Inside The $2.2B AI Research Accelerator | Turing

AI has eaten the internet, data labeling is so over, and $30 trillion of human work is on the verge of automation. Jonathan Siddharth https://www.linkedin.com/in/jonsid/⁠, Founder & CEO of Turing, joins Sourcery to break down the power shift in AI training — from commodity data labeling to expert research — positioning Turing apart from AI data providers like Scale AI, Mercor, & Surge. Turing has become a hidden force in the AI race, hitting $300M in ARR in 2024 (~3x YoY), achieving profitability, and raising $111M at a $2.2B valuation in March. That growth cements its position as one of the fastest-growing AGI infrastructure companies. Today, frontier labs like OpenAI, Anthropic, Meta, Google, Microsoft, Nvidia, & Amazon rely on Turing for the frontier data that pushes AI forward across the four pillars of superintelligence: • Multimodality • Reasoning • Tool use • Coding We explore Turing’s expansion into the enterprise, closing the “gap” – where Fortune 500s in finance, insurance, and pharma are racing to build proprietary intelligence on their own data, creating durable moats in the $30T knowledge work economy. PS Jonathan also explains how labs like OpenAI train models: • Pre-training on filtered internet corpora (Common Crawl, GitHub, books, video) • Post-training with supervised fine-tuning (human Q&A datasets) • Reinforcement learning (RLHF + verifiable domains) to align models with human preferences • Model-breaking data from Turing’s 4M+ engineers to close gaps and advance systems like GPT-5 1. Jonathan Siddharth: https://www.linkedin.com/in/jonsid/ 2. Molly O’Shea: ⁠https://x.com/MollySOShea⁠ 3. Sourcery: ⁠https://x.com/sourceryvc⁠ Brought to you by: • Brex https://brex.com/sourcery—The modern finance platform, combining the world’s smartest corporate card with integrated expense management, banking, bill pay, & travel. As a Sourcery Listener you get: 75,000 points after spending $3,000 on Brex card(s), white-glove onboarding, $5,000 in AWS credits, $2,500 in OpenAI credits, & access to $180k+ in SaaS discounts. On top of $500 toward Brex travel, $300 in cashback, plus exclusive perks (like billboards..) visit → https://brex.com/sourcery • Turing https://turing.com/sourcery—Turing delivers top-tier talent, data, and tools to help AI labs improve model performance—and enables enterprises to turn those models into powerful, production-ready systems. Visit: https://turing.com/sourcery • Carta https://carta.com/sourcery—Carta connects founders, investors, and limited partners through software purpose-built for private capital. Trusted by 65,000+ companies in 160+ countries, Carta’s platform of software & services lays the groundwork so you can build, invest, and scale with confidence. Visit: https://carta.com/sourcery • Kalshi https://kalshi.com/sourcery—The largest prediction market and the only legal platform in the US where people can trade directly on the outcomes of future events: https://kalshi.com/sourcery Follow Sourcery for the latest updates! https://www.sourcery.vc/ (00:00) AI Ate The Internet (00:49) Training superintelligence: the race to AGI (02:31) Viral tweet (03:24) What Turing actually does (04:43) The internet data is “used up” — where will new data come from? (05:34) Four pillars of superintelligence: multimodality, reasoning, tool use, coding (06:07) Automating $30T of global knowledge work (09:18) The $1B revenue opportunity (10:59) Why Turing is a research-first accelerator, not a data labeler (13:45) Jonathan’s Stanford AI Lab roots and founding DNA (17:57) How models are built: pre-training vs. post-training (20:14) RLHF, reinforcement learning, and “breaking the models” (25:19) GPT-5 and the myth of rapid takeoff (30:46) Safety debates and human-in-the-loop systems (34:53) Closing Enterprise Gap: finance, insurance, & pharma (39:23) Why proprietary enterprise data is the next moat in AI

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Molly O'Shea?

  300 million plus in revenue and profitable, 225 million in funding out of last valuation of 2.2 billion, and customers that include OpenAI, Nvidia, Anthropic, Google, Microsoft, Meta, Salesforce, Amazon, and I'm sure a lot more. Jonathan, what is Turing?

