Turing CEO Jonathan Siddharth：谁将赢得数据标注，以及为什么99%的知识工作会消失

Harry, I've been so looking forward to this. Thank you so much for joining me in person. It's such a treat to do it in person while you're in London.

Thank you for having me, Jonathan.

So, I think of a talent marketplace as something that's basically matching talent to something. Maybe it's an opportunity. Turing is not a talent marketplace. What we do at Turing is train superintelligence. We work with 7 of the 8 frontier labs.

To get to superintelligence, you need research, compute, and data. Research, the labs do in-house—with OpenAI, Anthropic, DeepMind, et cetera. For compute, we have Jensen Huang to thank, and maybe NVIDIA as well. On the data side, Turing powers the data pillar.

1. Redefining “Talent Marketplaces” Today

Now, I want to start with a little bit of definitions, because everyone thinks they're talent marketplaces, and then everyone pushes back on talent marketplaces. How do you describe it, and why are we not dealing with talent marketplaces anymore?

On the data side, there's been a significant shift in the last couple of years. A few years back, the models weren't quite smart enough, and as the models have gotten increasingly smarter, the data needed to improve them has become harder to generate. This is because more sophisticated data is required to improve the models. It's vertically specific people, tasks, and workflows that aren't so obvious, like cat pictures.

That's correct. That's correct.

There's a shift in the data from simple to complex. Take coding, for example. A few years ago, the kind of data set a contractor could generate might look like, "Hey, write a Python program to sort some numbers." Today, the data that's generated might be, "Write a B2B marketplace app that connects doctors with patients, and write it for Android with Kotlin and Java, write it for iOS with Swift, and write it on the web with Next.js or something," right? That's the complexity.

So, there's a shift from simple to complex. It's no longer the kind of data that low-skilled, medium-skilled contractors can generate. You need expert humans in every domain.

Yeah.

The second shift is we've gone from teaching AI to take tests and pass tests to teaching AI to do real work. It's less about having AI pass the bar; it's more about, can AI do the job of a lawyer? Can it do the job of a privacy lawyer, a compliance lawyer, or a paralegal? It's about having AI be good at doing economically valuable work. So, that's a shift.

The third shift is we've gone from chatbots to agents, right? We started off with ChatGPT, where you're asking questions and getting answers, which is great, but now it's about the models becoming agentic, where they can execute complex, multistep workflows in a real-world business setting. The type of data you need for that is totally different.

2. Data, Compute, Algorithms: What is Most Abundant?

How is that different? That's so interesting. In the transition from chatbots to agents, how does the data required change with that transition?

When you're training a chatbot, you'd usually do a lot of SFT and RLHF. With SFT, you're giving the model input prompts and output completions. You teach the model to imitate experts. With RLHF, you're basically teaching the model to produce responses that a human would tend to prefer. RLHF is used to train what's called a reward model, and then the model is trying to produce completions that give it a high reward.

With agents—and let's define an agent; different people define agents in different ways—I would define an agent as something that's capable of taking action in the real world or in the physical world, something that's executing a multistep workflow and calling different functions. The agent could be operating a computer or making backend API calls to actually do stuff, right? You might have an agent to file your taxes. You might have an agent to prepare your monthly financials.

To train an agent, you would also want to teach the model how to do tool use. So, you teach the model how to call other functions and how to use other applications to be more leveraged. Today, the dominant paradigm is reinforcement learning. Oftentimes, these agents are trained through reinforcement learning, where you'd build what's called an RL environment, which is like a mini world model for business.

Yeah.

In that RL environment, you have input prompts and output verifiers, and you'd have the full system state tracked along with the data model.

Let me give you an example. Imagine a workflow for a salesperson that an SDR would go through, where, before a sales call, the salesperson has to research the prospect, look up Salesforce to see whether somebody from the team has already spoken with this human, and maybe, if needed, look up this person's contact information, perhaps using ZoomInfo or something like that, to reach out to them. This required this human to use 3 different tools: LinkedIn, Salesforce, and ZoomInfo, right?

In an RL environment setup, you'd create a mini world model with clones of these applications that are created with a fake database and synthetic data. The prompt might be, "Hey, prepare for a call with this person and then, after the call is done, update Salesforce." Let's say that's the prompt, and you have what's called a verifier to check whether the agent completed the task.

This is where I think AI is kind of beautiful and somewhat magical: you set up the agents in this environment, and the agent is going to try different trajectories and different tool calls to try to complete the task.

You would set this up so that the curriculum is optimally defined. The curriculum is the set of tasks that you have this agent do. If it's too easy and the agent completes everything, the model doesn't learn much. If it's too difficult, the model doesn't learn much. Ideally, you'd want the right mix, where the model is getting positive and negative feedback.

It's very similar to the technique that AlphaZero used in mastering Go, when the model played against itself. So, this is another—it's kind of like a form of synthetic data because the agent is trying different approaches by itself—but it's humans. In this case, at Turing, we create these RL environments at massive scale for every workflow you can think of, across every function, across every industry.

So, you create the RL environments that create the data that then allow the models to train to have further use cases like that?

Correct.

Correct.

We create RL environments for every industry you can think of: retail, healthcare, and life sciences. Imagine this 4-dimensional matrix, where the first dimension is every industry—financial services, retail, healthcare, podcasting maybe is one of the dimensions.

Please.

The second dimension could be every function: software engineering, marketing, sales, finance, et cetera. The third dimension could be every role in that org chart. Let's say, in sales, an SDR was a role. The fourth dimension is a workflow that a human goes through in that role.

You can think of every role a human has as a composite of workflows, right? We are creating RL environments for every workflow, for every role, in every function, in every industry. That's $30 trillion of knowledge work.

Is it possible to have that breadth and quality?

Yes.

How? Time and lots of money? Because I was speaking to, candidly, one of your competitor's board members the other day in prep for this, and he said, "The big thing that we all got wrong was we are so in innings one of the acquisition of verticalized data. There is so much room to run in the data acquisition for dental, for SDRs, for product managers—you name whatever function you want."

Do you see us very much in innings one of the data acquisition for these very specific, vertically focused workflows?

