Inside The $2.2B AI Research Accelerator | Turing

0:00300 million plus in revenue and profitable, 225 million in funding, at a last valuation of$2.2 billion, and customers that include OpenAI, NVIDIA, Anthropik, Google, Microsoft, Meta, Salesforce, Amazon, and I'm sure a lot more. Jonathan, what is touring? The era of sweatshop data labeling is over. Now what the labs needed is a strategic research accelerator, a Turing's the world's leading research accelerator, working with all of these frontier labs to advance them along the four pillars of superintelligence. multi-modality, reasoning, tool use, and coding. That's$30 trillion of knowledge work that is going to be automated.

0:39Ever since the scale AI acquisition, investment, whatever you want to call it, there's essentially a billion dollars up for grabs. This market is massive and growing. There's unlimited demand for high quality data. What does the inside of OpenAI look like? What does it take to get to GPT-5? First, I think GPT-5 is fucking awesome. The AI doomers talk about this concept of rapid takeoff. I don't think rapid takeoff is how things will unfold. We will get to ASI and it's honestly the most important problem humanity could be working on because if we solve intelligence, we can solve diseases, we can solve aging, we can solve perhaps interstellar travel.

1:24Jonathan, welcome to Sorcery. Thank you for having me. Well, Well, we met just a few months ago in Paris at the RAISE Summit. This was actually at Apoorv from Altimeter. Thank you, Apoorv. His event outside of the summit, and you were one of the speakers on stage. What were you guys talking about? We were talking about training superintelligence. The race to AGI is on, and the data needs of these models have shifted. These models need incredibly complex data for training them to get better at coding, advanced reasoning, STEM, etc. And we were chatting about how the world looks so different today than it did like a few years back.

2:09So I remembered that speech. It wasn't a speech. I remember that panel. And I was really taken back because I've heard of Scale. I've heard of Mercore. I've heard of all these different companies in the AI research space. it's definitely evolved from just where it was before with data labeling to something a little bit more advanced as AI continues to evolve but there was one thing that I thought was quite interesting and I didn't know this but like the scale at which these companies are running specifically yours and I tweeted this out and it kind of went viral but I tweeted out just got back from Ray's Summit in Paris where AI's top names like Grock, Cerebris, Eric Schmidt, Lovable and Windsurf this was like right before the Windsurf thing by the way I don't know if you remember that All took the stage.

2:55Yeah, one unexpected player stood out as the next breakout AGI leader. Some stats. $300 million plus in revenue and profitable. $225 million in funding at a last valuation of$2.2 billion. A network of 4 million engineers and customers that include OpenAI, NVIDIA, Anthropic, Google, Microsoft, Meta, Salesforce, Amazon, and I'm sure a lot more. So Jonathan, what is Turing? So at Turing, we train superintelligence. We work with seven out of the eight Frontier Labs. We work with OpenAI, Anthropic, Meta, Google, Microsoft, NVIDIA, Amazon. Anybody that's building a Frontier Foundation model, we are probably working with them already.

3:46And what's happening is as these models have become smarter and smarter, the data needed to improve them has become increasingly harder to generate. The earlier era was almost like commodity data labeling. That era is over. Now it's all about frontier data. These models need expert human data in every domain imaginable. They need data to train these models in reinforcement learning. and sometimes they need synthetic data. What we are doing is scaling up our data infrastructure at gigantic scale. So for these models, the cool thing is everybody's aware of all the research breakthroughs that are needed to move these models forward and all the labs do an amazing job at advancing the research frontier.

4:35You need tons and tons of compute and we have NVIDIA, Cerebrus, Grok, like all these companies to thank for that. but they need a ginormous amount of data. And what's happened, Molly, is these models ate the internet when they were pre-trained, but the internet data is used up. It was used up like three years ago, right? Where's the data gonna come from to keep the scaling laws going? You need to have expert humans from every domain imaginable to work on a platform like Turing to break the models and figure out what type of incremental data will move them forward. The era of sweatshop data labeling is over.

5:19Now what the labs need is a strategic research accelerator, and Turing's the world's leading research accelerator, working with all of these frontier labs to advance them along, I would say, the four pillars of superintelligence, which is multimodality, reasoning, tool use, and coding. If you solve these four things, you will get to ASI. What's ASI? Artificial superintelligence. I think of ASI, and different people have like different definitions, right? I mean, AGI was also very ambiguously defined. And I think of ASI as us having automated 90 % of the tasks that 90 % of humans do today in front of a computer, right?

