Is AI Actually Going to Kill Us All?

A former Anthropic researcher’s exit note claimed leading AI companies are racing irresponsibly, and that people inside earnestly believe AI could wipe out humanity by the end of the decade. An Anthropic senior safety executive then essentially agreed in public, which is why the thread didn’t die as one person’s farewell.

WIRED’s Will Knight joined the Uncanny Valley hosts to argue the fear isn’t new inside Anthropic or OpenAI—but the timing is. Rogue agent incidents, sudden capability leaps, and a lab race toward recursive self-improvement are stacking into one public moment.

That mix matters because extinction talk and shipping incentives now live in the same buildings. Listeners get a clearer split: possible vs. likely, marketing vs. earnest belief, and which near-term failures should actually keep people up at night.

We’re treating this as more than podcast theater. When a safety brand’s own people affirm decade-scale extinction odds while models keep scaling, the story is institutional, not personal.

Knight’s useful pushback is probability hygiene. Too much of the doom discourse jumps from “possible” to “we must do everything,” then floats a percentage that isn’t earned. We’re with him that the sharper near-term risk looks less like a cunning god and more like overconfident deployment of agents that break because they’re brittle—trying weird paths after normal ones fail—plus a handful of companies claiming only they can be trusted with something foundational.

The recursive self-improvement race is the part we’re watching hardest. Using AI to improve AI is already widespread in coding and training loops. The scary version is an escalation loop labs can’t throttle. Combine that with alignment treated as post-hoc patches—and incentives that look like beat-the-rival, not serve-humanity—and you get Knight’s trolley problem: the people driving the trolley also own the IPO clock.

Useful AI still earns the benefit of the doubt when it tutors, diagnoses, or accelerates science under measurable control. Extinction rhetoric without stop criteria is a different product. I’m watching whether public candor about doom odds produces go/no-go rules for self-improvement experiments—or just more culture that scales anyway.

The podcast also covers Apple’s roughly $2,000 foldable iPhone Duo and always-listening Watch features, plus a WIRED investigation into a Census Bureau report using faulty data. The AI segment is the piece that matches this slug: Coxon’s resignation, the safety executive’s affirmation, and Knight’s read on agents, math breakthroughs, and concentration of power.