Labs are keeping their most capable cyber, math and bio models internal, warns lvwerra
TheZachMueller · x · 2026-11-10
Responding to Elie Bakouch's speculation that Anthropic may never publicly release a model better than Opus 5.5 at LLM training in the coming months, lvwerra argues the trend extends beyond AI research: the most capable cyber, math and bio models are already kept internal. Under the umbrella of safety and internal research, labs will keep more and more value-creating intelligence for their own purposes rather than releasing it.
More from AGI Musings
- Matt Pocock: agents supercharge Goodhart's Law — PR-count metrics now actively harmful — mattpocockuk · 2026-11-10
- Norway runs 2,000+ underwater AI laser robots delousing 50 million salmon — jjvincent · 2026-11-10
- PE investor refuses hourly-billed AI consultancies: 'This job costs $5,000 — whether it takes 5 hours or 2, we don't care' — alex_verem · 2026-11-10
- In the age of vertiginous AI ascent, the stiff-drink-first movie trope makes more sense — StewartalsopIII · 2026-11-10
- Outside the Western bubble, AI use is widespread and positive, even in official orgs — technollama · 2026-11-10
- 'An expert in every domain in your pocket': Reddit user calls out pure anti-AI resentment — chessboardtable · 2026-11-10