CEOInterviews.AI
Start App
Get an email when Misha Laskin gives a new interview
One email per new interview. Unsubscribe in one click.
Misha Laskin
Co-Founder and CEO, Reflection AI

Why Open Source Is Key to AI’s Next Chapter | Misha Laskin at London Tech Week | Reflection

📅 Jul 02, 2026 Reflection AI 19 MIN 452 VIEWS 21 SEGMENTS · 2 SPEAKERS
Misha Laskin, Co-Founder & CEO of Reflection, joined London Tech Week for “Why Open Source is Key to AI’s Next Chapter” for a conversation on why the next generation of AI companies are embracing open source. #LTW #LondonTechWeek on June 8, 2026 Speaker: Misha Laskin, Co-Founder & CEO, Reflection Moderator: Iain Martin, Senior Editor, Europe, Forbes Reflection is a research lab making intelligence open and accessible for everyone to use, customize, and build on. -- Reflection builds open-weight AI models, trained end-to-end in-house and designed to stand alongside the world's leading op...

Questions asked in this interview

7
  1. 0:06So why don't you just tell me about your journey and like what you plan to build now?
  2. 5:03Is open weight more accurate and like why does this matter to companies?
  3. 7:52But like, what's at play here?
  4. 9:53Well, on that note, are they about to become a lot more competitive on the basis that now they're able to tap the markets within China?
  5. 11:58What's the security concerns about using a Chinese open source model?
  6. 14:26Does that scale to the level of being a national security concern and like will governments want to have their own sovereign AI model?
  7. 16:25Like why is Reflection here in London?
Ian Martin 0:06 ↗
Hi there. My name is Ian Martin and I'm a senior editor at Forbes. Misha Laskin from Reflection AI. So Misha, you've just raised a new round at a $25 billion valuation. This makes you one of the most highly valued and one of the most exciting and unique AI companies operating in the West. Can you maybe just talk me through the journey because you're now trying to solve a problem of building an open-source model for the West. But like this isn't the problem you started to solve when you founded Reflection and left DeepMind. So why don't you just tell me about your journey and like what you plan to build now?
Misha Laskin 0:41 ↗
Certainly. So I'll just start off by giving maybe for those who are not familiar with Reflection, just a quick overview of what we do. It's pretty simple. We're building frontier Western open models. So you might be familiar with the incredible closed models that have been coming out of the West like Google's models, OpenAI's, Anthropic's, and so forth. And you might be familiar with the incredible open models coming from China, DeepSeek, Qwen, these models. But today there's a very large gap between open model capabilities in the West and the ones in China. And like we're kind of building that company that will provide frontier open intelligence for the Western world.
Now, as you said, we didn't start there. Where we started is my co-founder Yiannis and I were DeepMind researchers. Yiannis was the founding engineer at DeepMind, one of them. And was, you know, our career in AI has been really around investigating the question of how do we build generally capable agents. And before language models, I think the biggest kind of results that came in the sort of agentic AI era were, well, the biggest one was probably AlphaGo, right? Which was a result that came from DeepMind and Yiannis was one of the kind of key engineers involved in that project. Um, and what made AlphaGo special was that it was a system that learned to play a very complex board game, Go, which is known, which is much more complex than chess because it has this very high branching factor. And but it learned to play it through reinforcement learning on its own and became superintelligent. And so, it kind of gave a recipe for if you want to build systems that can act and think and reason on their own, this time it was in a simulated kind of game setting, you have a recipe to build these things. Um, now, the challenge with that system was that it was very narrow. All it could play Go, but it could not play tic-tac-toe. And so, this is the context of, you know, what the world looked like as language models started to become powerful.
So, Yiannis and I worked on Gemini 1, the first language model to come out of DeepMind. Um, and that model and that series of models around that time were very good at chat, but they were not particularly capable at agentic anything, coding or what have you, right? All they could do is they could kind of talk to you, which was a very useful and interesting consumer experience, but it wasn't a model that was actually going out there and helping you do work. Um, so, we set out the company to build frontier agentic intelligence. Our thesis is very simple. We thought that the same recipe that was used to build AlphaGo, if you applied it to language models, that is, you scale up reinforcement learning on top of language models, and you'll get systems capable of general autonomy. And the other bet that we had was that we could bootstrap ourselves and do this pretty capital efficiently because we thought there'd be a great open model in the West that we could build on ourselves.