• 0:19Speaker 01

  The era of sweatshop data labeling is over. Now what the labs needed is a strategic research accelerator, attaining to the world's leading research accelerator, working with all of these frontier labs to advance them along the four pillars of superintelligence: multimodality, reasoning, tool use, and coding. That's $30 trillion of knowledge work that is going to be automated.

• 0:39Molly O'Shea?

  Ever since the Scale AI acquisition, investment, whatever you want to call it, there's essentially a billion dollars up for grabs.

• 0:46Speaker 01

  This market is massive and growing. There's unlimited demand for high-quality data.

• 0:51Molly O'Shea?

  What does the inside of OpenAI look like? What does it take to get to GPT-5?

• 0:57Speaker 01

  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.

• 1:25Molly O'Shea?

  Jonathan, welcome to Sorcery.

• 1:27Speaker 01

  Thank you for having me.

• 1:29Molly O'Shea?

  Well, we met just a few months ago in Paris at the RAI's summit. This was actually at Apur from Ultimater. Thank you, Apur. His event outside of the summit, and you were one of the speakers on stage. What were you guys talking about?

• 1:45Speaker 01

  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.

• 2:09Molly O'Shea?

  So 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 Merckor, I've heard of all these different companies in the AI research space. Um, it's definitely evolved from just where it was before with data labeling to something a little bit more advanced as AI uh 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 Grok, Cerebrus, Eric Schmidt, Lovable and Windsurf, this was like right before the Windsurf thing, by the way, I don't know if you remember that.

• 2:53Molly O'Shea?

  All took the stage. Yet one unexpected player stood out as the next breakout AGI leader. Some stats, three hundred million plus in revenue and profitable. Two hundred and twenty-five million in funding at a last valuation of two point two billion. A network of four 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?

• 3:28Speaker 01

  So at Turing, we train superintelligence. Uh, we work with seven out of the eight frontier labs. Uh, 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. Uh, and what's happening uh is as these models have become smarter and smarter, the data needed to improve them has become increasingly harder to generate. Uh, the earlier era was um 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.

• 4:11Speaker 01

  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 you, 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. You need tons and tons of compute and we have Nvidia, Cerebras, Grok, like all these companies to thank for that, but they need a ginormous amount of data. And what's happened, Molly, is these, these models ate the internet when they were pre-trained, but the internet data is used up.

• 4:56Speaker 01

  It was used up like three years ago, right? Where's the data gonna come from to keep the scaling loss 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. Now 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.

• 5:39Speaker 01

  If you solve these four things, you will get to ASI.

• 5:43Molly O'Shea?

  What's ASI?

• 5:45Speaker 01

  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? If 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.

• 6:25Speaker 01

  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 I think of this like a four dimensional matrix. Think of every industry you can imagine. Financial services, retail, healthcare, life sciences, that'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. Uh, let's pick marketing, there's a CMO, there's a director of performance marketing, there is an SEO analyst. And fourth dimension, think of the workflow that a human performing that role goes through.

• 7:08Speaker 01

  Let's say we we pick a director of performance marketing. And they have to figure out how to allocate marketing budget across different channels. Solving for uh CAC, LTV, uh yield of the efficiency of the channel, scalability of the channel, etc. Right? Uh, this requires that human to do data analysis, uh to do A B tests. Uh, 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 uh funnel, right? All of this is going to be automated, right? This is like, if you look at those four dimensional ma- if you look at that four dimensional matrix, That's thirty trillion dollars of knowledge work that is going to be automated.

• 7:50Speaker 01

  And these models will improve to automate all of this. As long as we scale up compute and we scale up data. Uh, there is this joke among AI researchers. A researcher one says, hey, is it true that uh L L M's cannot uh reason well out of distribution? Meaning are L L M's uh just interpolating based on data and the training set? How will they generalize? Uh, 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. Uh, so humans are a lot more. What they do.

• 8:35Molly O'Shea?

