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

Reflection CEO Misha Laskin on open source AI models as a national security imperative

📅 Oct 05, 2026 Washington Post Live 24 MIN 1361 VIEWS 45 SEGMENTS · 3 SPEAKERS
Reflection CEO Misha Laskin speaks about the role of open models in a thriving AI ecosystem and national security and big tech’s “concentration of power” challenge. Recorded on Oct. 1, 2026 at The Washington Post’s AI Edge Summit. 00:00 - Intro 01:38 - Reflection origins 04:50 - Open vs. closed AI models explained (the real estate analogy) 07:35 - When will Reflection release its base model? 09:13 - Can Western open-source keep pace with China's models? 10:30 - Should frontier AI development slow down? 12:56 - Inside President Trump’s White House AI accord 14:05 - Should open-source AI be sub...

What Misha Laskin said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Misha Laskin, CEO of Reflection AI, argued that open-source AI models are a national security imperative because global AI adoption is currently built on Chinese models, including by American companies. He explained that open models enable ownership of intelligence, unlike closed models which are rented, and that this ownership is crucial for countries and enterprises. Laskin said Reflection aims to bring the open AI frontier back to the US and export it worldwide. He noted that the best open models today are Chinese, citing DeepSeek V3, and that this concentration is a geopolitical risk. He advocated for a multipolar AI ecosystem with many competing labs, arguing that distributed compute slows AI development responsibly. He also claimed that open models are safer due to distributed defense, citing the Hugging Face incident where a Chinese open model defended against a closed model's unintended consequences. Laskin said Reflection will release a model soon but gave no date.

Key takeaways

  1. Global AI adoption is built on Chinese open models, including by American companies, creating a national security risk.
  2. Reflection AI will release its first frontier model soon, built faster than any comparable model.
  3. Open models are safer than closed because distributed defense is more effective, as shown by the Hugging Face incident.
  4. Token share on OpenRouter flipped from 70% closed/30% open to 70% open/30% closed in six months, with most open tokens from Chinese models.
  5. The US should win by merit in open models rather than regulate Chinese models, as other countries will adopt the best available.

Numbers and commitments

FigureWhat it refers toTypeAt
70% closed, 30% open Token share on OpenRouter six months ago metric 22:03
70% open, 30% closed Token share on OpenRouter now metric 22:03
300 safety researchers Number of safety researchers across big labs metric 18:15

Chapters

  1. 0:00Open models as national security imperative
  2. 1:44Reflection's mission and founding
  3. 5:09Rental vs ownership of intelligence
  4. 7:22Model release timeline
  5. 9:03Competing with Chinese labs
  6. 10:47Slowing AI development via multipolarity
  7. 12:50Regulation and White House lunch
  8. 16:52Safety of open vs closed models
  9. 21:46Recommendation to President Trump

