Summary
One release owns this slot. Cursor open-sourced Mixture-of-Kittens (MoK), the mixture-of-experts training megakernel it uses on NVL72 racks, claiming 2.37x over the strongest public baselines and 1.41x end-to-end training throughput over its previous DeepEP-based stack, and six posts across four accounts carry it, including two from Cursor engineers and one from Hugging Face's Elie Bakouch. It is the most substantive open-source GPU artifact to cross this feed in weeks, and the interesting details are not the speedups but the engineering choices: full determinism, ring token buffers to remove CPU-GPU synchronization entirely, and mxfp8 rather than nvfp4. Second in weight is tinygrad, which paired a genuine architectural argument ("GPUs need a new operating system," because we launch kernels at what are really 256 independent processors) with a public offer to buy Intel's unsold DC Max 1450 stock and turn it into $5,000 DeepSeek V4 Flash boxes. Everything else is a step down: NVIDIA shipped Alpamayo 2 Super as an openly licensed reasoning model for robotaxis, the Linux Foundation opened comments on SAFE incident-sharing guidelines with 120-plus Open Secure AI Alliance members behind it, Pokee launched a 28B model claiming 10M usable context at $1 per million output tokens, and NIK published the sworn-declaration detail from Apple's preliminary-injunction motion against OpenAI. Signal density is low overall: roughly 46 of 79 posts are US, French and Middle East politics from five accounts that are AI-adjacent in name only, and the curated repost feed is empty again.
Posts
Cursor open-sources Mixture-of-Kittens, its MoE training megakernel for NVL72 (cluster of 6) (@cursor_ai, @cursor_ai, @cursor_ai, @eliebakouch, @ellev3n11, @ellev3n11 · blog · code). MoK fuses every part of the mixture-of-experts layer, the routing step where each token is sent through a small subset of specialized sub-networks, into one deterministic GPU kernel covering both communication and computation, targeted at NVIDIA's 72-GPU NVL72 Blackwell racks. The motivation is a measured bottleneck rather than a benchmark: while scaling Composer, their agentic coding model, Cursor found the MoE layer alone consuming more than half of end-to-end training time depending on configuration. Reported results are up to 2.37x over the strongest public baselines and, in production across tens of thousands of GPUs, 1.41x end-to-end throughput over their DeepEP-based stack. The details engineers picked up on are the ones that matter: bf16 and mxfp8 support, pull-based dispatch, configurable overlap granularity, full determinism, and no CPU-GPU synchronization via ring token buffers. Bakouch flagged the format choice, mxfp8 with no nvfp4 anywhere, which is the same axis MXAttention (08-01) argued about when it showed MXFP4's accuracy deficit against NVFP4 was two fixable numerical bugs rather than a property of the format. Cursor picking the open microscaling format for production training is a real vote in that direction. Full determinism deserves its own note: it is the property that makes RL and distillation training runs reproducible, and it is usually the first thing sacrificed for overlap. See GPU kernels for where this sits against the vendor kernel-agent stacks.
tinygrad: GPUs need a new operating system, and Intel should sell its unsold cards (cluster of 2) (@tinygrad, @tinygrad). The first post is the argument: we are stuck in a paradigm of launching kernels when the hardware is really 256 independent processors with their own synchronization and communication primitives, so the abstraction, not the silicon, is the limit. Read next to the Cursor release in the same hour, the two are the same complaint from opposite ends, since a megakernel is what you build when kernel launch boundaries are the thing costing you performance. The second post is an open offer to Intel to sell 5,000 DC Max 1450 cards for $1M, which tinygrad would turn into $5,000 DeepSeek V4 Flash boxes, framed as a cultural test. That follows directly from their 08-01 production profile, where they got DeepSeek-V4-Flash to roughly 245 tok/s sustained on two RTX Pro 6000 cards by composing W4A8 kernels, an fp8 KV cache and fixed-depth speculative decode. They already know the model runs on workstation-class hardware, so this is a supply question, not a feasibility one.
NVIDIA ships Alpamayo 2 Super under an open commercial license (cluster of 3) (@JensenHuang, @nvidia, @nvidia · blog). An open reasoning model for autonomous vehicles, released for commercial use under OpenMDW-1.1 so teams can inspect, fine-tune and deploy it, with 360-degree awareness, high-level driving decisions and automated reasoning labels. The pitch is explicitly about long-tail events rather than everyday scenarios, which is where reasoning-before-acting is supposed to earn its cost. Huang's framing is the strategic part: open weights advance safety and security, and robotics is the next wave. Consistent with NVIDIA's open-weights position (07-25), and the inspectability argument for a safety-critical deployment is the strongest version of that case anyone has made.
Linux Foundation opens comments on SAFE, a shared incident-findings standard for agentic security (cluster of 2) (@nvidia, @nvidia · blog). Timed to Black Hat, the Open Secure AI Alliance is now past 120 organizations and the Linux Foundation has published a Request for Comments on Shared AI Findings Exchange (SAFE), a proposed format for turning confidential agentic-AI incident findings into protection the rest of the ecosystem can act on. New members named include Akamai, Amazon, Cisco, Cloudflare, Cognition, CrowdStrike, LangChain, Microsoft, Okta, Palo Alto Networks, Perplexity, Red Hat, Uber and Visa. This is the missing coordination layer for the pattern the wiki has been tracking since the agent-intrusion timeline (07-29) and the agentic-CVE amplification stack (07-27), where the same agent capability that finds a bug also scales its exploitation. An RFC is not a fix, but a disclosure format is the precondition for one.