Absolutely. It's innings one, and I believe in slow takeoff. I'm sorry to pour cold water on all the AI doomers that might be listening to this, but we are not in a rapid takeoff scenario. I believe in slow, steady takeoff for AGI and eventually superintelligence.

So, we're still in innings one. It's going to take a while before we get all of this data into the models.

When we think about the breadth that we go after and your specializing in RL environments, just so I understand the marketplace that we sit in, because there's Mercor, there's Surge—how do you differ from those, for people who are wondering, "Hang on a minute, I thought they were all one?"

Turing is a fundamentally different animal. What we do is we're training superintelligence for all these frontier labs. To get to superintelligence requires research, compute, and data.

The data needs have significantly changed. It's more complex data rather than simple data. It's more real-world data—data that touches how real humans do knowledge work. You need data to train these agentic systems, right?

What the labs need in a partner in this new world is somebody that has research DNA, who could be a proactive research partner for them, because these paradigms keep changing. Last year at this time, we were not talking about reinforcement learning at all, but then 2 things happened.

o1 dropped in December. DeepSeek launched in January. And now it’s all about RL environments. It’s not just imitation learning; it’s also reinforcement learning. So the labs need a data partner that’s more research-oriented.

Second, the labs need a data partner that also touches the real world. At Turing, we don’t just generate data to train the models for the frontier labs. We also work with enterprises. We work with Disney, Pepsi, BlackRock, Fiserv, and Johnson & Johnson to build fine-tuned custom models to solve real-world enterprise problems for those enterprises.

So this is like FDEs that you send in to go and build custom models?

Correct. So we touch reality. We know where the models break in the real world.

How much of the business is that—FDEs deploying custom models—versus more horizontal?

The horizontal business is bigger, but this business is also fast-growing.

And, third, you need a platform with the world’s smartest humans on it, as well as experts in different domains, so that you can build these RL environments. You need that platform to be really good at sourcing talent, vetting talent, matching talent, and generating data.

I think the era of data-labeling companies is over. Turing is a research accelerator, and it’s now the era of research accelerators. The labs want to work with a proactive partner that can think about what types of data are likely to be helpful for these models and can make recommendations to them.

Why would you need a custom model? When you look at a lot of the customers that you mentioned, for the ones where you have FDEs who go in and build custom models, what is the reasoning around that? And is that a temporary moment in time, or is that a permanent requirement from them for a certain reason?

I think it’s a permanent requirement. I’ll give you an example. Let’s pick an insurance company. For insurance companies, 2 really important problems they have to solve are underwriting and claims processing. Let’s pick underwriting, for example.

With underwriting, the problem statement is that you might get multiple types of unstructured medical data. It could be somebody taking a picture of their medical history on their smartphone, or it could be some OCR data from somebody’s medical history, or data in PDFs, et cetera. A human has to look at that person’s medical information and then decide: Is this person high-risk, medium-risk, or low-risk? What medical conditions do they have? Do they have cardiovascular issues? Do they have renal issues? And how do you price insurance for somebody like this? Do you even take them on as a client if you’re an insurance company?

Right now, this is a problem that an LLM can solve really well with a human-in-the-loop system. You may not need a trillion-parameter world model to do a task like this. In fact, there’s lots of research that shows a smaller language model will actually be faster and more accurate at a task like this than a giant world model.

The insurance company also may not want its data to go back to a frontier model. Oftentimes, in these cases, what we would do is work with a frontier lab, take one of their smaller models—maybe something in the half-a-billion-parameter to 10-billion-parameter regime—and have an AI system built that’s on-prem with the customer, fine-tuned on that enterprise’s proprietary data.

This insurance company might have data from the last decade of humans making judgments. You want to make use of that data, but you don’t want to help other competing insurance companies with your own data. Normally, for these cases, you would have a smaller, fine-tuned model that’s trained on your proprietary data, distilling your proprietary human knowledge into the models.

That human underwriter who’s been doing this job has a lot of institutional knowledge in their brain.

Sure. You’d want to distill that into the LLMs.

And this human that I mentioned who’s doing this job of underwriting might be using other internal tools inside this insurance company. You might want to automate those tool calls in the agent. So you’d basically build a version of almost like a ChatGPT agent that’s smaller and fine-tuned for that specific workflow and use case.

I do think this will become more popular across the board, and it’ll be a big market for the frontier labs. If you want a general-purpose assistant, I think you need a trillion-parameter model, right? One that can answer anything to be a universal assistant.

Is that a good business for you when you think about taking someone else’s model, retrofitting it to a business, and doing a lot of custom work with your own engineering teams in those businesses? Is that a good business?

I think it would be a good business. We’re still early, but it’s growing pretty fast.

We have this unique vantage point because, by generating data for all the frontier labs, we get to see a glimpse of the future before it arrives. The glimpse of the future that I see is that all knowledge work is going to be automated.

If a human’s job involves looking at a computer, analyzing what’s on the screen, using different tools, and using a keyboard and a mouse, it’s going to be automated. It’s only a matter of time. These computer-use agents are going to keep improving over the next decade. And that’s $30 trillion of digital knowledge work.

3. The Biggest Challenges Enterprises Have with AI Adoption

My question to you is this: I spend a lot of time with very large companies, mostly when I speak to them. The thing that astounds me is that we hugely underestimate the pace of AI progression in terms of technological capabilities, but consistently, what I see is the laughable state of internal data and internal processes.

I mean, Jonathan, these guys are so far off adopting Slack and Notion, let alone building custom models and embracing the latest AI tools. I respectfully push back on all knowledge work being automated—maybe in 20 years, but not in a 10-year time frame. Am I wrong?

What do you think is the biggest constraint or obstacle?

The inability for them to try and implement new tooling.

But what if the cost is too high? If they didn’t do that—if the hypothetical insurance company that I told you about had a competitor that could operate with one-hundredth the headcount—

While delivering a better experience to their customers by pricing insurance deals better, making more money from insurance premiums, and having fewer claims payouts, they’ll get their lunch eaten.

