Turing CEO Jonathan Siddharth - The $30 Trillion Knowledge Work Market, Training Frontier AI Models and Building Stage Five Culture

Jonathan Siddharth is the founder and CEO of Turing, a $2.2 billion AI company that provides coding and reasoning data to train frontier models for OpenAI, Google, Meta, Anthropic and more. Turing’s mission is to accelerate superintelligence to drive economic growth. In this episode of World of DaaS, Jonathan and Auren discuss: How Turing creates expert data for frontier modelsWhy SaaS is dying in the age of AI agentsDisrupting the $30 trillion market for digital knowledge workBuilding a stag...

• 0:02Speaker 00

  Hello, fellow data nerds. Welcome to World of DAS. I'm your host, Orin Hoffman, CEO of Incubate and GP of Flex Capital. Discover more episodes, get weekly data as a service news, original content, articles on data and more at worldofdas.com. That's world of DAS, D-A-A-S, dot com. Hello, fellow data nerds. My guest today is Jonathan Siddharth. Jonathan is the co-founder and CEO of Turing, a $2.2 billion AI company that provides coding and reasoning data to train frontier models for OpenAI, Google, Meta, Anthropic and many, many others. He's a second time founder. His first company, Rover, was acquired by RevContent in 2016. Jonathan, welcome to World of DAS.

• 0:47Speaker 00

  Thank you, Orin, for having me. A fan of the show and excited to be here. I'm super excited as well. Now, you said like the bottleneck for AGI progress and just AI progress in general. It used to be compute and data, but now it's human intelligence. When did that bottleneck happen? And walk me through why human intelligence is the bottleneck now? So there's something very exciting happening as we're heading on this race towards superintelligence. So to get to superintelligence, you need three things. You need innovation in algorithms, model architectures, which the frontier labs do a phenomenal job of. You need innovation and scale up on the compute side. And power and all that stuff.

• 1:29Speaker 00

  Power, energy. Yeah, absolutely. And then you need innovation and scale up on the data side. So Turing powers that data pillar. And our mission is to accelerate superintelligence in a way that drives real economic progress. We wanted to show up in the world in terms of something that moves the GDP of the world. So we want to accelerate superintelligence in a way that drives real economic progress. And we do that in two ways. On one hand, we work with all the frontier AI labs to build high quality data, to help them improve their models, to do a wide variety of cognitive tasks. And on the other hand, we are working with enterprises. We work with some of the largest financial institutions

• 2:12Speaker 00

  in the world, media companies in the world, to help them build what we call proprietary intelligence, basically end to end AI systems to solve real enterprise Problems automate enterprise workflow. Is there a particular, so OpenAI has been talked about their hiring investment bankers to come in and go through the actual deal process or something. And other folks have brought in emergency doctors to help understand, are we talking about like that type of thing to get super high level? That is exactly right. I think of it, Orin, as a four dimensional matrix. Think of the first dimension as every industry you can think of, financial services, life sciences, healthcare, retail, et cetera.

• 2:52Speaker 00

  The second dimension is every function in an organization, software engineering, product management, sales, marketing. The third dimension is every role in the org chart in that function. For example, if we index into finance, there is the CFO, there's the director of FPNA, there's the head of accounting. The fourth dimension is every workflow that a human does. For example, a CFO goes through the workflow of preparing board materials or goes through the workflow of preparing monthly financial. And why can't we just watch their computer while they work? What if we think we don't? Okay, maybe we do. But this four dimensional matrix that I mentioned accounts for thirty trillion dollars of digital knowledge work, thirty trillion. This is my job, your job, every human's job in front of a computer.

• 3:35Speaker 00

  And if you add in robotics, we're talking a hundred and ten trillion dollar market. Let's look at the digital knowledge work, work humans do in front of a computer. A human looks at a computer, looks at the screen, analyzes data, is thinking, using different tools, using a keyboard and a mouse to drive action, like pixels in, act out. I see the fundamental building blocks here to be multimodality, your ability to process audio, video, image inputs, including textual input, reasoning, tool use and coding. My hypothesis is if you master all four, multimodality, reasoning, tool use and coding, you can do anything in front of a computer. What we are doing at Turing is massively scaling up the data factory that can generate the data to power the models to automate every cell in this four dimensional matrix.

• 4:21Speaker 00

  And you can think of a human doing a role as a composite of workflows. And you can think of a function as a composite of humans in that org chart. And you can think of a company as a composite. Of all the functions and we have these two wonderful learning paradigms for imitation learning and reinforcement learning both have very different data needs. So we've built a talent platform and on top of that we've built a data platform that powers generating data for imitation learning and RL and reinforcement learning for every cell in this four dimensional matrix. So we are hiring investment bankers, doctors, lawyers. We have to replicate their entire work environment. When we get into RL, if you're interested in RL, we can talk more about it later.

