OpenAI's next model solves ten decade-old problems in mathematics and theoretical CS
OpenAI published new results for ten problems that had seen no progress for at least a decade — spanning high-dimensional sphere packing, coding-theory bounds, a construction of non-sofic groups, a disproof of Connes's rigidity conjecture, arithmetic-circuit lower bounds, quantum parallel repetition, closest-vector-problem hardness (post-quantum lattice crypto), and three Erdős problems. The arguments were generated by an internal version of Astra, which OpenAI describes as its next major model, at a token cost of roughly $2,000 at Sol API rates; humans then prepared the manuscripts with the model, and every proof ships with a Lean 4 certificate plus a model-written narration of its reasoning. Simon Willison's read: the Lean certificates and paper are genuine transparency, but OpenAI hasn't published the prompts or said how many attempted problems produced nothing — so the survivorship rate is unknown.
Sources & depth
Three open letters split the AI industry over open weights and the pace of automation
Simon Willison published his side-by-side reading of the three open letters that landed in late July. Microsoft shepherded "Open Weights and American AI Leadership" (dated July 24, 235+ signers including NVIDIA, Amazon, Y Combinator, the Linux Foundation, and — as a later signer — OpenAI), which defends open-weight releases as essential to security research and US competitiveness. Anthropic, absent from that list, answered on July 27 with its own position: not opposed to open weights outright, but warning about misuse for cyber and biological attacks and calling for a crackdown on distillation operations. The third, "Pacing the Frontier" (July 28), is signed by 1,324 frontier-lab employees — including OpenAI's Jakub Pachocki, SSI's Ilya Sutskever, and Anthropic's Dario Amodei and Jack Clark — asking governments to support international coordination on deliberately pacing automated AI R&D.
Sources & depth
221,303 live credentials found in public Hugging Face training data
Truffle Security scanned every public dataset on Hugging Face — 7.6 petabytes across 187 million files — and verified 221,303 unique live credentials sitting in 6,003 datasets: 349 working GitHub personal-access tokens (223 with repo write access), 318 Docker Hub push tokens, 8,557 GCP service-account keys, 8,594 database logins, and 11,496 AI-provider API keys. One exposed AWS key appears in 1,131 datasets across 10,162 file locations, and they put a floor of $920,000/year on stolen inference from the OpenAI and Anthropic keys alone. The recommendations: scan corpora before publishing, rotate rather than delete leaked keys, and providers should offer bulk revocation — findings that land while the ecosystem is still cleaning up from July's Hugging Face intrusion.
Sources & depth
EU labeling mandate for AI-generated content takes effect today
From August 2, the EU AI Act's transparency rules require digital watermarks or labels on AI-generated images, audio, and text that could pass as authentic, with carve-outs for personal use and clearly artistic or satirical work. New AI systems entering the EU market must comply immediately; systems already deployed get a grace period until December 2. Non-compliance can draw fines of up to 3% of total annual revenue — and the rule targets the unlabeled AI content in advertising, publishing, and film that platform self-labeling hasn't touched.
Sources & depth