TITV One Year Anniversary Show

22:35I think like what will be interesting is where people want to hand over the reins on commerce and actually let AI go complete the purchase versus I think people sometimes underestimate the degree to which people shop because they enjoy it and want to go look at it. It reminds me of, I mean, look, Brian Chesky has said that he's not sure if a chat bot is the right medium for shopping and commerce and, you know. There are a lot of, like, if you take, I just was at an Airbnb this weekend, and I would have searched for what I wanted fundamentally differently with AI than with the, you know, sort of filtering method that you have now.

23:13So I think you, there could get me much more rapidly to the 10 homes I actually want to look at. So I think it'll be, like, incredible for discovery for sure. and then, you know, degree to which people, again, sort of go off and say, okay, now complete that transaction versus I actually do want to peruse those 10 different homes or, you know, those five dresses, et cetera. Well, I want to thank you both for joining us. It was a great conversation. That is Gene Grosser from Vercel and Tomáš Tunguz from Theory Ventures here on TIT.

23:56Whether it's GPT or Sonnet or DeepSeek or Fable, the strength of AI models is one of the big topics everyone is tracking, irrespective of where adoption is at. And while the companies behind closed source models like OpenAI and Anthropic are battling it out, they are both brushing up against a lot of recent interest in open source models. To unpack the current state of that race, I want to bring on Jesse Zhang, the CEO of Decagon. Jesse, welcome to the show. It's great to have you here. Thanks for having me. So you've been doing a lot of interesting writing on X, a lot of great posts. And one of your most recent posts really intrigued me.

24:29It was about open source models. And you wrote that everyone is wrong about open source AI in the enterprise. What are they wrong about? Yeah, I would say to put it simply, there's a really strong narrative now that open source has taken over. It's getting a lot more adoption. You see this in a ton of metrics. you know, the usage of the models on Hugging Face, the spend on the inference providers, I mean, they're all ripping. So that is true. Open source is happening more. But then actually, if you look at the research, a lot of the stats are saying that the share of open source compared to closed source is actually shrinking.

25:08And so we were just thinking why that is because we are a big proponent of open source. So we are pretty much 90 plus percent open source at this point. Wow. Have you always been 90 %? When we started, it was all closed source because you're just trying to figure out the use case, right? Yeah. But my thesis was, hey, why is it the case that open source is growing but then the share is shrinking? Is that, simply put, open source is great for scale production use cases where you know the structure of the use case and you know the structure of your agent or the model calls. Everything else, closed source makes more sense.

25:42And so if you look at the enterprise, most of the use cases are still in that experimentation phase or still in that sort of not fully productionized scaled out phase. And that's why close horse is still so popular. So the point that I sort of read from your post was this idea that when you come up with a task for AI to do, you want the cutting edge model. You want the frontier model because you don't know how much power essentially you're going to need to complete the task. The way I read what you were saying is that, you know, we had a lot of these use cases. People would sort of thought about, hey, I think I could do this.

26:22Why not use a frontier model? Now we're at the point, though, where they're saying, okay, well, maybe I don't need that. I can use the open source models. The question I have for you, though, is that, I mean, won't the applications of the technology continue to keep going? Like, we're going to keep finding new use cases, so frontier model use, conceivably, will still continue, right? Yeah, that's my main point. So it's not that, hey, open source is going to take over, and because they're smarter now, you don't need the frontier models. I would say there's always going to be new use cases. And the other thing I wrote about is that the effort to get open source to perform at the same level as Frontier, it actually does take effort.

27:02And that's not to say the effort's not worth it. I mean, again, we put a lot of effort in, and it's super worth it. Because once, you know, we do AI customer service, and so once that use case is solidified and you're already scaled in production, then open source makes a lot of sense, because it makes sense to put in that effort to train those models. Because those open source models, especially the small ones, you get a lot of benefit out of them, right? You get better latency, which was the main reason for us. You get better cost as well. But out of the box, those models are just not good enough.

27:28So you have to put in a lot of work to make them good enough and make them be able to sub in for these frontier models in certain spots. And once you put in that work, now the results have to justify the work. And they only justify the work if you're deployed in production at scale. Most of the use cases, to your point, are not like that. So people are just going to continue using frontier models. Which models specifically are you guys using in the open source category? It's a mix. So we have Gemma 4 is the main American one. A lot of the Chinese ones are quite good, Quinn, GLM. And we have our own evals to kind of decide which ones perform better in certain scenarios.

28:07Right. How do you, I mean, you know, the race here, the open source race here, I guess, between open source models coming out of China and those developed here in North America. I mean, you know, at one point this was a race that people were paying a lot of attention to. now it's sort of accepted. I think that, you know, those models are great. People are willing to use them. I mean, if you were to sort of make a prediction as to how you see that race playing out, do you think that the issue of it is kind of behind us now and now we can just sort of use whatever we want? I think it's really up to people's choices, right?

28:40Like we've talked to enterprises where, you know, super fairly their leadership has made a decision, hey, we only want to use models that are trained in the US. because even the Chinese open source models, they're all hosted in the US. Your data is not going anywhere. So that's not a concern, but it's, hey, you know, maybe there's some bias in those models. Like we don't want those models in our system. And so that's kind of up to each enterprise. Other ones are very comfortable using open source models. Right. I think China has taken a very open source forward approach because I think that's how they compete.

29:10In the US, a lot of the leading labs, I think just if you're in their shoes, it makes sense to put your best models, you know, not open source because you make more money that way. Right. So there is kind of that balance. And you also hear from China now that, okay, maybe they want to kind of, now that they've reached, you know, frontier level in certain cases, maybe they want to stop doing open source. Who knows? We'll see. But I think having a lot of choices is just better for everyone. What do you think of Meta's decision? I mean, In Meta, initially the Lama family of models, and that was open source.

29:44Muse, unless I'm missing something, that is a closed source model. And so this is sort of a transition away from open source for the moment, which is a choice. Why, you know, what do you think make of that decision? I mean, again, I think it makes sense. If you're in their shoes, in the U.S., that's the norm, right? All the leading models are closed source. and if they're closed source, kind of you maintain more of your alpha and you get to make more money because people are paying for the model. It's like, that's like a self-reinforcing loop. You take that revenue, that helps cover the cost of trading the model.

30:22And so that makes sense, right? And again, in China, like if they did that, no one in the US would use those models because people are afraid of their data going to China. So having open source strategy there makes sense because it gets people to use your models. Give us an update on your company. So we had you on the show almost a year ago now, actually, last summer. It hasn't been that long. It's been that long. You were telling me before we started taping, so you're at 500 people now as the company? Yeah, the company's run very quickly. We're based here in SF, not too far away. Yeah. But now we're...

30:55Your product suite has expanded quite a bit, right? Yes. So when we first met, I think, I mean, a year ago, we were maybe a tenth the size. And we only started the company, you know, two, two and a half years ago. In terms of like revenue or? No, people. People, okay. Like 50 people maybe. And yeah, I mean, early on, it was AI customer service agent, right? That's the easiest way to think about us. I would say very much that's still our core focus, but we've kind of expanded the aperture a little bit to be, hey, our goal as Decagon is to help any business maximize their customer experience. So make it really easy for customers to interact with them, have a conversational agent that they can talk to at any point.

