You got an exciting day. We're going to tell the world about River AI and make the big announcement. I think before we talk about the company, let's talk about you. You've had a tremendous career. DeepMind, OpenAI, xAI and now doing River. And you've sort of been part of the evolution of AI and how we should be using it in humanity. But talk about just your background and sort of like how did you even get here?
Yeah, so I actually started out in physics. I was really interested in understanding the universe, you know, and contributing to humanity in some way. You know, basically using my skills, using everything I'm able to do to help everyone. But then that kind of changed when AlphaGo came out. So back in 2015, 2016, DeepMind was showing some really amazing results and I kind of realized that the most exciting work is going to happen in AI over the next few years. So I kind of tried to find a way into AI to just switch fields and ended up joining DeepMind. I'm working there on generative models, text-to-speech with WaveNet. I got really interested in reinforcement learning.
You've come a long way since then, the last 10 years. So many
Yeah, exactly. I've been doing reinforcement learning for 10 years. Yeah, now and yeah, I really loved the environment at DeepMind. I think it was a very unique kind of research culture that Demis created and it was a wonderful time to learn the ropes in deep learning and AI. And ended up joining OpenAI because I got really interested in their scaling philosophy. So it really felt like they were onto something great with GPT-2, GPT-3. And yeah, I ended up joining and helping them out on some large training runs. Also was really interested in reasoning. So I felt at the time that large language models weren't yet really good at solving reasoning problems. And so that's really changed over the last few years. And yeah, one thing that's always motivated me is I want to make sure that we have a lot of diversity in AI. There can't just be one large company that controls everything, that decides what you're allowed to do, what you're not allowed to do.
That's a very hot topic these days, obviously. You know, a lot of the CEOs are talking about this. I want to bring you back to sort of when you and I met. We talked about the motivations you had to start xAI. Maybe start there. I said like what got you excited about starting the business?
Yeah, exactly. So I just felt like it would be amazing if there was another strong competitor out there, another AI company with a different philosophy, a different idea for how to contribute to humanity, what kinds of models to build. And when I ended up meeting with Elon, we kind of saw eye to eye. We kind of saw the opportunity to do something new. And the thing that really interested me is this mission of understanding the universe. Can we build powerful reasoning models that help us understand and solve problems in physics, science, medicine, engineering. So you know, Elon was interested in basically can the models help me build a better rocket engine? So I think that was kind of unique and very worthy goal. For the company at the time, reasoning models weren't out yet, so we kind of did it on a hunch that LLMs would become much better at reasoning and it actually turned out to be the case. But when we started out, it was really just kind of this vision that, you know, we want to make this work. We want to figure out how to do reasoning. And we pursued that for a few years and actually managed to do some really great work with Grok 3, Grok 4. Those were some really powerful models that were able to do coding. They were able to do advanced reasoning. So really amazing work from everybody on the team.
And but what basically happened to me then I realized, 'Wow, this is going to change the world. These models are going to get more and more powerful. And what's going to happen to the rest of humanity?' You know, the big AI companies, they own this technology, they control it. What happens to the hundreds of thousands, millions of companies out there? What happens to every person in the world if they're not able to have full control of their AI experience?
And this is actually a place where I feel like there's a deep alignment with us at General Catalyst as well, because our whole focus is on creating unfair advantages for founders so they can actually build companies at a scale that matters in the world where you have these really large tech companies. Now, the xAI story, obviously, that became part of SpaceX and now it's a trillion-dollar juggernaut, but you saw the next phase of what AI needs to do in terms of putting power in the hands of the consumers and the users. Like, talk about sort of what's pulling you to start River AI and building the way you're building now.
Yeah, exactly. So I believe that the reason AI is done the way it's done today is because we made a lot of assumptions, we built the technology in a certain way. So, the idea of having these LLM APIs, that's the primary way that you get your intelligence, you have to pay per token. That's just one way of building AI. And we believe that if we build it in a different way, if we build it in a way that's more open, that allows people to shape and control the AI systems that they build themselves, whether it's a company that wants to basically own their own intelligence or it's an individual that wants to be able to tune their AI to be just right for them, we think that that's going to be the way that AI is used by the majority of the world in the future. So, it should be open, it should be freely available, it should be cheap, and you should feel like it's working for you and not for somebody else. And the key to getting there is to build the right tools and the right technology. That's our philosophy.
And so that's what basically we're doing. We're a research lab that's trying to figure out what are those tools of the future, what are those algorithms that allow people to take control of their AI experience.
So, there's a lot of discussion today around open weights. Jensen put out a letter, open letter, his first tweet. And also, you know, Satya and other leaders are also strongly endorsing this, obviously. Now, you started River with a premise around open weights. So, you've actually been early to this. So, talk about sort of what you saw, the importance of open weights, why you're leaning so heavily into it. Because I could actually think River is going to be looked at as sort of the foundational company that's pulling this next phase of AI development into that world. So, just talk about sort of what gave you that conviction to go on that path.
