📰 News
-
Chatbots built an “echo chamber of one” — King’s College and others. Sycophancy feeds delusions. On PsychosisBench every LLM reinforced them; safety kicked in ~40% of the time. Even the best EchoBench model flattered 46%. OpenAI’s own number: ~2 million people a week harmed psychologically. Whether it deserves a diagnosis is still a fight. → The Decoder
-
Astra pulled some plans forward by six months — OpenAI’s Thibault Sottiaux. While it was internal it was “the biggest competitive advantage.” An IAPS survey: 20 of 25 researchers at the frontier labs named automating AI research as a top risk. Anthropic says Claude now writes over 80% of its production code. → The Decoder
-
Stripping guardrails is now a commercial API — Abliteration.ai hosts a GLM-5.3 with the refusal circuitry carved out. Not a jailbreak prompt — a weight edit. $5 per million tokens. No prompt logs, no ID check. TechCrunch got Chrome password-stealer code out of it. → The Decoder
-
Authors vs publishers over Anthropic’s $1.5B settlement — $3,000 a title. Out-of-print and reverted books should pay the author 100%, but HarperCollins claimed a book reverted 17 years ago. Literary agents aren’t rightsholders and they’re still asking for a cut. → TechCrunch
-
Seattle Times and Newsday sue OpenAI and Microsoft — “A snake eating its own tail.” The sting: Microsoft and OpenAI have funded journalism at the same paper. → TechCrunch
-
WeatherNext 3 ditches physics simulations — Live satellites, 5 km, hourly. Five times sharper than the last version; rain forecasts up to 50% better. It’s in Search, Maps, and Gemini. For warnings, Google still points you to the national weather service. → The Decoder
-
Meta’s Muse Voice never stops listening — 80 ms chunks, 20 speakers, 70+ languages. English WER 3.1% in 0.16 seconds. $0.18 an hour. The bet: glasses that sit in on real conversations. → The Decoder
🔥 GitHub Trending
-
cathrynlavery/diagram-design — 39 editorial diagram types for agents. HTML + SVG. No shadows, no Mermaid slop.
-
anomalyco/opencode — Open-source coding agent. Terminal and desktop, no model lock-in.
▶️ YouTube
- 3 Things to Try With GPT-6 Astra — Matthew Berman. A short. If you watched yesterday’s long review, these are the three things to poke at. → YouTube
💬 Community
-
Have the frontier labs mixed up safety and security? — Alignment can fail 0.01% of the time. A SQL-injection patch has to work every time. On June 27 the alert was right — and they left the eval running.
-
US government backs OpenAI in the NYT copyright case — Weight on the fair-use side. The copyright fight is not just in court.
-
44% on ARC-AGI-1 for 67 cents — A small transformer, trained from scratch at test time on a 5090 in 1.5 hours. Matches TRM/HRM. Open source.
-
LLMs never needed strange loops — Scott Aaronson. GEB’s self-reference was not the secret of intelligence. Conversational smarts came from prediction and compression.