Summary
The day was one dominating story across every slot: the US government's export-control directive suspending all foreign-national access to Anthropic's Fable 5 and Mythos 5, and Anthropic's response of disabling both models worldwide to comply. It opened in the morning as a five-post cluster reposting Anthropic's statement, deepened in the afternoon with @ns123abc's claim that an unnamed rival, not China, triggered the action plus Community Notes disputing the "jailbreak" framing, and spilled into the evening as French politicians (Philippe, Lalucq, Attal) reacted to AI sovereignty and @brivael ran a dozen-tweet anti-regulation thread. Around that, two real research signals surfaced via @bayesiansapien reposts: CMU's SusVibes benchmark (agents pass functional tests 61% of the time but ship insecure code in 80%+ of passes) and a Google paper pitched as ending the transformer era. The single strongest hardware datapoint was NVIDIA's AgentPerf debut, the first agentic-infrastructure benchmark, with Blackwell GB300 NVL72 running 20x more agents per megawatt than Hopper. A second industry thread emerged in the evening: a multi-state AG coalition opened a criminal probe and subpoenaed OpenAI days after its IPO filing. Everything else (SpaceX IPO, Elon's trillionaire milestone, French electoral debate, local-news leakage) is noise the keyword filter let through.
Posts
- US government suspends Fable 5 and Mythos 5; Anthropic disables both worldwide (cluster of 7) (@ClaudeDevs, @Scobleizer, @ns123abc, @kilocode · Anthropic statement) [morning + afternoon + evening]. Citing national security authorities, the directive bars all foreign nationals (inside or outside the US, including Anthropic's own staff) from the two models, so Anthropic disabled them for everyone; other Claude models unaffected. Anthropic says the trigger was a "jailbreak" that amounts to asking the model to read and fix a codebase's flaws, a capability it says is widely available elsewhere including GPT-5.5. See today's Global View.
- Action triggered by a rival, not China (cluster of 2) (@ns123abc · follow-up) [afternoon]. Claims an unnamed competitor set off the shutdown by asserting it could break Mythos's security. Community Notes add that Amazon researchers ran the work but WSJ never confirms they reported it to Commerce, and Anthropic calls them already-known minor bugs. Ties to the Glasswing/Mythos vulnerability story.
- French political reaction to model gating (cluster of ~12) (@brivael reposting Edouard Philippe, Aurore Lalucq, Gabriel Attal) [evening]. Officials are alarmed the US is gating its strongest models behind nationality, framing it as AI sovereignty; @brivael uses "l'affaire Claude" to argue the EU should scrap the AI Act, GDPR, and DSA. Policy core is real export-control precedent; the surrounding electoral debate is opinion.
- Anthropic leadership and IPO doubts (cluster of 3) (@Scobleizer) [afternoon]. Scoble questions how the IPO proceeds in this environment and says he can't see Dario Amodei surviving the week. Opinion, but a read on the mood around the company.
- Anthropic resets all rate limits (@ClaudeDevs) [morning]. Reset 5-hour and weekly limits for all users to soften being bumped off Fable 5 mid-workflow.
- Claude Build Day proceeds on Opus 4.8 (@ClaudeDevs) [afternoon]. Tomorrow's SF event (formerly "Claude Fable Build Day") is still on, now building on Opus 4.8.
- NVIDIA AgentPerf: Blackwell runs 20x more agents per megawatt than Hopper (@nvidia · NVIDIA blog) [morning]. The first agentic-AI infrastructure benchmark, built for workloads that chain dozens to hundreds of model calls. GB300 NVL72 leads with up to 20x more agents per megawatt than Hopper, a real efficiency-per-watt number for agentic serving.
- SusVibes: AI coding agents pass tests, ship insecure code (@bayesiansapien repost) [afternoon]. CMU benchmark of 200 repo-level tasks for top agents including SWE-Agent on Claude 4 Sonnet. Code worked 61% of the time but over 80% of passing code carried security flaws. Covered in the SusVibes summary.
- Google paper pitched as ending the transformer era (@bayesiansapien repost) [afternoon]. Frames a new architecture as escaping quadratic-attention cost while avoiding classic RNN amnesia. Repost text truncated; click through for the full claim.
- OpenAI under criminal investigation, served subpoena (@ns123abc · WSJ) [evening]. A multi-state AG coalition opened a probe and subpoenaed OpenAI on models, sycophancy, engagement design, health data, and activity involving minors, days after its IPO filing. Substantive signal on legal and safety risk attached to frontier deployment.
- Google Gemini-SQL2 hits state-of-the-art text-to-SQL on BIRD (@GoogleResearch) [morning]. Text-to-SQL system on Gemini 3.1 Pro claiming SOTA on BIRD, which measures execution-verified accuracy (the SQL runs and returns the right result).
- kilocode ships REVIEWS.md, repo-specific review standards (@kilocode · blog.kilo.ai) [morning]. An open-standard Markdown file defining a repo's review conventions; a Code Review Memory mode watches how you respond to PRs and proposes updating it. Same "compile the user's standards into a gate" pattern as today's TRACE paper.
- kilocode's Fable 5 "KISS" experiment (@kilocode · blog.kilo.ai) [morning]. Argues Fable 5 impressed at UI and front-end but over-engineered deeper tasks, with mistakes harder to spot because the output looks polished. Color on what "this powerful" meant in practice.
- Google Research: phone-cluster computing and dermatology AI (@GoogleResearch) [morning]. A low-carbon compute platform giving retired phones a second life, plus findings that dermatology AI may help laypeople name skin conditions. Google's sustainability and responsible-AI beat.
- NVIDIA: Jensen's "5-layer AI cake" (@nvidia) [morning]. Amplifies a Sequoia interview where Huang frames the shift "from retrieval to generation" as a multi-trillion-dollar opportunity across a five-layer AI ecosystem. Mostly vision-and-PR; the retrieval-to-generation framing is the throughline.
- magicsilicon: 8 of the top 10 companies design silicon (@magicsilicon) [morning]. A one-line chart: of the new top-10 companies by market value, 8 design chips, 1 manufactures (TSMC), 1 drills oil. A compact read on how the compute build-out reshaped the corporate league table.
- Anything: agents building and publishing apps end-to-end (@Scobleizer) [morning]. Scoble building a mobile app with the "Anything" CLI, which claims Claude Code, OpenClaw, Codex, and Hermes can design, build, and publish "with no human in the loop." Promotional and thin on detail.
- Kilocode article (@kilocode) [evening]. Opaque
x.com/i/article/link from the AI coding-tool account, no preview text. Click through to read. - tinygrad reacts to the Anthropic news (cluster of 2) (@tinygrad) [afternoon]. "lol at what point is this just comedy?" plus a tinybox plug. Reaction and promo. Skip.
- "Decompose and precondition" (@MillionInt) [evening]. A one-line aesthetic riff on matrix decomposition. Vibe, not content. Skip.
- SpaceX IPO and Elon trillionaire noise (@ns123abc, @brivael, @lexfridman) [morning + evening]. Heavy volume on Musk crossing $1T and SpaceX's IPO. No AI-research substance. Skip.