This DeepMind Vet Raised $2B to Open-Source Frontier AI

17:25Ioannis Antonoglou:So a world, for example, can be a terminal, right? Like you can have like a bash and then you type commands into the bash and then you just get the output of these commands. And then the whole point is just like for the model to just solve the task that you asked it to solve.

17:42Yannis:right yeah i've also and you know i guess it's kind of the same but i've also heard that environments can be data or data can be environments is that accurate or is that

17:52Ioannis Antonoglou:completely off base um i mean i i think it's like data and environments are kind of like quite uh similar and um it's also like you know what the terminology that like people use depends on their background. So like, you know, synthetic data, for example, like when people say synthetic data, and that like the model can learn through synthetic data, what that means is that the model generates data, and then it can actually like, you know, you do some form of filtering, and then you can learn on the data that you've selected based on the filtering. And, you know, this is kind of like exactly how reinforcement learning works.

18:32Ioannis Antonoglou:Like the whole idea of reinforcement learning is like trial and error as in the model tries something and then based on the outcome of like what you tried it either wants to just like do more of it or less of it and so like synthetic data and reinforcement learning like inherently they're like really kind of almost the same thing at some level of abstraction and what reinforcement learning tries to do is to just like be much more efficient in the way that it learns from this data it learns from this experience. So would you say that you fall in the camp

19:01Yannis:where you are more leaning towards Transformers as the path and scaling them up? Or are you deeply like, no, we need a new architecture at kind of like the Richard Sutton, Andre Carpathi kind of path? Where do you land in that?

19:16Ioannis Antonoglou:Yeah, so we actually, we believe in the Transformers. And I understand why people, and especially like people like Reed Sutton or Andre Kapathy feel that transformers have some inherent limitations that will stop us from going all the way to AGI. I feel like ultimately it also boils down to what or how you define AGI. And to us, it is a system that can interact with a computer and it can actually do things that people can do on this computer. And it's not something that might... It might not be something that... you know, for other people, it's like a super scientist or like, you know, something that kind of like understands everything.

19:59Ioannis Antonoglou:And we have like a more maybe kind of like pragmatic kind of definition of like what AGI is. So in order to just like build a system like that, I think that you can really go a long way with transformers. You might get to a point where like you need to tweak your architecture or just like make adjustments to it. But, you know, it's not – I don't think that, like, we need to kind of, like, scratch everything and go back to square one. Yeah, totally valid.

20:29Yannis:Yeah, I think that's really interesting. And it's kind of funny how problematic the lack of a consistent AGI definition has made conversation in the AI space along the way. It's interesting you'd call that out.

20:47Yannis:So, you know, there's kind of this concept of sovereign AI with countries having this idea to build their own AI models and ecosystems. What do you think that looks like in an open source context? Do you think that enables it or?

21:04Ioannis Antonoglou:I mean, that's a really good question. And I want to say that the previous generation of technology was built on top of the protocols, like the internet protocols that were developed in the US. And in a way, the US really shaped the infrastructure and the way that the internet operates at a global scale. even the fact that English is the language of the internet or the fact that the protocols have been defined by American companies and the biggest internet companies are actually out of the US so just being the place that defines the infrastructure layer of a new technology is extremely important and the only way to do that is by actually providing the open source solutions to that, like the internet is actually built on top of open source the most important packages like the protocols everything is open source and this is I think I'd like inherently important for any new technology and I think that AI infrastructure layer will be built on top of open models and currently like all the frontier open models are actually coming out of China which is kind of like you know really good for them you know I think I'd like they have the opportunity to kind of like really shape the AI infrastructure but at the same time you know like here in the US we should also proactively be doing something about that and just building open models that can be an alternative to the technology coming out of China and this is not what's happening there are no powerful advanced scale frontier open weight models coming out of the US and reflection is building them we want to just put them out there, make them accessible to everyone and ensure that, you know, there's like a solid AI infrastructure alternative, you know, that has been built in this country.