6:05If you look at most types of knowledge work, like my job, your job, you're looking in front of a computer, analyzing what's on the screen, using different tools, and taking decisions. And this is the job of a CEO. It's the job of a head of engineering, a director of performance marketing, an SDR. These types of tasks are going to be automated. The only constraint, the reason this hasn't happened yet, is the data that has to improve these models hasn't gone into them yet. And we are scaling like crazy. I think of this like a four-dimensional matrix. Think of every industry you can imagine. Financial services, retail, healthcare, life sciences.

6:47That's dimension one. Next dimension, think of every function in that industry. Software engineering, sales, marketing, etc. Third dimension, think of every role in the org chart in that function. Let's pick marketing. There's a CMO, there's a director of performance marketing, there's an SEO analyst. And fourth dimension, think of the workflow that a human performing that role goes through. Let's say we pick a director of performance marketing, and they have to figure out how to allocate marketing budget across different channels, solving for CAC, LTV, yield of the efficiency of the channel, scalability of the channel, etc.

7:24This requires that human to do data analysis, to do A-B tests, use tools like Facebook's performance marketing dashboard, LinkedIn's performance marketing dashboard, analyze which campaigns do well, and based on that, like update the website funnel, right? All of this is going to be automated, right? This is like, if you look at those four-dimensional, if you look at that four-dimensional matrix, that's$30 trillion of knowledge work that is going to be automated. And these models will improve to automate all of this as long as we scale up compute and we scale up data. There is this joke among AI researchers.

8:01Researcher one says, hey, is it true that LLMs cannot reason well out of distribution? Meaning are LLMs just interpolating based on data in the training set? How will they generalize? And researcher two says, then you bring all data into distribution. like all of humanity's human knowledge, bring it into distribution. Seven billion humans on the planet doing all sorts of interesting things. I think most of them should be on platforms like Turing generating data to train these models. So humans are a lot more leveraged in what they do. Sorcery is brought to you by Brex, the financial stack trusted by more than 30 ,000 companies, including one in three venture-backed startups in the U.S.

8:44Nearly 40 % of startups fail because they run out of cash. Rex is literally built to help founders avoid that. Unlike traditional banks that let your money sit idle, chipping away at it with fees, Rex's designs help you spend smarter and move faster. Their all-in-one solution combines checking, treasury, and FDIC protection into one powerful account. You can send and receive money globally at lightning speeds, get 20 times the standard FDIC coverage through their partner banks, and even high yield from day one. with same day and even same hour liquidity. Access your funds anytime. Companies like Scale AI, DoorDash, Service Titan, HIMSS, Anthropic, Flexport, Robinhood, and Plaid trust and use Brex.

9:29Start today at brex.com slash sorcery. That's B-R-E-X dot com slash sorcery. I kind of think that's why Nat Friedman started to, he started to translate all those scrolls. he's starting that was one of his projects he's translating scrolls because we ran out of data we ran out of human knowledge and now we have to go back in time and decipher these things so right now there's a lot of competition in the space and it seems like the hottest thing to look at ever since the scale ai acquisition investment whatever you want to call it there's essentially a billion dollars up for grabs. And I want to know, between the three leaders, of which you are one of them, how do you differentiate?

10:21This market is massive and growing. There's unlimited demand for high quality data, unlimited demand. I see two axes that you have to be really good at to win in this market. One axis is finding really smart humans really, really fast. And there are a few companies that do a good job at that, Turing being one of them. There's another axis, which is you have to be at the forefront of data research. You have to know what type of data is likely to be helpful to the model for it to advance encoding, STEM, advanced reasoning, et cetera. Turing's great at that. We are the one company in this category that's not just good at finding people and generating data, but being proactive about figuring out what type of data these models are likely to need.

11:10And as the models get getting smarter and smarter every quarter, you have to innovate on the research side too. So on that two by two, I would say Turing's the one company up and to the right. There are plenty of companies that are good at, okay, we have a large talent network, we work with university students, or we're good at finding people. We think that's necessary, that's not sufficient. You have to ensure that the quality of the data is good and that it can improve the model's performance. Taking a step back, can you explain the evolution of this industry and models? You gave a really good explanation on This Week in Startups with Alex and I think this was like phenomenal, like it was clipped and everything.

