Summary
A thin morning for research signal and a loud one for politics. There were no curated @bayesiansapien retweets, and the AI handle feed was dominated by the Anthropic-Fable governance saga and off-topic noise. The single substantive cluster is Kilo's "Product Week" recap (five shipped features in one day), which is the clearest social read on where agentic coding tooling is heading: review-standard files, parallel git-worktree agents, browser consoles, and direct MiniMax M3 token plans, all pointing at the same "learn-the-harness, meter-the-tokens" theme the day's digest tracks. The loudest non-cluster thread is governance: @ns123abc relaying that Anthropic is lobbying to get Fable 5 unbanned, plus reposts of a Satya Nadella long-form X article that could not be fetched (expired cookies). Two genuinely interesting individual signals: George Hotz (@tinygrad) floating that closed-model providers should be fully liable for outputs unless they release weights, framed as "real AI safety legislation," and HuggingFace's @eliebakouch openly wondering how much of model-improvement perception is just human psychology being reward-hacked. Chamath's claim (via @ns123abc) that a 1GW data center now costs $100B is the day's sharpest infra number. Everything else, Austin crime, a presidential birthday, LA mayoral politics, French political feuds, is noise.
Posts
- Kilo Product Week: five agentic-coding features in one day (cluster of 6) (@kilocode, blog.kilo.ai). Kilo shipped: (1) Coding Plans, buy a MiniMax M3 plan ($20/mo, ~1.7B tokens, 1M context) with your existing Kilo balance, no BYOK juggling; (2) REVIEWS.md, a repo-root file that teaches the review agent your conventions per-repo, the review-standard analogue of AGENT.md, the same "compile standards into agent gates" pattern as the 06-13 Cursor Auto-review and TRACE work; (3) Kilo Console, a browser UI to manage worktrees/sessions/settings without hand-editing JSON; (4) Agent Manager, run multiple agents in parallel, each in an isolated git worktree, to avoid the file-clobbering chaos of six agents in six tabs; (5) KiloClaw, a hosted OpenClaw with a morning-brief skill that pulls calendar conflicts unprompted. The connecting thread, and why it sits next to today's HarnessX and HarnessBridge papers in the digest, is that the harness (review rules, worktree isolation, skills, token plans) is being productized as the place value accrues. Their own framing: most engineers run 2-5 coding tools at once and every one is a moving pricing target (Roo Code shut down, Copilot went usage-based), so own the layer above them.
- MiniMax M3 token plans normalize "buy intelligence by the billion tokens" (@kilocode). A standalone follow-up contrasting Copilot's new usage-based billing ($10/mo buys $10 of metered API-rate credits) with MiniMax's $20-for-1.7-billion-tokens plan. The explicit pitch, "one plan is measured in dollars, the other in billions", is the consumer face of the token-cost-governance shift Meta and Nadella set enterprise-side on 06-14.
- George Hotz: make closed-model providers liable, or release weights (@tinygrad). Proposes that providers of closed-source models be 100% liable for their outputs (copyright, libel, fraud), with safe harbor only if they open the weights, pitched as the actual AI-safety legislation worth having. A pointed counter-frame to the week's Fable export-control story: where the White House gated weights as a risk, Hotz argues open weights should be the liability shield. Opinion, but a clean articulation of the open-weights camp's policy ask.
- eliebakouch: is model-improvement perception itself reward-hacked? (@eliebakouch). HuggingFace's Elie Bakouch admits he "kinda misses Fable" after using it one day and finds some Opus outputs dumb, but genuinely cannot tell if it is a real capability difference or his own perception being reward-hacked. A small but honest note on the unreliability of vibe-based model evaluation, the human-side mirror of the benchmark-saturation problem (Agents' Last Exam) in today's digest.
- Anthropic lobbying to unban Fable 5 (cluster of 3) (@ns123abc). @ns123abc relays that Anthropic is persuading the government to reverse the Fable 5 suspension, alongside meme-style "ITS HAPPENING" reposts (one of an old Elon "Macrohard Microsoft" tweet). Low-substance individually, but corroborates the Interconnects and Gary Marcus reads in the digest that the Friday shutdown is mid-negotiation, not settled.
- Chamath: a 1GW data center now costs $100B (@ns123abc). Relayed via video with no primary source; if accurate it is a striking jump in per-gigawatt capex and lines up with the Semiconductor Newsletter's "data-center power is now the primary scaling constraint" lead. Treat as an unverified but directionally-consistent infra datapoint.
- Satya Nadella long-form X article (opaque, cluster of 2) (@magicsilicon, @ns123abc). Two handles repost the same
x.com/i/article/...Nadella piece with "Just saying..." and "BRO LITERALLY PREDICTED THIS." The article body could not be fetched (expired X cookies), so the substance is unknown. Click through to read. - Scobleizer: Dario, robots, and "a lot of people will soon be out of jobs" (cluster of 7) (@Scobleizer). Robert Scoble's thread is mostly personal (a donated park, a Ferrari F1 win) but threads in speculation about what Dario Amodei will say to government officials and a recurring "pay and pay and pay" framing of AI's coming labor displacement and need for new social programs. Atmosphere, not signal, included for completeness on the Fable-week mood.
- Jerry Tworek: muon vs shampoo, and Claude vs Gemini in a Polish village (@MillionInt). OpenAI's Tworek jokes that his parents in rural Poland were debating whether Claude beats Gemini, and that he is waiting for their take on the muon-vs-shampoo optimizer debate. Light, but a real-world marker of how far frontier-model brand awareness has diffused.
- Off-topic / Skip. @AustinJustice (campus crime), @NICKIMINAJ (a presidential 80th-birthday note), @spencerpratt (LA mayoral politics), @heavypulp (gaming nostalgia), @TareqAmin_ (a Founders Forum thank-you with no concrete claim), @brivael (French political feuds, carried over from the prior slot). No AI substance.
Wiki: today's daily digest covers the harness-learning papers and the Fable governance thread in depth.