23:05Yannis:That's really interesting because you're right. All of that is coming out of China right now. And they're good models. You know, when you take a minute and you look at Quinn, you look at what can be done with Kimmy, those are really good models. But none of that's coming out of here. Most American companies that are doing anything seem to be focusing on really small models right now. It makes actually what you did unique because you raised$2 billion on the idea of open source. And that was something that was kind of unheard of a couple months ago, I think. I mean, this was coming seven months after you raised like at a$545 million valuation, right?

23:48Yannis:So what convinced investors to double down on you and open source? Yeah, I mean, I guess the thing that's really important in order to be able to build these

23:59Ioannis Antonoglou:models is to kind of have the know-how and help the people. So the reason why our investors really believed in us is because they saw that we had attracted really, really good talent from the top labs and they saw that we could actually attract more of them and ensure that we don't just like bring the people in but like they we have them like um you know they can work really nicely and productively together and um they really believe in the mission right like this is uh the the you you know the air market is crazy you know people get like uh crazy offers from like other places and uh what it boils down to is like whether people feel inspired by the mission and the vision of what they're building and you know we as reflection have a really strong and powerful message and vision.

24:50Ioannis Antonoglou:And that has really attracted some of the best people. And, you know, I think that like we will continue to attract the best people. And this is kind of like the thing that you need. You need the resources and the people to really make use of those resources.

25:04Yannis:I mean, I guess the question behind the question I was wondering is, do you think that the mentality of open source and how competitive it could be has changed in the industry. Like, do you think that, I guess more broadly, do you think that there is a chance that open source could even surpass the big labs, like, especially if you're financed enough to compete with them?

25:30Ioannis Antonoglou:Yeah, I mean, I guess like you're touching upon like the question of, you know, how do you make money to just like keep kind of like pushing to the frontier? Sure. And I don't see why open source cannot like be at the frontier. if something, there is an opportunity here to leapfrog to the frontier by just putting these models out there and working with other research groups in the world that can provide insight or they can just provide ideas that help you move faster. So usually open science and open research leads to faster progress. That's something that has happened over the centuries. That's why there are conferences.

26:12Ioannis Antonoglou:That's why there are like journals. That's why there's like peer review, right? That's an idea. Because like that's how you progress science. And this is not what's happening with the closed models. Now, at the same time, I think that there are like ways to commercialize open models. And especially like with, you know, in the enterprise market or the sovereign market where they need to have like on-prem solutions and they need to just like, you know, kind of like almost own the intelligence. So we have many partners. I'm not going to share our go-to-market plans here, but we have some really solid ways to commercialize the open models that will ensure that we continue to capitalize and have access to the resources that we need so that we can stay and shape the frontier.

27:04Yannis:Interesting. How do you balance, with regard to open source development, how do you balance safety with speed? I know there's a couple of interesting things to juggle along the way in open source. And I'm curious, like, what's your approach there?

27:24Ioannis Antonoglou:Yeah, I mean, of course, like, we take safety extremely seriously. we have like a team similarly to other labs that is looking into safety kind of like red teaming adversarial attacks like they really battle tests our models at the same time again it goes back to working with the community and just by just putting the models out there we also like want to encourage other research groups to just like think more about like safety and exactly what safety looks like at the frontier and like models that are really powerful. So, you know, I think that even if you look at the operating systems, right, like by having an open source operating system, like eventually you get like safer operating systems because like people just like try and solve the bugs and just like contribute to the source code rather than having them like closed.

28:16Ioannis Antonoglou:So we really believe...

28:17Yannis:More people are using it so they understand the code base and they're more incentivized.

