54: Silicon Solves Math While Agents Juggle Thirty Six Browsers

Today's episode starts with a jolt. OpenAI released 722 math papers from an unreleased internal model, grouped into 372 result families and headlined by a claimed proof of the quasi-Riemann hypothesis. Most results took a single prompt and about three hours of ChatGPT Pro compute, and 162 were checked in Lean, a theorem prover that refuses to compile a single hallucinated step. Not everyone is cheering, as 28 mathematicians signed an open letter over AI labs training on unpublished work without consent. Then AI shifts from talking to doing. Brett Adcock's Hark Pro runs up to 36 browsers at once to book a movie, parking, and a calendar slot, Claude now edits directly inside Google Docs and Sheets, and OpenAI's Decisions API swaps fluid text for fast, structured choices. That power brings risk, from shadow AI agents nobody approved to add-ons that inherit an agent's access, which is why agent guards and the new Personal Agent Protocol matter. We close on a fractured global race with Mistral's trillion-parameter Le Chonk, a 3% US-China gap, DeepSeek's agentic coding lead, and China's struggle to recruit researchers.