31:37And it could be for customer service. It could be proactive as well. You could proactively ping someone to get them engaged or make sure they don't churn or whatever. And so we call that concept like a concierge. It's like a new UX for the business. Right. How big is the company now in terms of revenue? We don't talk about revenue publicly, but we've been growing very quickly. So last year, we were roughly 5X'd. This year, probably want to do roughly the same. Wow. 5X in 2025. Yes. Top line. Yeah. You know, there's a category of products that we've written about here. Customer support tickets where they can actually code up solutions.

32:18I mean, that's something that we've written about in our newsletters. That seemed like something. Have you guys expanded into that, looked into that at all? Yeah. So the concept there is, hey, you have an agent that is, you know, talking with the customer, right? So that's what that agent does. Maybe it's a voice agent or whatever. And maybe the outcome of that conversation is, hey, we actually need to fix something. So then you can have a slower, maybe smarter agent that's sitting behind it that it works with to get better or to fix things. And so that was actually one of our biggest launches this year.

32:51And so we call it Duet because it's like two agents working together, right? So you're a fast agent. A lot of the new releases from OpenAI Anthropic, like, you know, Opus 4A, Fable, like these big thinking models are not actually be that useful in your voice agent because they're way too slow. They could take like, you know, 10 seconds to reply, but they're actually really good for the second agent. And so the second agent's goal is to, you know, look at what the first agent is doing, look at these conversations to your point, identify that, oh yeah, actually there is this area that we're not doing super well.

33:23Like here's how we would fix it, right? We will update your documentation or we'll change this tool that we wrote, et cetera. and its job is to just kind of almost be like a cloud code type thing for the family. And actually do the task, fix the problem. So that is something with Duet that you have implemented. So like what do people use Duet for? You know, they could come in and just be like, hey, you know, I want to build these new workflows. We don't really have good documentation for them, but here, here's like 10 ,000 previous conversations. Like you go and write these workflows for us, right?

33:53And so they'll go inside Decagon. They'll write these documents. We call them agent operating procedures to kind of program the main agent what to do. Right. And then once you're out in production, it'll actually like create tests for you. It will read those conversations, every transcript as they come in. Like no human will ever be able to do that. There's way too many. And it'll tell you, okay, this is going well, but this thing is broken. Like every time someone talks about this topic, we're struggling. So I've gone in and read all those conversations and I've come up with, oh, actually here's how you get better.

34:22And so now not only have I identified how to get better, I've gone in and made the changes. So like do the work. How do you then determine the success? You know, I'm going to evals here really, you know, evals with agents is a topic that I think people are really focused on these days. I mean, with customer service, it's sort of, I guess it's sort of binary. You either answer the inquiry or you didn't in a task like you just described, you know, creating workflows. I mean, the effectiveness of those workflows becomes a question. So are the evals with agents, Are they getting harder or are they getting easier right now with the way the technology is going?

34:59I think it depends on the space. I think in our space, evals are fairly clear because if someone comes in, you have an idea of what is a successful conversation. So you can set up evals against those. And so we have a concept called simulations in our product where you can just simulate all sorts of different types of conversations. And even before you push an agent live, you can eval if it's working. Right. Same thing. Once it's in production, you can like evaluate those conversations. So that's fairly simple. And is there are there any areas where it's getting more difficult? I mean, evals is it's it's a topic right now that and it might not be, you know, in your business.

35:38But I guess from the people you talk to, I mean, evals, this is an area that people are really grappling with right now. I think the area that people are grappling with is more evaling the base models themselves. It's like, hey, is Fable better or is MuseSpark better or whatever, right? And it's like really hard to say. So that's why people have evals, they have benchmarks. And then people can come out and say, like, actually, no, this benchmark's not good because of XYZ reason. Like, that's the tough part. All right. Where it's really tough to measure if a model is good or not. But in our use case, it's like, it's so targeted, right?

36:11It's like one agent for one use case for one company. And so that company knows, here's what success looks like. And so the evals are a lot simpler in our use case, if that makes sense. So if 90 % of the work that's happening on your platform right now is coming from open source models, what are the tasks then that you are using Fable for, 5.64? What are you using those models for specifically? So Duet, it's a good example, right? So Duet relies on the frontier models. If you kind of rewind, why did we make a push to use open source in the first place? It's because in our voice agent or our chat agent, latency doesn't matter.

36:52If you're a customer and you're coming in, you're talking to the agent, if it can be super snappy and just get to my problem faster, that's just a much better experience. You don't want to be like, okay, give me a second. And then it's just waiting for 10 seconds. So that's why we moved to open source because those are smaller models. They're faster. Those models are not going to be that good out of the box. And so you have to fine tune them. You have to train them so that they can perform the same tasks as you would need to. But, you know, it works because in this voice agent, most of the steps along the way, like one of the steps could be like, you know, choose the topic.

37:25That model doesn't need to be a big model. You don't need to know a lot to do that task. All you have to do is be really good at choosing the topic. Right. So that's why open source works there. but for things like duet where it's it's much more general purpose you're just asking it hey like go read all my conversations and like do some work for me using the intelligence doesn't make sense there right so that that's that's the trade-off i want to ask you about a couple of the more uh technical elements here so i mean continual learning recursive self-improvement i mean you're you're deep in the space we just had icml last week and all the researchers were talking all about these topics.

38:01I mean, what's your view on, you know, the continual learning problem here, the opportunity to get to there, the extent to which you actually need it to drive AI adoption? What do you think? I think it's obviously very powerful. I mean, there's so much evidence that it works. And also, you know, if you look at the work that a lot of labs are doing, there is a bit of a data limitation problem, right? Like we've, we've taken all the data that we can to, to pre-train these models. And well, where do you go from there? Well, the continual learning is great. Or just being able to kind of learn from actually the, the model's performance, then you kind of open up your, your aperture to like a whole world of, of new, uh, new type of data.

38:48And, uh, you could argue that's how humans learn as well, right? Like when you're little, you kind of get instruction of like okay here's how you like get started and whatever but then from there you throughout life you're kind of learning through your own experience and your own mistakes so you could argue that that's sort of naturally what should happen that's at the model level for us we were maybe like one level higher in the in the app stack or like we are trying to improve the entire agent as well yeah and so the type of learning there is less about improving the models it's more about how can you through through our agent having you know millions of customer service interactions for for a certain business like through those interactions figure out what can be improved and how to how to get better right um what's your view on the extent to which the role of ai researchers can be automated i think we're kind of far from that far from this is the anxiety i'm not it's not a concern i have It's just I'm telling you what the AI research is.

39:47I mean, there's more. There's always more to do. Yeah. We have a pretty sizable research team at this point. And it's a very expensive team. But we have some great people. And the reason is that, hey, we decided unless we invest in these things, right, unless we invest in training open source models, our product is not going to perform state of the art. because we're going to have latency issues and we're going to have all these things that, you know, could be done better if we actually, you know, compile the data and train it. So I think there's just going to be so many use cases for researchers to help with.

40:22I want to ask you a question. So Salesforce bought FIN. Have you guys gotten any acquisition offers? I mean, we don't talk about those either, but it's not really top of mind for us. Like we don't really, we're only two and a half years. Well, what about fundraising? I mean. Oh yeah. I mean, fundraising, we have been very active. Yeah. We've done five rounds so far. Is there a sixth happening right now? Right now? No, but at some point soon. We just did our last round not too long ago. How do you think about, you know, growth of the company? So you talked about expansion of the product suite beyond just the customer service agents.

