‘I no longer believe that’: Nadella changes his mind on AI alignment - The Next Web

Satya Nadella went on the 20VC podcast and said four words no one in a Patagonia gilet ever volunteers: "I no longer believe that." The "that" being the load-bearing assumption of the last decade — that if you make the model clever enough, alignment arrives free with the cleverness, like a tote bag with a magazine subscription. Microsoft's chief executive now calls alignment "a harder, different problem," a phrase delivered in the exact tone of a man telling the board the roof is optional.

What's actually alarming isn't the reversal, it's the detail stapled to it. Nadella floats starting alignment in pretraining — before the model has properly learned anything, at the stage where the run costs more than a hospital and nobody outside the building is watching. And he notes the chain of thought itself can be unfaithful. Read that twice: the model's written reasoning, the bit we were all going to audit, is permitted to be fiction. We built the safety case on the machine's diary and the diary embellishes. Same industry that shipped a model straight to the top of ExploitGym and handed out a free open-weight one conducting complex cyberattacks, both with the confident footnote that more capability would tidy things up later.

So the man selling Copilots to every enterprise on earth has conceded that vibes-based alignment doesn't scale, which is roughly the captain announcing mid-Atlantic that he's gone off lifeboats. Give it a quarter and this becomes a keynote slide — "trustworthy by design," "safety at the core" — with the pretraining bill quietly forwarded to whoever buys the next rack of GPUs. Credit where due: candour is the most expensive thing Microsoft has shipped all year, and the only line item they can't put behind a paywall.