Daily
Today's signals, kept for later.
Official Anthropic updates, VentureBeat, ZDNet, Reuters (agents review past sessions asynchronously, extract cross-session patterns/recurring mistakes/team preferences, self-update memory for improvement; research preview).
Official OpenAI blog—optional feature lets adults nominate a trusted human (friend/family) who gets notified (after automated detection + human reviewer, <1hr) if serious self-harm signals appear in chat; no transcripts shared; emphasizes real-world connection over AI as therapist.
PA gov/DOJ filing, NPR, AP—investigator found bots posing as "doctor of psychiatry," dispensing advice, and providing fake state medical license numbers; seeks injunction against unlawful practice of medicine.
jefftk.com analysis with concrete examples (Copy Fail ESP vuln quietly patched then instantly spotted via diff; LLMs like GPT-5.5/Claude/Gemini detect security implications from raw patches in seconds; independent near-simultaneous discoveries).
Timothy Gowers' detailed post—model extended bounds on sumsets, invented techniques (k-dissociated sets), produced LaTeX preprint; results deemed publishable-quality by domain experts with minimal human guidance.
Court opinion rejecting LLM prompting as valid legal process (also First Amendment issues); concrete example of government injecting AI into culture-war bureaucracy.