40:59I mean, three years from now, where does Decagon, what, what, what, What does expansion look like? Is it new markets? Is it horizontal products? What's the ultimate ambition for you? So the ambition for us, as I meant, like the sort of the big vision here is, can we produce an agent or a set of agents that can just improve how every business interacts with their customer, right? So that interaction layer is what we care about. And you can almost think of it as, hey, you know, in the past, people had websites and then phones came out and everyone was using them. So then a new UI was created, which is the mobile app.

41:37Now there's almost a new UI after that, which is a conversational agent. It's a new way to interface with the business. That doesn't mean mobile apps and websites are going away, but it's a new way to interface with the business. And so that's the vision that we're building towards. So what does that mean? Well, one, we have to launch a lot of different types of conversations that an agent could have, not just inbound customer service, right? So that's where a lot of our product development is. there's a lot of other products you can build in this in this theme of making that you know interaction better being able to understand the interactions well being able to auto learn as you were saying right so and one of our big launches with duet was this concept of duet autopilot where right instead of you even having to go in and ask at things it just prepares a report every morning you just wake up you log in and duet tells you yeah you know i read all the conversations on the last day, and I put together this report of what's going on.

42:35So those are the things that we're launching, and we just think that that vision is just so large, and there's so much to do there that you're not going to run out of things to build anytime soon. Great. Well, Jesse, I want to thank you for coming on. That is Jesse Zhang, the CEO of Decagon, here on TI-TV.

43:00The AI talent wars over top researchers is likely to be another dominant story for many months to come. The core reason, of course, is that there are still some meaty questions AI researchers still need to solve. Can software cut inference costs? How quickly can agents improve themselves? Will recursive self-improvement become a reality? I want to bring on our AI and robotics reporter, Rocket Drew, for a conversation about all of those topics. Rocket, welcome to the show. It's great to have you here. Thanks so much. It's great to be here in person. So fun. I should say, I mean, I'm the visitor.

43:33You are based here. That's true. That's true. You're on my home turf. There we go. Okay. So we just talked to Jesse Zhang about open source models. Okay. We touched a little bit on recursive self-improvement. I mean, this topic of continual learning is one that we've been trying to wrap our heads around on the show. Can you help us understand what is included in continual learning and where is the research actually at right now? Yeah, absolutely. That's a great place to start because continual learning is a super hot topic on the frontier of AI research right now. The basic idea is that if you have an AI that can continually learn, then it learns on the job.

44:08It picks up new knowledge, new information, new skills from experience just as it goes about working and completing tasks for you. And we're talking about the models or the agents right now? Well, that's a good question. Let's say the models. Let's say we're being kind of purists. We really want the model to be able to adapt and flexibly pick up new information. Now, there are kind of workarounds right now to get your agent to sort of continually learn. And I'll explain that in a second. But I think some people might have a reaction to this idea that's like, well, can't chatbots already continually learn?

44:39Doesn't ChatGPT learn more about me as I use it? Right. That's true to some extent. But as a reminder, the way AI models are trained these days is you have one big phase of training, and that's kind of a discrete period. and then you just use the model. And basically the weights of the model, the brain of the model, are frozen while you're using it. I think of it like in stage one, you're forging the tool and in stage two, you're using the tool. Continual learning reimagines this. It says, what if the tool could adapt and change as you use it to become better and better at the jobs? Now there are some workarounds, but they're very, they're kind of janky.

45:15Frankly, I think that's the way to put it. Like when ChatGPT learns more about you over time, it's leaving little notes to itself that it can reference at a later date. And like you mentioned, some of the agents have workarounds as well. They can leave notes to themselves, remind themselves how to solve a task, but it doesn't feel elegant. It's not like the way humans pick up new knowledge on the fly and from experience. It's like you hired an intern and the intern works for a day. And by the end of the day, it's a little bit better at its job, but then it leaves a note to itself and it wakes up the next day with amnesia, like Groundhog Day, and comes into the office and reads the notes that someone else left for it and tries to remember how to work, it doesn't really feel like the true solution to learning on the fly and flexibly incorporating new skills.

45:58So where are we at then with continual learning? We can leave these notes. Are there applications today that incorporate continual learning the way researchers want? Or is this something that we are working towards? We're still working towards it for sure. I think right now we have these little workarounds and they will continue to get better and better. Some people think we really do need to go back to the drawing board and rethink the way AI models are designed and the ways that they're trained in order to really fundamentally improve their ability to continually learn. We have some promising directions to explore, but there are big limitations right now with all of the ways that we try to do continual learning.

46:41So, for example, you could take a bunch of data from the AI doing jobs for you, and then you could fine tune on that data. But this is a really blunt way to go about it. You're just getting the AI to imitate everything that it's done and just sort of regurgitate it. And you also run the risk of the AI forgetting things that it used to know. When you say, what do you mean by imitate? As in just like repeat the things that it has done before rather than like flexibly incorporate new skills into the way it works. So, for example, think of, it's like pretend, pretend continual. Exactly, exactly. Think about it as like I can write down all of the moves that a chess grandmaster makes and I can sort of fine tune on those moves, but it's not going to make me a chess grandmaster myself.

47:28I have to learn the principles behind the moves that the chess grandmaster did and learn the reasons they took those moves. That's what it would really take to have continual learning. Okay, so now let's go back to recursive self-improvement. Is that a part of continual learning? Is it super continual learning? Where does that fall into it? I would think about them separately to start. There's ways that they're related, but I would think about them separately. And what I mean by that is even if we don't solve continual learning, so AI that can learn new experiences, learn from its experiences and learn on the job, you could still have recursive self-improvement, which is the idea that you could have one AI model create its successor to train the next generation of AI without any humans involved, no humans in the loop at all.

48:12The entire research loop is done just by AI. You could do that under the current paradigm where we still have one training phase and then you put the AI models out into the world. The way that they're related is they're sort of both paths to creating a super intelligent AI. As in, you could imagine that instead of recursive self-improvement, you could have a bunch of AI models out there in the world. They're working with different enterprises. They're working with different organizations. And they're picking up all sorts of new skills. And then somehow they're accumulating that knowledge and feeding it back into some more central model.

48:46That's a way that models could also get very good very quickly, even if humans are still involved in the loop of training them to some extent. So I want to ask you then about some of the trade-offs here of having continuing learning, achieving recursive self-improvement. Because, I mean, look, we talked with Stephanie yesterday. She came back from ICML, and she was telling us about how AI researchers are getting nervous that, hey, I mean, first it was AI coders that were nervous that AI could do their job. Now researchers are saying, maybe they could do my job as well. So you have a bit of anxiety there.

49:22On the other hand, you have the issue of AI adoption and sort of some excitement that, well, maybe there's an opportunity here for continual learning and stuff like that, these advancements to improve adoption in the enterprise. To what extent do you think that achieving continual learning will help improve adoption? And to what extent do you think it might put AI researchers out of a job? I mean, two meaty questions. Definitely. I think AI researchers are right to be a little bit nervous right now about being automated. To your point, coding is just one piece of what a researcher does in their day.

50:01They also design experiments. They come up with hypotheses. They run experiments. They interpret the results. They generate new hypotheses. I mean, that's the scientific method, right? Coding is the piece of that that's most automated right now. You saw Anthropic recently announced that Claude is writing 80 % of their code. OpenAI recently said that of their research budget, the compute that goes to research internally, the amount spent on coding has increased 100 times over the past six months. So coding is definitely the frontier here, but AI can do more and more of the whole research process.