• 5:01Speaker 00

  But it's such a fascinating area and these models are so data starved today because all the easy gains from eating the internet, eating all the books that humanity has ever created, that's done. That knowledge has already gone in. And companies like Turing are creating data with these expert humans working inside realistic work environments to generate data for these models for these different learning paradigms. And who knows what new learning paradigms will invent in the coming years. And if you think of just a simple consumer task, if I say to my assistant, hey, go buy some shoes for me. She knows how to go and go buy some shoes for me and she can go do it. If I say to an AI, go buy shoes for me, even if I tell exactly what shoes,

• 5:43Speaker 00

  I don't get the shoes. It doesn't know how to get my credit card, doesn't have to go log in and it's so simple. And so why can't even something like the basic stuff get done today? Are you into Star Wars, Arun? Yes. I was about to say your lack of faith in deep learning is disturbing.

• 6:02Speaker 00

  But you're asking a good question. The answer is not yet, but they absolutely will. Oh, for sure they will. But I'm so mad that it's not here now. Maybe by the time we release this podcast, it will be. I know it's happening fast, but. I'll get to your question of why these models are not able to do today. But let's back up a little bit to see how these models are actually trained at a high level first, and then we can see where the gap exists. These models first, there's the pre-training step where these models are fed the entire internet. And the model is building a base model that learns how to autocomplete. It's learned some rich representations of the world. And then there is post-training. In post-training, there's supervised fine-tuning and then. There's reinforcement learning where we hire human contractors to teach the model how to have a conversation,

• 6:47Speaker 00

  single turn, multi-turn, and then humans are generating data to train a reward model that helps the model know what's the type of completion a human will tend to like. And for something like code and math, which are these verifiable domains, the model is capable to do reinforcement learning where it can figure out which paths tend to result in a reward, which paths tend not to. For the example that you gave, it's actually not as simple. You ask the EA to go buy shoes. Think of all the implicit knowledge your EA has. Your EA needs to know what Orin's foot size. That's why memory is important, or having some access to knowing what are parameters of the shoes that you have to find. Second, ideally, the model also needs to know what are Orin's preferred brands.

• 7:30Speaker 00

  Maybe Orin likes Nike more than Adidas. A little bit of that would be good to know. But then the model needs to master tool use. In this case, Go to Amazon or wherever, go to browser, go do this stuff. Use Google, use Amazon. And that's where it's really terrible today. It just really falls down often. And a good human will also know that inside Amazon, this seller looks a little bit sketch. Not enough ratings, two star. This seller is much more reputable. You would do a lot of research and then maybe there is an Adidas store that's open near DC. If the model's really smart, it might then go and download maybe the Honey browser extension to see if there are some discount codes or coupons. And then it would need to know where does Orin live so that it can ship it there.

• 8:15Speaker 00

  Maybe have a conversation with the person who's selling the shoes to see if there's a sale coming up nearby. Maybe it's figuring out a lot of stuff. But everything I shared with you, Orin, again, goes back into the four pillars for super intelligence that I mentioned. Multimodality. This requires looking up images of shoes, understanding what's on the screen, talking to a human. That's multimodality. Reasoning, to be able to come up with a plan. What's my algorithm going to be? My algorithm is going to be, let's figure out Orin's shoe size. Let's figure out Orin's brand preferences. Let's figure out Orin's budget. Let's figure out how soon Orin wants it. You have to know what questions to ask. A good EA will acquire additional context before going off and doing things. So that has To be taught to the models. And then there is tool use that I mentioned. And then coding, you might wonder, why does coding need to happen here? Let's think for a second.

• 9:00Speaker 00

  I don't know, and this is a bit contrived, but maybe there is seller A and seller B. Seller A has lots of reviews, but maybe in recent times, the reviews have not been really updated. But like an up and coming, there's one person rising with like a high slope. Imagine like Yelp's rating algorithm to figure out what's really good. So maybe to write some Python code to figure out which seller to prefer with a complicated algorithm. Or what to buy. So the models cannot do this super well immediately right off the bat. Because in the internet tokens that these models have seen, this reasoning chain simply doesn't exist. Unless there was an E A that wrote a blog post on how do I get shoes for Orin? It doesn't quite exist. It's interesting because I found that just in the last month or so.

• 9:41Speaker 00

  Gemini has gotten really good at complex travel things. Complex flights, complex other things. I still don't think that the OpenAI or Anthropic models are yet there. So you can kind of see these things just slightly moving. The problem is for people like me, and I think for probably most of the audience that are using these AI tools is even every day as the AI gets better. Every day our expectations of the AI grows faster. So there's this delta every day of what we expect versus what it can deliver. So it's constantly underperforming because we expect it so much better than it is. You're exactly right, Orin, and I would say that today we are actually hiring EAs as well.

