The takeaway: A quiet overnight in AI, with one real exception. The United States is moving to reopen one of the hardest questions in warfare — how much of the decision to kill it will hand to a machine — even as the hardware reaches real battlefields. Elsewhere, the field mostly held its breath, and we'd rather tell you that plainly than manufacture a busy day out of a quiet one.

The one that matters: Washington reopens the autonomous-weapons question

According to a Washington Post feature on 20 July, the US administration is pressing the Pentagon to revisit the policies governing weapons that can select and engage targets on their own — testing, but not erasing, the long-standing requirement for human judgment over the use of force. It is not a decision to unleash anything; it is a decision to reopen the rulebook. And it lands at an awkward angle to Congress, where one bill would harden human control while a Senate committee's defense framework leans the other way.

What makes it more than a paper fight is that the machines already exist — a US startup has shown an AI-driven humanoid loading a mortar and has sent units to Ukraine, and even frontier labs supplying the underlying models are trying to fence off "fully autonomous" use. We've written the whole thing up: The Pentagon Revisits Its Line on Machines That Choose Targets, read alongside last week's Human Authority over Autonomous Weapons Act for the other side of the argument.

The Jacobian counterexample: more eyes, still no paper

The claimed counterexample to the 1939 Jacobian Conjecture — first floated over the weekend by mathematician Levent Alpöge together with the Claude model Fable — kept drawing attention over the past day, with fresh expert commentary and independent reproductions of its short symbolic checks. What has not appeared is the thing that would settle it: a written-up, refereed paper, or a credible refutation. Our position is unchanged from our write-up — genuinely interesting, widely discussed, not yet resolved. We'll update it the moment a paper or a debunk lands.

Otherwise, a held breath

Google still has not posted anything launch-shaped for Gemini 3.5; the reported delay stands. Moonshot's next open model, Kimi K3, still has its open weights pencilled in for 27 July. And a much-discussed claim of an AI perfect score at this year's Mathematical Olympiad remains unconfirmed by any official body — so we're not reporting it as fact.

On a quiet night, one true thing beats five loud ones. More as it breaks.