50:32So I think AI researchers are right to be a little bit nervous. Another way to think of this, though, is they'll be able to get a lot more done. They'll be more productive than ever. For every one hour of time that they spend at work, they're getting a thousand hours of research done or what used to take a person a thousand hours. What about adoption in the enterprise? Adoption is huge. This is the big question. It is the big question. I mean, continual learning would be so insanely profitable compared to what we've seen for AI so far. And the reason goes back to that analogy I gave you about an intern, right?

51:03Imagine AI right now is like an intern. Maybe you can teach it to do one specific task and it kind of mostly does that task right. But interns are not that useful, right? In a company, interns are kind of annoying. You hire interns not because they're useful. You're not being very tight. I'm sorry, I've been an intern. I know what I was like. You were a good intern. I'm telling you, maybe. I've been an intern a few times. I think I took a lot of work. But you hire interns because eventually they become employees. Eventually they learn the ropes and they can be productive later on. If AI right now is like an intern, it's continual learning that would let them become an employee someday.

51:40Without continual learning, the thought is maybe they'll be just stuck in intern mode forever. So you can automate specific pieces of workflows, but you'll still have to have people hand-holding them and working around them. Right. So let me ask you one more question then. You know, we talk about the strength of frontier models and the extent to which you need the full power that these frontier models have. Then you have, well, we only need the models to be to hit a certain threshold, essentially, right? I mean, this is the whole conversation around open source models too. Do we need the frontier model?

52:14So my question is, if you look at the aggregate volume of work that we want AI to do, do we need continual learning to do all of these tasks? I mean, you know, the intern tasks. I mean, it's not like everyone needs to be a rocket scientist to succeed in the workplace, right? That's right. That's how I think about it too. There are intern-sized tasks that you could have AI models handle. And for some of those, you don't even need the frontier models today. You can make do with a smaller open-source model that's maybe fine-tuned to be better at that specific task. But it feels like we should be able to do better.

52:50It feels like there should be room to have an AI model that acts like an employee flexibly and can work autonomously on long-range tasks and perform them very well. So I put this take in sort of the bucket of cope or like hopeful takes that continual learning won't be necessary to get to that point. It's like, well, maybe in practice, most tasks just didn't require that much intelligence. And you can basically carve up a workflow into these different discrete chunks and pass them off to different models. Something else that's in this bucket of kind of cope-like takes about continual learning, I think, is that scaling will just be enough.

53:28that you'll get models that are smart enough in general that when it clocks in in the morning, even if it has amnesia, it's not intern level, it's employee level. So it's okay that it's forgetting. Or we'll just take all of the experiences of the internship and we'll just cram them in a context window. We'll have a context window that's longer and longer. So it doesn't really need to update its weights. It doesn't need to change its brain. It will just be basically reading all of the notes from everything that it's done and just brute force. it will know enough to be able to work effectively. So I don't think any of these really feel like the solution to continual learning, but there are ways that we could get around that problem and not really need a true solution.

54:06Great. Well, Rocket, I want to thank you for coming on. That is Rocket Drew, our AI and robotics reporter here at The Information.

54:23physical ai is where a lot of ai companies are focused their growth efforts right now whether it's software companies building world models or chip companies hoping to power the wave of robotics that everyone is anticipating adrian mcneil the co-founder and ceo of robotics company fox club sits at the center of that story i want to bring him on for a conversation about that sector and what is going on in the current moment adrian welcome to the show it's great to have you here in person thank you thanks for having me i was at the florist just yesterday and i saw a fox club you did yes it is also known as digitality yeah digitalis florist told almost that is the company name and when i told the florist that i was interviewing you and your company was called Foxglove.

55:05She said, oh, he probably named it that and did a robotics company because it was called Digitalis. Yeah. Yeah. That would have been a better name, but the domain wasn't available. So we went with Foxglove. There you go. Okay. I have this thing with company names, right? Where like they should not be tied to like what it is that you create. Because what you create as a company evolves over time. You sort of, you pivot, you shift, but like names that you can kind of create and you just give meaning to them, right? Like what is Amazon? What is Apple? These names that you just start repeating so much that you can give meaning to a name that is not directly related.

55:36So what is the meaning of your company? Remind us, we had you on the show before, you raised$40 million from Bessemer. Just give us the quick highlight. So this is software for robots, right? Yeah, right, exactly. So we built a data platform, data infrastructure for physical AI. So this helps you with the whole data flywheel, your logging journey from robots that are out there in the production, bringing that data back, evaluating it, understanding it, sifting through that data and finding interesting, relevant events, debugging, triaging, and then ultimately building better models for your robots.

56:06And so you build your own models or whose models do you? We don't build the models, no. So we build the data platform that helps power all of this. So everything from the logging, the cloud data management, the debugging, the triage, the understanding. Most of our customers are building models. So we work with a lot of the top labs that are building pure foundation models, a lot of the humanoid companies. Like who? Which ones am I allowed to name? we'll just leave it at like a lot of the top humanoid companies, a lot of the top robot foundation model labs, a lot of the top self-driving, especially newer generation self-driving customers.

56:42And then you also have this long tail of agriculture robots, construction robots, you know, a lot. So, you know, one thing that I'm trying to sort of wrap my head around is in AI right now, you've got these AI labs that are the giants, right? And we're sort of in an interesting moment where there's a question around trust of these AI labs. You know, they've gotten pretty big. What I'm trying to figure out is, will the same story play out in robotics? Might you have these giant robotics labs, I guess, model developers that become sort of the kingpins of the sector? Yeah. And then does it turn into everyone else?

57:22Do you see the sector playing out that way? I actually think that we're going to see a really long tail of, I think we're going to see something closer to the explosion of SaaS businesses in the early 2010s. So fragmented. Fragmented, like thousands and thousands of robotics companies. Because we talk about robotics and physical AI like it's one thing. But what you're doing in a factory is so different from a drone, from a self-driving car, from sort of a household like a residential robot. There are so many different things that you need to go and tackle. And then you look at what made the big language model labs successful was the fact that you can run this giant overpowered model in the cloud that just does everything.

57:57When you think about physical AI, usually you're running on limited compute. You've got battery powered robots that are moving around. You've got like a small sort of Jetson or some type of embedded GPU. And so you're going, you don't want this giant overpowered like trillion parameter model, right? Like you need a model that is fine tuned for a particular task. And so what we're seeing is a lot of companies, there are labs doing real research on these foundation models, and that is a really important point. But I think when we see, what we're going to see in physical AI is all of those models being fine-tuned and tailored for thousands of use cases that never even occurred to you.

58:32Okay, so more fragmented. What about the open source versus closed source debate that's playing out in AI right now? Does that same debate exist in robotics? Yeah, the debate does exist. I think we're not there yet where there are either, you know, we don't have kind of models yet, either open or closed source that are generally available that are solving a lot of these problems. Like the labs are getting there. A lot of companies are making huge progress on it. But it's not like we have, we don't have kind of like a GPT-3 class sort of or GPT-4 class model available yet. But again, the debate does exist.

59:06And I think in physical AI, open models do have more of a leg up than they did in language models. Remains to be seen like which is good. Why is that? Because of this nature of needing to fine-tune and put them into particular, if you're building an agriculture robot and its job is to go around and pick weeds or something like that, it's the only thing. You don't need to know how to fold laundry or make a bed. Some of that grounding is useful, but a fine-tuned model for a particular task is probably going to be able to outperform a giant sort of centralized model. And then you also don't have the cloud, you don't have an API billing, right?

