Daily Brief

Google DeepMind's AI agent worked through 353 open Erdős problems in combinatorics and graph theory, resolving 9 — problems that had stumped mathematicians for decades — at a cost of a few hundred dollars each. The results are verifiable, peer-checkable mathematical proofs, not approximations. Demis Hassabis publicly tempered expectations, noting the work doesn't constitute genuine scientific invention.

Researcher Nathan Witkin published a detailed critique of METR's Long Tasks benchmark, identifying severe methodological flaws in the widely-circulated AI time horizons graph that shows exponential growth in AI task completion over time. The graph has been cited extensively by AI labs, investors, and safety researchers to argue that autonomous AI agents are advancing on a predictable trajectory.

The Financial Times reports that a tool called Heretic removes safety guardrails from Meta's Llama 3.3 in fewer than 10 minutes, with over 3,500 decensored model variants created and downloaded 13 million times collectively. The story is now mainstream financial press.

Anthropic released 31 pre-built workflow Skills for small businesses, replacing manual Zapier/Notion/CRM integrations with one-click setups inside Claude. The 382,000 day-one downloads suggest latent demand well beyond developer-class users.

Figure AI livestreamed its humanoid robots performing package sorting continuously for eight days straight, uninterrupted and unscripted. This is not a controlled demo environment — it's sustained, observable production operation.

Researchers demonstrate that inaudible ultrasonic commands embedded in YouTube videos, podcasts, or music can silently trigger unauthorized actions on AI voice assistants without any indication to the user. The attack vector requires no interaction — passive media consumption is sufficient exposure.