• 10:23Speaker 00

  So the problem you're talking about, trust me, it's getting solved over time because there are many EAs. We're distilling their human knowledge and skills into these frontier models. And the same reasoning chain that an awesome EA would go through to know what questions to ask you before starting off, what tools to use, how do I analyze results? It's all going to get baked in, it's only a matter of time. In that four dimensional matrix that I mentioned, I could have had one more dimension for consumer. And under consumer, there could have been travel or personal shopping or some of these other verticals. But it all boils down to multimodality, reasoning, tool use and coding. If you teach the models that, you get everything. To your second point, Orin, on how as these models keep getting better, our expectations. Of them keep growing.

• 11:06Speaker 00

  I'd say there is a flip side to that, too, which is when we work with enterprises, we see significant model capability overhang, where the models are actually capable of so much more than what enterprises are pulling from them today. If you pick an underwriter working at an insurance company, who has to process unstructured information and convert that information into a score for this person as high risk, medium risk, low risk, this is what the insurance premium needs to be. There's so much that can be done with good system prompting and user prompting. And then the next step being teaching the model how to use different tools. And the next step being maybe a good RAG system that the model could engineer its context to be better. The next level being supervised fine tuning, the next level using reinforcement learning to post train.

• 11:51Speaker 00

  There's so much stuff that these models are capable of, and we are still scratching the surface. So I'm excited for the reverse direction, too. I personally believe if all the frontier AI labs halted AI progress today for some crazy reason, there is still a ton of economic productivity will unlock from work. Absolutely. Even people I know are like really at the frontier of pushing their organization. We're still less than 10% of what we could be doing just with the current stuff. Just because it's all so new and we just haven't had the chance to retool our companies and retool our processes. One thing I find really cool, Orin, it's kind of like intelligence is an API. That's what this is, intelligence and API. But it's a weird form of alien intelligence that's different from human intelligence

• 12:35Speaker 00

  in that it's a composite of different humans in different domains. For example, when you're starting a tech company, let's say you've hired a designer to design your web app. Normally that designer will probably spike in design, but that's about it. But when you're working with a frontier model, you get the best designer's intelligence and the best UX web developers' intelligence. And then the intelligence of perhaps a psychologist or a marketing or persuasion expert. You get all of that baked into that API call. So it's amazing that we have this composite intelligence as an API. That's human intelligence in every dimension baked into one set of weights that we're getting, which is remarkable. It's kind of neat.

• 13:21Speaker 00

  That's super cool. One thing just really irks me is how bad recommendations are. You would think, okay, one thing we can do pretty well is recommendations. And it seems that Meta and Google, Google on YouTube, Meta on Instagram ads and stuff like that. They do a really good job on recommendations. And everyone else, and maybe because they're developing the technology natively, but everybody else from Amazon suggesting what we should buy to Spotify suggesting what we should listen to, to Netflix suggesting what we should watch. Everyone else, their recommendation system is absolutely atrocious, and they don't actually do a very good job today. It doesn't seem like it's that hard. Is it just because it's just not in their business interests to do it?

• 14:05Speaker 00

  Or why are the recommendations, even just ads that you see on the internet are so bad? Why have people not moved to make their recommendations better? You're right that the recommendations inside apps don't seem to have improved in any significant way. I've been pleasantly surprised by how good LLM-based recommendations are with in-context learning. So today, when I want a recommendation, here's a weird query that I do. I basically do in-context learning with a few short prompting. So with a frontier LLM, let's say using GPT five point two, I would give it examples of, hey, here are some cars that I like. Here are some cars that I don't like, or here are some books that I like. Here are some books that I don't like.

• 14:47Speaker 00

  Give me recommendations and give me recommendations that I otherwise might not have discovered. So I give all sorts of crazy constraints. And I've actually found those recommendations to be much better than exactly, as you say, recommendations inside these siloed apps. And there was one researcher I met at a frontier lab, who told me when they worked with one of the music subscription services, when they used an LLM to do recommendations, those recommendations were way better than just looking at features from the song. There's the old school supervised machine learning way of doing it. Come up with features, do supervised machine learning, use collaborative filtering or support vector machine. I think a lot. A lot of them are still just using collaborative filtering.