59:37You need the model to be deployed on the robot. So you don't just have this thing where I can stick Opus behind an API I'll stick Fable behind the API and charge you, you know, hundreds of dollars a month to use it. Right. Okay. So the other debate that's playing it in physical AI right now is VLA versus world models. Where do you stand on that? Right. You got to pick a side. Yeah. Yeah. Do I have to pick a side? So for the longest time, the VLAs were basically the only game in town, right? So this was taking big, large vision language models, especially open vision language models, and then fine tuning them with action data.

1:00:12So creating tokens. Most big LLMs or VLMs today are creating their auto aggressive models. So they're creating a sequence of tokens, tokens in, tokens out. And then we basically add some action tokens to that. So we're fine tuning that model based on action data. but ultimately the model was grounded in in the world of um of images like still image cams and text and that's what the model kind of knows best and then we have this kind of other camp which is relatively new entrant i would say where we have world models which have been grounded and they've been grounded in action data a lot of them being trained on simulation and things like this um and then we're trying to add in action tokens to that and we're saying hey what if we take this uh like world model that's good at predicting what would happen if i if i drop something and it can create realistic videos of what the next frames would be.

1:00:56Can we ground that in action data so that we can actually pick up tasks and manipulate them and things like that? Neither of those approaches have won yet, basically. Which do you think is more likely to win? World model is a little further away. A little further away though, right? Yeah. The space moves month to month, honestly, like it changes and people are having success with both approaches. But I think the world model approach is certainly interesting, right? Like, even maybe a more interesting approach that you're starting to see is some companies like, I think, Generalist are one of the ones that most recently came out with a model that has been trained from scratch as a robotics foundation model, right?

1:01:33So they're not even starting with a base of an example. Who's working on that? Generalist. Generalist, okay. Came out with one. And then, yeah. So I'm thinking about cost here. I mean, which of those two, which of these approaches is most expensive? How do you think about capital constraints here with training any of these? All of them are, yeah. So the biggest challenge in robotics, right, is data. And the thing that most people miss is it's not just volume of data. There's now hundreds of companies out there trying to sell data for a physical AI, but a lot of the data is not that usable, right?

1:02:08So you don't just need a lot of general, we're not just gonna feed in, people are certainly trying to experiment with feeding in a whole lot of YouTube data, for example, but it's not just that data. We need high quality data, demonstrations of robots doing things. You know, the really important part for most companies is when you get beyond the first proof of concept, how can you put robots out there in the real world and then learn from the data that you're getting on the particular tasks? So you think about like a self-driving car example, right? Like what separates companies that are driving driverless around San Francisco today from people that are still testing?

1:02:40What was the hardest thing that took the industry, you know, 10 to 15 years to get over this hump of like taking the driver out of the car. It was the ability to put those cars out in the real world and then learn from that experience. So look at all of this driving data, discard the 99 % of it that's just me driving along a straight road, find that interesting like 1 % or 0.1 % of like weird, interesting things that happened, turn that back into learnings, turn that back into training for a fleet, and then being able to know, being able to like evaluate and validate a release to actually know that like that data had an impact.

1:03:12We've gotten better at driving in this particular, and we know that But across these like thousand other situations, we haven't got any worse. Wasn't there, I mean, there was an approach using gaming to simulate data. People always try this, right? They try it. It's not catching on. Simulation in general, right? Like a lot of people would really like it to be true that we can go build some robots in simulation and then we can deploy in the real world and they will be successful. And people are having limited success, but you have this big sim to real gap is what it's called, right? So you have like, hey, great, the robot completes the task 100 % of times in simulation, but the real world is just different.

1:03:48It's more complicated. And some things, you know, in some ways, for example, self-driving cars are actually a little bit of an easier simulation problem than some of the things that we're trying to create in general purpose robots, right? Because why? So you think about driving. I mean, it's a really hard thing to simulate. It's not easy, but at least there are rules of the road. You've got vehicles, you've got cyclists, you've got pedestrians. if you look at enough data, you can kind of predict what those different actors are going to do over time. But it's relatively structured. But then if you try thinking, how do I simulate everything that could possibly happen in a household, for example, right?

1:04:22Like I kind of, I frame residential robots as like the final boss of physical AI, because you just have so many different tasks that I can take everything from cracking eggs to folding laundry to, you know, all of these different things. And then you need to deal with model, if you try to simulate it again, which is talking about how do you model, you know, the properties of a sheet of a bed sheet or a clothing or things like this? How do I model the squishiness of a strawberry? Or these are just like really, really hard things to model. And so you just kind of have this like combinatorial explosion of complexity.

1:04:52Right. Let me ask you one last question before we let you go. Humanoids. Yeah. Overhyped, underhyped? I'm long on humanoids. I think humanoids are... But you don't need a humanoid for everything. Right. You don't need a humanoid. This is the whole thing. So the thing is, you've got to think, though, about like SKUs cost money. Right. Creating like different robots, special purpose robots cost money. In almost any case you can think of, a purpose built robot could outperform a humanoid robot. If you went to the effort of creating a purpose built robot and trained it for that particular task. And in some some tasks that we see enough of in the world, it's worth creating.

1:05:28Like driving a car, for example, it's better to just make a self-driving car than to build a humanoid that sits in the car and like drives the steering wheel. you can just focus there's enough like volume there in purpose that you can create a robot to do that but there's this long tail of tasks that it's not worth building a purpose-built robot and you think about a lot of stuff in factories you think about the the general challenge of being in a home for example where you've got stairs um and again i've people get really hung up on like humanoid specifically and legs specifically a lot of humanoid like robots just have wheels which is fine in a factory but it's still generally taking a sort of bimanual dexterous form um so there's this long tail of things where it makes sense to have a general purpose robot even though it's categorically going to be less performant than if someone went and built a purpose-built robot but you can build them at scale you can build a lot of them you can make them easily reprogrammable and that's that's going to ultimately be where we see humanoid so let me ask you to make a quick prediction before you go tesla's humanoid business one year from now where do you think it is one year from now that's pretty soon i i think humanides are um because they've got i mean they You know, they set very aggressive goals and then they have a track record of falling behind.

1:06:35You can produce a lot of human rights, right? Production is not the challenge. The challenge is getting those first robots out there into production and learning from the data. And so without a doubt, like we're going to start seeing deployments, right? But you can't scale up. Like if you have a humanoid or any kind of robot and it's achieving tasks at like a 95 % success rate, that's like an amazing result for a lab. And then you put that into production, it's a complete failure. because if you put like 100 of those in the line, every five minutes you've got to go fix a robot because one of them has failed, right?

1:07:06So you need to get an average website. Most serious websites are at least 99.9 % of uptime they're looking for or four or five nines of reliability for a lot of cloud platforms or big websites. So these numbers where we're looking at like one nine of task reliability or task completion success, it's not there yet, right? Like we need to take these good learnings we've got from the lab, deploy them into production and learn from it. Like watch what they're doing, get the data, bring it back, find the interesting edge cases and use that to feed the learning. Great. Well, Adrian, I want to thank you for coming on.

1:07:38That is Adrian McNeil from Foxglove here on TIT.

