Reflection AI raises $2B to build the 'American DeepSeek' — open-weight frontier models trained in the US

Fun as always. We'll talk soon. Great to see you. Before our next guest joins, let me tell you about Adio. Customer relationship magic. Adio is the AI native CRM that builds, scales, and grows your company to the next level. CRM started for free, super intelligence. And we have another gongworthy guest.

Let's bring in Misha from Reflection. How are you doing? Boom. Hey guys, good to see you again. Good to see you again. Good to see you. You've been uh busy raising billions. Jordy has this habit of telling people when they come on and they do a great interview, we'll see you soon. But you delivered.

I think you're the first person that roll the tape. Jordy probably said we'll see you back. We got to we got to roll it back. I bet I said you called it. Knowing knowing the progress you've made and the progress you will will make. I bet you'll be back on here with more news.

But quickly give us a reintroduction to the company and of course give us the news. Uh a quick reintroduction to the company. Uh I'm Misha. I'm the co-founder and CEO of Reflection together with Giannis. Uh we started the company about a year and a half ago and uh we were formerly at DeepMind.

Giannis was uh one of the founding engineers at DeepMind. Uh contributed to a lot of projects uh like AlphaGo and Gemini and recruited a team of about 60 I would say uh researchers and engineers from Frontier Labs. uh and the charter of the company um has opened up since we last spoke.

We are we've raised this capital to really build out the frontier open intelligence um based in America and export it to the rest of the world. So before we get that is kind of the focus of the company. Yeah. Before we get into the details, how much did you raise and who did you raise it from?

Uh we raised uh in total of uh $2 billion from a syndicate of investors. Uh oh, this this has a gong hit from a syndicate investor. Let's go. Everyone's excited. Give it up for disruptive um DST 17 B Capital a bunch of uh existing investors as well like Lightseed Sequoia and CRV and so forth.

So it was quite um you know we're we're very grateful for the support from uh the syndicate. Okay. Get uh get uh even more kind of granular with what the focus is today. I my read on it is open source is the focus is that putting deepsee out of business. That's what I want to hear finally. That's right.

Uh yeah maybe the short of it is that uh right it's US deepseek. Yes. That's what we want. Thank you. Yeah. Frontier open weight models uh that we train here. Yes. In America and export uh to the rest of the world. So this is meant to be a global technology. We have a big presence in the UK. We have a team there.

And so this is not just you know even though it's American built, it's really built for the world uh more broadly. Can you share anything that you think you'll be able to uh do to outf fox deepseeek and GPTOSS?

We're having Dylan Patel from semi analysis come on next and he's talking about uh inference max he's benchmarking and I learned so much that you know it's not just the model it's how you run it the batching the the the the different GPUs sometimes Nvidia is better sometimes AMD is better like how are you understanding because I imagine it's not enough to just say it's American deep it's got to be better so what's your plan to actually beat them it's a really good question and I think it actually falls in two parts um First, actually just having an American compliant uh deepseek would go a really long way.

Yeah. Because a lot of enterprises are basically locked out from using those models because um of various uh legal marketing provenence data provenence risks that are associated with uh Chinese models.

So uh from a commercial standpoint, just having something that is as good but kind of compliant um and built here uh would be really powerful.

Uh but of course there is um you know an aspect that you want to uh leaprog and really be the leader in open intelligence across the world and we do have some tricks up our up our sleeve. Um obviously a lot of we can't share them yet but I mean hopefully eventually they'll be out. Yeah. Yeah. Yeah.

But you know we have um some great work happening on reinforcement learning within the team. Um the other thing that you know the Chinese labs don't have access to is obviously uh the same level of chips that American companies have access to.

And so what I think Deepseek did really well is co-designing their algorithms together with the chips they had access to. And so there's some really interesting stuff that you can do with co-designing algorithms with Frontier chips that are going to be accessible to us as well. Interesting.

Um quickly talk about the business model. I could imagine this turning into sort of like a Red Hat Linux play where uh there's an open source model but you're implementing it working with enterprise working with the government and there's a contracting piece a SAS layer on top.

Is that logical or or are you thinking more like you you nail open source and then you can do a closed source model sell API you could go and own the whole token factory the inference stack like where do you see the business looking in a couple years?

Um I think that uh the primary thing you need to set first is um how do you build the kind of open intelligence open models and the sets of tools around them for um you know you partner with some inference providers you uh you know set up um you know make it easy to customize things you make it easy to build agents out of these models and um ensure that that kind of spreads like wildfire.

So I think that having an open some sets of open models that are really fully permissive is really important. Um but the pull from a from you know for this kind of model really comes from large enterprise. Uh that's when does it you know make sense for you to move from closed to hybrid to open models.

Um, it's really once you're a very big consumer of intelligence and that's basically large enterprise sovereign and scaled up startups that are spending crazy amounts of money on closed APIs. Yeah.

And so the way you kind of commercialize it, yeah, you want you want them to be building on top of your models and there's all sorts of services and products that you can uh build out on top of it to effectively solve their problems end to end because just providing an openw rate model uh is not enough.

These things are very hard to customize. These things are very hard to build evaluations around. They're very hard to do anything useful with if you don't help a customer end to end.

