The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
OpenAI says an internal multi-agent system—stronger than GPT-6 Astra—produced a Lean-checked singularity proof for Navier–Stokes statements C and D, and it will not claim the Millennium Prize.
OpenAI Group PBC is putting up to $5 million—grants to $1 million each—behind independent studies of generative AI’s effects on teens 13–17, with applications due October 6, 2026.
Reducto collapses multi-stage agentic OCR into one full-page pass, cuts its own error rate 20%, and drops price to a flat cent a page—while keeping the weights closed and the evals internal.
OpenAI pushes ChatGPT Edu into two top j-schools and restates a years-long newsroom partnership map—access and training upstream, enterprise credits and playbooks downstream.
OpenBMB's 2.52B dense MiniCPM5-2B posts a 53.9 average across 34 benchmarks and ships Apache-2.0 weights plus training data—strongest where agents call tools, weaker on textbook knowledge.
Axis Robotics moves robot-demo collection into the browser and treats the dataset as a living product—not a one-shot lab dump—while keeping the big corpus gated and non-commercial.
MBZUAI's IFM ships six Apache 2.0 K2 Horizon models plus corpus, checkpoints, and code—and publishes its own reward-hacking audit that knocks the headline score down.
OpenAI, WAN-IFRA, and Ukraine's independent regional press association are packaging AI training, hands-on pilots, and API credits as wartime newsroom infrastructure.
Two more regional publishers join the copyright pile-on against OpenAI and Microsoft, seeking not just damages but destruction of models trained on their work.
H Company strips the vision tower and causal decoder from document retrievers—NeoMME's 260M single-tower encoder nearly matches a 3.75B ColQwen variant on ViDoRe v3.
Anthropic's $1.5 billion payout is clearing—and authors say publishers and even agents are grabbing shares the settlement rules never assigned them.
Meta FAIR and academic partners treat experiment selection—not idea generation—as the scarce resource, ranking candidate runs before any GPU hour is spent.