knowledge. There's a lot of like, you know, tricks and small and intuitions about where to set the hyper parameters and like, you know, ways to make these models work and to get the model training to work. The Chinese labs have been have been able to move so quickly and accelerate and make such fast progress, whereas some even like very talented U.S. labs like have made progress less quickly. And I just purely think it's because, you know, a lot of the secrets about how to train these models, you know, the secrets leave the frontier labs and make their way back to these Chinese labs.

I think the only way to model the future is that China has pretty advanced models. You know, the solace right now is they're not the best models. They're sort of like a half step behind, let's say. But it's tough to model what will happen when it's sort of truly neck and neck. we're very behind on energy production which is just pure regulation like that could be fixed in two seconds but you know hasn't been yet that's a huge problem i mean if you look at you know not that the past will be a predictor of the future if you look at what u.s total grid production looks like it's like looks flat as a pig and if you look at um you know chinese i saw that uh aggregate uh you know grid production it's like you know it's doubled over the past decade it's just like it's just this like straight up i saw that and it's astonishing it's i mean that's just a policy failure china just you know the vast majority of that is coal and coal's growing in china and um in the united states uh actually renewables have grown a lot but renewables trade off against the uh the sort of fossil fuels so we've sort of like done a done a transition of our of our um energy grid whereas they're just continuing to compound let's say we have this issue on

power production but we're advantaged in chips i think like net net we will come out ahead on compute um if you look at data i mean this goes towards a lot of the questions you've been you've been asking about but like i mean i think china is like fundamentally very well positioned on data um it's weird to say because it's obviously like you know we help all the american companies with data in china they can ignore copyright or other privacy rules and and they can sort of um you know build these large models without abandon and then and then the second issue is that um There's actually large scale government programs in China for data labeling.

There are, you know, seven data labeling centers, like in various cities that have been started up by the government itself. There's large scale subsidies for AI companies to use data labeling, a voucher system. In fact, there's like college programs because, you know, one of the interesting things is in China, like employment is such a large national priority that they like, you know, when they have a strategic area like ai they'll like figure out okay what are all the jobs and they'll like create these like funnels to um to to create those jobs and then we're seeing this in robotics data too where like there's already in china they're like large-scale factories full of robots that just go and collect data um and uh and strangely enough like even a lot of u.s companies today actually rely on data from china in training these like robotics foundation models long story short i think china likely has a data an advantage on data and then the algorithms um you know the us is on net much more innovative but if espionage continues to be a reality then like you know you're basically even on algorithms so um so it's hard to model but i think that probably like you know it's like 60 40 70 30 that the united states like has like in undeniable

I will continue to advantage, but there's like a lot of worlds where China just like catches up or potentially even overtakes. I mean, the scary thing for me is, you know, watching Optimus or YC has some robotics companies like Weave Robotics. And, you know, we look at those things, the software can be as good or better than anything coming out of China. But when it comes to the hardware, it's like bomb cost over here, 20,000, 30,000 bucks. Like you can't, you know, we can't even make like high percentage screws over here. And then over there, the same robot, the embodied robot, could be made for, I don't know, $2,000, $3,000, $4,000.

It's like you just walk down a street in Shenzhen and they got it. And so how do you compete against that at sort of that state level? The degree to which China is incredible at manufacturing, I mean, that's a very big problem. And it relates to defense and national security. It's a fundamental issue because on some level, defense and national security will boil down to which countries have more things that like can deter conflict or can can go into, you know, can can shoot other things down. Yeah, I don't think it's going to be fighter jets and aircraft carriers anymore. I mean, it's probably going to be, you know, this micro war of it's like hyper micro.

It's drones and embodied robots. and i mean yeah exactly drones and body robots cyber warfare the um cold war era um uh philosophy of like you know you build like bigger and bigger bombs um it's like the exact opposite of that it's actually like it's like the fragmentation and uh and and move towards sort of like you know smaller more nimble attributable resources um is the is the that's like one of the big picture trends i would say um and then the other big picture trend is just what we believe which is uh the move towards uh agentic warfare or agentic defense which is basically you know if you if you actually mapped out the what warfare looks like today or like what like the um you know the actual process of a conflict um you know if you look at russia ukraine or other conflict other conflict areas like the decision-making processes are driven are remarkably manual and human driven and it's just like all these just all these like very critical battle time decisions are made like with very limited information unfortunately um uh in these like very manual workflows and so it's very clear that that um if you used ai agents you would have perfect information and you would have uh immediate decision-making. And so we're going to see this huge shift towards

agent-driven warfare and agent-driven conflict. And it has the potential of turning these conflicts into these almost incomprehensibly fast-moving kinds of scenarios. And that's something that you guys are actively working on, right? Is there anything that you can talk about? I assume some of it is classified, but... Yeah. So one of the things we're doing is we're building this system called Thunder Forge with the Indo-Pacific Command out in Hawaii. It's responsible for the sort of the Indo-Pacific region. And it is the flagship DOD program for using AI for military planning and operations. So we're basically doing exactly what I said. We are we take the existing human workflow. The military works in what's called a doctrinal way or they're sort of like governed by the doctrine of this, like, you know, very established military planning process.

