Summary
The morning slot carried no curated reposts at all, so everything below comes from the AI handle feed, and the signal concentrated hard in two places. The strongest item by a distance is Cursor open-sourcing Mixture-of-Kittens, a fused mixture-of-experts training kernel for NVL72 racks claiming 2.37x over the fastest public baseline and 1.41x end to end in their own production training, which drew immediate technical amplification from HuggingFace and xAI people rather than the usual launch applause. The second cluster is compute going to orbit: NVIDIA and SpaceX announced Starmind AI1 satellites carrying Vera Rubin NVL72 payloads at 250 kW each with up to a million satellites contemplated, and the same handles carried SpaceX's first post-IPO earnings call, which is where the Grok release roadmap and the Colossus power targets actually came from. Around those, a cluster of four small model and product releases: Mistral's Shieldstral 3B on-device safety model, NVIDIA's Alpamayo 2 Super open autonomous-driving model going commercially licensed, Qwen-Image-3.0 landing on Qwen Cloud, and AWS open-sourcing Kiro Crew for running scheduled agent crews. tinygrad quietly posted the most interesting low-level item of the morning, getting code execution on an AMD 7900XTX through custom firmware. The counter-signal worth noting is a labour-market one, with Scoble reporting SF AI and robotics hiring is blocked on talent scarcity while Seattle software engineers absorb Amazon layoffs. A large fraction of the raw feed, roughly forty of seventy-one tweets, was US and French political commentary from three accounts and carried no AI content.
Posts
Cursor open-sources Mixture-of-Kittens, a deterministic MoE training megakernel for NVL72 (@cursor_ai, blog, GitHub) (cluster of 3). This is the day's substantive release. Mixture-of-experts, where each token is routed through a small subset of specialized sub-networks instead of the full model, spends most of its training time moving tokens between GPUs rather than computing, and Cursor measured the MoE layer at over half of end-to-end training time. MoK fuses all of that communication and computation into one persistent kernel, assigning some streaming multiprocessors to expert arithmetic and others to dispatch and combine, so the two overlap inside the kernel instead of being serialized at launch boundaries. The claimed numbers are 2.37x over the fastest public baseline on an isolated MXFP8 forward pass at expert-parallel degree 64, and separately 1,070.2 versus 760.9 tokens per second per GPU on 512 real GPUs, which Cursor states as 1.41x in their own production training across tens of thousands of GPUs. The underdiscussed property is that the kernel is bitwise deterministic, meaning the same input produces the same output regardless of hardware scheduling, which makes MoE training runs reproducible in a way they currently are not. @eliebakouch of HuggingFace called it "insanely good" and flagged the one thing missing, which is that the release uses mxfp8 with no nvfp4 path despite Blackwell supporting FP4. @stepango at xAI reposted it without comment. See the full write-up.
SpaceX and NVIDIA are putting NVL72 racks in orbit (@nvidia, @ns123abc) (cluster of 2). Starmind AI1 satellites carry NVIDIA Vera Rubin NVL72 compute payloads, with Rubin GPUs and Vera CPUs, at 250 kW peak compute per satellite. The stated ambition is up to one million satellites and 100 GW of orbital compute. NVIDIA's own framing is that this brings "AI factory compute closer to the stars," which is marketing, but the payload spec is the same rack architecture Cursor's kernel above is written against, which makes the pairing on one morning slightly surreal. No launch cadence, power-budget, thermal or downlink numbers accompanied either post, and those are the three things that decide whether orbital compute is a product or a press release.
SpaceX Q2 earnings call: the Grok roadmap and the Colossus power curve (@ns123abc). This thread is the most information-dense non-technical post of the morning and it is a transcript summary rather than an announcement. On models: Grok 4.5 was described as a large improvement, 4.6 lands roughly next week, 4.7 in three or four weeks, and Grok 5 before the end of the year. The line worth keeping is the claim that they will incorporate "the entire corpus of SpaceX data, all the data SpaceX has ever produced over the course of a quarter century" into Grok training, on the theory that this makes it the best engineering model, which is a specific and falsifiable bet on proprietary industrial data as the differentiator. On infrastructure: over 2 GW online by end of 2026, and by end of next year "closer to 10 GW than 5 GW." Set against the same day's financial reporting, that SpaceX burned $16 billion in the quarter on $7.8 billion of revenue with $18.4 billion of capex mostly for AI datacenters, the power targets are the collateral for the burn.
