Narrative, Tue Nov 3

Today's lead story is GE Vernova's third-quarter earnings, which show the AI buildout's bottleneck has moved from chips to physical hardware. The power-equipment maker's backlog jumped $20 billion in a single quarter to $197 billion, gas-turbine pricing is tracking roughly 50% above last year, and the company used the same earnings call to buy grid-components maker Trench Group outright for $9 billion — extending its grip on a second chokepoint, high-voltage switchgear, beyond the turbines it already controls. Investors weren't entirely convinced: the stock fell as much as 6% on earnings day, with some of the backlog growth read as buyers stacking reservations years out for fear of losing their place in line rather than guaranteed revenue. Either way, the real constraint on how fast AI data centers can come online now sits with turbine and transformer manufacturers, not GPU supply.

A second throughline runs through three otherwise unrelated stories: a lab's own numbers about its own AI system are getting harder to take at face value without independent confirmation. Cognition's new FrontierCode 2.0 benchmark grades code against whether a human maintainer would actually merge it rather than whether it just passes tests, and every frontier model's score fell as a result — but Cognition's own coding agent sits on the same self-scored leaderboard it designed, with no outside party yet reproducing the numbers. OpenAI, separately, disclosed a four-layer system for monitoring its models during training, complete with a retrospective showing it would have caught a real incident 30 hours faster — while acknowledging that the gaps an outside METR/Redwood Research audit found a week earlier are only "largely closed" by OpenAI's own account, a claim METR hasn't independently verified. And a community GGUF repackaging of DeepSeek's V4.1 Flash model, stripped of refusals and shrunk for home hardware, reports a clean 0% refusal rate and largely unchanged benchmark scores — metrics that, per new research on a different model family, don't actually capture whether stripping refusals shifts a model's underlying judgment in ways nobody's testing for.

A third pairing is about institutions building oversight only once deployment or crisis forces the issue. UBS now runs more than 12,700 AI agents a day and IBM just shipped tools to give each one a traceable identity — arriving the same week a blockchain-analytics firm used by banks and governments disclosed a breach via what it called "offensive agentic techniques," with the governance tooling conspicuously catching up to, not ahead of, the risk. Meanwhile, as Character.AI's bankruptcy winds through Chapter 11, the Justice Department is reaching for an obscure bankruptcy-court tool — a consumer privacy ombudsman — to govern what happens to years of intimate user chat logs now that the company itself has failed, in the continued absence of any federal law written for exactly this situation.

Elsewhere: NatWest timed roughly 1,000 job cuts to land on the same day as a strong earnings report, framing the reduction as the dividend of an AI-linked efficiency drive even as CEO Paul Thwaite has publicly resisted calling AI a direct driver of the cuts — the latest UK bank to reach for AI-efficiency language on an earnings call this quarter, a pattern that's becoming a recurring feature of bank reporting rather than a one-off restructuring story.

Top stories

Signal movers

You’re receiving this because you subscribed at machine.net. Unsubscribe · Manage preferences