The Signal — August 2, 2026
A thread runs through today's stories: AI is getting good enough at real technical work that the institutions around it are scrambling to keep up. OpenAI says its models cracked ten math problems that had sat untouched for decades, a court let Minnesota's ban on AI nudify apps take effect, and Apple started rate-limiting bug reports because too many of them are machine-written.
OpenAI publishes ten proofs to long-open math problems, with formal certificates
OpenAI released ten solutions to problems in mathematics and theoretical computer science, each one open for at least a decade with no progress on the main result, and in several cases far longer. The problems span high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, and lattice cryptography. The headline claims include a counterexample to Connes's rigidity conjecture and a construction showing that non-sofic groups exist. OpenAI says an internal model produced the work, and that the tokens to generate all ten would have cost roughly $2,000 at its API rates.
The part that makes this checkable is the paperwork. Each result ships with a Lean 4 formalization on GitHub, meaning a proof assistant has mechanically verified the logic, plus a walkthrough of how the model reached it. That matters because "AI solved a famous problem" claims usually arrive as a press release and leave as a correction. A Lean certificate is a much harder thing to wave away. It doesn't settle whether the informal framing holds up to mathematicians, and that review is still happening, but the formal core is there to inspect. This lands days after Anthropic reported Claude finding real cryptographic weaknesses, so the pattern of frontier models doing genuine research-grade work, not just demos, is starting to repeat across labs.
Sources: OpenAI · Simon Willison
A judge lets Minnesota's nudify-app ban take effect over xAI's objection
A federal judge denied xAI's request to block Minnesota's new law banning apps that generate non-consensual nude images, so the first ban of its kind in the country took effect on August 1. Judge Donovan Frank's reasoning was about timing rather than the merits. xAI filed its emergency motion on July 29, nearly three months after the law was signed and three days before it was set to take effect, and the court read that delay as a sign the harm wasn't actually urgent. The Minnesota attorney general made the same argument: a voluntary delay shouldn't manufacture an emergency.
This is a procedural loss, not a final ruling, and the two are worth keeping apart. xAI's broader claim that the law violates the First Amendment is still alive, with a preliminary injunction hearing set for August 19. But the practical outcome is that the ban is now in force while that fight plays out. It's an early data point on how courts will handle the collision between free-speech arguments and laws written specifically to curb AI-generated abuse, and here the law got the benefit of the doubt.
Sources: TechCrunch · NBC News
Apple caps bug reports after a flood of AI-written submissions
Apple added a submission cap and a 30-day cool-off period to its Feedback Assistant bug pipeline, telling researchers it was drowning in AI-assisted reports, according to the Financial Times. People who need more room can request higher quotas, so this is a throttle rather than a wall. It's a small operational change that points at a real problem: the same language models that can surface a genuine vulnerability can also generate a convincing-looking one that turns out to be nothing, and a maintainer still has to read every one to find out which is which.
Apple isn't first here. The curl project and the Internet Bug Bounty hit the same wall earlier, with curl's founder comparing the flood of low-quality AI reports to a denial-of-service attack on his own team. What makes Apple's move notable is the scale. When a company this size formally rate-limits vulnerability disclosure, it's an admission that AI-generated security noise has become a cost worth managing at the platform level, not just an annoyance for open-source volunteers. The tools that help find real bugs and the tools that bury triage queues in plausible junk are, for now, the same tools.
Sources: Financial Times · Techmeme
On the Editor's Desk
A couple we looked at and left out. MiniMax's new H3 video model is interesting, but the open weights it promised are still not out, so we're waiting until you can actually download it rather than covering a promise. AMD put out a fully open mixture-of-experts model trained on its own GPUs, which is a real story, except it landed a week ago and there's no fresh hook to justify running it now. And another batch of agent-benchmark preprints came through that read well but don't have outside verification yet.