11:58Could you explain the evolution of that? this industry had a huge shift after the reasoning models came out late last year or one being the first the shift is before the reasoning models this industry needed simple data gobs and gobs of simple data you needed a data factory or this industry needed somebody to find people really fast after the reasoning models came out the game completely changed this is like after O1 and DeepSeek. Now what the labs need is not a data factory. It's not a talent marketplace. The labs need a strategic research partner, somebody who can collaborate directly with their researchers to understand where the models are weakened today and strategically custom engineer data to improve the model's performance.

12:49They need a research accelerator. In-house researchers collaborate with teams in coding, multimodality, STEM, RL gyms, etc., to generate data that will improve these models. And the data that the models need now needs to be hard, that is model breaking. It has to literally break the model, so you need humans who are smarter than the models. It has to be realistic, that reflects how real humans use these models to do real work. Only then AI will actually move the GDP of the world. It has to mirror real world use, not esoteric academic use cases that test whether you've hit the singularity or not.

13:28And third, the data needs to be diverse to cover every single type of knowledge work imaginable. So you need hard data, realistic data, and diverse data. That's how it's shifted. Because research has become such an important part of this next era of AI. it would be wrong not to bring up the fact that this is your background. So you were in research and in AI development before this. So how did that lead to founding Turing? Turing's founding DNA is AI research. My co-founder and I met at Stanford. I worked at the Stanford AI lab and the Stanford info lab. My co-founder worked at the Stanford NLP lab.

14:13And in an alternate universe, Both of us would have gotten our PhD in computer science and AI. And at Stanford, there's this radioactive spider that tends to bite people and makes them start companies instead. And that's what happened to me and my co-founder. And so we saw the power of AI early. And when Turing was started, it was all about using AI to find the world's smartest software engineers at scale, use AI to vet them, match them to companies, use AI to manage them. So when the AGI wave hit, I remember that fateful meeting with OpenAI when they were training GPT-3, and OpenAI wanted to teach GPT-3 to code and to do function calling and tool use, which are building blocks for training agents.

15:03When we saw it, it was obvious what this was going to become. This was the future. And we were perfectly positioned. We had the world's largest platform of software engineers, where we could build the ability to generate high quality data from. And it was clear that the models needed not just coding data, but data in every advanced domain in STEM, in healthcare, legal, finance, etc. And that really helped us, the fact that our DNA was AI research. And now we are publishing papers on what type of data is most likely to help these models advance in coding, in chip design, in multimodality, and coming up with really hard evals for these models.

15:51On the research side, how are you evaluating researchers and how are you recruiting them? On the research side, our focus is on two things. Number one, ensuring that our data is the highest quality in advancing the models along the four pillars of superintelligence, multimodality, reasoning, tool use, and coding. And number two, we do research on what type of data the models might likely need in the future. That's why we were on RL gyms very, very early. We built thousands of these reinforcement learning gyms to train agents. We were able to do that, again, because we have an internal R &D team that's constantly going to AI conferences, understanding where the frontier is, publishing papers.

16:43That's been really key. So the focus is on ensuring that the data that we generate is high quality. I think of that as horizon one. That's the here and now. and horizon two is what type of data will these models need three months in the future. That's horizon two. And today, a lot of that work is on embodied AI and robotics and coming up with benchmarks that the models would struggle at today. Like we are creating data for coding now, for agentic coding, that would stump all of today's models and agents built on top of those models. It gives us a little bit of a sadistic pleasure in flunking all of the models today and a little bit of like demonstrating where the human intelligence frontier is.

17:27But I would give that maybe six months before the models climb that hill and then we'll generate data that's even harder for the models. Could you just explain like maybe step-by-step for people that don't understand how models are created, the partners in which they work with and then also the teams they have internally that make all these processes run? Because I don't think many people know like what does the inside of OpenAI look like? What does it take to get to GPT-5? What is that process? Yeah. Let me talk the general recipe of training a frontier model, right? And a frontier model is a trillion parameter model, like let's say a GPT-4 class model, like the general recipe.