  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 US. Nearly 40% of startups fail because they run out of cash. Brex is literally built to help founders avoid that. Unlike traditional banks that let your money sit idle, chipping away at it with fees, Brex'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.

• 9:21Molly O'Shea?

  Companies like Scale AI, DoorDash, ServiceTitan, Hims, Anthropic, Flexport, Robinhood, and Plaid, trust and use Brex. Start today at brex.com/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 started, 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.

• 9:56Molly O'Shea?

  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,

• 10:20Molly O'Shea?

  how do you differentiate?

• 10:21Speaker 01

  This market is massive and growing. There's unlimited demand for high-quality data, unlimited demand.

• 10:29Speaker 01

  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, right? 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. And as the models get getting smarter and smarter every quarter,

• 11:14Speaker 01

  you have to innovate on the research side, too. So on that two by two, I would say Turing's the one company up into 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 are good at finding people. Uh, we think that's like a, that's necessary, that's not sufficient. Uh, you have to ensure that the quality of the data is good and that it can improve the model's performance.

• 11:41Molly O'Shea?

  Taking a step back, can you explain the evolution of this industry and models? You gave a really good explanation on this weekend startups with Alex, and I think this was like phenomenal, like it was clipped and everything. Could you explain the evolution of that?

• 12:01Speaker 01

  This industry had a huge shift after the reasoning models came out late last year, oh, 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 O one 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.

• 12:43Speaker 01

  And strategically custom engineer data to improve the model's performance. They need a research accelerator. In-house researchers collaborate with teams in coding, multimodalities, STEM, RLGems, et cetera, to generate data that'll 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.

• 13:28Speaker 01

  And third, the data needs to be diverse to covers every single type of knowledge work imaginable. So you need hard data, realistic data, and diverse data. That's how it shifted.

• 13:41Molly O'Shea?

  Because research has become such an important part of this next era of AI,

• 13:49Molly O'Shea?

  it would be, 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?

• 14:02Speaker 01

  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. And in an alternate universe, both of us would have gotten our PhD in computer science and AI. And Stanford, at Stanford, there's this radioactive spider that tends to bite people and makes them start companies instead. And that's what happened, happened to me and my co-founder. Um, and uh, so we saw the power of AI early.

• 14:33Molly O'Shea?

  Um hum.

• 14:33Speaker 01

  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, right? So when um, the AGI wave hit, uh, or I I remember that fateful meeting with OpenAI when they were training GPT three. And OpenAI wanted to teach GPT three to code and to do function calling and tool use, which are building blocks for training agents. It, when we saw it, it was obvious what this was gonna become, right? This was the future. And we were perfectly positioned. We had the world's largest platform of software engineers. Where we could build uh, the ability to generate high quality data from.

• 15:18Speaker 01

  And it was clear that uh, the models needed not just coding data, but data in every advanced domain in STEM. In healthcare, legal, finance, etc. Um, and that uh, that really helped us, the fact that our DNA was AI research. Um, and now we are publishing papers on what type of data is most likely to help these models advance. In coding, in um, chip design, uh, in multimodality. Uh, and coming up with really hard um, evals for these models.

• 15:52Molly O'Shea?

  On the research side, how are you evaluating researchers and how are you recruiting them?

• 15:57Speaker 01

  Uh, on the research side, uh, our focus is on two things. Uh, 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, uh, 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.

• 16:25Molly O'Shea?

  Um hum.

• 16:25Speaker 01

  We've built thousands of uh, these reinforcement learning gyms to train agents. We were able to do that, again, because we have an internal R and D team that's constantly uh, uh, going to. AI conferences, understanding where the frontier is, publishing papers, that'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 and coming up with benchmarks that the models would struggle at today.

• 17:08Speaker 01

  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. But I would give give that maybe six months before the models climb that hill. And then we'll generate data that's even harder for the models.

• 17:33Molly O'Shea?

  Could you just explain like maybe step by step for people that don't understand how models are created?

• 17:41Molly O'Shea?

  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 what does the inside of OpenAI look like? What does it take to get to GPT five? What is that process?