28:20Ioannis Antonoglou:More people are using it. They like bugs just like surface like faster. there are like more contributors who try to just like fix the bugs so it's kind of like really just um you know try to fundamentally fundamentally we think that like people are good people like most people are good people there are like bad actors out there they're always like bad actors but like at a high level like you know most people are like good people who want to ensure that like these models are safe and we want to just like uh you know work with them to ensure that uh you know our models are like safer than other models yeah i agree with

28:55Yannis:your thesis like i i do think that it makes it safer and i think that there are maybe perhaps more risks from the big labs keeping their technology closed i just think that's it's good that they're publishing stuff but i mean i've always felt like there's there's really a double-edged sword to it almost in terms of either approach and it's a but i tend to be i i definitely lean open source. I think the ideas of what's possible when you put a more malleable model out into the world just gets so much more broad. And with that in mind, Giannis, what do you think needs to happen for open models to one day become the obvious choice over closed models for maybe a majority.

29:49Yannis:What do you think that tipping point is maybe?

29:53Ioannis Antonoglou:So, I mean, I think it really boils down to capabilities, right? Like you want models that are like really capable. And again, I really think that I don't see a reason why the open models cannot be the most capable models. You know, the most capable Chinese models are actually open. So I don't see why we can, we cannot have like, you know, frontier models, open models here in the US. And of course, there are other things, right? If you had two models, one closed and one open that are equally capable, I think most people would just choose the open one because it gives you more flexibility, transparency.

30:29Ioannis Antonoglou:You can just make it more secure or safe by just ensuring where it runs, how it uses your resources, what kind of resources you give it access to. So I think capabilities, extremely important, but all else being equal, open wins every day.

30:52Yannis:Yeah, it seems like what I'm hearing is that your company's thesis is that open source can eventually open or even surpass cloud from the speed that comes with open source when everyone's building on it and contributing to it per the open science idea that you shared earlier. Do you think that that's accurate? and if so, is it just a matter of scale at this point, of just scaling up the open science and getting it out there to as many people as possible?

31:20Ioannis Antonoglou:So yeah, I mean, I think it's like ultimately what you need, you know, just like build a frontier model. It's two things. One is talent and the other one is like resources. And, you know, in terms of like resources, you need to have like, you know, the do you need to have like the money you need to actually like stay at the the to have like access all the resources that you need in terms of talent you need um to just like have a an inspiring um kind of like mission like something that like people really want to work uh on and uh i think that our um mission of like open intelligence has uh really been resonated with like scientists.

32:01Ioannis Antonoglou:And this is why I think we have been so successful at like attracting some of the best talent in the industry. Yeah, I agree.

32:11Yannis:Well, you yourself are an engineer. So you're building tools. I assume you're building AI tools with coding AI tools, right? Yeah. Yeah. Okay. What do you look for in a coding agent that you use personally? And what's your favorite right now and why?

32:27Ioannis Antonoglou:Yeah, I mean, I guess like, one thing that I mean, first of all, I want to just like, you know, pause here and just like say that coding agents have had like the progress that we've seen in terms of capabilities over the past two years has been incredible. I think like every six months, you just like, you know, use your best kind of like coding agent tool and you just like get a completely, you're like, you're mind blown. I think like maybe we've grown like too uh i think that maybe now we're like too desensitized like you know a new world comes out it does like incredible things and you're like yeah but like does it go on its own not not yet totally i couldn't agree more covering the news on this every day couldn't agree more but uh it is uh it is insane like how far we've come like you know i've been i've been doing ai for like almost 14 years now and kind of like thinking where the industry was when I started and like where it's now like the coding agent is like on day and night.

33:33Ioannis Antonoglou:At the same time, I really, I mean, I like cloud code. I think that like it's a really, really good piece of technology. I think that like Anthropic has done really good work there, just like building that. So, you know, good news to them. um i've also like uh i've used cash flow in the past and that's that's also like something that i i found useful to me i feel like we're getting to the point we're not quite there yet but i feel like we'll just get there in the next like few months where uh the the like the programmer spends more time kind of like designing and just like architecting and less time about like really kind of writing the code.

34:16Ioannis Antonoglou:I mean, even today, you still need to just go into the weeds and just write the code. But these agents become more and more powerful. So I feel like we're not that far away from actually just being the conductors of an agent orchestra that writes the code. And that's what I want to see in the next year.