• 15:31Speaker 00

  I don't know why they are, though. I think it's just inertia. I agree. Maybe it's just they don't want to throw the computer at it. I don't know. Because collaborative filtering is cheap and they don't think people care or something. One of the other things that's going on right now, and to me, this is very interesting, is everybody, everywhere in that AI stack is running like crazy. Everyone is competing. Even if you think of Nvidia, they're just moving at such a fast pace. And they have all these really amazing competitors who are competing as them. They're all moving fast. If you think of OpenAI and Google and Anthropic, they're all just moving at such a crazy pace. And they could be irrelevant a year from now because it's just such a fast game. And even if you think of Turing, same thing, you guys are constantly moving up the stack.

• 16:14Speaker 00

  Maybe back in the day, we were trying to decide, okay, is this a cat or not a cat? Or if you think of a long time ago, when we're doing the supervised learning. And now, if you don't move super fast, you could be out of business in a year. How is it like just competing in this world? This is not like the old SaaS world. You can see me drinking a red bowl. It's 1:00 PM. Conventional wisdom might say, don't have caffeine afternoon, but whatever. I know that you haven't taken like almost any vacation in the last seven years. That's just the life we're in. If you want to run a company like this, it's insane competition. Insane competition. And when I think of how to reconcile that with Peter Thiel's view of avoid competition, competitions for losers, build a monopoly.

• 16:59Speaker 00

  Doesn't seem like there's anything like that in the AI world yet. Maybe Nvidia on the side or something. So at Turing, I think very intentionally about how to avoid competition and how to build compounding loops. And one loop we have is because unlike other data labeling companies, we are actually not a data labeling company. I see ourselves as a super intelligence accelerator like Nvidia. And what we do is we are in frontier AI labs as well as we are also in enterprise AI. Serving the frontier AI market teaches us where the frontier is and makes us an excellent partner for enterprise AI companies. And being in enterprise AI actually shows us where the model Break in practice, like your EA example, where the models actually break when they don't do the job,

• 17:43Speaker 00

  which informs us on what type of data needs to go to Frontier AI to help them improve. So one reinforces the other and we have this data loop that this platform is amplifying. Both of this is powered by our talent platform and our data generation platform. And I think the way you compete is by having like a really high learning rate. You have to learn faster than anybody else. You have to have an organization that has a high learning rate. Just to dive in your model, you have a person in some ways, or you have a system that's interacting with your corporate customer. Let's say it's a auto company or something over here. And somehow you need to like get that information from that customer, which could be an offhanded remark or could be watching you use the product or something.

• 18:28Speaker 00

  And then you have to get that to your other side where you're dealing with a frontier. That is not an easy thing to do. So it's a little simpler than that. For example, Lauren, one thing that we're doing is for some of our financial services clients, we are working on building systems that automate the end to end workflow of a private wealth manager. And even there, it breaks down into multimodality reasoning tool use coding. There's a sequence of tools that they're using and stuff breaks. And it breaks in really dumb ways. Maybe the model couldn't understand a PDF document where a table spanned multiple pages. Or it couldn't quite use a certain tool that a human private wealth manager was using. Or when it was looking up data from a dashboard, it was making some mistakes. Us knowing how these things break in the real world

• 19:13Speaker 00

  helps us figure out what type of data will these models need at a higher level of abstraction? What type of failure patterns exist that we can package up into data sets? Obviously not using any of the enterprise's data, but we know that there's a failure pattern with reading PDFs this way. Or when your output has to be a presentation, it makes these dumb mistakes. Or when you have to do some math to do DCF to value companies, you're making some mistakes. It teaches us what type of data gaps exist and helps us come up with data sets that actually help AI models do real world knowledge work well. There's this interesting paper, did you get a chance to read? The GDP ballot paper from OpenAI. I don't think so. So I'd recommend it. I'll text it to you. It's got a really good piece of work on evaluating how good the frontier AI models are today

• 19:58Speaker 00

  at automating real world knowledge work, where the artifact is something that a human expert actually creates, like a civil engineer creating a blueprint for the building they're about to build, or a movie production scheduler coming up with a schedule for how they are going to shoot a movie, real artifacts. And I think it's a good step in the direction of how do we get these things to work in the enterprise, to actually do real work. If I were a pessimist, I guess I'm not a pessimist, but my spicy take is last year, if you look at what did generative AI really transform in the enterprise, I would round it down to almost zero in terms of stuff that actually meaningfully changed the P and L of companies. And I don't think it's the model's fault. I think the models are insanely capable.

• 20:42Speaker 00

  I think there's a lot of first mile schlep, last mile schlep involved in building these systems to work. For every single workflow, you have to build a cursor for X that's designed for partial autonomy, have the right human back offs, have the right set of evals, and we are building that. So our goal is to like get these models to work in the enterprise, learn from that, use that to figure out what type of data can we sell to the frontier AI labs to help them improve. And as the models improve, we're going to go solve real world problems with it. So on one hand, we're like an input to frontier AI, and on the other hand, we want to drive super intelligence deployment. You've talked about how SaaS is dead. Walk us through what that means, what's the implications of that, and what comes next? There are two ways in which I think SaaS is dead.

