Every closed model is training its open replacement: does distillation kill the moat?
alex_verem · x · 2026-09-30
- The thread's core claim: every closed model is training its open replacement via distillation — frontier labs pay for the research, smaller open models learn from the finished answers at a fraction of the cost.
- Closed labs know it: Anthropic built a safeguard called "preserved thinking" into Fable and Opus 5.5 to stop API users editing Claude's reasoning to extract it, and watermarks every output.
- Case in point: TypeSafe launched the closed decision model Jev on Sept 15; within about a week, open versions appeared on GitHub, one built on Qwen with 7,000 stars.
- Limits: the student only learns what you ask the teacher, so closed labs keep an edge on brand-new tasks. But most real work runs on skills that already exist, and once a skill can be sampled, copying it gets cheap.
- The closing question: if a model's answers are enough to rebuild most of it, what is the moat?
More from AGI Musings
- Chollet: The Test for Human-Level AGI Is Passing ARC-AGI-(n+1) on Release — fchollet · 2026-09-30
- The cruel cycle: labs RL a skill, wrapper startups pile in, labs ship the killer — generativist · 2026-09-30
- Top interpretability expert Neel Nanda: don't rely on us to save you on the current trajectory — burny_tech · 2026-09-30
- Brundage: unheeded warnings should be the baseline assumption for all AI incidents — Miles_Brundage · 2026-09-30
- AI researcher Irina Rish: proceed with "caution without fear," not paralysis — irinarish · 2026-09-30
- OpenAI exec: deciphering ancient scripts is what superintelligence excites me most — BorisMPower · 2026-09-30