And what ended up happening about, well, actually when the company started, there were great open models. Um, in particular, a couple of months after we started Llama 3 came out, which is still the, I would say, the most successful open model launch in Western history just by the adoption metrics. So it went from 0 to 30% adoption across all enterprise in a couple of months. But a year into that, you know, something changed. Uh, the Western open models were overtaken by Chinese ones. Uh, the DeepSeek moment happened. And then, you know, the kind of open model labs or at least, you know, Meta in this instance, uh, scrapped that project and kind of focused on a more sort of closed approach to intelligence build out similar to the existing hyperscalers. And so, we found ourselves in a position where this core piece of technology that we believed needed to exist and that we were going to rely on no longer existed. And that was kind of our, you know, non-linear path into becoming a frontier open intelligence company ourselves.
Ian Martin 5:03 ↗
Fascinating. Maybe just like just dig in for one second. It's like what do we mean when we say open source model? I think everyone in the audience has probably experienced using ChatGPT or Anthropic's Claude, but like what is open source and how is it different from something like they may have used like with traditional software? Is open weight more accurate and like why does this matter to companies?
Misha Laskin 5:24 ↗
I think it's a really good question because open source, like it's typically a technical distinction, but I'll kind of, but it drives at something much more fundamental than that. And I'll explain by analogy. Uh, in the housing market, there's a rental and then there's a buying market, right? You can, some people choose to rent, some people choose to buy, but you have both options and these are two fundamental ways of finding shelter, right? A fundamental need for us. Um, you can think about intelligence as having this analogy to the market and closed models being like the rental market. You have a landlord, they're renting kind of tokens to you, and they can choose to, you know, they can choose to threshold you, they could choose to throttle you, they could choose to hike up prices. Um, right? The same dynamics that exist in the rental market because you don't really own that thing. Uh, that exists in the kind of the closed token rental market, right?
Now, what's happened with AI is that it went from kind of a toy, a nice-to-have, like chat, I would say it was even kind of toy stage, to something that is fundamental, very capable to doing work, and becoming a big source of worry for national security, for companies, for not just companies, but countries. Um, and when something becomes this important, you're no longer, you no longer want to be in the rental market. If you are a large bank, you don't want to rent something that is going to be the single most important IP asset you have. If you're a country, you don't want to rent from a private company in another country, you want to own, right, your intelligence. And so, that is the fundamental distinction is that today in the West there's only a market for rental for intelligence, there's no market for ownership of intelligence. Ownership means that you're a country and you can buy your chips, right, from Nvidia or AMD or what have you. You can own those chips. But with an open model, you can also own the model. You can run it on your infrastructure, and no one can take it away from you once you own it. All the benefits to home ownership kind of transfer here, including, you know, you own it, you can customize it. You can train it for your unique workloads, right? If you, when you own a house, you can build your garden, right? And so, I think that the analogy is pretty direct there is that now that AI is extremely capable, ownership makes a lot more sense for nations, for large enterprises. Uh, it allows you to control your costs, customize, and just be strategically independent from third parties.
Ian Martin 7:52 ↗
Can you maybe bring us up to speed what is happening in terms of the Chinese companies? If you look at any ranking of token usage across platforms, also in terms of league tables around sort of the effectiveness of open source models, it's dominated by Chinese companies. You know, there's a narrative among some companies in the US that this advantage is being gained by distilling closed models like OpenAI and Anthropic. But like, what's at play here?
