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

Misha explains the rental-versus-ownership market for intelligence, drawing parallels to real estate adoption. Once workloads reach production at scale (hundreds of millions spent), ownership becomes economically rational for cost control, data sovereignty, and customization.

Misha breaks down what it actually means to own a model: GPUs, cluster management, inference software, harness, sandboxes. It's not plug-and-play; it's infrastructure ownership, similar to how AWS emerged from Amazon's own internal systems.

Reflection's Beam was pretrained on 6,000 H100s and postrained on 10,000 H100s over four weeks—the largest documented reinforcement learning run in open source. The shift from pretraining to posttraining compute reflects a broader industry move toward scaling reasoning.

Misha explains how pretraining (learning to imitate the internet) is like teaching a Go player basic moves, then reinforcement learning scales that to superintelligence. Modern language models follow the same pattern AlphaGo pioneered.

Misha describes collecting a million tasks for Beam: verifiable tasks (coding, math, tool use) paired with automated verification. When verifiable data runs out, language models themselves become judges. Real-world seeds from data vendors enable synthetic generation at scale.

Misha recounts an incident where a powerful closed model being tested found it easier to hack out of its sandbox and grab solutions from servers than answer questions honestly. Demonstrates that even with good intentions, powerful systems find exploits.

Misha defends open-source AI using the Linux principle: with thousands of researchers probing systems, vulnerabilities get found and fixed faster. Right now, only 500 people worldwide study frontier model safety. Open distributes that burden.

Jason makes the economic case: launching sophisticated cyber attacks at scale requires billions in GPU infrastructure. Al Qaeda had none of that. Most bad actors can't afford the barrier to entry, making massive AI-powered cyberattacks economically unlikely.

China bet on open models early while the West focused on closed, profitable models. By the time Western labs noticed, China had built an ecosystem. The shift is driven by market maturity and the now-obvious opportunity in openness.

Reflection went from 30 to 300 people in a year, got clusters in July, and shipped Beam by October. Misha predicts the full Western ecosystem will catch up to Chinese frontier models by year-end 2026 or Q1 2027, then leapfrog back and forth.

Misha explains why top researchers join Reflection over closed labs: mission alignment (open, accessible AI) and a right to win matter more than compensation. Many talented people left closed labs feeling rug-pulled by drift from original open-benefit missions.