and you just convert that into you know a series of agents that work together um and and conduct you know the exact same task but it's just like all agent driven and then all of a sudden you you turn these like um very critical decision making cycles from you know 72 hours to 10 minutes and it kind of like changes it from um you know you know when you play chess if you play chess versus a human they just spend all this time thinking um you know you you know it's sort of this like slow game. Play chess against a computer. It's just like these immediate moves back. And it's like this sort of like unrelenting form of warfare.

I mean, some of it is like being able to see the chain of thought immediately is the most powerful thing. Yeah. Like, because, you know, I don't want the answer. I want to see how you got there. And then actually seeing the reasoning itself was so powerful. I mean, that's actually why the launch of that first DeepSeek was way more interesting because I think O1 had come out, but they hid the reasoning. And it's like, no, the reasoning is actually a really important part of it. And the only reason why they hid it was they didn't want other people to steal it, which they did anyway. I think that that's another like interesting thing about this space, which is that, you know, so far you could really model as like there's like advanced capabilities and you can try to keep those secret and you can try to keep those closed, but they open over time kind of no matter what you do.

Well, I mean, clearly, Alex, you've done a lot of incredible things and transformed your company multiple times, and you have all these deep matter expertise in many areas. You're clearly hardcore. Is there advice for the audience to be more like you? You know, I think that the biggest thing is you just have to really, really, really care. And I think it's like a folly of youth in some ways that when you're young, like almost everything feels like, you know, so astronomically important that you just like you try immensely hard and you care about every detail. you know everything uh matters just way more to you and i think um and i think that that trait is really really important and um you know it's like just in varying degrees for different people so i wrote this post many years ago called hire people who give a shit and it really is pretty simple you notice i notice you know when you interview people when you interact with people you can tell people who are just sort of like phone it in versus people who sort of like they like hang on to their work as like you know it's like it's like so incredibly monumental and forceful and important to them that they they do great work and it's sort of like eats at them when

they don't do great work and when they do great work they're sort of so satisfied with themselves and so there's sort of this like um the magnitude of of care and one of the greatest indicators of like a just like how much i enjoyed working with people or like frankly how successful they were at scale was really just this like what is what you know to what degree of their what degree their soul is invested in into um uh into the work that they do and so i think that that you know if you were to pick one thing that that probably is the sort of like unifier in some way it's like you know um i care a lot uh i care a lot about every decision we make at the company um you know i still review every hire at the company you know i i we have this process why where i approve or reject literally every single hire at the company. And so I care immensely. And then I work with all these people who care immensely. And then that enables us to really sort of like we feel much more deeply what happens in the business. And as a result, we sort of like, you know, we'll change course more quickly. We'll learn more quickly. We'll take our work more seriously. We'll adapt more quickly. And I think that that's been quite important to the success that we've had.

Alex, you were telling me a story recently that stuck with me about how quite recently, even when Scale was a very large company, you were personally hand-reviewing all the data that was being sent to partner companies and being basically the final quality control, like, that data point is not good enough. Yeah, exactly. I think a lot of founders would probably agree with this, but what your customers feel and when your customers are happy and sad, it really gets to you. And so when you have unhappy customers, it's like personally a very painful thing. Broadly speaking, we have this value at our company, quality.

fractal um and and i do believe that like high standards sort of like um they trickle down within an organization and um you know it's very rare that you see an organization where like where like standards um increase as you get lower and lower down in the organization you know most of the time when people realize their manager or their manager manager or their like director or whomever don't really care then they sort of like you know that that removes the sort of like the, like, deep desire to need to care. And so it's, like, incredibly important that high standards and this sort of, like, this deep sort of care for quality is, like, this, like, deeply embedded sort of tenet of the entire organization.

Founder mode, man. Founder mode. Man, we got to have you back. Thank you so much for spending time with us. With that, sorry we're out of time, but we'll see you next time. We'll be right back.