Questions asked in this interview

12
  1. 1:22What made you decide to leave and start your own business?
  2. 4:40... some of those geopolitical dynamics, but just to set the stage for this conversation, how many people here feel like they really understand what we mean when we say open systems versus closed proprietary models when we're talking about AI?
  3. 7:22When can we expect Reflection to do that?
  4. 8:46Are we talking weeks, months, years?
  5. 9:03Do you think you can move as fast as they do?
  6. 10:29We need to slow down development.' What is your response to those calls?
  7. 11:46What steps would have to be taken to more equally distribute that compute power among these labs?
  8. 12:50And I just am curious, from your perspective as someone building a lab in this area but who wasn't at that lunch, are you concerned that those conversations could create a regulatory moat that might disadvantage a business like yours?
  9. 13:57Is Reflection willing to take steps like that?
  10. 15:09Do you think your company's model should also be subject to those types of reviews?
  11. 16:30What's your message to those policymakers who are worried about that?
  12. 20:02But do you personally worry that you could be putting something out in the world that could then be modified by a terrorist group or hackers to do some real damage?
Narrator 0:00 ↗
A new divide has opened up in AI: open versus closed.
Misha Laskin 0:14 ↗
Global AI adoption is actually very much being built on top of Chinese models, including American companies.
Our intention is to build a new frontier lab that drives the frontier of knowledge in artificial intelligence, but doing it in an open way.
Our reason to exist is to really bring the open AI frontier back to the US and export this technology to the entire world, rather than importing it.
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.
Interviewer 1:18 ↗
Thank you so much for being here with us today.
Misha Laskin 1:20 ↗
Thank you, Cat, for having me.
Interviewer 1:22 ↗
Well, we have a lot to get into today, but just to start the conversation, I wanted to talk to you a little bit about your decision to start this company. Before you were there, you were at DeepMind, which was widely considered one of the most advanced AI labs in the world at the time. What made you decide to leave and start your own business?
Misha Laskin 1:44 ↗
I think that, you know, just a quick introduction on Reflection: we are building American open models, and we'll discuss why that's important, you know, what is the difference between open and closed. One thing I'll say is that AI, I think everyone here has noticed how fast it's moved. And when we started Reflection, we actually had a different bet than, you know, the one we ended up pursuing.
And it was that my co-founder and I, we were at DeepMind, we led a lot of the reinforcement learning work for the first Gemini series of models, Gemini 1, 1.5. And if you recall those series of models and early ChatGPT, they were good at chat, but they were not actually good at anything agentic, which is where in the world now these models are great at agentic work, coding, and that was not the world then. So we thought that there was a scientific breakthrough that needed to happen through this way of learning called reinforcement learning. And that's really why we set out to build Reflection at that time. And we believed that we could do it because we believed there'd be great open models that we could build on as a company and therefore operate a lot more cost-effectively than having to, you know, go and start everything from scratch.
Now, what happened was that, one, the reinforcement learning stuff started working much faster than we thought, not just in our lab, but other labs as well. That's when the reasoning models started coming out. And the other thing that happened that really changed the course of the trajectory of the company was that the American open model kind of frontier receded. Like there was a period in time when both the best closed and open models were coming from America. And that changed with DeepSeek, with DeepSeek V3. You know, China really took the global stage in terms of producing the best open models, and that's actually still the case, you know, up until, I mean, this point, that's still the case. The best open models in the world today are Chinese.
And so we saw two things that are about to happen that, you know, made us kind of confident that we need to step into this mission. One is that there's going to be some geopolitical issues if you have a, you know, monopoly-type of world where only one country is producing great open models, because other countries will adopt that. And two, that there will be an unprecedented concentration of power among closed intelligence, you know, in the States. And fundamentally we thought that a much better, much more beneficial way for AI to develop is in a multipolar way, in a competitive way, where there are not just two labs that are, you know, sitting on the best intelligence, where anyone can own it, and it's not just one country that's producing great open intelligence, two, hopefully more. We thought that multipolarity is really important.
Interviewer 4:40 ↗
And I want to get into some of those geopolitical dynamics, but just to set the stage for this conversation, how many people here feel like they really understand what we mean when we say open systems versus closed proprietary models when we're talking about AI? Show of hands. So it's a pretty good number in this room, but I see some hesitation and a few hands down. So I was hoping you could walk us through, just in simple terms, what you mean when you talk about open models.
Misha Laskin 5:09 ↗
