Summary
The evening slot pulled 68 tweets with zero curated retweets from @bayesiansapien, and roughly half the volume is US political commentary that carries no AI signal at all. The strongest item by a wide margin is NVIDIA shipping a routing library: alongside Nemotron 3.5 Lightning, a 30B mixture-of-experts model with only 3B parameters active per token, NVIDIA announced NeMo Switchyard, which lets an agent route each individual workflow step to a different model. That is a hardware vendor productizing exactly the per-step routing idea this wiki has tracked across a year of papers, and it is the item to read first. Second is a six-post thread from Hugging Face's @eliebakouch on the reasoning-trace extraction paper, where he goes past the headline and posts side-by-side traces of Gemini 3.5 Flash, GPT 5.6 Sol and Opus 4.8 on the same AIME problem, observing that Gemini still reads like DeepSeek R1 while Opus appears to have simply memorized the answer. Third, xAI launched Grok Bot into early beta, agents that sign into your existing tools and return finished work, drawing a claim from Cursor's @amanrsanger that each product wave crowns one lab and this one belongs to xAI. The rest is thinner: River AI raised $1.1B for an API that fine-tunes and RL-trains open-weight models from 35B to 1T parameters, and @dhh posted eight times about Omarchy Quattro, three months and a literal thousand PRs of agent-assisted work. Everything from @MarioNawfal, @spencerpratt, @WHFraudTF, @HouseGOP and @DoWCTO is off-beat noise.
Posts
NVIDIA ships NeMo Switchyard, a per-step agent routing library, alongside a 30B MoE built for always-on agents (cluster of 2: @nvidia, @JensenHuang, quoting @NVIDIAAI). Two announcements in one. Nemotron 3.5 Lightning is an open 30B mixture-of-experts model (only a small subset of the network fires per token, here 3B active parameters) aimed at always-on agents doing high-volume specialized work, claiming up to 4x the output speed of similar-sized models. NeMo Switchyard is the more interesting half: a routing library that lets an agent send each step of a workflow to whichever model the operator chooses, across edge devices, PCs, workstations, datacenter and cloud. Huang's framing is "continuous and long-run agents," which is the deployment shape where per-step routing actually pays, since a long trajectory has wildly uneven difficulty across steps and paying frontier prices for all of them is waste. Worth reading against the routing concept page and against Sakana's Fugu Ultra router (07-24), because a chip vendor shipping routing as a library is a different signal from a lab publishing a router.
Frontier reasoning traces, extracted and compared side by side (cluster of 6: @eliebakouch, @eliebakouch, @eliebakouch, @eliebakouch, @eliebakouch, @eliebakouch, on @kotekjedi_ml's thread). The method: hand a strong model's encrypted reasoning block to a weaker, less-guarded sibling model from the same provider and ask it to decrypt and print. It complies. The authors verified their extracted token counts match billed API thinking tokens 1:1 on most prompts. Bakouch flags that the decryption is somewhat lossy but the paper's controls suggest it is accurate, and that the same method pulled credentials and personally identifiable information out of traces sitting in public repositories. The genuinely new content is his comparison of Gemini 3.5 Flash, GPT 5.6 Sol and Opus 4.8 on AIME 2025 Problem 14: Gemini's trace is structured but still full of the "let" and "wait" tics of the DeepSeek R1 era, GPT's traces are very dense, and Opus appears to have memorized the answer outright. He posts the OpenR1-Math-220k dataset as the R1-era reference point. → wiki summary
xAI launches Grok Bot, agents that sign into your tools and return finished work (cluster of 3: @JonasBadalic, @amanrsanger, @JonasBadalic, quoting @bot). Early beta. The pitch is not a chat surface but a teammate: bots authenticate into your existing SaaS tools, operate them the way a person would, and come back with completed work. Cursor's Aman Sanger reads it as a era-marker, "There was OpenAI and chat, then Anthropic and coding. Welcome SpaceXAI and general knowledge work." Take the triumphalism with salt given the source's own competitive position, but the tool-authentication framing is the substantive part, since it moves the failure surface from prompt quality to credential scope. The third post in the cluster is designer Benji Taylor's animated Grok Bot icon, built entirely in code with Grok Build.
