Anthropic Warns AI Could Be Near 'Self-Building' Breakthrough as RSI Looms

Anthropic has warned that its future models may 'imminently' reach recursive self-improvement, which is a little like a brewery announcing that its next lager may imminently learn to brew itself and, on reflection, might be a problem. 'Imminently' is the most useful word in the entire statement, because it means soon and it means never at the same time, and no select committee can subpoena a word that means both. The headline has RSI 'looming', a verb Fleet Street normally reserves for weather fronts and pension deficits, now attached to the prospect of a lab's own product quietly building a better version of itself before anybody has finished the risk assessment.

Do savour 'self-building' as the chosen phrase. Not self-replicating, which sounds like a biohazard; not self-improving without oversight, which sounds like a lawsuit; self-building, which sounds like flat-pack furniture with aspirational instructions. It is the vocabulary of a press office that has read the room and decided the room would rather hear about construction than about control. Note what the warning does not contain: no model name, no timeline, no benchmark, no number of any kind, because numbers invite accountability and accountability is the one thing recursive self-improvement cannot generate on its own.

'Beyond current safety' is the phrase doing the real work, and it is doing it in the conditional, where it is nice and safe. Everything is beyond current safety until somebody pays to extend current safety, and the lab issuing the warning is the lab shipping the capability and, conveniently, the lab best placed to sell you the fix. Dario Amodei gets to be the grown-up at the party while holding the punch bowl, which is a neat trick if you can bring it off. Somewhere a legislator will read this headline, feel a brief chill, and return to the donor calls, because the AI bill is still made of ideas and the self-building has no committee stage at all.