• 17:58Speaker 01

  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 four class model. Like the general recipe, this is roughly what all the all the labs follow, right? So there are two large, 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 the base model. And all the base model knows is how to autocomplete stuff. Pre-training is done completely inside the frontier labs.

• 18:38Speaker 01

  And during pre-training, you feed the models a subset of the internet. It's a filtered subset of the internet. There are these data sets like Common Crawl, C4, GitHub, Archive. You also feed a lot of books. You might you might feed videos. You you feed all these, let's call it humanities, knowledge corpus. You feed it to these models. Uh, and the model is basically doing the task of language modeling, which is learning how to predict the next token, right? Uh, this is a model that just knows how to autocomplete stuff. An example of a pre-trained model is something like G P T three. We haven't yet built chat G P T, this is just G P T three, right? At the, as the result of pre-training.

• 19:24Speaker 01

  To this model, if you say something like, uh, Molly runs sorcery, it'll then respond by saying, uh, Molly is, uh, interviewing Jonathan for the next episode. That's how the model will respond. It'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 G P T,

• 20:04Speaker 01

  uh, which lays out this, uh, process in more detail. I'd encourage your readers to listen to it, to read the paper.

• 20:12Speaker 01

  In post-training, this is when G P T three becomes chat G P T. You teach this base model to become a helpful, honest, harmless assistant. And 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, uh, that teach the model how to answer questions. Uh, 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.

• 20:52Speaker 01

  At this step, if you ask the model, uh, uh, 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? Might be the next, might be an appropriate response that you teach the model to do. You teach the model how to respond, right? In the case of coding, a contractor might create a question like, hey, help me, 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.

• 21:29Speaker 01

  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, right? It's humans showing the model question-answer pairs, so the model learns. Again, this is still language modeling. The model is still learning how to predict, but now you've aligned the model. If you ask the model a dangerous question, and there is this, 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 wanna train the model to decline to answer, right? That gets done in this step two, this supervised fine-tuning step.

• 22:11Speaker 01

  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. For example, if the question is, tell me more about Molly Shea and sorcery, and there's one set of responses that is maybe super verbose and has, is not easy to consume, whereas there's another response that's like in crisp bullet points, and maybe that hasn't, that links to the conversation we are having, maybe a human contractor maybe picks that, oh, that, that response is better.

• 22:53Speaker 01

  This RLHF step is used to train a reward model, and the reward model teaches, 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. Humans may, for 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 is more modular. But those preferences are learnt in RLHF.

• 23:33Speaker 01

  And once you, 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. Right. So it's, and as the model improves, you'll you'll figure out where the next, where the model breaks next, and then generate data to improve the model there again. It's a loop between, you you first do pre-training, pre-training is done relatively infrequently. And then post-training is a highly iterative process. You 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 the model is weakened.

• 24:16Speaker 01

  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 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,

• 25:02Speaker 01

  that particular trajectory that the agent took to come to the right answer gets reinforced. And 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 AlphaZero days, when DeepMind was teaching.

• 25:27Speaker 01

  Their Go playing algorithm. The first version of uh uh Google's uh AlphaGo system uh learned from expert humans, which was quite good. The next generation, which was even better, just played against itself uh and figured out which paths tended to it winning and which paths tended to it losing and learned from that. And there is this uh uh move thirty seven in one of the games uh 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 uh sort of the best human experts. And I'm super excited for AI in the enterprise.

• 26:11Speaker 01

  When an agent takes a path that is so novel and so interesting and so different uh that really exceeds superintelligence. Uh 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 uh exceed us in many interesting ways.

• 26:30Molly O'Shea?

  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. From 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/sorcery, spelt S-O-U-R-C-E-R-Y. That's Turing.com/sorcery. I think that's a really interesting point, and it's also a point that scares people a lot.

• 27:12Molly O'Shea?

  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 Zacks had not too long ago, like right after the GPT. G P T five release and there's 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 um seemingly incremental.

• 27:57Molly O'Shea?

  I don't wanna offend anyone, seemingly incremental improvement to the model. Um but I'm curious from your standpoint, what did you make of that moment in that tweet that David Sacks 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, you know, we have real crazy moves happening.

