Nathan Lambert pushes back on Anthropic: 'Open dangerous, closed safe' is a false dichotomy
natolambert · x · 2026-09-30
Allen AI researcher Nathan Lambert published a critique of Anthropic's safety blog post implying that Z ai's open GLM models pose unique misuse risks. He argues documented cyber abuse is actually more associated with closed models, and that open weights and safeguarded closed APIs are both close to easily misused — the real picture may be 'open unsafe, closed unsafe.' Stronger capabilities in closed models could matter more for net harm if guardrails are porous, while unguarded open models help diffuse cyber readiness across the economy. He concludes the blog is reasonable in a narrow line but serves as an effective tool for reinforcing a particular safety worldview.
Related event: Anthropic's Open Model Safety Claims Spark Backlash(3 posts)→
More from AGI Musings
- Michael Levin's argument: LLMs may harbor abilities and goals far beyond language probes — danfaggella · 2026-09-30
- EigenGender: Friend-Network Hiring in AI Safety Is Bad Because of the Field's Monoculture — EigenGender · 2026-09-30
- AI Safety Hiring Debate: Friend-Network Screening Is Harmful Only Because of Field-Wide Monoculture — EigenGender · 2026-09-30
- F. Chollet: Hype That Frames AI as 'Killing' Things Makes Public Backlash Inevitable — fchollet · 2026-09-30
- Bill Gurley cites Tetlock: generalist forecasters beat AI-doomer specialists — kevinnbass · 2026-09-30
- Lean creator Leonardo de Moura on AI proofs: the Collatz exploit shows verified checkmarks can lie — Machine Learning Street Talk · 2026-09-30