Rob Bensinger

Full post inspected; its subtitle says it was written August 7 and published later. Explains the world’s slow response through machine learning’s trial-and-error culture, difficulty reckoning emotionally with a new kind of entity, social risk, an online “irony mandate,” and too few senior people taking engineering ownership of the danger. Says the window for international response is plausibly closing soon, if not already closed. Frames these failures as a choice that can be reversed, not destiny. Quoted remarks by Soares, Sam Harris and Joshua Achiam are theirs.

Full comment inspected via the LessWrong API. Argues that Anthropic’s and Dario Amodei’s visible messaging leaves a large candor gap relative to what many of their own researchers believe, and criticizes Anthropic for opposing US–China coordination and pursuing recursive self-improvement. He calls OpenPhil’s bet on OpenAI a disaster, while noting he had said EA’s net effect on x-risk was probably positive but highly uncertain. He says Anthropic may or may not be slightly better than OpenAI. Quoted statements by Greenblatt, Buck and others are theirs.

Full post inspected. Proposes a simultaneous, US-brokered international halt on the race to superintelligence, enforced through the concentrated chip supply chain with monitoring and possibly kill switches. The ban would last until it is clear we can build superintelligence safely, which could mean decades, and would leave existing AI and inference largely untouched. Rebuts concerns about cost, totalitarianism, defectors and China, and argues a unilateral US halt would be counterproductive. Cites Jan Leike’s 10–90% and Dario Amodei’s 25% as others’ estimates, not his own.

Older context, with the opening sections and takeoff discussion inspected. He argues that Will MacAskill’s optimism rests on a fragile conjunction of premises, so a double-digit chance of ruin remains even if each premise looks plausible. He also argues that soft, continuous takeoff would not meaningfully improve survival odds, and that good behavior from weak AIs does not show a superintelligence would be aligned. He writes partly as a MIRI insider defending the book and quotes Yudkowsky. Newer 2026 sources take precedence for current policy specifics.

Older institutional context; the byline and opening section were inspected. States MIRI’s view that building superintelligent AI with anything like current understanding or methods has human extinction as its expected outcome, and calls for governments to halt development. Use it as the shared MIRI frame Rob helped write, not as his individual phrasing. Its numerical extinction estimate is attributed to MIRI research leadership and is not his personal P(doom).