• 28:23Speaker 01

  Yeah, let me unpack this. I think there are there are two parts to your question. There's one part which is, are we slowing down? Um like, are things getting gonna be more incremental going on? That's that's one part to your to your question. And then there's another part, uh which is, uh how should we think about safety uh in these AI systems? So let me uh take that separately. Uh first, I think G P T five is fucking awesome. I think it's awesome. I think we've just gotten used to magic. Uh that maybe we are not fully seeing the impact uh of uh of what's just out there. And I'm not just saying that because uh we we work with the company that that built it. I think G P T five is awesome.

• 29:07Speaker 01

  I think one thing is real, though, which is the AI doomers talk about this concept of rapid takeoff. Um 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 gonna be steady, continuous progress. Every step of the way, we're gonna keep moving forward. Um and it's gonna be great, and it's great for a few reasons. Uh the reason I think rapid takeoff will not happen, and I know why the doomers believed in rapid takeoff, is because G P T three felt like magic, right? It was like fucking magic.

• 29:48Speaker 01

  It was magic because there were a lot of one-time jumps in there. Num- 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 G P T three, right? Uh, and G P T four, again, uh, in all of these and all the models like G P T four, like Gemini two point five pro, Grok two, Grok three, uh, Claude. It's the same formula. We all ate the internet and that was a one, huge one time jump. And we, uh, we put in all of humanity's knowledge, all the books that you own, the licenses of were fed into the models, right? So humanity's collective intelligence was eaten in, in, in that, the, in the G P T three, G P T four era.

• 30:33Speaker 01

  And we, um, scaled up compute. Some of the early gains in compute where we also got that, right? And computers takes time to build. You gotta solve energy and you gotta 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 is 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?

• 31:15Speaker 01

  I'm a huge optimist in humans and human potential. I think we humans will solve bigger, more interesting problems instead of the more mundane stuff that we do today. I believe I will be a hundred X more productive. If I'm a hundred X more productive, maybe I'll run a hundred companies in parallel. Elon might run six hundred companies in parallel, right? So I think the solution is simple.

• 31:39Speaker 01

  Keep hiring the best researchers on the planet, keep scaling up compute, keep scaling up data. We will get to A S I 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, right? It's gonna be, it's gonna be awesome.

• 32:01Molly O'Shea?

  On the safety side, I think the narrative has changed a little bit. Cal she actually has a market out on this. Of will models pause research for safety reasons and?

• 32:16Molly O'Shea?

  This goes until twenty twenty seven. And so right now, it's at about nine percent chance. But it was trending upwards of like forty, not too long ago. It's really shifted and it's taken a nosedive in the last like week or so. So why do you think that is? Why do you think the narrative on safety is changing?

• 32:35Speaker 01

  People have realized that the best way to deploy these systems is with humans in the loop. And with humans in the loop, a lot of the safety issues go away. Um, you have to design these systems for partial autonomy. Uh, people's models for mental models, when they think about AI, is self-driving cars. I feel like the self-driving car industry did a lot of damage to AI in one particular way, which is it, uh, because with self-driving cars, ninety-nine percent accuracy is not enough. That that last one percent really matters. Uh, but with a lot of AI systems, if you're automating the job of an investment analyst, or if you're automating the job of a CEO or a CFO,

• 33:19Speaker 01

  uh, you don't have to be, uh, a hundred percent, um, accurate, as long as there is a human overseeing the results of the model. Um, so I think, uh, we know how to build good safety harnesses around these systems, uh, with, uh, a human verifying the results of the model. Uh, so I think a lot of the risks, many of the risks are, were somewhat exaggerated. Uh, and we have good techniques to address safety in the post-training, uh, phase. Um, you can teach the models to decline to answer questions that could be dangerous in the CBRN, um, side of the, side of the, uh, space, uh, chemical, biological, radioactive, nuclear.

• 34:03Speaker 01

  So we have some good techniques to, um, uh, contain the negative behaviors of these models. Uh, safety is important, but I think it was, uh, the, some of the risks were over-exaggerated, uh, and they could be solved with really good engineering. I think what's the bigger risk is we don't deploy these systems quickly enough.

