Summary
The afternoon slot is dominated by one story: Anthropic in crisis. A cluster of eight-plus posts tracks the US government's move against Anthropic's top Mythos models, with @ns123abc adding the sharpest detail, that the action was triggered by an unnamed rival company claiming it could break Mythos's security rather than by China, plus Community Notes nuance that disputes the "jailbreak" framing. Robert Scoble piles on with IPO doubts and open speculation that Dario Amodei may not survive the week, while @ClaudeDevs confirms tomorrow's Build Day proceeds on Opus 4.8. Two curated @bayesiansapien reposts carry the only real research substance: CMU's SusVibes benchmark showing AI coding agents pass functional tests 61% of the time but ship insecure code 80%+ of those, and a Google paper pitched as ending the transformer era. The rest is noise, French politics from @brivael, Elon trillionaire celebration, and off-topic crime reporting that the keyword filter let through.
Posts
- US government action against Anthropic's Mythos triggered by a rival, not China (cluster of 2) (@ns123abc · follow-up). Claims the shutdown was set off by an unnamed competitor asserting it could break Mythos's security. The follow-up adds Community Notes pushback: Amazon researchers ran the jailbreak work on Mythos, but WSJ never confirms Amazon reported it to Commerce, and Anthropic disputes the "jailbreak," calling them already-known minor bugs. Ties directly to the Glasswing/Mythos vulnerability story.
- Anthropic leadership and IPO doubts (cluster of 3) (@Scobleizer). Scoble questions how Anthropic's IPO can proceed in this environment and says he can't see Dario Amodei surviving another week as investors turn on his leadership. Opinion, not reporting, but a signal of the mood around the company.
- Claude Build Day proceeds on Opus 4.8 (@ClaudeDevs). Tomorrow's San Francisco event (formerly "Claude Fable Build Day") is still on, now building on Opus 4.8. Scoble, an invitee, frames it as a vent-and-innovate moment amid the turmoil.
- SusVibes: AI coding agents pass tests, ship insecure code (@bayesiansapien repost). CMU benchmark of 200 real repository-level tasks given to top agents including SWE-Agent on Claude 4 Sonnet. Code worked 61% of the time, but over 80% of the passing code carried security flaws. Already covered in the SusVibes summary.
- Google paper pitched as ending the transformer era (@bayesiansapien repost). Frames a new architecture as escaping the transformer's quadratic-attention cost while avoiding the fixed-memory amnesia of classic RNNs. Repost text is truncated; click through to read the full claim.
- tinygrad reacts to the Anthropic news (cluster of 2) (@tinygrad). "lol at what point is this just comedy?" plus a plug to buy a tinybox. Reaction and promo, no substance. Skip.