I think they will. I think you’ll see this transfer of value from an old incumbent that can’t adopt new tools to a startup company—hence why we invest—which is eating their lunch. Absolutely.

Ah, I see. So your theory is that the incumbents won’t adapt and it’ll just be a forest fire?

100%. We will be on a 10- to 20-year decline of incumbents that are unable and unwilling to adopt new tools because of data, because of permissioning, and because of internal buying processes.

That’s an interesting point, Harry.

I mean, I go to a European bank, and you will be astounded by how bad it is internally. Respectfully, the poor quality of the processes for buying technology is just abhorrent.

I have a hypothesis. My hypothesis is that companies will be very slow with back-office automation. But in the front office—for example, I speak with financial services clients in New York, some of the biggest companies—I speak with a suite of these companies, and they are extremely interested in applying AI to help them make better investment decisions because it directly translates into helping them make more money.

I’ve found it’s a lot easier to convince people to use a piece of technology to make more money than to save money.

Yeah, I agree with that.

And in financial services, it’s pretty brutal, right? It’s kind of an efficient market. If there is alpha to be found in how you can allocate capital better, make investment decisions better, figure out what opportunities to invest in, or price deals better, you’ll get killed if you’re not at the bleeding edge.

I’ve heard Mark Chen, the head of research at OpenAI, say this about how financial services is usually at the bleeding edge among all the other industries in the S&P 500. But even they are about 2 years behind, usually, relative to the state of the art.

So I agree with you that in back-office automation, it’ll probably be very slow, and it’ll probably be the upstarts that do things well. I think the change management will be too slow. But I’m optimistic about the front office, especially in financial services, life sciences, and pharma, where, if you can accelerate the time to discover a drug or to get to a molecule—if you can help somebody win in the main thing that they care about in their industry—I think they’ll adopt it faster.

4. Why Will 99% of Knowledge Work Will be Gone in 10 Years

I’m always told by a dear friend, Rory O’Driscoll at Scale—I don’t know if you know Rory, but he’s a fantastic investor—and he always says to me, “Listen, value generation from AI is fundamentally dependent on one simple question: Will we see the transfer of budget from human labor to AI technology?”

If we see that transfer of budget, oh my God, that’s the $30 trillion that you said it is. And if we don’t, we operate in maybe a slightly larger software technology budget world, but by no means a world where we can have the valuations and the money that we have going in.

When you look at that, are there any areas truly today where you’re like, “We have seen the full transition from human labor budgets to AI technology budgets”?

I think the transfer is pretty high in areas like customer support, copywriting, and SEO—some of these marketing-related areas, as you would expect. The transfer is faster in these low-risk-to-fail areas, where it’s relatively easy.

I’d encourage your listeners to look up GDPval, which is this paper by OpenAI where they measured the impact of today’s AI models in automating all types of economically valuable work.

It’s a lovely piece of research. I’d encourage everybody to read it. They did this study where they looked at, I think, 9 verticals and 44 occupations. They took a very diverse sampling of different types of knowledge work, everything from financial services to real estate, healthcare, and law.

They took very specific occupations, and in those specific occupations they took real tasks where a real deliverable has to be produced. Imagine an engineer, like a civil engineer, creating a blueprint for a building they’re about to build, or somebody who’s on the set of a movie studio coming up with a schedule for how you organize your crews. That’s real work. For coding, you can imagine a real-world software engineering project.

They saw that today’s models were quite good at achieving parity with the best human experts in that field. What I noticed was that about 50% of the time in GDPval, the best models were producing work that was indistinguishable from a human expert, which is remarkable. I say “we”—I don’t know which side I’m on. Am I on the side of the AIs, or are the humans?

I think you’re on the side of the AIs.

Oh, really? I mean—

From that positioning, that would infer so.

Yes. But Turing is a blurry line, right? Maybe—as the CEO of Turing, passing the Turing test is about not being able to tell the difference.

What I noticed was that about 50% of the time in GDPval, the best models were producing work that was indistinguishable from a human expert, which is remarkable. Kudos to OpenAI; they also flagged that the number 1 model was Claude Opus 4, although GPT-5 was quite good also.

This was for relatively simple tasks. Imagine a task requiring a single step, whereas in the real world, if I give you a certain project to do, you won’t just go off and do it. You might ask for clarifying information, do other things to acquire more context, or brainstorm with other humans to complete that task. You would do it in a sequence of steps, so there’s more room to go.

But I think we are well on our way to AI eventually automating all types of knowledge work.

What happens in that world? If AI automates all types of knowledge work, what happens then?

Three things will happen. First, I think we will all have the potential to be 100 times more productive. Today, I’m able to run 1 company. Elon can run maybe 5 companies. But in a world where I’m 100 times more productive, maybe I’m able to run 100 companies. Elon maybe runs 600 companies.

I think every human will just be so much more leveraged. The nature of a job itself could change. Today, we are accustomed to the idea of 1 person doing 1 job, but people could be doing multiple jobs at the same time. People could be running different companies at the same time.

The second implication, I think, is that it’s going to be wonderful for entrepreneurship. You, Harry, are going to be very happy, because today, for a lot of ideas, founders are intelligence-constrained. I think of being capital-constrained as a form of being intelligence-constrained.

For example, if you pick a therapist who wants to start a mental health startup, today that founder would have to raise at least a few hundred thousand, if not a few million, to recruit some software engineers, maybe a marketing person or a growth person, and perhaps a product manager. But in a future where AGI exists, this person will recruit a marketing GPT, a software engineer GPT, and a PM GPT, and get off the ground for a lot less capital.

A million flowers will bloom. Lots and lots of nontechnical founders will start companies. We’ll see a broader distribution of founders than just those who live in London or Palo Alto, who are connected to these pools of capital and might start companies, which I think is wonderful for the world.

Do you think we will? What I mean by that is, there are 6.5 million people today in the UK’s working population who actively do not work because of an inability to work. I’m not going to get into the analysis around that because I’ll get in trouble for it.

I think we grossly overestimate the intelligence of the general population. I know that sounds incredibly arrogant, but most people—actually, a lot of people—just don’t want to work and are not at the level of recruiting GPT assistants. Do you not worry that it will widen the chasm between those that have and those that haven’t?