• 34:24Molly O'Shea?

  Hmm.

• 34:24Speaker 01

  Um, I mean, there was that MIT study that showed how ninety-five percent of Gen AI pilots fail. I'm not surprised by that result at all. Um, they fail because oftentimes, uh, firstly, there is a big data gap.

• 34:40Speaker 01

  To solve real world enterprise workflows, you need a real world enterprise data. And that data doesn't exist on the internet. Uh, to have that data, you have to hire people doing those workflows. Uh, and train the models the right way, which is, which is what we're doing. So there's an enterprise data gap. And there's also an enterprise, uh, expertise gap. Uh, you, there's a lot of first mile schlep and last mile schlep to make these systems actually work well. And there's work to do. And the bigger risk is an insurance company or a company in financial services doesn't adopt or make the best use of these models quickly enough. And if you don't do it quickly, their competitors will, and whoever does it really well will win.

• 35:23Molly O'Shea?

  Okay, but how are you helping close the enterprise gap?

• 35:27Speaker 01

  With enterprises, the biggest gap is making sure that you are building custom models, which could be an off the shelf model that you take. But you fine tune it on your proprietary data.

• 35:40Molly O'Shea?

  This is what you're doing with enterprises?

• 35:42Speaker 01

  That's right.

• 35:42Molly O'Shea?

  Okay.

• 35:43Speaker 01

  That's right. Um, with enterprises, uh, the old way of thinking was, uh, you would take this off the shelf model. Or you give everybody access to chat G P T or Claude and call it a day, right? That's the SaaS era thinking, right?

• 35:57Molly O'Shea?

  Yeah.

• 35:57Speaker 01

  Plug and play.

• 35:58Molly O'Shea?

  Mm hmm.

• 35:59Speaker 01

  But the best enterprises know that it's not about integrating general intelligence. It's about building proprietary intelligence. You'd wanna take a model, maybe from built by one of the frontier labs. But you'd wanna fine tune that model on your proprietary enterprise data. Your data is your mode. And fine tune it on your proprietary human intelligence. Uh, you might automate your proprietary internal workflows. It'll probably be a custom model that's on prem. It'll probably, probably be a smaller model. Oftentimes we see a half a billion to a ten billion parameter model. That's the regime, not a trillion parameter model. And this will be faster, more accurate.

• 36:43Speaker 01

  You'll have full control over governance, confidentiality. You don't want your model to help your competitors. So that's what you wanna build. And but the constraint for these models is step one, you first have to prepare your data. Enterprises have lots and lots of unstructured data, but you wanna convert it into structured data into a format to fine tune these L L Ms. You'd wanna platform to make sure you can distill your human expertise into the models and suck up like your uh your proprietary data uh into a way to uh fine tune these models. And you'd wanna build the right workflows designed for partial autonomy. You'd wanna have good evals for your specific workflows. You'd wanna solve safety um.

• 37:26Speaker 01

  And the and the enterprises need a neutral evaluator to help them pick the right models, pick the right systems to build this, and that's where we come in. The same technology that we built to help the frontier labs get to A G I. We are building a version of this for enterprises to build proprietary intelligence to build their own uh internal chat G P T or or systems built on top of chat G P T. Uh and that's gonna transform. And I'm very bullish for Fortune five hundred companies that have uh good data collection practices. As long as uh you you at the very least maintained good

• 38:09Speaker 01

  unstructured data from your prior business interactions with your customers. You have an opportunity to win because your proprietary data is your moat. As long as you distill it into, as long as you fine tune L L Ms based on that knowledge and you uh the you, most companies have really amazing human expertise inside their organizations. They just have to transfer that expertise from those human minds into those uh machine minds. Uh and and that will give them an edge. It'll it'll actually give them an edge over a startup starting from scratch. If you're Geico. Um. As long as you do a good job, if you're GEICO, Progressive, or any of these insurance companies, as long as you do a good job of

• 38:52Speaker 01

  fine tuning your custom models on your underwriting data, on your claims processing data, and building systems around it to automate underwriting and claims processing, you can win. But if you don't do that, there could be an upstart AI company that starts today, that does this from scratch, that has a good data collection feedback loop. That company could, could, could jump you, right?

