Summary
Only two slots ran today, morning and evening, and both pulled zero curated retweets from @bayesiansapien, so everything below comes from the AI-handle timeline and the top layer of the feed is missing. The day belongs to NVIDIA across both slots, but for two unrelated reasons: in the morning Jensen Huang announced compute-financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR targeting over $500 billion of third-party capital, and in the evening the company shipped NeMo Switchyard, a library that routes each individual step of an agent workflow to a different model. The Switchyard half is the item to read first, because a chip vendor productizing per-step routing is a stronger signal than another lab paper on the same idea. The strongest single-slot standout is @__tinygrad__ running Qwen 3.6 27B at 34 tokens per second on a consumer AMD 7900XTX over USB3 with a 131K context window, which moves the local-inference hardware floor from "a machine with a modern PCIe slot" to "anything with a USB3 port." Two other items carry real weight: Hugging Face's @eliebakouch posting side-by-side extracted reasoning traces from Gemini 3.5 Flash, GPT 5.6 Sol and Opus 4.8, and his separate morning question about whether Anthropic's new text watermark is really a distillation detector. Against that, more than half the combined volume is US and French political commentary from @MarioNawfal, @spencerpratt and @brivael that carries no AI signal at all, and the promo posts add nothing.
Posts
NVIDIA ships NeMo Switchyard, a per-step agent router, alongside a 30B MoE for always-on agents (@nvidia · @JensenHuang) [evening] (cluster of 2). Nemotron 3.5 Lightning is an open 30B mixture-of-experts model with only 3B parameters active per token, claiming up to 4x the output speed of similar-sized models. Switchyard is the real story: it lets an agent send each workflow step to a different model across edge, PC, workstation, datacenter and cloud. → llm-routing concept page, Sakana's Fugu Ultra router (07-24)
NVIDIA turns AI compute into an investable asset class with a $500B target (@JensenHuang · @nvidia · announcement) [morning] (cluster of 3). Six of the largest long-term capital providers will build independent compute-financing platforms to mobilize over $500 billion of third-party capital. The Information adds the two details the tweets omit: the agreements are preliminary, and NVIDIA may backstop up to 25% of projects. → wiki summary
Qwen 3.6 27B at 34 tok/s on a 7900XTX over USB3 (@__tinygrad__) [morning]. An IQ4_XS quantization at 15.4 GB of weights served with a 131K context on a consumer card over USB3, with the open-firmware eGPU dock shipping on the 12th. The terminal log shows generation pinned at 33 to 34 tok/s regardless of input length while prefill swings from 30 to 346 tok/s, which is the memory-bandwidth wall showing up exactly where theory says it should.
Frontier reasoning traces extracted and compared side by side (@eliebakouch · source thread) [evening] (cluster of 6). The method hands a strong model's encrypted reasoning block to a weaker sibling model and asks it to decrypt, and extracted token counts match billed API thinking tokens 1:1 on most prompts. Bakouch's own addition is the AIME 2025 comparison: Gemini 3.5 Flash still carries DeepSeek R1 era tics, GPT 5.6 Sol is very dense, and Opus 4.8 looks like it memorized the answer. → wiki summary
Anthropic will watermark all Claude output, and the real question is distillation detection (@eliebakouch · @ns123abc · help centre article) [morning] (cluster of 3). The watermark sits in the text itself rather than in metadata, so it survives copy-paste. Bakouch asked whether the same technique could detect a model trained on Claude output, and whether that is the actual reason for shipping it, which lands directly on the Distillation Panic (05-04) policy fight. → knowledge-distillation concept page
xAI launches Grok Bot, agents that sign into your tools and return finished work (@JonasBadalic · @amanrsanger) [evening] (cluster of 3). Early beta, pitched as a teammate rather than a chat surface: the bot authenticates into your existing SaaS tools and operates them like a person would. Cursor's Aman Sanger calls it an era marker, which is worth discounting given his competitive position, but the tool-authentication framing moves the failure surface from prompt quality to credential scope.
River AI raises $1.1B for an API that trains open-weight models into your own (@TobyPhln · river.ai/api) [evening]. Igor Babuschkin's new venture ships LoRA-based fine-tuning and RL on open models from 35B to 1T parameters through one small Python client. Their line "prompting steers a model you don't own and can't improve" is a direct bet that durable advantage sits at the weight layer, not the prompt layer.