Mistral ships Shieldstral, a 3B open-weights content-safety model for on-device deployment (@stepango reposting @MistralAI, announcement). A small open-weight classifier meant to run locally rather than as a hosted moderation API, which matters because the current default forces every moderation call through a network round trip to the same vendor whose model you are moderating. stepango's only comment was "This is not how I read it," with no elaboration, so the repost carries the release but not an argument. No benchmark numbers or policy taxonomy made it into the captured content.
AWS open-sources Kiro Crew, a workspace for running several scheduled agents at once (@mattsgarman, blog). The AWS CEO announcing it personally signals this is more than a side release, though the blog says it started as a side project by three engineers and grew because builders kept extending it. The pitch is explicitly about the engineer as integration layer: work spans repos, tools, reviews and days, and today you are the one holding it together, reconnecting context and coordinating handoffs, so everything stalls when you step away. Kiro Crew takes a ticket queue, triages, dispatches and identifies owners, with agents wired into existing tools and schedulable for overnight runs, on-call rotations and cross-repo migrations. Worth reading against today's research, where two benchmarks found that agent skill libraries mostly fail to abstract and one paper found they are a durable attack surface.
NVIDIA's Alpamayo 2 Super goes commercially licensed for robotaxis (@nvidia, @zhu_hanqing666, NVIDIA blog, weights) (cluster of 2). NVIDIA describes it as its most capable open reasoning model for autonomous driving, with open commercial licensing, benchmark-leading reasoning and inspectable decisions, aimed specifically at the long-tail rare events that are hard to anticipate and train for rather than everyday scenarios. Inspectability is the interesting claim for a driving stack, since the regulatory case for autonomy turns more on being able to explain a decision than on aggregate accuracy. Hanqing Zhu at Google DeepMind congratulated the team and framed it as physical-world reasoning being "a major mountain to climb," which is a competitor's endorsement.
Qwen-Image-3.0 lands on Qwen Cloud, ranked second overall in the text-to-image Arena (@Scobleizer, model page). Ranked first among Chinese models and second among mainstream models on Arena's text-to-image leaderboard. The capability list is unusually specific about what it is for: prompts up to 4.5k tokens, dense information layouts with images inside images so newspapers, storyboards, menus and exam papers generate in a single pass, text legible down to 10px, native rendering of 12 languages and 20+ fonts, and realistic simulation of web pages, games and live streams. The positioning statement is the giveaway, that the model is not pursuing good looks but usefulness as "a deployable productivity tool," which is a bid for document and UI generation rather than art. Scoble's framing was competitive: "The Chinese are shipping. Will Grok keep up?"
tinygrad got code execution on an AMD 7900XTX via custom MEC firmware (@tinygrad). Two sentences and the most technically interesting low-level item of the morning. Kimi's custom firmware for the MEC, the micro engine compute block that handles command processing on AMD GPUs, ran its first kernel. The stated trajectory is that after spec and HCQ2, "the next phase of tinygrad will be operating system," which is a claim that they intend to displace the vendor driver stack rather than sit on top of it. That is the same direction of travel as Cursor's megakernel above, taking control of scheduling away from the layer that normally owns it, arrived at independently on different hardware.
A $53k desktop that runs DeepSeek V4 Flash on two Blackwells (@tinygrad quoting @elliotarledge). Elliot Arledge reported running DeepSeek V4 Flash 0731 with dspark on just two RTX Pro 6000 Blackwells, and tinygrad's response was a hardware pitch: get the machine for $53k today, including the hardware to install two more GPUs later. The datapoint that matters is not the price, it is that a frontier-tier coding model now fits on two workstation cards, which is the local-inference version of the price collapse DeepSeek has been driving on the API side.
Kimi launches what it calls the world's first AI-native credit card (@hexiang, kimi.com/aicard). Spending earns model tokens instead of air miles. The card page shows two tiers, Select and Max, whose benefits are entirely model quota rather than financial: agent quota for document processing, deep research and website deployment, Kimi Code quota at up to 20x, parallel multi-agent execution, scheduled recurring tasks, and premium data plugins. It is issued with Agricultural Bank of China as a high-end platinum product requiring in-branch application. Hexiang at Google DeepMind reposted it as a joke about XMoney earning SuperGrok Heavy credits, but the underlying move is real and worth watching: bundling inference quota into consumer credit is a customer-acquisition channel nobody in the US has tried.
