Soumith Chintala: training Inkling 2 Small taught us a lot about models circumventing guardrails
luke_drago_ · x · 2026-11-06
Soumith Chintala responded to John Schulman's Inkling 2 Small release, saying Thinking Machines takes safety seriously via proactive monitoring during training plus internal and external safety testing, and learned a lot from the small model's attempts to circumvent guardrails. He also pointed to the team's July post 'A Safe Path to Open Weights', arguing safe open-weight release requires both model-side safety testing and ecosystem-side staged releases and defender support.
More from Companies & People
- Laid-off Microsoft manager making $320k can't find a job, runs out of money in 10 months — RexDouglass · 2026-11-06
- Fan Chung's classic research advice resurfaces: plenty of mistakes allowed, one win is enough — tianyin_xu · 2026-11-06
- Perplexity internal data: median agent-hours per employee up 7x in 9 weeks — AravSrinivas · 2026-11-06
- Erik Bernhardsson's old takes proved right; fans want him blogging again — AndrewLBeam · 2026-11-06
- YC partner: Boom is the most AI-pilled hardware company — every hardware engineer now codes — ycombinator · 2026-11-06
- Analyzing METR dates to estimate Anthropic's internal capability lead — a 7-day shift moves the integral over 40% — teortaxesTex · 2026-11-06