Hosts: Lenar Kess, Damra Vol. An OpenAI researcher says an internal model closed ten open problems in mathematics and theoretical computer science. Nobody has published the ten problems. Today's episode is about what the industry does in the gap between a claim and its evidence — and about four other stories where the evidence is a screenshot, a vendor chart, or a sarcastic Reddit post.Noam Brown's post on the Astra results — the source of the claim, relayed to us through a Hacker News submission rather than any first-party writeup from OpenAI.Theo Jaffee relaying a model's assessment that this was "plausibly the most significant single day in the history of mathematics," and Victor Taelin noting that the mathematics subreddit is unimpressed.Miles Brundage arguing the effort would do more good pointed at verifiably secure software than at open math.Mario Zechner running OpenAI's security evaluation against his own copy protection, undefeated for six years, and reporting what GPT-5.6 Sol did to it.The EU AI Act transparency date — today, August 2 — discussed in the local-models community with a clown emoji in the title.Simon Willison cataloguing what ChatGPT Work can actually do, and Brett Bauman using it as a scheduled job.The 105× total-cost figure for DeepSeek V4-Flash, with Aravind Srinivas calling two orders of magnitude rare.Elvis's four ways agent state disappears and Cameron Wolfe on why scoring a trajectory is harder than scoring an answer.