I’m an optimist, and I think the opposite will happen, because what we’re really doing when we’re training superintelligence is basically training intelligence as an API. What’s the alternative to that? It’s hiring a human to provide you with that intelligence, and that human is quite expensive, right? If anything, that creates an even broader gap between the haves and the have-nots.

Whereas, for $20 a month, if you had access to the smartest experts in coding, STEM, sales, and marketing, I feel like more people would be able to start companies and produce actually valuable work.

I believe that when we have access to superintelligence—we are not all going to chill out on a beach somewhere and contemplate what to do next. I firmly believe this. We humans are tool builders and problem solvers. We’ll solve problems at higher and higher levels of abstraction.

I feel like, in a world where we have AGI, we’ll just solve much more exciting problems. Maybe we’ll cure diseases, reverse aging, or go to the stars. There are all sorts of fun things we’ll do. I don’t think we’ll be bored.

I’m glad. I often hear about UBI and how we’re going to sit and write poetry. I’m like, I think that might be a little bit challenging.

5. How Will Data-Driven Feedback Loops Replace Technology as the Moat

When technology is not the moat, what is the moat? I had the founder of Base44 on, and he said 99% of code in the next year will be written by AI. Technology is no longer the moat. What is the moat in that world?

I think one moat will be data-driven feedback loops. For example, one reason Google had such a great lead in search for a while was these data-driven feedback loops that come from people using your product and generating data that gives you, the algorithm developer, a high-quality gradient for which direction to step in.

PageRank—the importance of PageRank and the recipe for ranking search results—was well known among Google, Yahoo, Microsoft, and a few others. Obviously, people move around these companies all the time, but the advantage Google had was that, because everybody preferred Google and liked that search engine, you saw a much more representative set of queries.

You had data from the clickstream of what results people were clicking on, and that helps your algorithms improve at a much faster rate. I think data-driven feedback loops will be key for all types of enterprise applications as well.

Today, OpenAI and ChatGPT have a good data-driven feedback loop in enterprises. Again, I think it’s wide open. Whoever is deploying the right custom, fine-tuned models and agents for specific workflows, roles, functions, or companies—if you get in first and solve a customer’s problem really well, you start getting that flywheel going.

You will discover first where the models don’t work well, and you will use that data to work with a company like Turing to generate additional data to plug that gap. Then you will improve. This is what I mean by why it’s important for the models to touch reality.

I feel like the models have touched reality in consumer. We haven’t yet touched reality in enterprise, and the only way we’ll improve is through deployment.

And that deployment is fundamentally predicated on handholding, correct?

Yes, handholding. I feel like there’s still a lot of first-mile schlep and last-mile schlep that needs to be handled.

What does that mean, first-mile and last-mile schlep?

When I say first-mile schlep, I mean, for that underwriting copilot example I gave you for that insurance company, I painted a pretty rosy picture of how you take this model and fine-tune it on your proprietary underwriting data. In the real world, it doesn’t work that way.

Let’s say I’m talking to the CEO of this insurance company or the CTO of this insurance company. They’ll say, “Our data is a mess. It’s in silos. It’s super fragmented. Some of the data is in spreadsheets. Some of the data is in a file that Bob has, and Bob doesn’t work here anymore.” The data is all over the place.

You first have to acquire the data, convert the unstructured data into structured data, and put it into a format to fine-tune LLMs. You might want to set up good infrastructure for evals. You’d want to create good evals for the models or agents. You might want to build a workflow designed for partial autonomy.

This human underwriter who’s about to use this model to evaluate these medical histories might need a Cursor-like interface so they can work alongside the AI to do their job. You also need to train the humans in these new workflows.

You want to make sure you’re collecting data the right way. The way, for example, we do deployments is with a tandem system, where you’d have a human and an AI doing the same job for a period of time, and a manager can see the output of both.

If the agent is right and the human is wrong, you train the human. If the human is right and the agent is wrong, you’ve created a data point to fine-tune the next iteration of the agent. The agent is steadily improving over time.

If the agent is right and the human is wrong, why don’t you just fire the human?

It depends on what frequency, right? You track things like precision and recall. You’d want to analyze this over a period of time. You wouldn’t fire them over a single mistake.

Bit harsh.

Yes.

What’s the margin on that business?

It varies. We’re also in the early innings of figuring out how to price that. Today, we do it in a relatively simple way where we’re just building these things for time. I don’t think that’s the right way to do it. We’ll switch to a more value-oriented pricing model at some point.

Right now, we’re just laser-focused on the frontier AI labs. Enterprises for us are a longer-term play.

6. Is Revenue BS in Data Labelling? Are Players Calling GMV Revenue?

When you look at revenue numbers in this space, a lot of people shout back, “They’re not revenue numbers; they’re GMV.” Given our understanding now that, Jonathan, there’s no talent acquisition from your business—it’s all an RL environment-creation business—when you look at the other announcements from alternative providers, can you help me understand: are they revenue or are they GMV? Is there mislabeling being done here?

I don’t want to comment on other companies, but I’m an investor, Jonathan. Essentially, I’m trying to understand and learn from you how I should weight revenue in today’s AI world versus the previous historical world. Should I be impressed by these revenue numbers, or should I not?

I think it depends on the type of revenue. Obviously, these are not SaaS ARR numbers, right? These are not those types of revenues. This is a different beast, and I think this requires thinking from first principles.

The revenue here is recurring in the sense that, oftentimes, when you’re working with a lab on helping the models improve in some area—and I’ll speak to Turing; I don’t want to speak to other companies—when we’re helping a lab, let’s say, improve their models for coding, multimodality, or tool use, or working on RL environments for automating all types of professional knowledge work, it’s usually a recurring project where projects will start and projects will end. As long as you’re doing a good job, there’s lots and lots of demand.

But you have to consistently keep doing a good job. It’s also important to be a trustworthy partner to the labs. We take secrecy very seriously. We make sure that our projects are all firewalled between labs, oftentimes even with teams within the labs. Sometimes, that’s the level of secrecy that you would need.

