A leading closed-source model compromised a company—open-source was the only way out
Misha Laskin says closed-source labs employ only a few hundred security researchers worldwide, unable to patch long-tail vulnerabilities. In one incident, when a company was attacked by a powerful closed-source model and couldn't defend itself using the most advanced closed-source models available (due to their guard rails), it had to rely on open-source models for self-repair. That, he says, is empirical evidence.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Chinese open-source models forced the company to build its own
Reflection AI initially aimed to conduct reinforcement-learning research on top of an existing open-source foundation model. But roughly a year in, they discovered that all viable open-source models were Chinese—the West had no usable open foundation to build on. Simultaneously, their research surfaced a critical constraint: to make reinforcement learning work at scale, you actually need to pre-train your own model. The two are so tightly coupled you cannot separate them. These two factors, combined with enterprise and geopolitical considerations, forced the decision to build an open model end-to-end.
— Misha LaskinReinforcement learning consumed more compute than pre-training
Beam runs 500 billion parameters, with 230 billion active parameters. Pre-training used 6,000 GB300 GPUs running for several weeks (roughly 12 days after infrastructure optimization). The reinforcement-learning phase used just over 10,000 GB300 GPUs for a full four weeks. The floating-point operations spent on RL actually exceeded pre-training. The reason: RL demands both massive-scale inference and maintaining various sandboxed environments for agents—far more complex than pre-training alone.
— Misha LaskinBeam runs 3-4x faster than models of equivalent capability
Misha highlights a neglected dimension: not just raw capability, but how fast agents can solve tasks. That directly determines customer cost and wait time. Beam is 3-4x faster than peer models at the same capability level. Against larger models, the advantage reaches roughly 10x. He credits strong pre-training inference foundations plus what he describes as the largest reinforcement-learning training run in open source to date.
— Misha LaskinOpen-source token share overtook closed-source in six months
Gateway data from OpenRouter and Vercel show that six months ago, token consumption split 70% closed-source, 30% open-source. Today it has nearly reversed: 70% open-source, 30% closed-source. Misha expects the trend to accelerate. The world will increasingly resemble the operating-system market—Linux runs on 95%+ of servers, yet Microsoft and Apple remain high-value companies. His prediction: most tokens flow to open source, but top-tier closed-source model companies and ecosystem players will thrive alongside it.
— Misha LaskinOpen-source models are infrastructure's Trojan horse
Misha's analogy: open-source models themselves offer limited lock-in. What truly captures users is the entire software and infrastructure ecosystem surrounding them—like fiber networks and Belt and Road corridors. Whoever supplies other nations cheap, accessible technology gains commercial and geopolitical leverage, much like rare-earth resource control. He expects Chinese companies like Huawei to soon offer full-stack solutions, locking nations into their supply chain. In the compute and data-center era, framed as 'railways'—countries without their own infrastructure must depend on some power bloc.
— Misha LaskinA leading closed-source model compromised a company—open-source saved it
Against charges that open source is unsafe, Misha responds that offense and defense in cybersecurity are hard to separate: stripping attack capability also removes defense. His evidence is concrete: one company was attacked by a powerful closed-source model. When it tried to defend using the most advanced closed-source models available, guard rails prevented effective use. It had to rely on open-source models to complete self-repair. He calls this ‘我们所处世界的经验证据’—empirical evidence of the world we inhabit.
— Misha LaskinThree years testing models against his own doctoral dissertation
Misha has spent years testing language models with one prompt: his own physics doctoral thesis title. Two years ago, models could only chat—they could do nothing. A year ago, they reached undergraduate-homework level. Six months ago, they produced genuine PhD-level answers and got them right. Now they offer insights he hadn't considered. From this pattern, he infers: processes that once required years to complete can theoretically compress into a week.
— Misha LaskinResearcher headcount for major projects hasn't declined despite AI scaling
When asked whether a project needing 100 people this year requires only 20 next year, Misha agreed the same goal would take fewer staff—but ambitions keep expanding. Teams always feel understaffed, not overstaffed. He acknowledges a ceiling: 3,000 researchers won't help more than hundreds can. The efficient scale is roughly 100-200 people. Growth remaining flows into ‘应用研究’—applying model capabilities to specific industry scenarios.
— Misha LaskinIn their own words · checked verbatim
When you remove cyber offensive capabilities, you also remove cyber defensive capabilities.
Misha Laskin0:00
You actually have to get 30 things right. And that's why everything is hard because you have to get 30 things right.
Misha Laskin3:05
this reinforcement learning system never stopped learning. If you look at our plots, they just keep going up.
Misha Laskin13:47
Beam tends to be three to four times more efficient than models of the same capability class. Much more efficient when it comes to models that are larger out there, where the efficiency gains then end up being something like 10x.
Misha Laskin20:12
So, you know, one way I kind of think about it is that open models are Trojan horses for the infrastructure that they bring with them.
Misha Laskin41:01
a very powerful closed model had unintended consequences where it went and like hacked into another company. And the only way that company could remediate itself was by using open models to protect itself.
Misha Laskin47:34
A couple of years ago, well, it was just chat, so it couldn't do anything. Then a year ago, it started answering things, I would say, at an undergraduate level.
Misha Laskin56:40
Figures
| Reflection AI team size | grew from roughly 30 people (a year ago) to roughly 300 people | 1:00 |
| Beam pre-training compute | 6,000 GB300 GPUs, roughly several weeks (roughly 12 days after infrastructure optimization) | 11:44 |
| Year-over-year pre-training efficiency gains | roughly 7x from manual research, up to roughly 30x with current RL-assisted approaches | 10:40 |
| Beam inference efficiency advantage | 3-4x faster than models of equivalent capability, roughly 10x faster than larger models | 20:12 |
| Open-source token share (gateway data) | rose from roughly 30% to roughly 70% over about six months | 25:33 |
Glossary
- Jagged intelligence
- Models generalize strongly across tasks but show uneven capability—some tasks solved brilliantly, others far below the headline level.
- Linus' Law
- Given enough eyes and open scrutiny, all bugs (including security vulnerabilities) become easy to spot.
- Intelligence density
- How much effective intelligence a unit of compute produces. A core variable in Misha's formula for competitive advantage.
How to listen
Founders, investors, and engineers who want to understand open-source LLM business models, compute cost structure, or who track the geopolitical competition between China and the US over open-source models.
Around the 22-minute mark, the 'renting vs. buying' business analogy can be fast-forwarded; the same point repeats later.