Summary
The morning feed is overwhelmingly one story: the US government issued an export-control directive to suspend access to Anthropic's Fable 5 and Mythos 5 for all foreign nationals, and Anthropic responded by disabling both models for every customer worldwide. This is a cluster of at least five posts (ClaudeDevs, Scoble x3, ns123abc x2, kilocode's thread) all linking to Anthropic's statement, with the announcement card reading "Statement on the US government directive to suspend access to Fable 5 and Mythos 5." The tone ranges from disbelief (Scoble: an entrepreneur's app was switched off mid-build) to schadenfreude (ns123abc: "be anthropic, publish essay asking govt to block dangerous AI, government blocks, only YOU"). Anthropic frames it as a misunderstanding over a narrow jailbreak and says it is working to restore access. The strongest non-Anthropic signal is NVIDIA's AgentPerf debut, the first agentic-infrastructure benchmark, with Blackwell GB300 NVL72 running 20x more agents per megawatt than Hopper, a genuine Tier 1 hardware data point. Google Research had a productive morning (Gemini-SQL2 text-to-SQL at state-of-the-art on BIRD, phone-cluster computing for low-carbon compute, dermatology AI), and kilocode shipped REVIEWS.md, repo-specific code-review standards. The rest is SpaceX-IPO and Elon-trillionaire noise from brivael and ns123abc that does not bear on AI research.
Posts
US government suspends Fable 5 and Mythos 5 (cluster of 5) (@ClaudeDevs, @Scobleizer, @ns123abc, @kilocode · Anthropic statement). Anthropic's statement: the US government, "citing national security authorities," issued an export-control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the US, including Anthropic's own foreign-national employees. The net effect is that Anthropic had to abruptly disable both models for all customers to comply; all other Claude models are unaffected. The announcement card (a stark black card reading "STATEMENT ON / The US government directive to suspend access to Fable 5 and Mythos 5") was reposted identically across the cluster. Anthropic says the government's concern appears to be a method of "jailbreaking" Fable 5 that essentially consists of asking the model to read a codebase and fix its software flaws, and that it validated the demonstrated capability is "widely available from other models (including OpenAI's GPT-5.5)" and used daily by defenders. ClaudeDevs adds that new sessions fall back to the user's default model or Opus 4.8 and existing Fable 5 sessions end with an error. Scoble reports an SF entrepreneur whose build was killed mid-run and argues the order "hurts American national security" because companies were using Mythos to close their own security holes. ns123abc relays Commerce Secretary Lutnick's letter to Dario Amodei: export controls on any location outside the US, all foreign persons within the US, license required for export, re-export, or domestic transfer. See today's digest Global View.
Anthropic resets all rate limits (@ClaudeDevs). Amid the disruption, ClaudeDevs reset 5-hour and weekly rate limits for all users, presumably to soften the blow of users being bumped off Fable 5 onto other models mid-workflow.
NVIDIA AgentPerf: Blackwell runs 20x more agents per megawatt than Hopper (@nvidia · NVIDIA blog). NVIDIA and Artificial Analysis launched AgentPerf, billed as the industry's first agentic-AI infrastructure benchmark. The motivation: an AI agent chains dozens to hundreds of model calls, using tools and gathering context until a task is done, and existing benchmarks were not built for that workload. In the first published round, the Blackwell Ultra NVL72 platform leads, with the GB300 NVL72 delivering up to 20x more agents per megawatt than Hopper. This is a real efficiency-per-watt number for agentic serving, the hardware-side counterpart to today's agent-efficiency papers.
Google Gemini-SQL2 hits state-of-the-art text-to-SQL on BIRD (@GoogleResearch). Google introduced Gemini-SQL2, a text-to-SQL system powered by Gemini 3.1 Pro, claiming state-of-the-art on the competitive BIRD benchmark, which measures execution-verified accuracy (the generated SQL not only looks right, it runs and returns the correct result). The thread stresses that complex business contexts make accurate SQL generation hard, and that improved SQL understanding feeds Google's broader data services.
kilocode ships REVIEWS.md, repo-specific review standards (@kilocode · blog.kilo.ai). Most AI review tools run a generic rubric on every repo and miss what a team actually cares about. kilocode's REVIEWS.md is an open-standard Markdown file (like a README, but for a review agent) that defines repo-specific conventions, and the attached image shows an example with error-handling rules ("flag any unhandled promise," "require typed error returns") and style rules ("don't nitpick line length," "enforce named exports"). A Code Review Memory mode watches how you respond to PRs (dismissing nitpicks, acting on error-handling comments) and proposes updating REVIEWS.md to match, even opening the PR. This is the same "compile the user's standards into a gate" pattern as today's TRACE paper in the digest.
Google Research: phone-cluster computing and dermatology AI (@GoogleResearch). Two lighter items: a blog on "phone cluster computing," a low-carbon compute platform that gives retired phones a second life and avoids new raw-material extraction; and findings that dermatology AI tools may help laypeople name skin conditions and decide next steps. Both are Google's responsible-AI / sustainability beat rather than core research signal.
NVIDIA: Jensen's "5-layer AI cake" and SpaceX IPO (@nvidia). NVIDIA amplified a Sequoia interview where Jensen Huang frames computing's biggest shift in 60 years as a move "from retrieval to generation," a multi-trillion-dollar opportunity across a five-layer AI ecosystem. Separately NVIDIA congratulated SpaceX on its IPO, noting a near-decade partnership from the first DGX-1 in 2016 to a custom DGX Spark handoff at Starbase. Mostly vision-and-PR, but the retrieval-to-generation framing is the throughline.
magicsilicon: 8 of the top 10 companies design silicon (@magicsilicon). A one-line observation with a chart: of the new top-10 companies by market value, 8 design silicon chips, 1 manufactures them for the other 8 (TSMC), and 1 drills for oil. A compact statement of how thoroughly the compute build-out has reshaped the corporate league table.
kilocode's Fable 5 "KISS" experiment (@kilocode · blog.kilo.ai). In its suspension thread, kilocode linked an earlier experiment arguing Fable 5's failure modes are subtler than older models': it impressed at UI and front-end (one-shotting a GitHub clone from a screenshot) but tended to over-engineer deeper tasks, with mistakes that are harder to spot precisely because the output looks polished. Relevant color on what "this powerful" actually meant in practice.
Anything: agents building and publishing apps end-to-end (@Scobleizer). Scoble is building a mobile app with the "Anything" CLI, which claims Claude Code, OpenClaw, Codex, and Hermes can now design, build, and publish through it "with no human in the loop." A data point on the autonomous-app-builder trend, though promotional and thin on detail.
SpaceX IPO / Elon trillionaire noise (@ns123abc, @brivael). Heavy volume on Elon Musk crossing $1T net worth and SpaceX's IPO debut. No AI-research substance. Skip.