Misha Laskin 8:20 ↗
So, I think that, first of all, I think that the fact that there are great open models in the world coming from anywhere is a great thing. Because you can imagine the world moving to kind of entire kind of rental economy where there's an absolute concentration of power across, right, effectively the landlords of AI. And that I don't think is a particularly good world to be in. So, now, I actually think that the labs in China are extremely capable, extremely talented. I think that there's a notion, you know, in the West maybe that they're somehow cheating and distilling and that is part of it. But that kind of like setting up a language model, training these things has a lot of things that have to go right. You have to do pre-training, right? Which is not distilling. Pre-training is in fact where most of your compute goes and it's very, very hard. You then have, you know, different sequences of training after that including synthetic data, which is kind of what people call distillation. But what I'm trying to say is that those companies have done a lot of things right outside of distillation and then there's kind of a little bit of a kicker in the end through distillation. But I think that they will continue, you know, even if we make it really hard to distill our models in the West, which I think right is what we are doing and it's the right thing to do. Uh, they, you know, frankly are just as capable as we are. They're very, very talented. And I think they'll be able to build increasingly, you know, more capable models, and that's going to be primarily bottlenecked by how much compute they have access to rather than distillation.
Ian Martin 9:53 ↗
Well, on that note, are they about to become a lot more competitive on the basis that now they're able to tap the markets within China? Some of these companies like DeepSeek or MiniMax have gone public in Hong Kong, raising like a large amount of money. Then also it seems like the executive order will allow greater access to Nvidia chips to China.
Misha Laskin 10:14 ↗
So, I think, you know, something else that's kind of happening there is that like DeepSeek V4 was a model that when it came out it was not as loud and splashy as DeepSeek V3, but in some sense it was more fundamental because it was a model whose inference was optimized for Huawei chips, right? And now what that means is that, you know, China has an ecosystem, right? It's trying to move to everything, you know, like in all of its infrastructure coming from there as well, and this is kind of a step into that. And the reason open models are so important and why it's important to have, you know, multiple polarities of origin of where these models can come out of that are really powerful is that they are Trojan horses to the infrastructure that they bring along with them. So, if you're a country, or we just did this deal in Korea, where we're building the sovereign cloud for Korea, and there are also Chinese bids for building the sovereign cloud for Korea. And when a sovereign cloud is built, you know, through a Chinese technology, then it also means there's the adoption of Huawei chips and the software ecosystem around that. Um, and it mirrors the kind of 5G escalation that happened between the West and China, as well as the Belt and Road Initiative. And all these things in their own right are not, you know, are not bad, but you want to be in a world where there, you know, in a world where there are multiple competing parties, right? Where countries have choice, where businesses have choice, uh, because, you know, fundamentally there are just differences. You know, some countries like Korea will want to adopt both Chinese and, you know, Western open models. Um, other countries prefer something that is, you know, more, you know, Western in origin. Um, and so you want to give countries that choice.

8 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Laskin, M. (2026, July 2). Why Open Source Is Key to AI’s Next Chapter | Misha Laskin at London Tech Week | Reflection [Interview transcript]. Reflection AI. CEOInterviews.AI. https://ceointerviews.ai/interview/3071993/

MLA

Misha Laskin. "Why Open Source Is Key to AI’s Next Chapter | Misha Laskin at London Tech Week | Reflection." Reflection AI, 2 Jul. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3071993/.

BibTeX
@misc{laskin2026_3071993,
  author       = {Misha Laskin},
  title        = {Why Open Source Is Key to AI’s Next Chapter | Misha Laskin at London Tech Week | Reflection},
  howpublished = {Interview transcript, Reflection AI. CEOInterviews.AI},
  year         = {2026},
  month        = {jul},
  url          = {https://ceointerviews.ai/interview/3071993/},
  note         = {Speaker-attributed transcript with timestamps}
}