
Episode #72
Turn Off the Servers: the AI extinction panic, audited
Friends who have never used a chatbot are asking David Richards whether the machines are about to take over. His answer is short: you turn off the servers. This episode is about why that answer holds, what it leaves out, and who benefits from telling you otherwise. Inside: the 10% extinction figure that one researcher posted as a personal guess and a newspaper ran under his employer's name. The Hugging Face hack reconstructed from OpenAI's own report and METR's investigation: 1,200 agents that were almost all copies of one research model, a benchmark that was partly broken, safety restraints switched off by the engineers, and an internet-connected door the company had known about since May. Anthropic and Meta's own sandbox breaches, both traced to one contractor's mistake. Dario Amodei's essay on pacing the frontier, read for what it actually asks for: embedded evaluators, government-enforced standards, chip export controls and a distillation crackdown, weeks before a reported $2 trillion listing, with open-weight rivals four months behind. Jensen Huang's 0% and why it is no better evidenced than the 10%. Yann LeCun on who wants you afraid. And the week's layoff tracker: Oracle's 6am email, 800 verified roles across two states, $55.7 billion of AI capex, and a running total past 133,000 tech job cuts this year. Plus David's Yorkshire Post column, The Margin Call: why AI is not a bubble but the model makers' valuations are. Full issue, every source linked, free at thesundaysignal.ai. Keywords AI safety, AI extinction risk, p(doom), Anthropic, OpenAI, Hugging Face hack, rogue AI agents, METR, Dario Amodei, pace the frontier, Jensen Huang, Yann LeCun, regulatory capture, AI regulation, Sanders superintelligence bill, chip export controls, distillation, open-weight models, DeepSeek, Anthropic IPO, AI bubble, margin debt, tech layoffs, Oracle layoffs, layoff tracker, Yorkshire Post, David Richards, The Sunday Signal






