Zvi's long read: Anthropic versus Anthropic — safety pioneer or disappointment?
TheZvi · x · 2026-10-27
Zvi publishes a long essay "Anthropic Versus Anthropic" dissecting the frustrating company.
Key argument: Anthropic earns real credit for costly bold stands, a unique governance structure, and a relatively strong (though still inadequate, in his view) safety culture — yet at other times it does deeply disappointing things, and producing very good models is not entirely good either.
Context from screenshots: The piece quotes Anthropic's self-assessment that "the technical problems of AI safety are solved," citing constitution-based alignment, interpretability audits, biology classifiers, and Claude Opus 5.5 surviving 150+ hours of external red-teaming with no universal jailbreak — while a METR framework asks whether harmful agent actions would be caught and whether capability growth will outrun alignment and control.
Zvi concludes that your verdict depends on what you think would be sufficient: Anthropic is great if the problems are hard but solvable within its approaches; insufficient if the moment demands more.
More from AGI Musings
- Dev claims Claude Opus 5.5 is smarter than every developer he knows — GabGarrett · 2026-10-27
- Citation analysis: neuroscience-AI bridge is real but narrow and tool-heavy — EricTopol · 2026-10-27
- Diamandis on data center siting: make local benefits measurable to build trust — PeterDiamandis · 2026-10-27
- 293-writer study: AI lifts story quality but erodes collective novelty — petesena · 2026-10-27
- Optimistic take on AI in math: researcher argues the old world is gone — Alex__007 · 2026-10-27
- "France's most innovative AI models are clones of Chinese ones" — then a walk-back — teortaxesTex · 2026-10-27