• 21:27Speaker 00

  The first way is, so if you think of typical SaaS software that most users use, it was built in an era where complex software was hard to build. In many cases, AI powered applications were hard to build. Imagine if you were building a chatbot to do sales support on your website. You wouldn't dream of building it yourself because the company that's building it probably has to raid Chris Manning's lab at Stanford to hire some computer science NLP PhDs. They'll spend six months collecting training data. They'll do some feature engineering. They'll come up with some different models that they'll test. They'll probably do some active learning and continuously keep updating these models. It'll be a big R and D effort. And it's amortized. The benefit is for customers of that company to just plug that in and use it.

• 22:11Speaker 00

  But now with all applications being AI applications or benefiting from AI in some way. And the intelligence being pushed down into the foundation model layer, the weights of the foundation models people build. It's become easier than ever to build LLM applications wrapped around these models. So as the cost of creating custom software tends to zero, I think more and more what we'll see is software engineers working at companies, building these internal tools themselves, building lots and lots of custom software. It'll be easier to build. They can customize it however they want. And there'll be less of a need to buy third party software. So that's one.

• 22:53Speaker 00

  Second, I also feel like a lot of SaaS software was designed in an era of human consumption, when humans are navigating menus and GUIs and taking action. But in the era of computer use agents and tool calling agents, where an agent is using these software. I don't think these tools are really designed for that. I mean, computer use agents will be like a good bridge, but I think we'll see a lot more tool calling agents that automatically call the right APIs to take action. So I think the current way in which these products are built is going to change. And there's so many of these SaaS companies out there. What would your advice be to these CEOs and product leaders at these companies?

• 23:37Speaker 00

  Is it just to give up and join a company like Turing? If you look at the space, on one hand, the models are becoming more and more agentic. So every frontier model is becoming more agentic, where it will take action, execute complex multi-step workflows in a real world business setting by itself. Especially if it's a computer use agent or a tool calling agent, it'll start operating your computer and being you. So some things will get blitzed by the model layer. And then there is the other side where my hypothesis is these enterprise companies will have smaller models fine tuned on their own proprietary data that distill their own human intelligence into the models with companies like Turing, helping them build end to end AI systems.

• 24:21Speaker 00

  Their own internal versions of chat GPT for like automating different workflows. So that will happen. We'll arm them as well, the enterprise AI companies. So that leaves the middle layer, like when somebody is going outside for like a point solution for one specific agent to automate it. It feels like a red ocean to me. Just when I was driving on the one on one, like I see five billboards for AI to automate sales or like automate your SDR. It's just a red ocean. So I would take a leaf out of the P two T L playbook. You know, Peter very well. If I were channeling Peter, my guess would be like, he probably doesn't look at a pitch that says agents or an AI agent for X, he probably would filter it out. One of the things about these SaaS companies, some of the big ones right now, and maybe it's like a short term way of shoring up revenue.

• 25:03Speaker 00

  They're not allowing agents to interact with them. They're staying away from things like MCP. They're trying to double down on that model. Maybe because they have to meet their quarterly goals, et cetera. But it does seem like you can only hold a dam so much before it bursts. I saw this interesting company called Humane One. It's a company based in Saudi Arabia. Where I saw a demo of a system where their approach is basically to own the system of record. If you really think of a SaaS application, it's like a UI layer wrapped around a database in most cases, Salesforce, Jira, stuff. But if you controlled the ground truth data, could you have a collection of agents that just interface with the human?

• 25:44Speaker 00

  And what would that clean slate approach look like? For example, if I were looking to kick off a rec for hiring for reinforcement learning software engineer. Today, what I would do is I would talk to our head of recruiting. They'll say, hey, do you have a JD? Create a JD and then they'll have it. They'll talk to a recruiter. They'll spin it up. They'll start a search. But in the future, the way it could work is I could talk to an agent and say, hey, I need a RL engineer. It'll ask me, okay, sure. Let me look at Turing's knowledge base and write a custom JD for you. And the agent already knows what type of RL. Work we do, we build our own environments for automating enterprise workflows. So it could create that JD and then it could ask me, hey, can I update the website with this? Done, boom, it's updated the website with this. And then it could say, I'm scanning your email now.