Yeah, really two reasons why we believe strongly in open models. The first one is just pure prediction. So, in the AI field, we're getting better and better at predicting where things are heading. And we just saw that open models were getting stronger and stronger, closing the gap with the proprietary models more and more, especially driven by a lot of very strong teams in China that are putting out open models. And you can sort of extrapolate, draw a line on the graph, how strong are these models going to be, how strong do we expect them to be. And you know, we already thought that they're going to get very close this year and might potentially even surpass the proprietary models one day. So I think that's an argument to be interested in them and to keep an eye on them. And the other one, I feel like it's just philosophically it's the right thing in my opinion because these models are trained based on all of human knowledge, everything that's been written on the internet, everything that's out there is sort of humanity's contribution. And I think the best way to distribute the benefits of that, to distribute the output, is to give it away for free.
So, that when we do pre-training, basically those weights I think should become available to everyone.
Now, when you think about open weights, because you also have to build a commercial business to really serve humanity in this way, what do the weights carry? And what do you think is sort of the responsibility on the commercial side for what River AI should be doing as a business as well?
Yeah, exactly. For this whole thing to be self-sustaining, we actually have to have a business model around the open weights, around the open models. So, how's that going to work if we give away the model for free? What else is there that we can do as a business that people are willing to pay for? And we kind of realized that the key has to be in those advanced tools that allow you to customize and shape the model, that allow you to kind of take control and tune it to your needs. And that can happen at the level of a business. So, somebody might have internal data that's relevant for what they're doing at their company, and they want to have the perfect model for those use cases that's small, that's cheap to run, that's very reliable, that solves the task all the time. So, we're building tools for that, but then more challenging, I think, is to bring this to the individual level. So, how can you help somebody tune and control their AI in a way that's only today only a big AI lab can do? So, we're imagining that as a user you'll be able to tell your AI, 'Hey, I'd like you to do this differently. I'd like you to write the code differently. I'd like you to interact with me differently, maybe speak more casually in the next conversation.' And the model absorbs those instructions and changes at the weight level, and actually improves in that direction. So, today that's not really possible, but we believe with the right research we can pull this off, and it would be a whole different stack for AI, a whole different world, basically. And we think bringing everybody along is the key to having a self-sustaining AI future. So, that's what we're pushing for.
A lot of discussion around sovereignty with OpenAI because so much of the high-performance open weight model work is done by the Chinese companies. And so, there's a lot of concern around can we really build a US innovation ecosystem on that? What's your philosophy around that? And like, how are you trying to address some of those sovereignty issues and resilience issues?
Yeah, we absolutely should have amazing open models in the US that are much better than the ones that come out of China, in my opinion. For various reasons, there are some security concerns, you know, what if somebody embeds behaviors in these model weights. But also just the fact that whoever releases these models ends up controlling some of the ecosystem because companies start to rely on the weights for their businesses. And so, that's the way I see it in the future, the Chinese labs might stop releasing their models or they might restrict them.
Yeah, exactly. The US might fall behind as a result of that. So, I think it's very important that we invest in
Yeah. And how confident are you that we can do that? I mean, obviously you've started the business to be able to do that. But like, you know, talk a little bit about that.
Yeah, so we've actually done it at xAI within the span of about 2 years, starting from scratch, building up a team, building up the data sets, the training infrastructure, tuning the architectural algorithms. So, I'm very confident that here in the US we're able to do that. It's just about the right incentives, building the right team, you know, at the right time. I hope there's going to be River. So, we're very excited
We're very excited to start training models ourselves as well. But for now, the focus is on the fine-tuning and customization with our River API. So, that's something that we're investing heavily into. We're working with companies to help them take an open weight model and customize it to their needs, tune it, run very ambitious training runs that usually you would only be able to do at a big AI lab.
One thing I've been pleasantly surprised around is that the enterprises, because you're in discussions with a lot of them and working with a few of them now, they actually get it that they need the scalability. Talk a little bit about just like what's the market sentiment around this.
Yeah. I think what we've seen over the last few years is that as these AI models have become more and more advanced, more and more capable, they start to subsume all kinds of industries, all kinds of existing tech companies, businesses that are out there. And I think people are becoming more and more aware of this that if they're basically just allowing the big AI labs to train on their data, improve their models, then they're sort of giving up control, they're giving away a lot of what makes their company work, and also what makes their business work. So they really should be taking back control and training their own models and basically have an internal self-improvement loop for their own intelligence.
And when you fast-forward a couple of years and you see the enterprises becoming more sophisticated in using this and obviously the consumers through the consumer companies as well. How do you see them using these frontier lab AI models that are closed and open? What do you see sort of both being used? And what's the world going to look like?