34:34Yannis:Boyd, when you just think about the difference in the last year, it's come a long way. And there are a lot of new players. it's become a competitive part of the industry even.

34:46Ioannis Antonoglou:Yes, yes. Like there are like many, many players out there. And I think this is because like there's almost like a discovery that we need to do. Once you have like these powerful agents, what does coding look like in this new world, right? Yeah.

35:10Ioannis Antonoglou:we need to kind of like almost like rethink how we just like code and what coding looks like and then you need you can just like think of different ways and different paths of doing that and they're like yeah many kind of like different companies trying to just like almost do slightly different things at the same time you know most converge to coding agents as like the solution like in some way, in some form. So I don't know, I think it's too early to just like say exactly what coding will look like in even like three years from now. Oh yeah.

35:49Yannis:I guess how will your personal interests and how you use coding, you know, maybe it's still, this is still like totally blue ocean and you don't know yet, kind of guide how you build at Reflection AI. you mean i mean in the sense of yeah like how how how will like i guess if if without giving away the secret sauce because i know yeah i don't want i don't want to have to put you on the spot to give away any secrets but but as far as what you think is missing or or how you like to code how will that focus what you decide to build and what you're building right now yeah i mean for us again

36:30Ioannis Antonoglou:And like, it's all about having agents that you can really trust that like when you ask them to do something, they'll just like go and do it. And they'll just like self-correct. Like even now, you just have like these agents, they will try to do something. They will just like trip over. They'll just like go into circles. Like, you know, they might kind of like waste tokens on like dead ends. They're just like not as efficient. so you know having like stronger agents is like definitely like the the thing that's needed for any uh for any future i think like in terms of but you know you know we know exactly what you need to do there like we just need to make these agents like more reliable and like uh and better now there's like uh what i'm trying to say is that like there's a another question imagine that you had access to extremely powerful agents what is the best way for you to interact with them what's the best way what's the best ui for you to just like uh coordinate them and to delegate tasks to them and just keep track of like what they've done and what they haven't done like what is the best way to run your team of uh agents and i think i've seen you know different ways of like approaching the problem but like uh i think it's early days like uh once we there's like a lot of discovery on the kind of like user AI interaction interface.

37:48Ioannis Antonoglou:What's the best way to, you know, work with these agents?

37:53Yannis:Totally, yeah. And I've seen a bunch of different combinations, and I don't know if we're 100 % there yet. I mean, I like the terminal experience a lot, actually. I was intimidated by it for a while, but I've come to really, really love the terminal first. But I think there's also, you know, There's also a time and a place for the IDE version. And maybe there's a higher level abstraction too where you're almost like watching them build the UI in real time or building the infrastructure in real time. And it's like visualized in some way. I don't know. I mean, that's a higher level than maybe you meant.

38:29Ioannis Antonoglou:Exactly. I feel like this is... Yeah, I feel like this... Once we have the really powerful agents, we need to rethink what's the best way to just like, you know, guide them and, you know, see what they're doing.

38:44Yannis:As far as like the actual sizes of the models that you're going to build, are you thinking of building things all over the spectrum? Are you just going straight at the frontier level? Or will you make stuff that people could run even locally on their own computer?

38:59Ioannis Antonoglou:We want to build a family of models. it's also important that every time you just decide to release a model it needs to go through a whole process so we need to just be we haven't really finalized exactly what models will just drop all the models will just release in the next year but it's going to be different sizes so not just one big one

39:28Yannis:nice That's cool. Something I could run on my computer and code with, perhaps. Yeah. Maybe. Maybe. We'll see. We're both open source nerds and tend to chase whatever's new and download it and see what we can run. We're always tinkering. So let's step away from coding for a second. For someone in, let's say, marketing, sales, finance, what's the first thing you think agents will be doing for them from an open source model? Like what's that killer use case that open models can handle that maybe isn't smart with closed labs?

40:13Ioannis Antonoglou:That's a good question, actually. I think that at some level of abstraction, both open and closed models will just be able to do about the same tasks. They'll just be able to solve the same problems.