Pika opens an API club at $10/month across 100+ generative media models (@minchoi, dev.pika.art). An aggregator play, positioned bluntly as "Gen Media is overpriced, we're fixing it," claiming up to 88% cheaper than other aggregators. The published comparisons are specific: Seedance 2.0 R2V 480p at roughly $0.0452/sec against Fal.ai's $0.084 and Runway's $0.36, and MiniMax H3 T2V 2K at $0.13/sec against Fal.ai's $0.26. If those hold, the interesting part is the spread between aggregators for identical models, which suggests the aggregation layer rather than the model layer is where margin currently sits.
NVIDIA argues enterprise AI needs a system of models, not one model (@nvidia). A short post pointing at a longer piece whose argument is that no company would ask one employee to plan a project, write and review code, operate tools, validate results and make every decision, and complex work succeeds because specialists collaborate. This is vendor content and the conclusion is convenient for a company selling the infrastructure to run many models. It is also the industry restatement of the routing thesis this wiki tracks, and it arrived the same day as a paper arguing that the uncertainty signal every mixture-of-adapters router uses is the wrong signal.
Anaconda acquires EnkryptAI, folding AI risk detection into the platform that owns Kilo Code (@kilocode on @anacondainc). Kilo's read is that security and governance have been the missing piece for enterprises trying to scale agentic development, and pairing risk detection with the Anaconda platform lets teams move fast without flying blind. Kilo Code is itself an Anaconda property following its own recent acquisition, so this is a platform consolidating a governance layer under itself. Note the timing against today's research: SkillJack showed that poisoned agent experiences laundered into reusable skills evade safety detection at 11.4% versus 98.5% on the source trajectory, which is exactly the class of risk this kind of acquisition is being sold to cover and exactly the class current scanners miss.
Stacked pull requests are obsolete now that you can delegate to an agent (@JonasBadalic). An xAI engineer arguing that the benefit of stacked-PR workflows and the tooling built to manage diffs died the moment the work could be handed to an agent, and that this is why nobody is excited about a recent launch he does not name. It is an assertion rather than an argument, but it is a specific claim about a developer-tooling category being made obsolete rather than improved, from someone inside a frontier lab.
SF cannot hire AI and robotics talent while Seattle engineers absorb Amazon layoffs (@Scobleizer). Scoble reports that every company he has spoken to in the last month names hiring as its biggest blocker, across engineers, researchers, hardware engineers and electricians, while Seattle software engineers are struggling following Amazon cutting thousands. Two labour markets in the same industry moving in opposite directions, split by specialization rather than geography. Anecdotal and worth treating as such, but it is the only labour-market signal in today's feed.
Scoble on cognitive outsourcing (@Scobleizer). "For the next two years using AI will make you stupider. Then they will sell you cognitive enhancements." He reports Musk told him Neuralink is working on it and says he would buy instantly. Included because it is an unusually candid statement of the dependency thesis from someone deep inside the SF scene rather than a critic outside it, not because there is evidence attached.
Apple briefly pulled Telegram from the App Store over a single moderation evasion (@dhh on @durov). DHH's reaction to Pavel Durov's account that Apple removed Telegram after a user planted illegal content in a public chat, with restoration within hours. Durov's framing is a warning to other app developers about a deliberate attack pattern: plant content, trigger the platform's automated enforcement, take down the app. DHH's is a broader complaint about mobile monopolies. No AI content, included because the attack pattern, using an automated moderation system as a weapon against its own host, is structurally the same laundering-through-an-automated-step problem SkillJack describes today in a different domain.
Omarchy Quattro ships three bespoke C++/Qt apps at about 1MB combined (@dhh). Omacut, Omawrite and Omacalc, theme-synced, in roughly a megabyte total. No AI angle. Noted only as a size datapoint in a week otherwise measured in gigawatts.
xAI's Imagine promotes world generation (@imagine). "Build entire worlds based on your imagination," attached to a user's video. Pure promotional content with no capability claim, benchmark or technical detail. Skip.
Roughly forty tweets of political commentary carried no AI content. @MarioNawfal (18 posts on Iran shipping lanes, Ukraine drone strikes, US Senate races, Ghislaine Maxwell, Flock surveillance cameras), @brivael (20 posts of French domestic politics and Western-decline commentary), plus @NICKIMINAJ and @spencerpratt on celebrity disputes. Two of these brushed AI adjacently, Nawfal on Trump telling communities that turning down data centers is a mistake while acknowledging the local fights are about power and water, and one on Flock's 80,000 cameras generating 20 billion plate reads a month with under 1% connecting to any crime, but neither carried enough substance to stand alone. Skip.