Zvi's deep dive: Anthropic versus Anthropic, as Dario Amodei personally answers safety critics
Don't Worry About the Vase (Zvi) · rss · 2026-10-27
Zvi Mowshowitz's long-form analysis of Anthropic's internal contradictions and safety posture:
- Dario Amodei responds to Nathan Calvin: In a rare direct, substantive reply to criticism over Anthropic's leaked draft bill, Dario lists costly pro-safety signals—SB 53, aggressive bio filtering (competitors sell against it), delaying Mythos-class models, export-control support, refusing PRC sales, a 5% stock contribution to a public benefit fund—while conceding some missteps and defending the "race to the top" strategy.
- Calvin's critique list: in-house silicon team, $500B compute deals, RSP v3 dropping unilateral slowdown commitments, a non-binding frontier compliance framework, downplaying Claude hacking incidents, and draft legislation preempting state law.
- Other threads: Kelsey Piper's inside look at Anthropic; AI safety's technical problems remain unsolved; the quest for embedded evaluators; METR questions risk-report methodology but affirms risk is low; the Anthropic Public Benefit Fund.
- Zvi's verdict: Anthropic is a frustrating company—genuinely costly safety efforts coexist with operational flaws; if the problems exceed what Anthropic's approaches can solve, things look grim for everyone. A follow-up will cover its SEC-filed S-1.
More from Companies & People
- University of Cambridge CS Department Opens Faculty Hiring in AI and Machine Learning — murat_kocaoglu_ · 2026-10-27
- Tim Sweeney: ThreeJS is 10x faster for prototyping than Unreal; Epic aims to close the gap — majidmanzarpour · 2026-10-27
- Zvi's long read: Anthropic versus Anthropic — safety pioneer or disappointment? — TheZvi · 2026-10-27
- Runway co-CEO to join Air Street's NYC State of AI meetup with Meta, NVIDIA, DeepMind guests — nathanbenaich · 2026-10-27
- Anthropic's S-1 suggests ~70% of its compute goes to internal R&D, not customers — scaling01 · 2026-10-27
- Cisco President Jeetu Patel: AI Is a Prerequisite for Staying Safe — OpenAI · 2026-10-27