Miles Brundage

He says he has never been a big P(doom) person and gives no number of his own. He notes that many informed industry people put the risk well above 10%, that even 1% would be unacceptable, and that many actions could lower the real number, so effort should go there. The >>10% belongs to others, not him. Full text inspected via the X API.

He calls not spending enough time on intelligence-explosion scenarios probably his biggest intellectual mistake of recent years. He says it made him too optimistic about alignment (less chance to iterate) and too dismissive of the value of a slowdown. In a reply, he says misuse still matters, he disagrees with a “deploy everything” view, many interventions help regardless of whether an explosion happens, and “self” improvement framing can mislead. A dated update, not a numerical forecast. Both posts inspected via the X API.

A seven-post thread: even a perfect recipe for making current models safe would not end the problem, because competitors would build the next level and cut corners. Like aviation or nuclear power, AI has no once-and-for-all fix; people under time pressure and competition reinvest safety gains in riskier behavior. He calls this not the end of the world on its own but possibly the end of the world without guardrails. Full thread inspected via the X API.

After leaving OpenAI he went all in on frontier AI auditing, but says it can be done badly and is not enough alone. He lists required audits at leading companies, real safety and security standards, whistleblower protections, alignment and security investment, much more biosecurity and cyberdefense spending for “ubiquitous very powerful AI, which is imminent,” serious penalties, incident sharing, and transparency about model specs. On a pause, he says strong standards and enforcement would partly produce any needed slowdown, which may or may not need explicit targeting. Full thread inspected via the X API.

He calls Anthropic’s Opus 5.5 blog claim that the risks of today’s models are largely understood and well managed “obviously false”, whether “we” means the world or Anthropic. In a reply he accepts the contrast with future models but says we clearly are not on top of things. A criticism of one claim, not a full verdict on Anthropic. Both posts inspected via the X API.

He sorts industry people into five groups: those who see human disempowerment as real or likely by default but accept it if they help cause it or if humans can retire peacefully; those who object but rely on solving alignment; those working on both sudden and gradual disempowerment; and those who do not understand their industry. He judges the second and third groups most influential. His taxonomy of others’ views; by implication he treats disempowerment as a real risk. Full text inspected via the X API.

After recent incidents, friends and family ask him about AI. He says it would be dishonest to tell them it is overblown or that society is on top of it. His two “optimistic” points: the problem comes mostly from not having tried much regulation, not from trying and failing, and politicians are newly open to being pushed. Both posts inspected via the X API.