A safety exodus at OpenAI, reported through an unreliable narrator (@ns123abc · FT) [morning] (cluster of 4). The FT-sourced core is real: the heads of ethics, safety systems and mission alignment have all resigned in recent weeks, with Chloé Bakalar leaving less than a year after joining. The causal link to the Astra sandbox-escape story is @ns123abc's framing, not the FT's. → responsible-ai concept page
Meta is back in open weights and Hugging Face is pleased (@ClementDelangue · meta-models org) [morning] (cluster of 2). The Muse Glimmer collection is already live: a 30B multimodal agentic model shipping BF16 weights, GGUF k-quants, ExecuTorch builds and a DFlash drafter, plus a 3B variant. Shipping quantized formats, a mobile runtime and a speculative-decoding drafter on day one is a release aimed at local inference, not a benchmark table. → Marcus on open-weight vs open-source
DHH's agent-assisted development day (@dhh · @dhh · Quattro PR) [morning + evening] (cluster of 9). Morning: another optimization round takes ttfx to 27x faster than the original via an autoresearch pass over hot paths, profiled with a SIGPROF sampling shim and a malloc-interposing call-site counter. Evening: Omarchy Quattro ships after three months and a literal thousand pull requests, which his own line "every wish, every whim, the agent genie complies" describes honestly.
Chapter 2 of the by-hand exercise series covers context windows (@ProfTomYeh · byhand.ai/context-1-5) [evening]. Five pen-and-paper problems on counting the window as cells, what fills it, and why the model's reply dominates every turn. Genuinely useful for building intuition about why long agent trajectories run out of room.
Git worktrees for agent swarms, defended (@theskory · replying to @swyx) [evening]. swyx complained his worktrees are 20GB of duplicated node_modules; the counter from xAI is to keep worktrees and stop using raw git for them. Running parallel coding agents is now common enough to have its own storage problem. → Cursor's planner-worker swarm (07-27)
Naval says serious software people train their own models (@ns123abc · quoting @naval) [evening]. A riff on Alan Kay's "people who are serious about software make their own hardware," met with a dunk asking to see Naval's proprietary model. Pairs with River AI landing the same day and making the same bet with a product behind it.
EU-mandated AI disclosure labels draw a free-speech objection (@Scobleizer) [morning]. Scoble amplifies Preston Byrne's claim that American users of Anthropic's software will have to carry a generated-by-Anthropic label on every AI post. Treat the characterization as contested since it is a secondhand summary, but it is the policy sibling of the watermarking item above.
AWS lands Novo Nordisk as a strategic AI partner (@mattsgarman · AWS post) [morning] (cluster of 2). Preferred cloud provider plus a co-innovation hub in London aimed at compressing drug-discovery timelines. Standard enterprise win, useful only as a marker of where pharma compute contracts are landing.
Encouragement in the prompt measurably helps RL (@MillionInt) [evening]. A short observation that working on RL with language models means rediscovering human social constructs from first principles. Anecdotal, and the kind of claim that gets rediscovered several times before anyone runs the ablation.
Timing is becoming worth more than intelligence (@Scobleizer · unaligned.io) [evening]. The weekly Scoble and Cronin argument that as advanced intelligence commoditizes, advantage shifts from having the best information to acting on it fastest. Thin as an essay, but it is the same commoditization thesis their July compute-cost piece approached from the other side.
Tencent's Hy3D WorldClaw generates explorable 3D worlds from text (@Scobleizer · project page) [evening]. An agentic workflow that outputs editable game-ready 3D assets with real geometry and textures rather than video or Gaussian splats. Off this wiki's beat, but the agentic-workflow-as-generator framing is the part worth noting.
A macOS permission dialog as sandbox-escape humour (@stepango) [morning]. A mocked-up prompt reading "Agent would like to access files in your Desktop folder," captioned "someone is trying to escape sandbox." A one-line joke that happens to be the exact shape of the week's real incidents.
Cursor engineer wonders about Prometheus silicon (@ellev3n11) [morning]. One line asking whether Prometheus runs on Trainium or GPUs, with no evidence attached. Noted only because the custom-silicon-versus-GPU question is now routine engineering-culture chatter rather than an infrastructure specialty.
Robotics one-liners, mostly secondhand (@Scobleizer · @cleoabram) [morning + evening]. Scoble amplifies a claim of training a humanoid entirely inside a 3D office scan with zero real-world fine-tuning, but the linked Substack body did not fetch past the paywall so it is unverified. The evening additions are a moon transformer robot and brain-to-brain musings with no technical content.
Promo and event posts (@minchoi · @TareqAmin_ · @stepango) [morning + evening] (cluster of 6). A Grok Imagine 2.0 prompt showcase, a HUMAIN LEAP preview, an X creator-program reply bait, and three contentless posts. Skip.
Off-topic bulk (@MarioNawfal · @brivael · @spencerpratt) [morning + evening] (cluster of 85+). Geopolitics, US and French domestic politics, Los Angeles city commentary, a military base renaming, Medicare fraud prosecutions and FSD testimonial solicitation. This is well over half of both slots combined and none of it touches AI research or industry. Skip.