Summary
The day's strongest cross-slot cluster is the Anthropic Fable 5 governance saga, surfacing in both morning and evening: a federal class-action fraud suit over Max plan usage limits, rumors of an ID-gated US-only return, claims of a government directive behind the shutdown, and Anthropic reportedly lobbying to reverse it. The other recurring thread is Kilo, which turned the Fable outage into a routing pitch across morning and evening, plan with the strong model and implement with a cheaper one for 59% less, alongside a "Product Week" recap of five shipped agentic-coding features (review-standard files, parallel git-worktree agents, MiniMax M3 token plans). Those two threads are the real signal; everything else is thin. The single-slot standout is the evening's Hugging Face plus Cursor engineers dissecting the loss curve and learning-rate schedule of Mistral's next model ("le chaton fat"), a rare concrete pre-release read. Two sharp individual takes are worth noting: George Hotz arguing closed-model providers should be fully liable unless they release weights, and eliebakouch openly wondering whether model-improvement perception is itself reward-hacked. The afternoon slot was near-empty, and a large volume of @brivael French political commentary, SpaceX valuation reposts, and assorted off-topic noise can be skipped.
Posts
- Anthropic Fable/Mythos shutdown saga (@ns123abc · WSJ) [morning + evening] (cluster of 9). A federal class-action accuses Anthropic of marketing fraud on its Max 5x/20x plans (one user hit 15% of weekly allowance in a single 5-hour sprint). Surrounding posts speculate a US-only ID-verified return, claim a government directive triggered the shutdown, and relay that Anthropic is lobbying to reverse it. Mostly rumor and meme, but the lawsuit and WSJ coverage are real.
- Kilo: plan with the strong model, implement with the cheap one (@kilocode · blog) [evening]. Head-to-head: Fable 5 wrote the better plan (9.1 vs 8.3), but both models passed all 15 acceptance checks implementing it, with GPT-5.5 at $6.30 vs Fable 5's $16.66. A concrete phase-level routing result. See llm-routing and the Kilo routing audit.
- Kilo Product Week: five agentic-coding features in one day (@kilocode · blog) [morning] (cluster of 6). Kilo shipped Coding Plans (MiniMax M3, $20/mo, ~1.7B tokens, 1M context), REVIEWS.md (a per-repo file teaching the review agent your conventions), Kilo Console (browser UI for worktrees/sessions), Agent Manager (parallel agents each in an isolated git worktree), and KiloClaw (hosted OpenClaw with a morning-brief skill). The connecting theme is productizing the harness as where value accrues.
- MiniMax M3 token plans normalize "buy intelligence by the billion tokens" (@kilocode) [morning]. Contrasts Copilot's new usage-based billing ($10/mo of metered credits) with MiniMax's $20-for-1.7-billion-tokens plan. The consumer face of the token-cost-governance shift.
- MiniMax M3 calls the draw (@kilocode) [evening]. Of five models and one human predicting Qatar vs Switzerland, M3 alone took the draw correctly, reportedly weighing form over the betting line. Anecdotal but on-brand for the multi-model thesis. Relates to MaxProof MiniMax M3.
- Mistral's next model loss curve ("le chaton fat") (@eliebakouch, @ellev3n11) [evening] (cluster of 2). HF and Cursor engineers note the learning-rate schedule matches DeepSeek V2's timing, except the second phase looks like a slight decay rather than constant. Hints at a pending Mistral tech report worth tracking.
- Grok Build agent dashboard (@JasonBud, @ns123abc, @milichab) [evening] (cluster of 3). Grok Build adds a dashboard to manage multiple agents at once (one user running ~5 in parallel); a Warp integration lets a SuperGrok subscription drive Grok models inside Warp.
- George Hotz: make closed-model providers liable, or release weights (@tinygrad) [morning]. Proposes 100% liability for closed-model outputs (copyright, libel, fraud), with safe harbor only if weights are open, pitched as the real AI-safety legislation. A clean articulation of the open-weights policy camp.
- eliebakouch: is model-improvement perception itself reward-hacked? (@eliebakouch) [morning]. Bakouch admits missing Fable after a day and finding some Opus outputs dumb, but cannot tell whether the capability difference is real or his own perception being reward-hacked. An honest note on the unreliability of vibe-based model evaluation.
- Chamath: a 1GW data center now costs $100B (@ns123abc) [morning]. Relayed via video with no primary source. If accurate, a striking per-gigawatt capex jump consistent with the "power is the scaling constraint" read. Treat as unverified.
- The rise of World-Action Models (@Scobleizer · NVIDIA blog) [evening]. Moritz Reuss surveys WAMs (robot policies on pretrained world-model or video backbones) as a second recipe for robot foundation models. Robotics-adjacent, low priority, but a clean overview.
- HUMAIN Fabric and Dr. Nora AlTwairesh recognition (@TareqAmin_) [afternoon]. HUMAIN's CEO ties Dr. AlTwairesh's work to HUMAIN Fabric, a sovereign governed data foundation. Recognition post, no technical detail. Reposted in Arabic.
- Satya Nadella long-form X article (opaque) (@magicsilicon, @ns123abc) [morning] (cluster of 2). Two handles repost the same
x.com/i/article/...Nadella piece; the body could not be fetched (expired X cookies), so substance is unknown. - Jerry Tworek: muon vs shampoo, Claude vs Gemini in a Polish village (@MillionInt) [morning]. OpenAI's Tworek jokes his rural-Poland parents debated Claude vs Gemini. Light, but a marker of how far frontier-model brand awareness has diffused.
- SpaceX valuation and revenue reposts (@ns123abc, @brivael) [afternoon]. Musk's ~$1T-revenue-by-2030 claim and a $2.2T overnight market cap. Business-of-tech, not AI. Skip.
- Scobleizer threads, AlignedNews/AhaCreator promo (@Scobleizer) [morning + evening]. Mostly personal atmosphere plus a vendor demo of an influencer-marketing AI agent. Skip.
- Off-topic noise (@brivael French politics and crypto, @AustinJustice, @NICKIMINAJ, @spencerpratt, @heavypulp, @TareqAmin_ Founders Forum thanks) [morning + afternoon + evening]. No AI substance. Skip.