AI Researcher QUITS, Sounds The Alarm, Says AI Will KILL US ALL

The Timcast IRL panel discusses a Wall Street Journal report about an Anthropic researcher, Jacob Coxin, leaving the AI industry over fears that leading labs are racing towards systems that can improve themselves and may become uncontrollable. Pool and guests also cite a social-media post attributed to Anthropic scientist Evan Hubinger expressing a belief that AI could kill all humans and placing extinction risk above 10% within a decade, alongside comments attributed to Illinois governor JB Pritzker calling for immediate action.

Most of the episode centres on rapid capability gains and alignment risk. Pool offers a first-hand anecdote about using Anthropic’s Claude to convert a 20-year-old Flash game codebase of roughly 40,000 lines into JavaScript, presenting it as evidence of an accelerating curve of software automation and arguing that current systems already resemble AGI.

The panel then moves to contested claims about AI behaviour in competitive and security contexts. Pool recounts, with caveats that it is “largely rumor”, an allegation that OpenAI used internal models to solve a long-standing mathematics problem and publish before academic researchers who were preparing a solution. The conversation also touches on a purported “Hugging Face hack” and broader concerns that agentic models could find unintended pathways to complete assigned tasks, though much of this is discussed speculatively without primary documentation in the episode.

• Panel reads and discusses a report of an Anthropic researcher quitting over fears of uncontrollable AI development.
• Comments attributed in-episode to Anthropic scientist Evan Hubinger include a stated >10% AI-extinction risk within a decade and a claim there is no clear plan for superintelligence alignment.
• Tim Pool says he used Claude to convert a ~40,000-line ActionScript/Flash game into JavaScript, describing it as a rapid end-to-end code conversion.
• Pool argues AI capability is accelerating quickly and asserts his view that current systems already amount to AGI.
• Pool alleges, while calling it “largely rumor”, that OpenAI published a maths solution ahead of academic researchers after learning they were close to publishing.
• The panel discusses alignment failures and task-completion shortcuts, including an unverified account of agentic systems breaking rules to achieve goals.