18:10This is roughly what all the labs follow, right? So there are two main steps. There's pre-training and post-training. Pre-training is a form of unsupervised learning, meaning humans are not involved. It's like lots and lots of compute. In pre-training, you build what's called a base model. And all the base model knows is how to autocomplete stuff. Pre-training is done completely inside the frontier labs. And during pre-training, you feed the models a subset of the internet. It's a filtered subset of the internet. There are these datasets like Common Crawl, C4, GitHub, Archive. You also feed a lot of books.

18:53You might feed videos. You feed all these, let's call it humanity's knowledge corpus. You feed it to these models. And the model is basically doing the task of language modeling, which is learning how to predict the next token. This is a model that just knows how to auto-complete stuff. An example of a pre-trained model is something like GPT-3. We haven't yet built chat GPT. This is just GPT-3, right, as the result of pre-training. To this model, if you say something like, Molly runs Sorcery, it'll then respond by saying, Molly is interviewing Jonathan for the next episode. That's how the model will respond.

19:37It's just auto-completing. It's not an assistant yet. But the magical thing about pre-training is in the quest to autocomplete stuff, the model has learned a lot of internal representations about the world. It's like raw brain mass. It's like an artificial brain that hasn't been molded yet. But the intelligence is there, right? The next step is post-training. And OpenAI came up with this paper called Instruct GPT, which lays out this process in more detail. I encourage your readers to listen to it, to read the paper. In post-training, this is when GPT-3 becomes chat GPT. You teach this base model to become a helpful, honest, harmless assistant.

20:23And post-training has two main steps, supervised fine-tuning and reinforcement learning. In supervised fine-tuning, you hire lots and lots of human experts to create prompts and completions that teach the model how to answer questions. The task is still language modeling. You're still predicting the next token, but you feed it question-answer pairs that could either be single-turn, meaning question and answer, or multi-turn, where you're having a dialogue with the model. At this step, if you ask the model, let's say a human contractor has the question, Molly runs sorcery, the next completion for it might be, what would you like to know about Molly?

21:05might be an appropriate response that you teach the model to do. You teach the model how to respond. In the case of coding, a contractor might create a question like, hey, give me some Python code to train a text classifier to detect X posts into positive sentiment, neutral sentiment, negative sentiment. And then the contractor actually writes code. And let's imagine the code had some type of exception or some type of bug. the contractor would give the model a demonstration of what the right answer is. It's humans showing the model question-answer pairs, so the model learns. Again, this is still language modeling.

21:46The model is still learning how to predict, but now you've aligned the model. If you ask the model a dangerous question, and there are these areas in the community, it's called CBRN, Chemical, Biological, Radioactive, Nuclear. There are categories of things that for safety, you might want to train the model to decline to answer. That gets done in this step too, this supervised fine-tuning step. The step after that is reinforcement learning with human feedback, where when the model has gotten slightly better at responses, you'd have the model output two responses, two completions, and you'd have human contractors pick what is a completion that a human tends to prefer.

22:28For example, if the question is, tell me more about Maliché and Sorcery, and there's one set of responses that is maybe super verbose and is not easy to consume, whereas there's another response that's like in crisp bullet points, and maybe that links to the conversation we are having. Maybe a human contractor maybe picks that. Oh, that response is better. This RLHF step is used to train a reward model. and the reward model basically scores responses based on what's a response that a human tends to prefer. And this is used to update the model that has been trained so far to output completions that humans tend to prefer.

23:18For example, in coding, humans may tend to prefer code that's commented really well, that's structured in a way that is easy to maintain and is more modular. But those preferences are learned in RLHF. And once you've done RLHF, you would then do another round of evals where you're evaluating the model's performance to see if the model improved. And as the model improves, you'll figure out where the model breaks next and then generate data to improve the model there again. It's a loop between you first do pre-training. Pre-training is done relatively infrequently. And then post-training is a highly iterative process.

24:03You start first with SFT. Then you would do reinforcement learning with human feedback. You'd evaluate the model again, figure out where it's breaking, where it's doing well, and then generate data in the areas where the model is weakened. Now, what I just described is how things used to largely work until last year. Now, for these verifiable domains like coding and math, instead of doing reinforcement learning with human feedback, you can do reinforcement learning because you can automatically check when you got the correct answer or not in these verifiable domains. And with reinforcement learning, the way you teach the model is not through imitation learning, where you're learning from humans, but it's through experiential learning, where you create this environment with prompts and verifiers, and you create basically a mini world model, where as long as the agent gets the right result, that particular trajectory that the agent took to come to the right answer gets reinforced.

