Kevin Xu: distillation buys labs only 1-2 months, and xAI and Meta did it too
kevinsxu · x · 2026-10-07
Kevin Xu pushed back on the claim that only Chinese model labs engage in adversarial distillation:
- Understanding how high-quality data functions in post-training, he argues, implies distillation yields at most a 1-2 month advantage.
- It's not true that only Chinese labs have violated terms of service to distill: xAI and Meta have both done so to varying degrees, though Chinese labs are more aggressive with the tactic.
- Only a lawsuit can settle how much distillation matters. He'd welcome an Anthropic vs Alibaba suit, but thinks "IP theft" is a weak hand since model outputs aren't IP — noting that argument has quietly given way to the "national security threat" framing.
More from Companies & People
- Meta's chief AI officer: an agent swarm can outperform a team of 100 engineers — garrytan · 2026-10-07
- a16z data: Only 3 US states where Anthropic users outspend OpenAI users — Polymarket · 2026-10-07
- AI PM roadmap rebuilt as a subway map: engineers and consultants need different paths — aakashgupta · 2026-10-07
- 13 most common reasons AI pilots fail, from skipping process mapping to ignoring exceptions — vasuman · 2026-10-07
- Yoav Artzi tips his hat to COLM: research stands on the shoulders of giants — yoavartzi · 2026-10-07
- Early Cursor engineer departs, recalling the demo that led from LLM tricks to Mars AI — msfeldstein · 2026-10-07