So I think that there's a lot of opportunity for commercialization, but you really need to be the core intelligence that others are building on before you can really um be useful at the next layer as well.

uh why uh why do you think the dialogue around open-source open AI models went from you know up to a fever pitch people demanding it then they release it and then now uh you don't hear it talked about really uh at least online in the timeline much at all clearly clearly and and and and I would say like what what I'm trying to understand is like clearly there's massive demand for open source models But I have a feeling that developers would like to be leveraging uh the technology of a company like Reflection who's dedicated to open source and and and dedicated to commit, you know, and and really committed to it.

Whereas it's hard to, you know, we we had Sam on today. We've had a bunch of people on from OpenAI. It's hard. It It'd be hard for anyone at OpenAI to say like open source is our top three priority, right? Uh maybe it's in the top five. Exactly. Exactly.

It's it's really hard for you to both be the world's open model and open intelligence provider and it for for it to be the number two, number three or number five thing, right, that your company is focused on. And the reason is that what what matters is capability.

You want highly capable open models and the only way to get that is if your commercial incentives are fully aligned with open intelligence as the first and primary thing that you're doing. Um now when you release something like this you can't just release the model.

I mean I think that that's one part of it and then you know inference providers can take that model and optimize their stack around it. But these models are so big and hard to do anything with unless you are an expert that you really need to help with that as well. Uh the models seem to be exhibiting spiky intelligence.

Where are open-source models particularly best or demanded to be best like the the customers of open of of open models what do they want to do that uh might not be as relevant in a closed source ecosystem?

I could imagine that agentic payments is maybe not the hottest thing in open source models or or IMO level math that might be maybe that's really important in open source but what what is unique about the customer of the open source model what do they want it to be best at yeah there are basically two things that as customer uh you are looking at looking to do and achieve when you when you adopt an open model um the first thing is suppose you have good performance on something from a closed model but it's ludicrously expensive which is very common then you want to drive down right the cost while keeping the performance so you want to customize the model for those tasks uh the other way around is that yeah you have some finicky data distribution that was not represented when the closed model was trained and the closed model is spiky but not on the data that you need it to be good at and so then you want to drive performance on that and so then you want to post train and customize for that so it's really you're customizing for driving extra performance or you're customizing for driving down the cost, but it's really important to have control over both.

Yeah, that makes a ton of sense. Uh well, thank you so much for coming on the show. Congratulations on the huge raise. I'm sure we'll see you back here in a couple months. Yeah, I I uh I'm so curious. I I imagine you guys are thinking about uh how you can create your own Deep Seek moment. So, uh looking forward to it.

When the time is right, come back on. We'll pump it. Looking forward. Play that eagle sound. Thank you. Have a great rest of your day. Thank you so much for having me. Of course. Great to great to catch up. Talk to you soon. Congrats to the team. How did you sleep last night? I woke up way too early.

I woke up at I woke up. I'm physically unable of a 38. Oh no. This is so bad. I have not gone. You're You're still back to back with with a 90. What you got? I got a 79. I got a 69. You beat me. There we go. You've beat me almost every day this week or maybe the last two or three. I got I got So you got your three Pete.

You got your three Pete. Got cocky. So play play some side effects. Do something. Let me also tell you about public. com investing for those that take it seriously. They got multiasset investing industry leading yields and they're trusted by millions. Please take seriously. Did you see that?

Uh marginal revolution is calling for Vitalic Buterine, the co-founder of Ethereum to win the Nobel Prize in economics. Really? I think that'd be very very cool. It's a wild card. Tyler Cowan's mentioned it a few times that Vitalik would be kind of the outside the box pick.

But in terms of advancing economic theory, designing Ethereum, I mean, it's it'd be remarkable. Uh, and uh, and he also advocates for Robin Hansen, the father of prediction markets to win the Nobel Prize. I don't know. We'll see. Um, it'll be it'll be fun to track the Super Bowl for economics grads, I guess.

Um, I want your take on this. Luke Kawa is quoting the Pepsi CEO. I think fiber will be the next protein. Consumers are starting to understand that fiber is a benefit that they need. Uh we we we put creatine in everything. We put protein in everything. Is fiber the next thing caffeine and everything?

I mean you why not just add them in? So the thing here is that uh fiber gummies. Is that a thing? Lollipop already like leaned heavily into Okay. So it might already be happening and maybe the Pepsi CEO is a little behind the times game. He might also be Did Pepsi buy Lollipop's competitor? I don't know.

I mean, Kyler Scandlin in the in the reply says, "Didn't we already do this with the rise and fall of Fiber One? " Um, so I don't know. Maybe maybe it's too late. Pepsi Co. acquired Poppy. I think Poppy includes fiber. Yes. Oh, I have a reaction for you want to keep you want to keep going on that.

You have anything else? No. Okay. Um, I have a reaction to Rune who put us in the truth zone.

So, on a previous show, we said that we said incorrectly that during the Alph Go game between Lisa Doll and and DeepMind, um, Alph Go dropped the 37th move, move 37, that iconic moment that kind of scrambled Lisa Doll's, uh, brain. Uh, we told the story such that move 37 happened.

Lisa Doll was so racked by it that he stepped outside to smoke a cigarette. Apparently, that's not true. Apparently, he smoked a cigarette before move 37. And so, it's just more cinematic to tell it that way.

I think it's uh I think it's actually maybe even more dramatic because potentially move 37 was so crazy that he couldn't even bring himself to