• 39:16Molly O'Shea?

  Mm hmm.

• 39:16Speaker 01

  Uh. But that company will still need time to acquire that data. If you're GEICO or Progressive, you have access to that data. So it's your battle to lose. Uh. But to do that, you have to act fast, uh, to make sure that you are, uh, you own your proprietary intelligence and you don't outsource it away.

• 39:34Molly O'Shea?

  With over three hundred million dollars in revenue, I'm sure this number has changed, because that was a couple of months ago. But who are your customers? You can talk categorically, you don't have to tell me their names. I know some of these things are sensitive, but like, what ways in which are you helping these different industries deploy their own models?

• 39:56Speaker 01

  So we are working with, um, uh, almost everybody that's building a foundation model. Uh, we are working with, um, eight out of the nine frontier foundation labs. And now we've expanded to start to work with enterprises in financial services, in insurance, uh, in pharma, uh, and we are rapidly expanding into more verticals.

• 40:24Molly O'Shea?

  So this is like investment banks, this is maybe private equity funds, these are like real financial institutions. That's pretty interesting.

• 40:32Speaker 01

  That's right. And if the job involves analyzing data and making, um, relatively, uh, quantifiable decisions, like verifiable decisions, uh, it's a great use case. And for investment banks, private wealth management, uh, firms, um,

• 40:53Speaker 01

  These other companies in in capital markets, if the job is to like uh figure out good investment opportunities and you're looking to find alpha, uh this is AI is sort of is is like it'll be uh an existential risk to ignore. Mm hmm. Because um now with this technology, you can ask a question like uh hey, how many times did the CEO of this public company talk about this particular topic uh in the last twelve months and how has that changed, right? In the pre L L M world, this question would have been uh would have been a very sad piece of work for an intern or like some analyst to do for like many, many hours.

• 41:38Speaker 01

  But for L L Ms, this is like nothing, right? So L L Ms are really good at synthesizing gobs and gobs of information uh to come up with um useful signal, especially if they are fine tuned on the right data sets. So if you're an investment firm and you have a history of evaluating investment opportunities and making decisions and maybe you decided to um buy a stake in a company at a certain price, or maybe you decided to pass on a company and you probably wrote an investment memo outlining your decisions, that's incredibly valuable data to fine tune in L L M, right? And you own that data, that's your edge, your data is your mote.

• 42:21Speaker 01

  And I think like the best firms are already doing this uh to help an investment analyst make better decisions faster and eventually with significantly fewer people.

• 42:35Molly O'Shea?

  Sorcery is proudly sponsored by Carta. Carta is transforming the private marketplace, connecting founders, investors, and limited partners through software purpose-built for private capital. Trusted by more than sixty-five thousand companies in over one hundred sixty countries, Carta's platform of software and services lays the groundwork so you can build, invest, and scale with confidence. Carta's fund administration platform supports over nine thousand funds and S P Vs, representing nearly one hundred eighty-five billion dollars in assets under management, with tools designed to enhance the strategic impact of fund CFOs. For more information, visit carta.com/sorcery. That's C A R T A dot com slash S O U R C E R Y.

• 43:21Molly O'Shea?

  How do you give conviction in your investors, in outside entities, that this is sustainable, durable revenue? The whole image of Merckor, Surge, Scale AI, is that this is service-based revenue and it's not long-term. And we all know, the VC world investors, they love long, sticky revenue. But how do you position this to investors? And how do you think about revenue yourself?

• 43:55Speaker 01

  So Turing's an interesting company in that we are a lot like Nvidia, in one sense, in that we are a picks and shovels into the AGI industry.

• 44:09Speaker 01

  These, as AGI continues to advance, they're going to need lots and lots of compute and lots and lots of data. So that's going to continue to grow for the foreseeable future. I think we are still a significant distance away from automating everything. If I asked you, how much of enterprise workflows have we automated today? You'd probably say next to nothing.