Pokee-Isaac 28B claims 10M context at 28B parameters (cluster of 2) (@Scobleizer, @Scobleizer · console). The claim is all three of small enough to self-host, very long context and real agentic performance in one model: 28B parameters, 10M tokens of usable context, strong tool use and multi-step execution, deployable in a VPC or on-premises, at $0.15 per million input and $1 per million output tokens. "Usable" is the load-bearing word and no benchmark accompanies it, so treat this as a launch claim rather than a result. It is worth logging anyway because it lands on exactly the constraint SemiAnalysis's Kimi K3 primer (08-04) quantified, that a B300 node holds only about 3.25M tokens of KV budget after weights. A 28B model with a 10M window either has a very aggressive KV cache story or a much narrower definition of usable than the number implies. Click through and check what they actually measured.
Apple's preliminary-injunction motion against OpenAI, with sworn declarations (cluster of 4) (@ns123abc, @ns123abc, @ns123abc, @ns123abc · OpenAI statement). Apple has moved for a preliminary injunction plus forensic supervision and expedited discovery, backed by nine sworn declarations and a 28-page memorandum, alleging an eight-year Apple employee now at OpenAI exploited an authentication bug to take trade secrets on at least five separate occasions between February and April 2026 while already working for OpenAI. NIK calls OpenAI's public rebuttal, which the afternoon slot covered, thoroughly deceptive and says the timeline is being reported too generously. Sworn declarations are a materially higher evidentiary bar than a complaint, so this stopped being a press-release fight today. Contested claims on both sides, so click through and read the filings rather than either party's summary.
Kimi Delta Attention worked out by hand in Excel (@ProfTomYeh · seminar). Yeh implemented Kimi Delta Attention in a spreadsheet for this week's seminar on Kimi 3, which also covers attention residuals, and the same series has already hand-worked Gemma 4's alternating global and local attention and Qwen 3.6's long-context scaling. The delta rule family is the linear-attention line the wiki tracked through Gated DeltaNet 2 (05-24), which decoupled erase from write so the recurrent state could forget and store independently, and a worked numerical example is the fastest way to see why that separation matters. Useful companion to attention mechanisms and to the Kimi K3 architecture primer.
dhh on Claude Code refusing the ~/.agents/skills convention (cluster of 2) (@dhh, @dhh · docs). DHH is annoyed that Claude will not look for skills in
~/.agents/skills, and when told he can symlink it, points out that the docs acknowledging the workaround is the insult. Small as a complaint, real as a signal: agent tooling is at the stage where a cross-vendor filesystem convention would compound, and every vendor keeping its own path is a tax paid by anyone running more than one agent.Grok Build moves to Grok 4.5 (@MarioNawfal · cli). xAI's coding agent now runs on Grok 4.5 with a native subagent view, Plan Mode integration, more commands it will execute without asking, and easier navigation of long responses. The framing worth noting is the permission boundary, since expanding the set of commands an agent runs unsupervised is the same trade the SAFE proposal above exists to clean up after.
Kilo Code's JetBrains plugin built on split mode (cluster of 2) (@kilocode, @kilocode · blog). A native Kotlin plugin split across frontend and backend modules using JetBrains' split mode API, so the UI runs locally while indexing, terminal and file access run wherever the code lives. Not a webview and not a VS Code extension in a frame, which is how most AI coding tools handle JetBrains. The remote-development case is where agent tooling actually gets deployed in enterprises, so this is a more consequential engineering decision than it sounds.
S&P 500 hits an intraday record as Amazon crosses $3 trillion (@MarioNawfal). The index touched roughly 7,700, with over 84% of S&P 500 companies beating earnings expectations this quarter and Microsoft, NVIDIA, Meta and Alphabet all up on the week. Big Tech is doing most of the lifting, which keeps the compute capex story funded for another quarter regardless of what the research says.
An ESP32 voice AI build (@minchoi · repo). Code for putting an AI assistant on an ESP32 microcontroller. Small, but the microcontroller tier is where anything claiming to be edge AI eventually has to run.
"Load-bearing" as the next em dash tell (@mlevchin). Levchin asks whether "load-bearing" has become the new giveaway phrase for AI-written prose. Half joke, but the stylometric arms race between LLM writing tells and the people learning to avoid them is real and self-defeating for anyone hoping detection stays cheap.
Scoble on the AI experience gap (@Scobleizer · newsletter). The weekly free newsletter with Irena Cronin argues human experience, judgment and lived context become the competitive advantage as models absorb knowledge work. No evidence attached, and it is the most common shape of AI opinion writing right now, but click through if the framing is useful.
Vocci AI ring, an "AI wearable that captures your context" (@Scobleizer). Titanium ring, always-on ambient context, tag it inside your chat app. Sponsored-shaped post with no measurement. Skip.
Google Research booth activities at Deep Learning Indaba (cluster of 2) (@GoogleResearch, @GoogleResearch). A WeatherNext cyclone-forecasting demo and a quiz with prizes, both scheduled events at a conference booth. Skip.
US, French and Middle East politics, Anduril teasers, tennis, Lollapalooza (cluster of 46) (@MarioNawfal, @brivael, @spencerpratt, @HouseGOP, @AustinJustice, @heavypulp, @stepango, @MillionInt). Strait of Hormuz negotiations, Syrian troop movements, German party-ban petitions, Ukrainian conscription footage, a Paramount op-ed, plus assorted memes. No AI content. Skip.