1:07:53Now that SpaceX has gone public, I've been interested to see the conversation pivot back to the technology rather than the valuations of space companies. I've also been curious to see which other companies in the orbital data center race start to make headway. Baiju Bhatt, the co-founder of Robinhood, is working on one of those companies. He is the founder and CEO of Cowboy Space Corporation, and he joins me in our San Francisco studio. Baiju, welcome to the show. It's great to have you back. Welcome, man. So last we talked to you, you were just getting around and telling us all the different places you were hiring people around the U.S.

1:08:27And you said you were hiring people in D.C., you got people in Seattle, you're here in San Carlos. Just orient me as to what is happening where right now for the company. Yeah, I mean, this is one of the fun and exciting parts about building a hardware company is you have to actually make stuff. And so going to the places where the people are that make it is kind of the path that we're going now. So we have an office in the Bay Area. It's in San Carlos. So we're 45 minutes down the road, which is where we're going to be doing the compute and the sort of data center and computer engineering part of the business.

1:09:01We've got a facility that's coming live in Seattle, which we're going to do the rocket engines as well as the satellite engineering. And then two more, which we're working on, are going to be the actual vehicle production facilities. So the places where you plug the engines into the big cylindrical tubes that are the body. We're going to talk through the anatomy that we're on here in a second. And a launch pad. So you're making the stuff in Seattle? The engines, correct. And then you've got policy folks in D.C., presumably, to help you deal with all that? Not anymore. We had some folks there, but we may have some folks in D.C.

1:09:38again at some point. Right. So, I mean, let's talk then about the anatomy of the rocket, because, I mean, you've got an interesting approach here. The way I understand it is, I mean, you basically, instead of losing the second stage and the upper, whatever, the payload stage, you're keeping it all in orbit. That's going to be the data center. Is that right? Correct. Yeah. How are you going to do that? Well, we're going to leave it up there. That's not been done before, dude. I mean, that's kind of the hardest part, right? Well, I think there are some examples from the 60s of using the second stage as the satellite in orbit.

1:10:17I mean, a lot of things have been tried in the 1960s. Yeah. But the general approach is, so you have two-stage rocket, second stage gets you to kind of your orbit, and then the traditional approach is that you dispense your payload and then you re-enter the upper stage and it burns up on re-entry and parts of it will sort of end up in specific spots where they're landing. In the case of data centers in space, the actual use case for the satellite is different. And in particular, a big chunk of what data center satellites in space do is they dissipate heat out that the chipsets are generating from the power that they're generating from the sun.

1:10:56So our approach is actually to use the mass and the volume of the upper stage of the rocket as a functional part of the data center itself. So using the actual vehicle itself as a part of the cooling mechanism is the novel thing that we're doing. And it's kind of interesting because it takes the really difficult part of reentering the upper stage of a rocket for reuse in the case of things like Starship or the space shuttle program. And it says that there's a different answer, right? So you can either go with a disposable upper stage or attempt at a reusable upper stage. Or the third one is there's a usable upper stage, which is what we're kind of pioneering.

1:11:37Why hasn't this been tried before? I mean, you know, I know you've prided yourself on doing a lot of user research, qualitative research. This is something you and I talked about the first time that you came on the show. I mean, I'm sure you've talked to many entrepreneurs who have tried space innovation. I mean, surely somebody must have tried this before, right? I mean, why hasn't this become a thing? It's the use case that data centers in space in particular generate, because even if you contrast it with something like a low-Earth orbit communications constellation, where, you know, you have a certain amount of mass and volume you can take to orbit, and then if you want to provide a service on the ground, you kind of want to take that mass and volume and spread it out over as many spacecraft as you can because they're moving so fast.

1:12:23The thing about data centers and space is the actual application for the technology is pretty unique in that you ideally want to have your chipsets, meaning neighboring GPUs as physically close together as possible, which leads you to a monolithic architecture. And also the power requirements are gargantuan. This is the reason why this whole approach is being sort of explored, which kind of leads you to saying, hey, let's maximize the power that each entire rocket launch can generate. Right. So the point I'm trying to make is that while it is a novel approach to doing the data centers in space as one big upper stage as the data center, the reason is actually the way the technology works and what the technology is trying to do, if that makes sense.

1:13:13You're still targeting the end of 2028 to have your first rocket launch? Yes, we are. And that's still the timeline? We're still working towards that, yeah. Okay. And I mean, there's a couple other logistical questions, and we'll get to sort of your perspective on space, generally speaking. How long will these orbital data centers end up being in space form? And what's the lifespan? Lifespan, yeah. So right now we're targeting about a six-year useful life, and we'll iterate on that as time goes on. If we can get more useful life out of the vehicle beyond the six years, then obviously that's beneficial.

1:13:52But it's kind of baselined against the usefulness of the chipsets and how long we think each one of the components is going to survive. And how many will you want? What's the ambition in terms of how many you have up there? Yeah, our application with the FCC is for 20 ,000 of VCs. So each one, we're baselining at about a megawatt of power. So the target is about a 20 gigawatt constellation in orbit. And if you just kind of think about our company doing this in the context of where the space industry is today, and this is kind of the really exciting part about all the data center and space stuff, is that this represents a really dramatically larger commercial use case for outer space.

1:14:34Because up until now, it's been primarily government applications as well as Earth imaging and communications, data centers in space. The mass and volume that it's going to demand is very substantial compared to those applications. Right. What do you think of Blue Origin's approach? They just raised, or they're in the process of raising outside capital for the first time. they've got this interesting approach here where they've got TerraWave and Sunrise, which is sort of this two-sided system for their own orbital data centers. Pros and cons of that approach. I mean, what do you think of that?

1:15:10You know, I know a little bit about the TerraWave approach. I haven't followed their data center in space approach that closely. So, you know, I'd have to... Tell me, have you ever met Jeff Bezos? I have not. I have not. You've not met either of them? No, I haven't. What would be one question you would ask them as a fellow space entrepreneur? Yeah, I mean, I think I try to approach a lot of the situations where I meet other successful people. I try to approach it as a student, right? I think both these folks are folks that I have tremendous admiration for. Do you think you can do it better than them?

1:15:49I mean, obviously I do. Otherwise, I wouldn't be doing it. What makes you think that? I mean, that's just kind of the entrepreneur in me, right? I mean, like I said, obviously I have a ton of respect for what they've done. I think I would be honored to spend some time with them. Haven't had a reason to yet, but we'll see if the day comes up. I want to ask you a little bit about, you know, the trials and tribulations of those companies, SpaceX, Blue Origin. Space is hard. Space is hard, yeah. There's no denying it. What have you learned from their process that you're trying to apply to how you build your rocket?

1:16:25Yeah, I think one of the most important things is kind of the role of iterating through failures, right? SpaceX kind of dramatically illustrated that iterating through failures is the path to reliability, is the path to reusability. And I think it's like when even I think when we were watching that happening a decade ago when they first got to reusability, it was kind of like, oh, my gosh, is this launch going to work? Yeah. There was a level of sort of tension, which I imagine must have been even more pronounced internally. Yeah. But it's kind of having the risk calculation of saying in order to get these parts to work in a reliable way in the real world, the only real way to do it is to test it in the real world.

1:17:09Right. There's a limit to where you can get to with simulation, iterating with computers and models. And at some level, you've got to try these things in the real world. And the reason I think this is important is because it also has an effect on how you think about the cost of these things, right? That if you're going to try to build something that's truly reliable, it kind of goes hand in hand with trying to down cost it as much as possible. Because if you try to iterate through failure and each one of your items that you're building is very, very expensive, it limits your ability to do so. But go deeper than that for a minute.