• 44:36Speaker 01

  If I think of a zero to 10-point scale, on the consumer side, we are maybe at a three. On the enterprise side, we are maybe at a 0.25, at best. So there's a huge market ahead that's massive and growing on the data side. So we're very excited about that. The thing that we've discovered, and this is one way in which we are different from like the traditional data labeling companies, like the ones that you mentioned, is we also work with enterprises to help them take advantage of AGI.

• 45:12Speaker 01

  And for enterprises, they don't need just data. They need somebody to help them build end-to-end AI systems in the first mile and the last mile. I think that's going to be an even bigger market potentially that we are perfectly positioned to. I think of it a little bit like we are working with the Formula One teams in the Frontier AI Labs. But there's the car companies where there's a trickle-down effect of some of this technology. Like, we are now building RL gyms for giant financial institutions as well, right? And that's a massive market. Like, if you look at the services market, it's a trillion-dollar market, right?

• 45:53Speaker 01

  And we are talking about thirty trillion dollars of knowledge work that's about to be automated. And enterprises need all the help they can get to not just build the right models, but to build the right systems around these models to automate workflows. So I think Turing will probably replace companies like McKinsey-Bain, BCG, and every services company like Accenture, TCS, Wipro, Infosys, like all of these companies doing vanilla services. That's going to be automated with these agentic systems. And we are perfectly positioned as the company that's helping the Frontier Labs build smarter and smarter superintelligence. We are simultaneously helping enterprises take advantage of these superintelligent systems,

• 46:37Speaker 01

  and one helps the other. If you're just a pure data vendor, you're not touching reality. You're just selling data to labs. But because we also try to build stuff on top of these models, we touch reality. We know where the models break when you try to solve underwriting. When you, one of the use cases that you told me was after you record an interview, you try to pick the right clips from those interviews that are likely to go viral. In the future, you'll probably use like a Turing model that is like fine-tuned for your own use case to do that. That's an example of something that actually touches a real-world use case. So the fact that we are also helping enterprises helps us do a better job serving the Frontier AI Labs.

• 47:21Speaker 01

  So talking about revenue, revenue has become really public.

• 47:26Molly O'Shea?

  Because all of a sudden, there's a billion dollars up for grabs um with the Scale AI acquisition, Aqua Hire, whatever you want to call it, investment round. How do you think about revenue and your competitors?

• 47:39Speaker 01

  So this is a massive market that's growing super fast. And the interesting thing is, and there are a few companies in the space, there's companies like uh Serge, Mercor, Scale AI, of course, uh being the company that started uh in the space early.

• 48:00Speaker 01

  This unlimited demand for really high quality data. If you look at the competitive landscape, uh I see three categories of companies. There is the pure play data labeling companies that are recruiting talent marketplaces. Uh, we are none of them. We are a research first data accelerator. Uh, that sets us apart. Of course, we are really good at finding really smart people and we built a platform that helps us generate data really efficiently. But what sets us apart is our research first posture. Old school data labeling is over and the era of recruiting or a talent marketplace is also over.

• 48:43Speaker 01

  These la- labs need more than just data thrown over a wall. They need more than just higher smart, hi- higher fifty PhDs in physics, chemistry, math, biology. They don't need just that. What they need is a true partner that can understand their research objectives in improving their models for coding, for math, for STEM, for reasoning, for agentic workflows. And they need a partner that can go uh that can help them figure out what type of data is likely to be helpful and do things before being asked. Uh, that sets Turing apart. I would say our, most of the other companies in the space fall into either being a data labeler or a people finder.

• 49:27Speaker 01

  And Turing in is a research first accelerator. We are a research accelerator for the frontier labs.

• 49:33Molly O'Shea?

  Hey, it's Molly. If you enjoy. Interviews, check out our newsletter sorcery dot V C, where we deliver a once a week top deals and tech headlines email, and also go deeper on our podcast interviews. Subscribe to Sorcery today. And don't forget to subscribe to the podcast on YouTube, Spotify, Apple, or wherever you listen. Link in description to sign up.