• 26:29Speaker 00

  Whenever you get anybody apply for this role, I've also pasted it on your LinkedIn. Again, tool use, it's maybe gone to my LinkedIn and pasted the job. When somebody applies, it'll automatically archive the email and the agent will tell me, hey, Jonathan, there's a person ready to interview. Shall I have them take the first screening test? And they could take the first screening test. And after that, hey, this person is ready to chat with you. I've set up something on your calendar. Do you want to talk to them tomorrow? Yes. In this whole interaction, again, it's the same for building blocks or multimodality reasoning tool use and coding. I didn't use greenhouse. Heck, I didn't even use a whole recruiting function. I just spoke to an agent. As long as it's maintaining a database state of candidates who's in process. You have to have a data schema that's tracking who's applying, what are the different jobs? Today, I'm sure for many companies, it sits in crazy spreadsheets or a greenhouse or an ATS.

• 27:12Speaker 00

  This is what I mean by SaaS is going to die. That greenhouse lever type company, I don't know how they'll survive. If somebody goes direct to the source and owns the data layer. And everything that I just said, you can apply it to sales. You can apply it to JIRA, like project management. You could apply it to so many different things. If you control the data layer. So that company humane one is doing some interesting work where their approach is if somehow you could have access to the data layer, then it's just data and network of agents and then the user. And that's it. You don't need any SaaS software in the middle. And you could have your own custom on-prem version of all of this. Absolutely. Now Turing's a profitable company, three hundred million revenue. By the way, the disclaimer, which I forgot to say at the beginning is I'm a Turing investor, very happy Turing investor, by the way.

• 27:56Speaker 00

  Why have you decided to run it profitably? Was that a conscious decision? How do you think about that as going forward? So we hit profitability a while back and we used that to demonstrate to ourselves that this is a business with great unit economics. We are not profitable today. Now again, stepped on the gas to start going. Now you're starting to accelerate again. Now we know we can, we can do it. We had to prove it to ourselves. Why did you have to prove it to yourself? Why was that a psychological thing? Or what was it about that that you had to do it? So my goal with Turing is to build an enduring company that could last for decades. I grew up admiring companies like Google, SpaceX, Tesla, Apple, these generational companies that just last forever.

• 28:40Speaker 00

  So I'm not a fan of raise for the next round, and then the next round, and the next round, and then try to sell. I want to build an enduring and generational company. So I had to make sure that the company's unit economics were sound and that we could actually do this. That was the reasoning, because it wasn't clear. This industry is so new and so nascent and changes so fast. I think we wanted to do that test. Now we are stepping on the gas. Our goal here is to accelerate superintelligence in a way that drives real economic progress. The GDP of the world has to improve. And we are fortunate and privileged to work with all the best frontier AI labs that are working on this incredible technology

• 29:19Speaker 00

  and working with the best enterprises in the world that are incredibly forward-leaning in adopting it and having that flywheel between both. And I think, like I said, it's like a $30 trillion market in terms of if you look at services in general, anybody doing any type of knowledge work in front of a computer. I'm also incredibly motivated, Orin, of the idea of unleashing humanity's untapped potential with superintelligence. I feel like AI is the equivalent of the autonomous Iron Man suit from like the more recent movies where you could have digital surrogates of yourself that could go and do different tasks. Eventually, we may have physical surrogates. I feel like if you had 100x the productivity, if I had 100x the productivity, we're not going to chill in a beach and drink Mai Tais.

• 30:01Speaker 00

  We'll probably start to run 100 companies. Today, I'm running one company, 100x more productive, maybe I'll run 100. Elon might run 1,000. Companies. You may already be. Yeah, you may already be. A random question to you about branding. So I love the name Turing. I like it because it's both a name after a person like Tesla is a name after a person, a famous person who invented really exciting, interesting things. And Turing is also a verb. It's also describing something that we talk about the Turing test and maybe now that's somewhat obsolete. I love that name. Was there a conscious thing in the branding when you're thinking about naming the company in the beginning?

• 30:40Speaker 00

  Absolutely, there was. And I love the name because A, Alan Turing is regarded as the father of computer science, the father of artificial intelligence. The name is synonymous with artificial intelligence. And especially when you think of the Turing test, it touches a flavor of AI where you don't know whether you're interacting with a human or an AI. And I think that's the future that we're going to head towards, especially if you want to accelerate superintelligence to achieve real economic progress. I think in the future, we're going to get to a point where when you're getting something from somebody, you won't know whether you're getting it from a human or an AI. In some sense, the AI is going to make the human better. The human is going to make the AI better with a continual learning, continual fine tuning loop over time.

• 31:23Speaker 00

  And we're going to get intertwined. I love that as a nod to Turing with these foundations in computer science. So I like the nod to the Turing machine. It has all the things going for it, kind of like Tesla. It's two syllables. It's six letters. It's simple. It's easy. People have heard of it before. I like these names, especially names are actually really important sometimes in companies. And often you see these companies way over complicate things. And I feel like sometimes the company itself could be way better if they just had a better name. My first company did not have a good name. I did go through the loop of changing names a couple of times. It didn't quite feel right. This, I remember where I was. I was driving on Page Mul Road in Palo Alto and Turing was number one.

