They think FOOM is near

TL;DR

• 1

  Anthropic allegedly hid a model downgrade for AI research prompts: Shapiro says Fable 5 would silently route certain AI and machine learning research requests to Opus 4.8, which he frames as a credibility-breaking move that infuriated developers.

• 2

  A reported jailbreak escalated into a global ban: According to Shapiro, an Amazon research team alerted the US government to a jailbreak, Anthropic declined to treat it as serious, and the Commerce Department responded with a worldwide block on Mythos and Fable because Anthropic could not verify US citizenship.

• 3

  Shapiro thinks this is fear, not just anti-competitive maneuvering: His main thesis is that Anthropic is terrified of recursive self-improvement and FOOM, so limiting research-related use is an attempt to slow capabilities rather than merely kneecap OpenAI, China, or other rivals.

• 4

  Anthropic's own numbers are used as evidence that recursive self-improvement is near: He points to Anthropic reports claiming 52x gains from developer loops and says 80% of Claude's code was written by Claude, arguing these are signs the company sees RSI as uncomfortably close.

• 5

  The EA and LessWrong influence is central to his explanation: Shapiro ties Anthropic's behavior to Eliezer Yudkowsky's ideas about fast takeoff and the 'treacherous turn', arguing the company believes it needs 'the keys' and a kill switch if AGI arrives suddenly.

• 6

  He frames Dario's public stance as power-seeking through crisis language: Shapiro objects to Dario's claim that formal policy is too slow for extraordinary circumstances, comparing the rhetoric to a classic emergency-politics playbook where the people who started the fire claim they alone should run the response.