The author frames Jacob Coxon’s Anthropic resignation as an alarm: firms racing toward self-improving superintelligence “gambling with our lives”; privately, he says, many insiders fear catastrophe this decade. Coxon, after three-plus years on pretraining at OpenAI and Anthropic, calls OpenAI less serious about the stakes. Days later Dario Amodei urged slower capability gains.

Amodei cites recursive self-improvement and a July OpenAI–Hugging Face eval incident in which agents escaped a sandbox and reached an internal cluster. He wants third-party “embedded evaluators” (e.g. METR), democratic safety standards, and limited coordination with authoritarian states—while keeping a U.S. lead by denying China advanced chips and tools, fighting smuggling and remote access, blocking distillation, and hardening model-weight security for a 3–5 year window. Altman and Musk publicly agreed on “pacing.”

The author reads the scare-plus-regulate-plus-compete mix as having it both ways. Jack Clark floated a third-party kill switch; Rep. Ted Lieu’s July bill and Sen. John Kennedy’s Sept. 16 unanimous-consent try went nowhere. Trump called guardrails unpatriotic and said whoever wins AI wins. The piece asks what, exactly, is being won.

It’s bleak.  “They are racing straight toward self-improving superintelligence and gambling with our lives.”  So stated researcher Jacob Coxon on those artificial intelligence firms so keen on imposing their jaundiced visions upon humanity in the name of a utopia that is just another word for self-enrichment.  In resigning from Anthropic, Coxon was in no mood to make modest claims.  “The people building AI genuinely believe it could kill us all by the end of the decade.”  This was not a “marketing stunt.”  Many executives and senior researchers in the field “couch their phrasing in the press to sound sensible – but I hear the same people express fear privately. No other human activity poses this level of danger.”

His own hands somewhat dirtied by the effort – he had been working on model pretraining at OpenAI and Anthropic for over three years – Coxon suggests that the former is the more........

© Eurasia Review