• 32:03Speaker 00

  And the number two name was 10x or something like that. It's also one of these things where like the gap between number one and number two is so vast. Number two and three were so lame compared to the number one. And I'm really happy we got the name too. I love it. All right. Two personal questions that we ask all of our guests. What is a conspiracy theory that you believe? I don't think AI will automate people out of a job, like in the sense of all of these, that $30 trillion of knowledge work that I've mentioned, like with multimodality reasoning, tool use and coding, you'll get there. I think it'll all be human in the loop systems. And I think people tend to conflate superintelligence with like self-driving cars where the car's driving itself, no humans required. I think humans will always be required.

• 32:46Speaker 00

  I don't know. That's a conspiracy theory. What's a really conspiracy theory they believe in? In some ways that is. It's an anti-conspiracy theory, right? Which is really good. All right. Last question we ask all of our guests. What conventional wisdom or advice do you think is generally bad advice? Micromanaging is bad. I think it's important for leaders to be in the details. I think there's a conventional wisdom that when you're running a startup company, you hire a great team and get out of the way. Or you build an exec team and operate with leverage. I think you should build a great team, push them hard, but be very close to the ground. The truth lies with people writing code and people talking to customers. You should always be talking to customers or being close to the people who write code or write code.

• 33:26Speaker 00

  So not operate through layers, but operate flat. And if somebody doesn't say that you're micromanaging or like, hey, can you back off? Maybe you're not leaning forward strongly enough. And obviously, if you have 2,000 people, you gotta pick your spots or something. I'm sure like a Tesla, yes, Elon Musk is quote unquote, micromanaging, but he's jumping into a particular thing. Is it finding there's this point in the supply chain that we've got to jump in? Or is it just walking around and seeing things and then just jumping in? Is it picking the little things or picking the very big things? Or how do you think about that with your own time? Cause you can't jump in everything.

• 34:04Speaker 00

  It's exactly what you said. It's usually in a company that is a 99-1 or a 90-10 in terms of the most important bottleneck that needs to be solved. And I think the CEO should be hands-on with the team to solve it. They don't be being in the details on everything. That doesn't scale. That means you're parachuting in to something and you need to quickly get up to speed and then you. Move to this next thing somewhere else. That's exactly right. I read this about Elon somewhere that every week in one of his companies, he picks the most important problem that needs to be unblocked or solved, and he works with the team to solve it. It's more like a T. You're not going deep in everything. It's like you're going very deep in the one thing that matters. For example, right now, I have three things that I'm focused on. Actually, it's two too many. But this week I have three things that I'm gonna solve, and that's all I think about.

• 34:50Speaker 00

  One of the other things I think these really good CEOs are doing is, if you think like the things in the company, there's these really, really big things, and it could be like that big unblock, or it could be something super strategic. And, okay, that's a great place where the CEO can be. Then you have all the middle things in the company that need to happen that are important. And then you've got these very little things. What I find is like the good CEOs are obviously doing the big things. They're never doing the middle things. And then they jump into the little things, because sometimes that's a very good way. If you see something going wrong with the little thing, that's a very good culture thing and other way of, okay, let's fix that, because that's indicative of something else somewhere else. I like that framework, kind of like a barbell, and I also think of it the same way.

• 35:35Speaker 00

  I sometimes focus on some of the little things, not because of that one thing. I think of the generating function that generated that one little defect, and I'm using that as a spot check to fix the generator. Because if you don't pay attention to that, slop can creep in across the board. So you kind of have to, the mistake. It's like that classic Van Halen contract where they talk about the green M and Ms. And so if you can just look into this one thing to inspect this one thing, then you realize, oh, wow, that means that so many other things weren't followed. One last thing, Arun, there's like a recent revelation that I had is, I think it's important for CEOs to enjoy themselves, enjoy the act of operating and executing on the company's mission. I think when you enjoy your day to day, work is fun. This is like, what a blessing to be working on this.

• 36:20Speaker 00

  I think it's infectious and it sets the right energy and tone for the company versus the CEO feeling like, oh, man, this is a grind. Let me grind through this. Let me endure. I used to only optimize for like outcomes. I gotta make sure we hit this revenue milestone, or this profit milestone, or this growth milestone, or this valuation milestone. Now I also optimize for my inner enjoyment. I want to do this while having fun doing this. And does that mean there's a certain things you just make sure you don't do? Or there are certain types of people that I need to make sure I have around me? What is that? Optimization mean in practice? So it means every day I do a check in to see what was energizing today, what was draining today? And I get curious about why this was draining.