By the way, that's actually a lot more hands than were up, you know, when similar questions are asked, I would say, even six months ago, 12 months ago. So it's been really interesting seeing how openness has become important over the last half year, I would say. But I can posit a pretty simple analogy in terms of how I think about closed versus open models, and it's not a this or that, it's a this and that. And the analogy is to the real estate market, which is in real estate there are two ways to, you know, buy property. You can rent it, right, or you can own it. You can rent an apartment or you can own one or own a house. And typically, right, the trajectory of a consumer is that when you're younger, you're out of college, you start renting.
Then as you get older, you get more financial security, you have a family, ownership becomes important for various reasons. So you can think about what's happening in AI is that it's a, I mean, the market has only been around for a few years, three years or so since ChatGPT was released, and a young market starts in rental. And closed models is you're renting intelligence. It doesn't belong to you. It's not an asset that belongs to you. You're renting it. It's easy to use, and that's great. But as you mature, and this is maybe not as an individual, but as a corporation, either you're a startup that started small and was renting and now you're really big, you're spending a billion dollars on AI, you might want to own that, you know, asset at that point. Or you're an enterprise or you're a public sector agency where you need to control your intelligence for various reasons. You're a country. You're a country. You don't want to be constantly renting intelligence from someone else. You want to build your own data centers. You want to put your own models in there.
As AI has matured and become both a national security imperative, not just for the United States but other countries, and as enterprises have matured in their adoption of AI, ownership is all of a sudden important. And so what we're seeing is that, I would say, even the transition from a rental to an ownership market, but the birth of an ownership market in intelligence. And ownership is powered by open models because by definition when it's open you can own it.
Interviewer 7:22 ↗
And I want to talk to you a little bit about where Reflection is today. You've raised billions of dollars to build this open-source lab, yet while we keep seeing new models from companies like DeepSeek, you still haven't released your main base model. When can we expect Reflection to do that?
Misha Laskin 7:40 ↗
We are releasing a model soon. I think the way to think about it is that the amount of capital expenditure, it's kind of like building rockets. The amount of capital expenditure required to build one of these things is very high. You need a lot of GPUs, you need a lot of data, and you need a lot of critical talent. And from the day that you decide to go out and, you know, build a rocket, you know that even, you know, that that's going to take time even if you're moving as fast as humanly possible. So, you know, over the last year, we've set up a frontier intelligence lab. We have folks, you know, have joined us from OpenAI, from DeepMind, from Anthropic. I think there's a lot of researchers are starting to see the kind of dangers of concentration of power and geopolitical dangers from having this kind of monopoly, right, in the world, and are drawn to this mission. And so as a result, we've been scaling and have been building very strong models and are excited to share some of those soon.
Interviewer 8:46 ↗
And when you say soon, that can mean so many different things in the tech industry. Are we talking weeks, months, years?
Misha Laskin 8:54 ↗
We're talking soon.
Interviewer 8:59 ↗
All right, not making my job too easy today.
Misha Laskin 9:01 ↗
But AI moves quickly.
Interviewer 9:03 ↗
Yes, that is for sure. And on that point, you know, we see how quickly the Chinese labs are moving and shipping open models. Do you think you can move as fast as they do?
Misha Laskin 9:15 ↗
I would actually posit that, you know, the rate at which we set up, you know, this first model is one of the fastest, if not the fastest, a frontier-grade kind of model has been built. So I think that the Chinese labs move very quickly. I think that it's completely possible for Western labs to compete. And I think that we're moving into what we'd like to see is an ecosystem in the United States and the West that is as vibrant as the one in China. There are five labs right in China that are all kind of leapfrogging each other, you know, producing great models. And you want to see that in the West as well, right? You don't want just one company or two companies. You ideally want as many as possible. And multipolarity is, right, it's very good because there's no single point of failure. And it's very good for the customer because there's no single company that amasses so much power that has a lot of leverage over the customer.
Interviewer 10:29 ↗
And when we talk about the landscape in the West, there are several large labs where we've seen the CEOs, particularly Anthropic first, come out and say, 'We need to slow down the pace of the frontier. We need to slow down development.' What is your response to those calls?

17 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, October 5). Reflection CEO Misha Laskin on open source AI models as a national security imperative [Interview transcript]. Washington Post Live. CEOInterviews.AI. https://ceointerviews.ai/interview/3071988/

MLA

Misha Laskin. "Reflection CEO Misha Laskin on open source AI models as a national security imperative." Washington Post Live, 5 Oct. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3071988/.

BibTeX
@misc{laskin2026_3071988,
  author       = {Misha Laskin},
  title        = {Reflection CEO Misha Laskin on open source AI models as a national security imperative},
  howpublished = {Interview transcript, Washington Post Live. CEOInterviews.AI},
  year         = {2026},
  month        = {oct},
  url          = {https://ceointerviews.ai/interview/3071988/},
  note         = {Speaker-attributed transcript with timestamps}
}