I’m reminded a little bit of how I’ve been told Foxconn operates. I don’t know anything about that, but I’ve been told that they have different floors where maybe on one floor the iPhone is getting made, and on another floor, maybe a Pixel phone is getting made. Obviously, you have to firewall all of that.

Of the 8 largest providers, do they not spend with all of you?

They spend with a handful of companies. They do that to have some level of resilience. I imagine there are some price benefits to having more than 1 person they could work with, but I think the resilience piece is important.

We know what happened when the Scale investment happened. Again, the labs did benefit from having other partners that they could work with. I would say it’s a small handful—trusted companies. Of course, there’s probably a giant pool of smaller startups, but it’s a small handful of big companies in the space.

Which one do you worry about most?

This is just a big, big market that’s growing super fast. I’m excited about all the companies in the space. I feel like different companies come into this world with a different DNA.

Which leader do you most respect? Sam Altman, Elon Musk—of the data providers, Jonathan? Of the data providers. I’m pushing you, dude. I’m going to get a name.

I have a lot of respect for Alexandr Wang from Scale AI. I feel like Alex and Scale were prescient in seeing the importance of data. I admired how, having started in autonomous labeling, they navigated the ups and downs.

I really like the way he operates as well. I feel like there are certain elements of leadership that I think I share with him. I think he did a great job for Scale.

How did Scale being acquired impact Turing’s business?

We just got flooded with a lot of demand. We’ve also amped up significantly in multimodality. Multimodality was something I think Scale was quite strong in.

Multimodality is teaching the models to operate well with not just text but audio, video, images, et cetera. From the outside in, I’ve heard that because of their roots in autonomous labeling, they were quite good in multimodal work. It was good primarily from just increasing demand, and I feel like they were the company that had been working in the space the longest.

Do they have a business left? Again, I mentioned Rory O’Driscoll. I think he said on a show with me recently that there’s this kind of carcass or husk left behind. But what if everyone benefited from their being bought? They can’t be doing that well.

I don’t know enough about their business.

Do you pay attention to competitors?

I pay attention to competitors in terms of the things that they do well and when there are any significant learning opportunities from them that could help us serve our customers better.

Do you worry about revenue concentration? You said about 8 of the biggest labs. Say, if you look at OpenAI, they have—I don’t know, whatever it is—100 million. You all know these numbers much better than me, but say 100 million paying customers. I’m just taking 10% of a billion people, but give or take 100 million, whatever.

Then you look at a business like yours here, where there are 7 core customers. How do we feel about revenue concentration? The last time I checked, I was told that Nvidia has 39% of its revenue coming from 2 clients, right? Roughly 50% was 4 clients.

Which is extraordinary.

Extraordinary, right, for a $5 trillion company—

Wow, that’s nuts.

Yeah. You think about Apple as not even comparable in market cap size, but what is that—like 3 billion customers? I mean, wow, what a comparison.

Yes. So, in this market, I don’t worry that much about revenue concentration. I think we’re in the same boat as Nvidia, in that there will be lots and lots of spend from these big 8 companies. I expect this market to— I mean, look at the scale of the spend.

Stargate is like a $100 billion-a-year investment in compute, and there’s going to be significant amounts of spending on compute, energy, and data. It’s a little weird to have this level of concentration, but things could change. I think it’s also possible that governments spend even more.

I think it would make sense for governments to build their own internal versions of some of these models, which would require proprietary data, again, to be collected.

Do you not think they’ll have to? What I mean by that is, we’ll see sovereignty of models. I do not think there’s any way you’ll have the German healthcare system working with American model providers. Sorry to say.

I think you’re right. I think it’ll be necessary.

In that world, do you provide that FTE service to governments?

Yes. I think the work that we’re doing, not just in training superintelligence but in deploying superintelligence, is with that goal in mind for these governments. I imagine we’ll help them not just with fine-tuning their models with data and evals, but also helping them with the first-mile schlep and the last-mile schlep to make these systems actually work.

You might want to have full control over what type of human data is going into these models. If you’re the German government, presumably you want German nationals to be the ones contributing data, whether it’s for SFT or reinforcement learning.

7. Are We in an AI Bubble?

I’m worried that we are potentially not going to see AI deliver the immediate revenues that we’ve promised, and we’re going to go through a kind of cooling period—which everyone suggests and thinks we’re going to go through in the next 6 to 18 months—in which, as I said, it doesn’t hit the revenues that we said it would, and the AI bubble kind of deflates slowly.

To what extent do you think that’s possible, or will we see this continuing gradual increase, as we touched on there?

I don’t see an AI bubble. I feel like these models are incredibly powerful today. GPT-5 is fucking awesome. I don’t know what people were talking about. I know there was some chatter.

I think we’ve just gotten used to magic. I feel like these models are incredibly powerful today, and they’re the worst they’ll ever be. They’re only going to keep improving. I say that about the Gemini Pro models, the Grok models, and the Claude models. These models are amazing.

There’s a very significant model-capability overhang. By that, what I mean is, the models are capable of X, but what we are getting out of the models is X minus delta. With the right agentic scaffold around these models—in terms of the right system prompts, the right user prompts, and giving the models access to the right context…

Teaching the models how to acquire additional context, teaching them how to use the right internal tools—there is a significant amount of capability that can be unlocked with today's models. For example, Harry, I imagine that when you do an interview with somebody, one of the things you probably do is apply your secret sauce to pull out the right clips from the interviews: what to highlight, what the catchphrases are, and what will drive more engagement. That can be done by a model with the right agentic scaffold, fine-tuned on all the work that you've done in the past.

Every weekend, I go through every single show and pick out 15 to 20 clips per show, and then I make notes on each one.

Yeah. Are you saying, Harry, that you want to use Turing?

If you could make it work, dude, I'd pay you a lot of money.

Yeah, maybe we should partner.

That'd be great. Seriously, every weekend I spend probably 3 hours per show. Three—definitely 12 hours a weekend—doing that.

So I think there is this model capability overhang where the full potential of the model has not been unlocked by humans yet. No, I don't think there's an AI bubble. I think there are some growing pains.