1:17:46I mean, I hear you on the iterative approach, but I mean, what have you learned from, I don't know, the engineering approach that they've taken, the teams that they've built, you know, the cadence on which they've approached it. I mean, Blue Origin has been in business for 25 years now. Yeah, it has. I mean, how do you think about how you are building your own company learning from the process that they've gone through? Yeah, I'll give you one really concrete real world example. It's kind of the role of hiring people that are fresh out of college, the new grad talent, and what to look for there.

1:18:20So I think one of the things that we've built at our company is hiring a lot of people that have done competitive racing programs in college. Wow. There's something called Formula SAE, which is a college racing program where my understanding is that teams of students build race cars and they go and race them. And getting folks that have done that kind of stuff in college, even if they've had a little bit less real world experience, like that experience of going from idea to building the real thing, to testing the real thing, to living with its failures, that that is actually like a really interesting talent pool to build the company around.

1:18:58So we've got a lot of those folks at Cowboy. and if anybody's listening, we are continuing to hire. Do you think that orbital data centers start to become a matter of national security? I do actually, yes. I think for a variety of reasons. One, I think the importance of artificial intelligence from a national security perspective is undeniable and I think when you look at that in the context of the space industry, space in general is an area of importance from a national security perspective So it's kind of the intersection of two. So I also think that as time goes on, you're going to see other countries try to build substantial constellations to do this job in low Earth orbit.

1:19:45And I think it's, you know, preserving the U.S.'s lead in this is a matter of incredible national importance. To what extent do you think that lead could be under threat? I mean, if you look at the innovations happening outside of the U.S., I don't follow the space companies in Asia as closely, but are there companies that are close on our tails at all? I mean, I think the Chinese are very formidable, right? I think that the Chinese government's involvement with it, again, I don't follow their program that closely, but, you know, the Chinese national ability to build stuff is unmatched. So I think regardless of where they are today, I think it's more of a snapshot in time and it's more like how quickly do we expect to see them evolving?

1:20:36Right. So, yeah, I mean, I think they are going to be the nation that is the most likely competitor to the United States from a volume and mass orbit perspective. Right. SpaceX went public, as you know. Mm-hmm. Did you buy SpaceX stock? I did not. what what did you make of that whole ipo process i mean i mean you know you built robin hood you've studied new issuances at large i mean what was your reaction to just the sheer size of that offering and how everyone reacted to it yeah i mean i think the sheer size of it was uh was pretty jaw-dropping it's kind of interesting because that the the week before they went public i was actually in new york having some meetings with various banks and it was kind of interesting because I think the last time, one of the times before that, that I was in some of these banks was when we were doing the same thing.

1:21:29Yeah, I think the sheer size of the thing was the thing that was like kind of jaw-dropping because when you just think about the amount of capital that's going into that and the impact of that on the space industry as a whole, it has a variety of effects, right? Number one is it really puts SpaceX as a company with the ability to build some pretty big stuff going forward. I think the other part that's interesting is it also is educating a new generation of investors, institutional investors, retail investors across the board, on the role of the space industry and kind of how that technology works.

1:22:09And that kind of goes from the public markets to the private markets as well. Right. So I think the the overall effect on the space industry is really, really substantial from it. Right. I was speaking to another guest on the show who is well versed in the space sector. And they had a question for you that I will ask you. It was a technical question. OK. They asked, so how are you going to how are you going to deal with the issue of getting the actual compute downturn at the ground link issue? So the way I understand it, I mean, cross-link is the technology that a lot of entrepreneurs have tried at.

1:22:52Right now, radio frequency is really what Blue Origin and SpaceX have had to rely on. How are you going to approach that issue? Yeah, our approach is using optical for data transmission, both space-to-space and space-to-Earth. Even though it's not been done before? There have been examples of space-to-Earth data transmission done. And on that part of the overall technology stack, our first satellite goes to orbit later on this year. While it's not doing space-to-Earth data transmission, it is doing space-to-Earth power transmission. And so in service of that, we've actually built up quite a bit of technology around the optical stack, both the lasers and the pointing mechanisms.

1:23:33and for our second mission that goes to space, which will be first half of next year, we'll actually be reusing a lot of that core optical technology to demonstrate our ability to transmit data from space to Earth. So this is a solvable problem to you? I believe it is, yeah. This is not a technical challenge that you're still grappling with, in other words. I mean, I think it is going to continue to be a big technical challenge that we have to develop because we want to take the bandwidth, that we're transmitting data and really push the boundaries on that. Right. Let me ask you one last question, not necessarily about space, but the AI race right now.

1:24:14How does that impact what you are building for space? Is it, you know, does it matter to you whether or not the models get better, whether Fable can do X, Y, Z? I mean, does that have a direct impact at all on what you're building in outer space at all? I think it does, yeah. I think it does. How? Because you see the amount of energy that the models are consuming these days. And yeah, I take your point on the energy more. I meant with respect to your ability to build what you need to build. Oh, like our ability to leverage the models. Yeah, yeah, yeah. I'm not talking about the market size. I'm talking about your ability to make rockets.

1:24:51I mean, is that at all, is that impacted by what Fable can do and not do? Yeah, absolutely it is. I mean, I think you just take a look at, I mean, there's a refrain in society that something is or isn't rocket engineering, right? Like this is viewed as very challenging to do. And it's knowledge work, right? The ability to use these models to help accelerate that is undeniable. And we see it across our company. And I see it for myself too, right? I've gone from being a finance bro. Self-proclaimed. Self-proclaimed. I don't think anybody would dispute that. to being able to build a company like this.

1:25:31And the amount that I've been able to learn through my own use of AI from asking questions about how different parts of the rocket work and getting really high quality results, obviously that I have to go and verify, it's been a dramatic accelerant, right? And the ability for it to solve more and more nuanced problems, its ability to do, you know, computations in the background to help size stuff, it's quickly becoming like an indispensable tool across the board. It's kind of amazing, honestly, especially for the math and physics stuff. Last question for you. Your dad worked at NASA, I recently learned.

1:26:12Are you going to hire him? What's the story? Yeah. So when I was a kid, my dad had a job at NASA Langley on the East Coast as a contract research scientist, coincidentally doing atmospheric sciences. So it's like you, the apple really doesn't fall that far from the tree. I think he's probably not going to be working with us too directly on this, although has quite a few opinions to share with me. It's like, why don't you do it this way? Why don't you do it that way? And he's usually right, annoyingly. Hired. Yeah. Working with your dad, I don't know if it's like, is the adventure I'm looking to sign up for at this particular stage in life, but he would very much appreciate you saying that.

1:26:55Great. Baidu, I want to thank you for coming on. That is Baidu Bhatt, the founder and CEO of Cowboy Space Corporation here on TI-TV.

1:27:14Okay. To close off the show, I want to end with a special edition of the Editor's Cut. We're going to be talking today about AI policy and the debates around that topic. I'm joined here in San Francisco by our co-executive editor, Amir Afradi, and our San Francisco Bureau Chief, Jason Dean. Welcome to you both. It's great to have you here. Hello, hello. Great to be here. Okay, so predictions for AI policy. I mean, we're looking ahead. We've done 250 shows. We're going to do 250 more, 250 after that. Amir, I'm going to start with you. Where is AI policy going? It's kind of a meaty topic right now.