• 37:05Speaker 00

  Is that like a calendar audit where at the end of the day, you're like, that meeting sucked? It's an energy audit. It's like a calendar, but calendar audits I do for time. This is an energy audit, but it's a very similar concept. So I actually track it in a spreadsheet. What was my enjoyment score today? And you remember that at 10:00 PM, you remember what overall your energy score was or something? Cause maybe at 11 AM it was great, but at 2:00 PM it was bad or something or? I'm sure it's not perfect, but the prompt I ask myself is what was my overall enjoyment level today? One means something, two means something, three means something, four means something, five means something. And I write it down. It's a vibe thing. So I'm going off of what I remember and I'm okay optimizing for that. And I think of what was my peak moment? What gave me the most joy today? Which meeting? Like what was I doing?

• 37:49Speaker 00

  And I also think about what was the most draining and I use that to get curious about what happened. Is it because this project is off track for this reason? Or do I not think this person is scaling? So I use it to like, get curious about how do I edit it and fix it? So every single day I'm taking a step forward to making things better. And it puts me in the headspace of being a little more present. So some people I'm sure look at their breathing and stuff. I instead look at my enjoyment levels. Okay, I'm enjoying this meeting. For example, right now I'm enjoying this meeting with you. So. That's awesome. I love it. There's a loop in my head that's going, Orange's conversation was really great. So I tried to do that and I feel like it fuels me. Because running a company is a little bit like in the NBA, you want the championship season. But you shouldn't just optimize for the time that you're lifting the trophy at the end of the year.

• 38:34Speaker 00

  I used to do that and what would happen is we had a great quarter and I'd be happy for like an hour. Two seconds. And then think of the next quarter.

• 38:42Speaker 00

  But now what I do is I'm like fucking enjoying the game. This game is awesome. Yeah. Oh, Serge is doing this. Murka is doing this. That's interesting. Turing is doing this. That's great. I like what we're doing. I like what OpenAI is doing. It's operating with a sense of. Joy and a sense of innocent wonder. I love that and it's infectious. So it's a virus. The other one's a virus too. If you're not enjoying it, it can be a virus pretty quickly as well. But just like laughter is a virus, people having fun is a virus, just getting excited, having wonder, all those types of things is when you're around those types of people. And of course, we take cues from looking up as well. So a tough day or, you know, whatever. It's both like a happiness kind of thing. You said, but there's an energy and

• 39:26Speaker 00

  maybe that's why you're down on those Red Bulls and stuff, but there's an energy level that's so important from the leader. That's right. Orin, have you read this book, Levin Rings? No. So it's a book about Phil Jackson, who coached the Bulls to like six championships and then Kobe in fact for like five. So he talks about this lesson from tribal leadership where there's five stages in culture. There's a stage one culture, stage two and stage five, and a stage two culture is superior to a stage one culture and so on. In his mind, the stage one culture is a group of people characterized by a collective belief that life sucks. And that's like some of these drug gangs and other street gangs, et cetera, sometimes have that. A stage two culture is my life sucks.

• 40:09Speaker 00

  You can think of it like the office. People have a sense of apathy. A stage three culture is a culture where I'm great and you're not, characterized by lone warriors. And he talks about Kobe and the Lakers where one person is like you against the world. A stage four culture is a team which is we are great and you're not, where the team will rise to meet the level of their adversary. And he spoke about many of the Lakers championship teams were like that, but you're defined by your competition. And he said, there's a rare stage, a stage five culture, where the team is characterized by a sense of innocent wonder and a feeling that life is great. And he bulls Chicago 98. That was like the last dance bulls. My goal for touring is to be that. The stage five culture,

• 40:53Speaker 00

  life is great, where you're not thinking about the competition, you'll give a high five to your competition, but you'll meet them on the battlefield and enjoy beating them. That's awesome. I love that. This is amazing. This is a great place to end. This is amazing. Thank you, Jonathan, for joining us. World of Das. But I follow you at J O N S I D D on X. I encourage our listeners to engage with you there. This has been a ton of fun, super great. I really appreciate you coming on. Great. Thank you, Orin. And for your listeners, if they want to learn more, they can go to touring.com and check us out too. Absolutely. com as well.

• 41:27Speaker 00

  Thanks for listening. If you haven't already done so, please subscribe to World of DAS wherever you consume your podcasts, YouTube, Apple, Spotify, and more. And please help us get discovered by leaving a review and check out worldofdas.com. That's worldofdas, D-A-A-S dot com. And of course, connect with me on Twitter, at Auren, that's A-U-R-E-N. Would absolutely love to hear from you.