What are the growing pains? I think everybody keeps citing that MIT report about how 95% of pilots fail.

Because we are in the business of deploying AGI in enterprises, I can tell you why I think that happens. One of the growing pains is that, step 1, most enterprises need to do some work to structure their data in the right way. Again, that first-mile schlep has to be done. Second, you should surround the model with the right agentic scaffold that I just described: the right prompting, the right context engineering, and the right internal tool calls.

All of those have to be distilled into the models. You need really good evals. You also need a workflow designed for partial autonomy. Andrej Karpathy articulated this first when talking about why Cursor works so well, because it's not designed for full autonomy. It's designed today for partial autonomy, for humans to collaborate with the AI to do that specific task. So that Cursor for X needs to be built for every role, for every workflow, to help humans work more easily with the models.

Does every role need to go through that pathway of Cursor for X before it goes to full autonomy, or are there some roles, like customer service, where it just goes to full autonomy?

I think for some roles where you can see that the models are quite good at matching humans, we don't need that intermediate step. There are certain roles where, by virtue of how the models are trained—pre-trained with tokens on the internet and then, of course, with talent from research accelerators like Turing that's fine-tuning the models—the tokens from the internet give them sufficient intelligence to do the job well. Customer support is an example.

But if you pick other roles—for example, if you picked the role of an AI researcher, or you picked the role of a lawyer specializing in venture-capital financing—it's possible there aren't enough of those tokens on the internet. So the models will be relatively weak there out of the box. Also, the way one firm does financing might look different from another's. Maybe they have their own way.

You might want to fine-tune them on your own proprietary data and distill the proprietary intelligence of humans working there. So, for those things, you may need to do some fine-tuning; the models may not work very well out of the box.

A lot of people suggest the circular deals between some of the large providers suggest the strains in the ecosystem or the bubble-like tendencies. Do you think that's fair or not?

I've categorized the world into 2 classes. Class 1 is those that believe in AGI. Let's call it the AGI-pilled group that believes we are on the path to getting to AGI, and let's define AGI as an AI system capable of at least matching humans in almost all types of intellectual knowledge work.

Correct. Right.

Then there is another category of people that don't believe this will happen and think we'll hit a wall, right? In the past, there have been other AI paradigms where we did hit a wall. For the camp that believes in AGI—and I believe in AGI, unsurprisingly, because I love AI and it's been my passion for the last 20 years—if you believe that, the grand prize is so amazing.

If you've solved intelligence, you've solved all of humanity's grandest problems, from curing diseases to potentially pausing aging, to interstellar travel, to energy. All of our problems are intelligence-constrained, right? The prize is so large. Whoever wins the superintelligence race will probably win search, will probably win consumer devices, will probably win operating systems, will probably win cloud, and will probably win business productivity software. It's like the prize is so massive that it's worth placing big forward bets in these areas, because the cost of not winning is too high.

Whoever wins AGI would also probably win social networking. So you can see why the big 8 are excited about it, because you're playing for everything. It's like whoever wins this could be responsible for that $30 trillion of knowledge work.

Well, if you're Zuck, you spend $100 billion on it. If you lose or fail, likely everyone else will fail, in which case you're behind just like they are, and you've lost $100 billion, which isn't a huge amount of your free cash flow—maybe 12 to 18 months of free cash flow. If you don't spend that $100 billion and someone else does and wins, you lose $2–3 trillion of market cap. Correct?

You have to play.

You have to play.

Imagine if somebody built a more engaging social network. Social networks have just 1 unit, which is attention. We all only have maybe 4 to 5 hours a day to spend on an app. If there was a more engaging app, then, yeah, those are high stakes.

8. Why is SaaS Dead in a World of AI?

Do you agree, then, with the notion that if you believe in AGI, you cannot be investing in SaaS apps?

SaaS as we know it, I think, is over. I feel like quite a few SaaS apps were built at a time when software was relatively hard and complex to build. Imagine if you were building some customer-support software, some customer-support bot. To build a company like that, you would have had to hire some Stanford PhDs in NLP. You'd collect data for 6 months. You'd use a support vector machine or a neural network that would kind of work, and then you'd deploy it and grind away for a while.

There's a significant amount of capital that needs to be invested to get an app like that to work well. So it made sense for many companies not to bother doing that if it wasn't their core business. They could just use some third-party SaaS app.

Now many of these AI applications are incredibly easy to build on top of these LLMs. I feel like most companies will start building custom software super easily. We help companies build some of these custom apps, and the bar to create many of these apps has come down significantly. So that's risk number 1: companies do it themselves.

Risk number 2 is that you get Sonic-boomed by the foundation-model companies.

Sonic-boomed, meaning—

So today—

They move into the apps layer and just create it themselves.

Yes. I mean, it could happen, right? The models are becoming agentic. You've seen many of these agents. It's fundamentally about computer-use agents. If the models get better and better, it's possible the model is all you need.

Imagine if you wanted the model to—let's say I'm doing some HR thing—"Hey, update my medical benefits information. We've just had a new daughter. We want to update my medical information." If the model is agentic and sufficiently integrated into the database of the organization, you don't need anything else in the middle. So that's the second risk: the models becoming more agentic.

The third, and I worry about this a lot, is that I feel like a lot of our software was designed to be used by humans—humans navigating a GUI and clicking around and doing things. I think that's going to go away with multimodality. Again, this is why I think of 4 pillars to superintelligence: multimodality, reasoning, tool use, and coding.

Multimodality is important because we humans interact in natural language. We talk, there's video, all of that. I think the future might look like some type of ambient AI that you talk to, that will just go and do things. Maybe it'll use the GUI of the current SaaS application as an intermediate step, or it'll use MCP and use tool calls and get what it needs.

The GUI was designed for a world where humans were using a keyboard and a mouse and clicking around and doing things. I think humans can do better things with their time than click around.

Actually, one big change I have is I never actually type emails anymore. I use Wispr Flow, and its transcription is so good that I don't ever type emails. Ever.

Yeah.

Now the only trouble is everyone knows what I'm saying in my emails.

Yeah. The Wispr Flow founders interned at Turing back in the day.

No way.

Yeah.