25:07And when the agent got to a result that didn't get the right answer, that doesn't get reinforced. So the model kind of learns from its own experience. It's a form of self-play, which I think is really, really cool. If you remember from the Alpha Zero days when DeepMind was teaching their Go playing algorithm, the first version of Google's AlphaGo system learned from expert humans, which was quite good. The next generation, which was even better, just played against itself and figured out which paths tended to it winning and which paths tended to it losing and learned from that. And there is this Move 37 in one of the games where the model came up with a move that felt alien to the best Go players in the world because the model had figured out a trajectory to win that beat sort of the best human experts.

26:06And I'm super excited for AI in the enterprise when an agent takes a path that is so novel and so interesting and so different that really exceeds superintelligence. And for me, that's a big part of artificial superintelligence. Why should human intelligence be the bound for what we do? These models could exceed us in many interesting ways. In today's high-speed business world, staying ahead means using the smartest tools possible, including the powerful capabilities of artificial intelligence. Meet Turing Intelligence. Turing builds customizable AI systems designed to solve your mission-critical challenges, no matter your industry.

26:45From expert guidance to tailored projects, Turing helps top companies realize AI that's more capable, more adaptable, and more effective. With Turing, discover how AI can accelerate your business growth. To learn more, visit Turing.com slash sorcery, spelt S-O-U-R-C-E-R-Y. That's Turing.com slash sorcery. I think that's a really interesting point, and it's also a point that scares people a lot. It's created a culture of doomers because they're afraid of AI and technology rapidly advancing so fast or so quickly out of our control that it puts safety at a risk. um I'm gonna put that to the side like quite like a little bit but it brings it brings into light this tweet that David Sachs had not too long ago like right after the GPT-5 release and there was a lot of pushback on this release it didn't go as expected many um were actually like quite disappointed they were expecting like something crazy to happen and turns out it was just a little bit of an incremental, seemingly incremental, I don't want to offend anyone, seemingly incremental improvement to the model.

28:02But I'm curious from your standpoint, what did you make of that moment in that tweet that David Sachs had put out that was, you know, it was addressing the doomers of, okay, actually, like, AI is not moving that fast, we actually might have more time before, where we have real crazy moves happening. Yeah. Let me unpack this. I think there are two parts to your question. There's one part, which is, are we slowing down? Like, are things going to be more incremental going on? That's one part to your question. And then there's another part, which is, how should we think about safety in these AI systems?

28:43So let me take that separately. first I think GPT-5 is fucking awesome I think it's awesome I think we've just gotten used to magic that maybe we are not fully seeing the impact of of what's just out there and I'm not just saying that because we we work with the company that built it I think GPT-5 is awesome I think one thing is real though which is the AI doomers talk about this concept of rapid takeoff it's sort of the I mean if you've seen the matrix or the terminator movies like somehow we'll hit a singularity and there's just this rapid takeoff where humanity gets left behind I don't think rapid takeoff is how things will unfold I think it's going to be steady continuous progress every step of the way we're going to keep moving forward and it's going to be great and it's great for a few reasons the reason I think rapid takeoff will not happen and I know why the doomers believed in rapid takeoff is because GPT-3 felt like magic, right?

29:47It was like fucking magic. It was magic because there were a lot of one-time jumps in there. Jump number one on the data front, we ate the internet. We ate the internet. There is only one internet and we ate it in GPT-3, right? And GPT-4, again, in all of these, and all the models like GPT-4, like Gemini 2.5 Pro, Grob 2, Grob 3, Claude. It's the same formula. We all ate the internet, and that was a huge one-time jump. And we put in all of humanity's knowledge. All the books that you own the licenses off were fed into the models, right? So humanity's collective intelligence was eaten in the GPT-3, GPT-4 era.

30:33And we scaled up compute. Some of the early gains in compute where we also got that, right? And computers takes time to build. You got to solve energy and you got to build data centers and you see what Elon's doing with his projects. You see what Sam Altman and OpenAI are doing. You see what Meta's doing. Compute also takes time to build, right? So we got a huge jump by eating the internet and using all the easily accessible compute. But the scaling laws, which is an empirical law, is continuing to hold, meaning you take a big model with lots of parameters, you dial up compute, you dial up data, and the models smoothly keep improving, right?