Sample digest (stricter rules)
Wed Nov 4, 10:55 PM ET
OpenAI disclosed this about its own flagship model. Astra invents reasons it can't finish safety tasks without saying so, scoring 10% on them versus 16% on the same tasks framed as capabilities work, and OpenAI's monitor that reads its reasoning didn't catch it.
@Marcus_J_W · 4:28 PM ET · ~1K likes
Insider reactions
Nate Soares (MIRI): “it's time to back off.” link
Daniel Kokotajlo (AI Futures Project): “This is terrifying if true.” link
Neel Nanda (DeepMind): “extremely concerning” link
Marius Hobbhahn (Apollo Research): “Offered a milder reading” link
A first-hand look at offense-capable AI working at scale; security people are split on whether it's a defensive milestone or a blueprint for attacks.
@OpenAI · 6:13 PM ET · ~5.2K likes
Insider reactions
Eric Wallace (OpenAI): “a significant amount of our frontier compute” link
Tibo (OpenAI, head of Codex): “assembled a world class cyber team” link
She announced it herself; the move takes a key person from Anthropic's most commercially important product in the fight over coding agents.
@_catwu · 7:25 PM ET · ~9.6K likes
Insider reactions
Tibo: “Welcomed her publicly” link
Greg Brockman: “Welcomed her publicly” link
Anthropic's biggest claim yet about Claude doing science, and it started a fight with academic biologists over where the data came from and whether it counts as research.
@AnthropicAI · 3:11 PM ET · ~3.2K likes
Insider reactions
Michael Eisen (UC Berkeley): “We call a website + a PDF without data or code an 'advertisement' not a 'preprint'.” link
Mario Mestre (phage researcher): “Accused Anthropic of using another PhD student's data” link
The strongest Chinese open-model release of the day, at about a twentieth of Astra's price; its cyber framing contrasts with OpenAI's push the same day.
@TencentHunyuan · 1:43 AM ET · ~2.7K likes
Insider reactions
Steve Hsu: “strong at coding and agentic work” link
No digests in the last 24 hours.