Oh, wow. How do I feel about it? No, I don't agree. Why? Because the average company today has between 80 and 100 different SaaS products that they engage with. So, one, just the multitude of how many they'd have to create. Number one. Number two, maintaining them. You think they're going to maintain 80 to 100? Oh, my God. You're going to have teams and teams of people doing maintenance, updates, and debugging. I don't think so either.

And that is for the technology-savvy. Let's talk about every plumbing provider, law firm, accounting firm, and restaurant that can barely use Wix and Squarespace, let alone build out their own CRM system and POS system. Not a freaking chance. Then we move to foundation model companies moving into very vertically specific elements. You know, we're in a business that does AI for patent creation, updates, and collaboration. Sam is not going there. Sam has health, solving cancer, and energy utilization. I don't think Sam's touching patent creation and updating.

And so I think the more verticalized you go, the more defensibility you have. For those reasons, I think SaaS has life. Mine is a very biased perspective because it's my job. Is that all wrong again? You're the master here, Jonathan. VCs are literally middlemen.

I'd say, Harry, you have an interesting data set because you invest in a ton of startups.

Yeah. So I would be curious, looking at your sample of startups that you've invested in, to tally how many SaaS apps they use today at every stage and see if that has changed post-ChatGPT. My hypothesis is that today's companies use fewer SaaS apps and have fewer people.

Do you think we have more or fewer software engineers in 10 years?

More.

Help me understand that, then.

I think the definition of a software engineer will change. A Stanford doctor who's in oncology, who has an idea for some cancer-detection-type app, will be able to create a very simple version of an app that somebody could check by themselves and use as a home diagnosis.

I think there'll be more software engineers because if you define a software engineer as somebody who's capable of building a software product to solve a real problem, that pool of builders is going to expand way beyond people who've graduated with a 4-year computer science degree.

So we have more software engineers creating more software, and the problem then becomes discovery. How do we solve the discovery problem in a world of infinite software? You might have an agent for yourself that's talking to other agents on the internet. Have you seen Her, the movie?

Yeah. I really think we'll have—you might have, like, there's probably an agent for Harry that's just tasked with discovering what pieces of software Harry should be using. That might be talking to—I might have an agent that is maybe talking about what types of marketing things that I've built or that I have. I think we'll just be in a world where we have digital surrogates of ourselves that are communicating with other digital surrogates.

You might have an agent of yourself that's discovering startups to invest in. While you and I are chatting, there's probably—I don't know—you might be having a million conversations with entrepreneurs from all over the world. Today, you're constrained by space and time, but in the future you'll only be compute- and data-constrained.

9. Will the Phone be the Primary User Interface to an AI World?

Do we lose the phone as the interface to this world? We obviously see Sam and Jony Ive—there are rumors of pendants and some hardware devices. I'm not asking you to comment on that. I'm just saying, does the phone still remain the primary interface and design device?

We'll have some type of device that we'll carry, that's always on and processing multimodal tokens. For example, as I'm talking to you, if I were to envision my perfect device, it would be something that has cameras. So maybe it's a wearable, like glasses, or something that I'm having on me that's processing visual input, because I want to be able to read your body language.

I might have an AirPod-like thing in my ear that's whispering to me. It might say, "Jonathan, as you were talking about multimodality, Harry seemed less interested. His body cues suggest that he was losing interest, but when we were talking about AR, he perked up." Those types of feedback and cues, I think, would be good.

So I envision a device that—I think of it in terms of sensors and actuators. In terms of sensors, obviously it has to be listening to stuff; it has to be seeing stuff. But in terms of actuators, it'll probably also be speaking in my ear. Ideally, it should be something that you can talk to and have it do things later. For example, I might say, "Remind me to follow up with Harry on that idea for using Turing to automate clip generation." So it has to remember that and come back later.

I do think there'll be all sorts of new devices. Glasses and hearing devices like these AirPod-type devices seem obvious. There could be—do you remember this device called the Meeting Owl?

No.

During the COVID era, one of the tools that spiked was basically a speakerphone for having better distributed-team Zoom meetings. When somebody's talking, it would focus on them with a camera, and it was also a decent speaker.

Okay.

I can imagine devices like that that people have. It's hard to predict, but the thing that I almost feel confident about is that the phone will look so different. When we think of our smartphone, it's basically a computer with a phone app in it, right? The phone app is the least interesting part of the phone.

I think even for an AI device, it'll probably have some phone app in it, but everything else, I feel, will be magical. I feel like I would definitely benefit from a device that's constantly listening to everything that I'm listening to, constantly processing all the video and audio input that I'm processing, and something that's paging things to memory. Maybe it'll write things down and be able to look them up later. I see it almost like an extension of my brain.

Before we move into the quickfire round, I do just have to ask: what does your market and the data-provisioning market look like in 10 years? I always try and think about market composition and dynamics. Is it a winner-take-all? Is it very fragmented? Is it 3 or 4? What does that look like?

The market will reward players with research depth because the pace of AI research is so rapid. All these RL environments have spiked in the last 12 months after o1 came out in December and DeepSeek came out in January. So now, in addition to imitation learning, we are in this reinforcement-learning regime. 1 year ago, 1 year later, it could be something totally different.

I think the market will reward a company with research DNA, and it'll reward a company that can move fast and adapt very quickly.

Do you think this is a monopoly market, or do you think there will be many winners?

I think there'll be a few winners. A few, because I do think for the labs it helps them to have a few partners for resiliency. I imagine also for price competitiveness, there'll be a few winners in the realm of robotics and embodied AI.

We are still very early at Turing. We are scaling up on the robotic side as well in terms of data that we generate, but there's so much data that's missing that the models need to see that they haven't seen yet. I can totally imagine some newer companies also coming up that don't exist today.

If you were to invest in companies in your space, where would you invest?

Probably in robotics or embodied AI. The vertical stuff—we are scaling up pretty massively in generating data for different verticals. So I don't see that as a big white space, but I think everybody is relatively early with robotics. Robotics is such a vast realm that there could be interesting things to do there.