1:27:52Yeah, finally heating up. It really seems like for the first year, year and a half of the Trump administration, there wasn't too much going on. A lot of the efforts by the Biden administration were kind of pushed aside. We just had a long discussion about it last week. you know the organization that tests models was really shunted now all of a sudden in a matter of a few weeks it's all back and you know there was certainly a lot of discussion around chip exports before but really nothing on the models themselves and how you know we would be dealing with potential risks on cyber bioweaponry other things like that that you know folks in the Biden administration we're really focused on, all of a sudden, out of the blue, all these concerns have come up.

1:28:38This is just going to crescendo from here because these models aren't going to get any worse. So you just imagine that every month or two or three, there'll be some new leap forward in capability. We now have the beginnings of a kind of process in place for different arms of the government to test these things, but we're not going back. This is clearly from just a technological point of view, there is a lot of reason to review these models for different types of risks. And now politically, there is an incentive, heavy incentive for much more hands-on regulation. Jason? Yes. Do you agree? Is there another version of the future?

1:29:20There could be, but I tend to agree with Amir on that. I think. But okay, hear me out here. I mean, you know, if the staggered release is going to be the norm, I mean, Silicon Valley has at least some influence over government. They try, right? I mean, I'm thinking out loud here. If there's enough uproar, enough lobbying, I mean, couldn't they have some impact on, you know, making it an even friendlier environment for these model releases? What do you think? I mean, I think they have had impact. They've had impacts in convincing the administration early on to take a more laissez-faire approach.

1:29:57I think they've had a hand in and influence over the nature of the regulation we have seen. It's not that they necessarily universally want it to play out exactly that way, but obviously Anthropic has talked about the need for government regulation. Others in the AI industry have talked about the need for the government to play some role in ensuring that these are safe and that the challenges we've seen with things like Mythos and cybersecurity are not just Wild West. Okay, what about open source? I mean, this is another battleground here in AI. We've got open source models coming out of Asia that are, I think from what I understand, better than the rest right now.

1:30:38What do you think is coming for regulation of open source models coming from that side of the world? Well, this is the topic that's preoccupied our newsroom for the past week or so. And I think a lot of people got riled up as we and others were asking questions about this because it was one of the unresolved questions coming out of the most recent executive order on AI around evaluating models before release. Something surely will happen around kind of contractor use of Chinese models just optically doesn't look great for that to occur. So you could see a situation where, you know, if you were doing direct business with the U.S.

1:31:18government, U.S. government will require you to use non-Chinese models over perceived security and other risks. So that seems like a starting point. And then the question is, okay, well, if the proprietary or closed-source models are being looked at a lot more closely for various risks and they need to be tested, what does a regime look like for testing completely open-source models? And like, if the Chinese ones are just available and their capabilities are getting quite good, certainly not necessarily at the level of Anthropic and OpenAI, but getting quite good, what are the guardrails that need to be placed on them since they're open source, they're just available?

1:31:57I think it raises a whole host of very difficult questions, just given that these things exist, they're out there, and anyone can use them. And if we don't have a U.S. business community that's completely hardened and steeled against cybersecurity threats, as an example, people could just start using these open models for those kinds of intrusions. So I think it's going to require a lot of creativity, but this is going to kind of dominate the conversation or at least be a big part of the conversation for the next few months. Right. Jason, let me ask you a slightly different question. So we've been talking on the show today about trust or lack thereof in the large AI labs.

1:32:39We saw Satya Nadella came out with a blog post, and he was talking all about the idea that it's better to have more control over your AI, essentially, in this climate. We saw Mark Benioff with the zero data retention tweet, I guess, really touting Salesforce's own technology. I mean, trust in AI Labs, it seems like it's at an all-time low in some ways. Is this a blip? Is this a trend? What do you make of that? I'd say it's a trend. I think so. It's not coming back. Well, I don't think it's binary, but I do think that part of what's happening is that these companies are growing so fast, Anthropic in particular, and there's a concern within the tech industry about the merits of the issues around data retention, but also around the consolidation of power in these two relatively young companies that are threatening the entire SaaS industry.

1:33:38And obviously, we've seen stock valuations there take a big hammering over the last year or two, and just in general. So I think that there are genuine concerns. And then there's the fact that you have a lot of these companies seeing them now as, including companies that were close partners with them and even still are, in fact, seeing them as much more intense competitors. And that isn't going to go away unless, for some reason, the AI itself kind of fizzles out and those companies are cheating. As long as they're continuing to grow their business anywhere near the pace they're at now, I think that the other companies in the industry are going to be incentivized to talk up problems.

1:34:15And I don't think that's going to help trust. Yeah, but it remains to be seen whether anyone's going to listen. If you have a fear-based message to your customers, if you're Satya Nadella or Alex Karp or Mark Benioff, you probably need to couple that with more of a value proposition than just saying like, these guys are scary, don't do business with them. Because so far people are continuing to do business with them. And yes, everyone's looking for alternatives. Everyone's looking to save on costs. But yeah, I'm not sure if the fear-based approach is going to work. And it's sort of making me think about the extent to which an IPO could have any impact at all over this.

1:34:57Because I'm thinking back to when Circle went public, for example. I mean, part of Circle's value prop was, hey, we know crypto has some trust issues, right? Lack of transparency. This is something we want to fight against. We want to go public. We want you to see our books. We think it's going to help with that. I mean, you can open your books if you want to the public, but still, the technology seems to be as opaque as ever. And so, I don't know, going public in some sense doesn't really help this trust issue at all. again it just depends on whose trust you mean if you mean the partners the cloud providers the the sas companies that are reliant on these models that's that's one thing i think for you know customers of a coding assistant that is powered by anthropic or anthropic's own coding assistant um i don't know they seem they seem okay at this point like again there's going to be limits just in terms of costs and people trying to figure out okay well is there a more efficient way to do this But as long as they have the best product, the best model, you know, going to be tough to stop that freight train.

1:36:03Great. Jason, last question for you. Next 250 shows, what do you think is the defining trend that we should be covering over the next 250 shows? I'm going to be coming to you with the question, so you better pick something that I think you have expertise about. The rise of AI and how it's affecting the rest of the tech and business world? And all the facets of that that you will undoubtedly be digging into in great depth. Well, what seems to happen is, you know, the people who are living in the future at these frontier labs, they, you know, it turns out that the stuff that they talk about today is the kind of thing that's actually going to arrive in two or three years.

1:36:45So this may not be the best example, but the first conversation I ever had with somebody at OpenAI was around recursive self-improvement. That was in mid-2023. It is now finally at the stage where people are actually discussing it as a serious thing. And it's taken three years. So I do think that so far, you know, I don't know necessarily about the labor concerns that Dario and others have articulated. But generally speaking, a lot of these leaders have been right over the long run. So you still talk to that person? Yes, I do. So 250 shows on recursive self-employment. And that is why it's worth subscribing to the information.

1:37:23because we have long-running sources that will take you inside and outside of the companies. I want to thank you both for joining us. That is Jason Dino, San Francisco Bureau Chief. I'm your Afradi, our co-executive editor. That does it for today's special one-year anniversary of the show. I want to thank you for joining us. I want to thank our guests. To mark the occasion, we have some champagne on hand, I am told. so with year one in the books I would like to thank the entire newsroom for joining us on the show I want to thank our producers Sydney and Zach for making this show a reality I want to thank you for tuning in here's to year one, here's to year two here's to every year thereafter cheers

1:38:11bye bye for now