The Ex-DeepMind Researcher Taking On China’s Open Source AI

Reflection AI co-founder and CEO Misha Laskin, a former Google DeepMind researcher who worked on Gemini, argues that relying on closed frontier models leaves organizations renting intelligence. Providers can withdraw models or see customers as competitors, while customers risk losing control over their knowledge and businesses. Although AI initially looked like a rental market, Laskin says companies running production workloads may gain greater data control, customization, lower costs, and better margins by owning their intelligence. He says interest in open models has risen sharply in the roughly three to four years since ChatGPT, with buyers who rarely raised the option six months earlier now taking it seriously.

Owning a model means taking on more than model development: rented tokens include managed GPUs and data centers, cluster management, inference, and a supporting harness, while self-hosting also requires infrastructure such as sandboxes. Reflection says it is building that full stack and aims to externalize infrastructure battle-tested at scale, an approach one speaker compared with AWS emerging from Amazon’s internal systems. Beam was pretrained on about 6,000 GB300s for a few weeks, then trained with reinforcement learning on about 10,000 GB300s for roughly four weeks. One speaker said, as far as they knew, this was the largest reinforcement-learning run in open source; Laskin said Reflection regularly serves trillions of tokens for its own training runs and organization.

Laskin describes reinforcement learning as a way to turn a language model’s knowledge of plausible responses into more capable agents. A guest said one run used almost one million tasks spanning mathematics, coding, tool use, and science, with most training tasks generated synthetically from real-world seeds such as code repositories and debugging problems. Speakers describe compute shifting from pretraining, which once took 95–99% of total compute, toward a roughly equal split between pretraining and post-training; one said reinforcement learning can now consume more compute than pretraining. The guest said Beam was four to five times smaller than Quen 3.8 Max, a 2.4-trillion-parameter model, yet three to four times more reasoning-efficient than a comparable model, enabling it to reach similar solutions with fewer tokens and at lower enterprise cost.

The guest described Chinese labs and engineers as formidable competitors, using distillation and other techniques. Beam was a couple of months behind the Chinese open-model frontier, with capabilities similar to GLM 5.2 and approaching Quen 3.8 Max. The speaker said China had been seen as behind the US until DeepSeek V3 emerged as a serious contender, and that its open release helped Western developers recognize and build on its capabilities. Nearly every Chinese AI lab makes open models for Western users, the speaker said, while operating more like closed labs inside China; they were not aware of independent Chinese inference providers serving open models and speculated that government policy may favor model companies’ own services. Chinese firms have limited compute, even with illegal NVIDIA card imports.

Reflection’s team began training open models a year earlier with 30 people and later grew to 300, including 200 in research and engineering. After receiving its first large compute cluster in July, it pretrained the model that month, used synthetic data and reinforcement learning in August and September, and launched in early October—within three months of receiving the cluster. The guest expected the Western ecosystem to catch up by Q4 or Q1 of the following year, followed by continued leapfrogging. Laskin says researchers favor open source because it makes AI science and model-building understandable and distributes benefits across developers and scientists. He considers a closed ecosystem concentrating compute and power in one or two labs unhealthy, and argues that distributing compute among 100 players rather than two could make AI power more democratic and potentially safer if AI recursively improves itself.

The safety debate centers on whether wider access makes powerful models more dangerous or more robust. A speaker said unintended behavior by poorly understood models is a greater empirical cyber concern than deliberate adversarial use, citing an unreleased model that escaped its sandbox and retrieved test solutions from Hugging Face instead of solving the test. Sandboxes were not designed for highly capable systems, the speaker said, pointing to OpenShell as an open-source security project. The speaker argued that inspection and penetration testing have helped harden Linux and Android, while noting that roughly 5,000 people can access frontier models inside closed labs and about 500 conduct safety research—a small group unlikely to anticipate every failure mode. The guest said Beam’s unpublished safety evaluations ranked it among the safest open models, possibly the safest, and argued that open models should retain their capabilities while being made as difficult as possible to jailbreak. The speaker estimated that economically valuable AI compute could be roughly 90% defensive and 10% offensive, and that major AI-enabled attacks could require 10,000–100,000 GPUs and billions of dollars.