40:30Ioannis Antonoglou:Now, I think that there are many solutions, many products. Coding, I understand the industry because I do a lot of coding myself, so it's much easier. But for other industries like marketing, to be honest, I'm not an expert. So I'm not quite sure exactly what the right cases are there. Well, I'll give you an example.

40:57Yannis:And you can tell me if you think this is directly correct. The thing that I think is missing from both a capability standpoint, and you probably wouldn't want to do this with a closed model, is the personal assistant that was promised where it's just like, it actually knows everything about you. And I've heard a lot of discussions about like, oh, you don't want to leave this platform or that platform because you have all of your memories in there and that memory is this big lock-in. But I could see having a closed model on my computer, like if I'm working even in a job perspective, I could see like having something that knows a lot about me and all the projects I'm working on and having that all stored locally, but being able to interface with all of the different applications that I use on that device.

41:46Yannis:Do you think that that is likely with where we're going? Like where you could actually be able to store things, store all of that context and information locally?

41:56Ioannis Antonoglou:Yeah, I mean, I guess like we are not there in terms of like computation of being able to run the most powerful models locally. Like, you know, you cannot, even if you had access to it, you cannot run GPT 5.1 or like, like locally. You just, you cannot do it, right? Or like, you know, the Gemini 3, you cannot just like run it locally. So there is like a limitation in terms of computation.

42:27Ioannis Antonoglou:But besides that, I feel like once we are in a place where you can just like run really powerful models locally, I don't see a reason why you wouldn't be able to have all the data in one place, right? Now, at the same time, we also see that smaller models become better and better and better. So there are definitely cases where I can see a smaller model that can be run locally doing a lot of this work. Yeah. Okay.

42:53Yannis:Yeah. I have one last question here before we let you go. I know we're getting tight on time. And we touched on this a little earlier, but I'd like to dig in a bit deeper. What does it mean to you to have American open source AI in the race and not all of it be concentrated only in China? Yeah, I mean, I think that, like, it is important that, like, there are alternatives out there.

43:26Ioannis Antonoglou:there is a question of transparency and what's happening with the Chinese models, how they've been trained there are things that have come out that they censor certain things so they might have certain cultural biases encoded in them and I think that having an American alternative that is more transparent, it encodes different values it's going to be a net positive for the world and for America

44:01Yannis:Do you think that it's possible to build an open source AI open American AI that also will appeal to other countries like countries in Europe or countries in Asia and how could we do that?

44:16Ioannis Antonoglou:Yeah, absolutely I think it's not going to be just we build it in America but we don't just build it for America we build it for the world and the way to do that is just like be transparent with like, you know, how you build the model, you know, give it out. There is an understanding that like many of the, many of the kind of like values that like this model encodes kind of like shared across other countries, not just like America. So, and at the same time, by just like releasing the model, the model can always be fine-tuned and ensure that like other countries can use it like with, they can further train it on their own data and a lot of like their cultural nuances of like this particular country will also like be understood by the model.

45:04Ioannis Antonoglou:So absolutely, we're building a model that we want to just make it accessible to everyone in the world and, you know, just ensure that like also people can customize it further. Excellent.

45:20Yannis:Well, Giannis, thank you so much for joining us today. It's been great to have you on. Where can, I suppose people want to learn more about what you're doing at Reflection and keep up because there's going to be a model coming sometime.

45:33Ioannis Antonoglou:Yeah, I mean, of course you can always go to our website. We have many job openings, so really hiring across the board to just bring incredible engineers and scientists to work with us on building the most powerful open-weight models in the world. So that's www.reflection.ai you can we have like a Twitter presence X presence where we just like post a lot of like updates on what's happening in the company so yeah

46:06Yannis:cool cool well go check them out they're doing some really neat work over there everyone we hope you enjoyed today's show if you haven't already please take a moment to like and subscribe so we can continue to bring conversations like this to you every week join more than 600 and some odd thousand readers who pick up the neuron every single morning to get their latest AI news. But that's all for today. So until next time, farewell humans.