Reflection AI signs $1B compute deal with Nebius to train frontier open models, valued at $8B

the on his like side. It's kind of like a bang bang play. I want Spain to win.

Bang bang.

Okay. Well, our next guest is in the waiting room. We'll bring in uh Giannis from Reflection. He's a co-founder, president, and CTO. It's his first time on the show. Welcome to the show, Giannis. How are you doing?

I'm doing all right. Thank you so much for the invitation.

Thank you so much. Hopefully, we didn't pull you away from the World Cup. I'm not sure if you've been following, but uh it's one zero in case

hasn't even seen a minute.

No, he's locked in.

But uh since it is your first time on the show, I would love uh brief introduction on yourself and then I'd love to hear about the news and the deal with Nebas.

Absolutely. So, first of all, again, thank you so much for having me. Um I'm Janis. I'm the co-founder of Reflection. Uh before that, I was one of the founding engineers of deep mind joined really early and I spent most of my career there working on deep reinforcement learning research. So anything from DQN which is the first deep reinforced learning agent to ever exist to Alph Go, Alpha Zero, New Zealand RHF for Gemini before I left and uh left with Misha to start reflection where we're building frontier open models and we want to ensure that um open intelligence remains open and accessible to everyone.

So I imagine demand for the business is through the roof because you just signed this big new deal. Uh what's actually driving it? what are the what are the enterprise use cases that you're seeing who who are the key customers what's the shape of the customer base uh and then why now with this particular deal and why Nebius as a partner

yeah so I mean let me just like start by saying that uh you know building frontier open models is uh requires two things right requires like a lot of incredible talent and we've been uh extremely privileged with uh the fact that like many of the best people in the industry have actually joined us to work on frontier open models and closely uh with me. So just like really like for that at the same time the other thing that you need is uh compute right like this is what fuels uh AI research like that's that's kind of like the the the thing that like keeps the researchers busy. Uh so uh to this end we actually like uh uh you know signed this uh billion dollar deal with Nibius to just like get the computer that we need and that's

congratulations. Thank you. It's amazing. Um, I'm I'm I'm interested in uh if we can shift to history for a second, just your reflections on um no pun intended, sorry. The the uh just the progress in computer use. Uh recently uh I saw a demo of Codeex and 5.6 six soul playing Slay the Spire, this card game that I played a lot of and I know how hard it is and it sat there for five hours and played the daily challenge. Uh, and I'm wondering if are we ahead of your timelines in terms of generalization? What is your overall uh thesis on progress? Where have you been surprised based on what you obviously had a front seat to a preview of uh years ago?

Yeah, that's a really good question. I think that um uh it's actually like you know I've been in the industry for like 15 years and I've actually like seen the progress that we've actually made in the past 15 years and uh you know every time that we felt that we there was a wall or there was like something stopping us we just kind of like overcame it uh almost uh immediately. So it's kind of like really and you know this is like the best time for anyone to be doing like AI research and uh uh you know I think that I'm not surprised anymore. I think that like I've seen so many things in the past like 50 years that like it's really hard to be surprised. Uh sometimes this exponential growth or exponential curve it's like really hard for the human mind to just like fully uh understand it. Uh but uh you know just things just change extremely fast and we should just be really adaptable and really understand that uh you know AI is still on this exponential curve and incredible things are ahead of us. But once you're on once you're living on the exponential, you sort of internally like emotionally operate on the second derivative. And so you're sort of like yes, this is as expected once you've internalized it fully. But uh yes, I I I agree that it's shocking. Can you can you talk about the advantages and disadvantages between uh Chinese uh labs that are making open models versus American labs like reflection that are making open models because I imagine it's like you know you might have better access to compute maybe more capital but uh potentially

uh more restrictions around you know things like distillation that could provide advantages in some way but how do you how do you think about uh that uh competition.

Yeah. So, you know, one thing is that like it's uh it's really OPEC exactly like what the Chinese labs are doing like on the ground. Um you know, some things we know is that like maybe uh there hasn't been as much of a respect to IP and uh you know restrictions in terms of like use of data uh that you know as an American company of course we are like fully compliant. Um at the same time there are like there've been accusations I don't know if they're true or not but you know from like US labs that the Chinese labs are actually doing uh distillation at like an industrial scale from their models and uh that's not of course something that like any uh western lab would ever do. Uh and you know there are like definitely gains in terms of access to compute or like you know things that we we can get access to like the latest compute. Um at the same time I've also heard that like the Chinese labs uh you know have found ways to bypass maybe the the restrictions of compute like via getting comput from like other places. Um I think that like we have uh definitely the uh advance of like having uh you know the talent here in the sense that uh you know the frontier labs the close frontier labs are based in the United States and like many of these people have actually been to the frontier they've seen the frontier and many of them have actually chosen to join us now and just like work on our models so that's definitely a benefit um

uh yeah I think like u this is this kind of like the the world as it is. But you know there I don't think that like the fact that we are a western lab and we might not we cannot cheat means that like we won't just build the most powerful open models. I think like it's actually the opposite. The fact that the closed labs also like never cheated and they are the frontier means that like we have the talent we know we have the knowhow and we now have the compute of like all these deals to ensure that uh we you know we catch up to the uh Chinese open frontier and after that our ambition is to just like close the gap between closed and open and ensure that like frontier intelligence uh is accessible to everyone. Um, awesome. In the in keeping with uh bringing access to frontier intelligence to everyone, um, your former colleague Dennis Hassabis from Deep Mind uh, just today uh, is advocating for a US frontier AI standards body and he specifically says that he wants to uh, apply rules and uh, and uh, um, tests uh, submit models for testing uh, that would even include open- source models. Have you grappled with that? I mean, you're a large company, very successful, eight billion dollars as of the last round, and and yet as a uh as someone who can uh be a little wary of overregulation, I I don't want to slow a an exciting company down that might not have the resources to staff a huge lobbying group in Washington DC to get open source models approved. So how are you grappling with the idea of as models get more powerful uh deepening your relationship with the United States government?

Yeah. So I think like it's important uh as reflection we also like always want to engage and to ensure that like the models are um safe they're like well received and they actually meet standards. So just like do whatever is required from us. Um I don't want I want to ensure and I think like many people in the community feel the same that like there is a frontier open model in the US market. This is kind of the bedrock upon which research uh in research institutions around the country are building their solutions. Uh this is uh the tools that like developers and like early stage startups use to just like build new products and uh kind of like drive innovations. So we need to just like be uh we need we need to just ensure that there is like room for um frontier open intelligence in the United States and in the western world. Uh and I feel like this would be just a net positive for the society as a whole. Uh so you know always here to engage in any way we can just like uh uh from and also like express the voice of people who believe in open in the open ecosystem. Yeah, I think my nightmare scenario would be some situation where you're slowed down, but international firms are not slowed down by some weird quirk. And so I I hope that whatever happens, we at least get a level playing field for everyone, both closed source, open source, but also internationally. You don't want to adv advantage a geopolitical rival by accident. Um, but that makes a ton of sense. And uh I'm sure you'll uh navigate it all flawlessly. Thank you so much for coming on the show. Congratulations on all the progress and really great to meet you.

Have a great rest of your day.

You too and thanks so much.