Wired reports that major US financial institutions are blocking outside bots from accessing accounts, and warning customers that any costly mistake an agent makes is theirs to cover.
A joint statement says Levent Alpöge released a Lean library of about 17 million lines, that it compiles, and that the theorem it proves matches Mathlib's definition of the Riemann hypothesis. The proof passed Lean's kernel and several independent checkers, and the communities say they have full confidence it is correct.
Ethan Mollick(Bluesky): Argues it is no coincidence that the people proving Millennium Problems with AI are great mathematicians.
Hacker News(discussion): More than 1,100 points and 900 comments.
A researcher planted unique, trackable secrets in prompts sent through AI services and found that the data does not stay private, Ars Technica reports.
President Trump said companies that build data centers outside the US would face the tax. The Verge notes the threat comes as states and cities around the country are placing restrictions on new data center applications.
MIT Technology Review reports that the evaluation nonprofit's new report finds some AI capabilities still improving along an exponential trajectory, while models do less well on other kinds of tasks.
Levent Alpöge says he found the proof while working with Claude and that the method extends to the generalized Riemann hypothesis. The 167-year-old problem is one of the Millennium Prize problems. TechCrunch reports that a competing proof from a human mathematician has surfaced as well.
Levent Alpöge(quoted by Simon Willison): “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.”
Ethan Mollick(Bluesky): Says the hypothesis may have been solved with Claude's help, and links the paper.
Hacker News(discussion): More than 2,800 points and 2,000 comments on the paper.
Signal is restricting screenshots in its Android and iOS apps, describing AI surveillance as an existential threat. The Verge reports the change blocks assistants such as Gemini and Siri from reading what is on screen.
OpenAI describes a layered approach that combines model training, safeguards, detection and enforcement to keep others from extracting its models' capabilities.
Ethan Mollick(Bluesky): Suggests the end of easy distillation from OpenAI and Anthropic models may explain slower progress in open-weights models.