Summary
One genuinely good item in sixty-two tweets, and it lands squarely in routing. Kilo Code ran an A/B on its Auto Model router against hand-picked models on the same backend build, and the router won on cost by more than 2x while producing a functionally identical service, which is the third time this quarter Kilo has published numbers on the plan-versus-implement model split. The other real signal is Dario Amodei's post stating flatly that Anthropic has never advocated banning Chinese open-weights models, which arrives three days after NVIDIA organized an industry letter on the same question and reads as a direct response to being named in it. Cursor launched a ₹649/month India plan built around Grok 4.5 and Composer, with the notable disclosure that India is now its third-largest market at 3M developers and the highest agent-requests-per-developer of anywhere. Everything else is filler: no curated reposts today, and roughly thirty-five of the sixty-two tweets are @MarioNawfal geopolitics and @brivael French domestic politics that the AI keyword filter waved through.
Posts
Kilo Code's Auto Model router beat hand-picked models on cost, not correctness (cluster of 5) (@kilocode · blog). Same backend task, three setups: Auto Model routing frontier-for-planning then efficient-for-implementation cost $1.26, GPT-5.6 Sol everywhere cost $2.90, Claude Sonnet 5 everywhere cost $2.47. All three passed the same 29 of 30 logic checks, so correctness was a wash and the router's only wins were price and being the sole run that matched the fixed API contract exactly. This is the cleanest confirmation yet of Kilo's own plan/implement split finding from June, which argued the expensive model earns its cost during planning and almost nothing during implementation. See also their earlier model-task routing audit and the LLM routing concept page. One caveat: n=1 task, and a single 29/30 tie is not evidence that the cheap model never degrades quality.
Anthropic states it has never advocated an open-weights ban (@ns123abc · Anthropic). Dario Amodei's post says protectionist bans on Chinese open-weights models would not address his actual national security concerns, and that open-weights models without dangerous capabilities are a public good. The timing matters: this lands three days after NVIDIA's industry letter on open weights, and reads as Anthropic declining to be cast as the lab lobbying for restriction.
Cursor Start, a ₹649/month India plan (cluster of 3) (@cursor_ai · @amanrsanger · blog). Grok 4.5 and Composer, UPI payment, plus cloud agents, iOS, MCP servers and hooks. The disclosure buried in the launch is the interesting part: India tripled to over 3M developers in a year, is now Cursor's third-largest market, and runs more agent requests per developer than any other country. A frontier-model-free tier priced at roughly $7 is a bet that Grok 4.5 and Composer are cheap enough per token to make agent-heavy usage sustainable at that price.
Kilo's weekly model roundup, four new models in a week (@kilocode). Framed as a question about whether the frontier-to-open-weight gap is closing. No data attached, but it is the same release-churn premise that motivates the Auto Model experiment above: six major releases in a month makes manual model selection a standing chore.
Anthropic's book-scanning defense, as a shitpost with a real citation (@ns123abc · court order). A joke about buying millions of physical books, stripping the spines, scanning every page, and separately training on pirated copies. The follow-up tweet links the actual Bartz v. Anthropic order, which is the document where the destructive-scan-is-fair-use, pirated-library-is-not distinction was drawn. Comedy with a primary source attached.
Do AI glasses need a camera? (@Scobleizer). Reacting to a Gurman report that Apple may ship AI glasses without a recording camera, Scoble argues the phone in your pocket already has three better cameras and a 3D sensor, so the glasses may not need one. Speculative, but the framing is right: the interesting design question is what sensing the glasses must own versus what they can offload.
Polygres pitch: turn your entire database into a context window (@Scobleizer). A YC application video repost. The one-line claim is worth noting as a naming of the pattern, though there is nothing here about how it actually works.
Grok 4.5 praised for math work (@MarioNawfal). A vibe review of Grok 4.5 on programming language theory math, tagging @Grok and @xAI. Reads as promotional. Skip.
X Money rolls out to US Premium subscribers (@stepango). Payments landing inside X for Premium and Premium+ in the US. Not AI, but relevant to the everything-app strategy that funds xAI.
Warner Bros / Paramount merger op-ed (@spencerpratt · Guardian). Cumberbatch, Wong and Cumming ask the UK government to block the merger; the repost accuses them of selective outrage. Media consolidation, not AI.
Geopolitics, French politics, and off-topic bulk (cluster of roughly 37). @MarioNawfal posted fifteen items on a M7.1 Kumamoto earthquake, Israeli buffer zones in southern Lebanon, Ukrainian drone strikes on Iranian ships in the Caspian, an Oman proposal to jointly administer the Strait of Hormuz, a leaked White House plan to pay states to host nuclear waste, and a DEA assessment that Operation Southern Spear has not reduced cocaine supply. @brivael posted nineteen on Ursula von der Leyen, the AfD platform, a Sarah Knafo book launch, and Musk appreciation. @ns123abc added a few shitposts, @heavypulp three image-only posts with no text, and @AustinJustice a thread on Austin light rail costs. No AI substance. Skip.