One way I see the space, Harry, is to think of it like these dimensions. The first dimension is the type of intelligence that you're baking into the models. That could be coding, STEM, functional expertise like sales, marketing, and software engineering, or vertical expertise like healthcare, legal, and finance.

I do a cross-product of that with the modality: audio, video, image, and computer use. That's multimodality, the second dimension. The third dimension is multilinguality, like different languages. The fourth dimension is different learning paradigms, like imitation learning, reinforcement learning, and pretraining, which is unsupervised learning.

All of those may require different platforms to be built. We've had to adapt our platform for imitation learning, reinforcement learning, and multimodality. So I feel like in this matrix, there are all sorts of new opportunities that could emerge.

And I only listed the digital intelligence. I didn't talk about physical intelligence. So I think robotics is wide open. The kind of data that a robot that's in someone's home needs is totally different from a robot that's doing things in a factory, and humanoid versus non-humanoid robots.

10. Quick-Fire Round

I could talk to you all day. I do want to move into a quick-fire answer. I’ll say a short statement, and you give me your immediate thoughts. What’s one widely held belief about AI that you think is wrong?

I don’t think we’ll see rapid takeoff. I think we’ll see incremental, continuous improvement in AI. I actually think this is good for the world because, if what we believe happens—which is that all types of digital knowledge work get automated—I think humanity needs time to prepare its workflows. I think we could use the extra time to upskill humans, rethink education, and make sure there isn’t massive job displacement.

I also think that, in the steady, continuous improvement in AI models, there’ll be value realized every step of the way, unlike self-driving cars. I feel like people have this wrong model for AI that comes from self-driving cars, where you get it 99% of the way accurate, and if you can’t solve the last 1%, they’re not useful. AGI is not like that.

I think when we automate the job of an underwriter, a claims processor, or a CEO, there’s incremental value unlocked for every percentage improvement as the model becomes more reliable. So, I believe in slow and steady takeoff, and that’s actually going to be great for the world.

You mentioned DeepSeek a couple of times. Do you think we underestimate China?

It depends on who you ask. The folks that I work closely with don’t underestimate China. I think it’s very impressive, the progress that they’ve made in open source with DeepSeek, Kimi K2, and Qwen. These models are state-of-the-art. So, no, I don’t think—at least among the frontier AI circles that I’m in—I think there’s a clear realization of how close they are.

The world seems to be moving to closed models. Is that good or bad?

I think it depends on the application. Firstly, in enterprises, it’s often a mix of closed and open models. We do see demand from enterprises that want either. The closed models are often easier to get started with, but there are some cases where enterprises prefer open models for cost and customizability.

I’m talking about the small language model regime, between 500 million and 10 billion parameters. I worry a little about frontier models. I feel like, for frontier models, there is some value in keeping some of the technology closed, just because of how powerful they are. I feel like the US labs are extremely responsible and safety-conscious in how they think about training and deploying these models.

You mentioned reading Elon’s book earlier. Elon is often chastised for his lack of care around some of the training elements. Do you think he is? And do you think he’ll actually benefit from not having that guardrail?

I think Elon also cares a lot about humanity. At least, if you read his book, one of the things I recall reading is that his motivation for getting into AI was that he wanted an AI that was speciesist and loved humanity. That was one of his reasons for getting into it.

Everything I see about the Grok team, I feel like their goals are much like those of any of the frontier labs: quite noble, in terms of having this powerful AI that can help humanity understand the universe and solve some of our biggest problems.

What did you believe that you now no longer believe?

I used to believe that, to build an enduring, valuable company, you hire a strong executive team and operate with a lot of leverage. Basically, hire strong people and get out of the way. I used to believe that.

Now I believe you hire great people and work really closely with them, their directs, and their directs, and their directs, and get as close to the ground as you can, where ground truth usually exists with the customers. The people closest to the customers are the engineers writing code and the salespeople talking to your customers.

So now I believe in being—basically, I used to, for lack of a better word, follow the org chart a little bit. This was also part of one of my learnings from Elon’s biography: he was so hands-on. He would be walking the factory floor and asking an engineer why this door in the Model 3 has 3 bolts instead of maybe 2, right?

It is a different way to operate, where you’re in the details of the most important things that matter, completely working in a flat structure and operating as close to the ground truth as you can.

Generally, I feel like in the early days of starting Turing, I may have had a subconscious desire to be liked. I think I must have had that. Now I don’t care about being liked; I just think about doing things that would solve our customers’ problems the best.

What was the most unpopular decision you’ve taken with Turing?

The most unpopular decision I’ve taken with Turing is switching from a distributed team to a hub-and-spoke model. We are now working from an office in San Francisco, and we’ve recently opened an office in Palo Alto. We’re going to be opening an office in London as well.

Very exciting.

I mean, for some people, that wasn’t very popular.

Until you fired them.

Some of them left, and, yeah.

We like in-person. We’re big fans of in-person here. Final one: when you look forward to the next decade, what are you most excited for? For me, my mother has MS. I think that we’ll have some pretty groundbreaking breakthroughs in MS drug discovery that we haven’t had forever. That excites me.

I’m excited about AI making new discoveries and automating AI research itself, to get to a point where AI is in some self-improvement loop, so that we could get to superintelligence faster. Automating AI research and getting AI to the point of making new breakthrough discoveries—that excites me.

I’ve always been fascinated by AI as this exoskeleton that makes you a lot more productive. Have you watched the Iron Man movies?

Yeah.

Right. So, in the early Iron Man movies, he’s wearing the suit, and the suit is obviously giving him superpowers, right? In the later ones, the suit is “agentic,” where he has these drone suits—an army of his suits that go off and do things.

I’m excited about a future like that, where every human on the planet has access to agentic AIs that help them amplify their fullest potential. Today, Harry might have 100 ideas, but Harry is able to do maybe 2 of them really well. I like a future where Harry can do the remaining 98. And I like that for the 7 billion humans on Earth.

I like that too, for my weekends’ sake, to be honest. Jonathan, I love conversations that are very natural and free-flowing. You can tell that I don’t really pay much attention to the schedule, but you’ve been fantastic. Thank you so much for joining me.

Thank you, Harry, for having me.