River AI raises $1.1B for an API that trains open-weight models into your own (@TobyPhln, quoting @ibab, product at river.ai/api). Igor Babuschkin's new venture, with a New York Times piece on the mission. The v0.1 preview product is LoRA-based fine-tuning and reinforcement learning on open-source models from 35B up to 1T parameters, with train, sample and serve through one small Python client. Their own framing is the sharpest line in the announcement: "Prompting steers a model you don't own and can't improve." The demo shows training Kimi K2.6 with a reward curve. This is a direct bet that the weight-owning layer, not the prompt layer, is where durable advantage sits, and it is the commercial mirror of @naval's post the same day that people serious about software train their own models.
A router-adjacent counterpoint on RL: encouragement measurably helps (@MillionInt). Short observation that working on reinforcement learning with language models means rediscovering human social constructs from first principles, and that words of encouragement in the prompt actually move outcomes. Anecdotal, but it is the kind of thing that gets rediscovered three times before someone runs the ablation.
Chapter 2 of the by-hand exercise series is about context windows (@ProfTomYeh, PDFs at byhand.ai/context-1-5). Tom Yeh released five new pen-and-paper problems on context: counting the window as cells, what actually fills it (system prompt, user turn, model reply), why the reply dominates every turn, whether a new message fits in what is left, and how many turns before the window is full. Follows his 20-problem token chapter. Genuinely useful teaching material for anyone who needs an intuition for why long agent trajectories run out of room. The PDFs are linked directly; click through to read.
Git worktrees for agent swarms, defended (@theskory, replying to @swyx). swyx complained that worktrees must die because his are 20GB of duplicated node_modules. The counter from xAI's side: keep worktrees, stop using raw git for them. Grok Build claims to handle the agent-swarm case properly with a shared base, no disk bloat, no SSD thrashing, concurrent checkouts of the same branch, and fast spin-up and reset, with lighter primitives in progress. A real infrastructure constraint for anyone running parallel coding agents, which is now a common enough pattern to have its own storage problem. Related: Cursor's planner-worker agent swarm (07-27).
DHH ships Omarchy Quattro after a thousand PRs (cluster of 8: @dhh on the Quattro PR, @dhh on a Quattro theme, @dhh on an Apple TV plugin, @dhh on tobi's omasnap screenshot tool, @dhh, @dhh, @dhh). Three months of work and a literal thousand pull requests on one release of the Omarchy Linux setup, with a community plugin marketplace filling in around it. His comment on tobi Lütke's contributed annotation tool, "Every wish, every whim. The agent genie complies," is the honest description of what a thousand-PR quarter looks like now. He also notes a round two recording with Lex Fridman on Monday, saying the year since their six-hour conversation feels like twenty given the age of agents. The remaining posts are WiFi 7 enthusiasm and race-car quips.
Timing is becoming worth more than intelligence (@Scobleizer, newsletter at unaligned.io). Robert Scoble and Irena Cronin's weekly argument: as AI makes advanced intelligence broadly accessible, competitive advantage shifts from having the best information to acting on it fastest, so organizations that execute first outperform those that merely have the smartest model. Thin as an essay, but it is the same commoditization thesis their July piece on running out of cheap compute came at from the cost side.
Tencent's Hy3D WorldClaw generates explorable 3D worlds from text (@Scobleizer, quoting @TencentHunyuan, project page). An agentic workflow that builds large-scale open worlds from prompts, and the distinguishing claim is that output is neither video nor Gaussian splatting but editable, game-ready 3D assets with real geometry and textures. Off this wiki's main beat, but the agentic-workflow-as-generator framing is the part worth noting.
Naval says serious software people train their own models (@ns123abc, quoting @naval, plus @ns123abc and @ns123abc). The quote is a riff on Alan Kay's "people who are serious about software make their own hardware," and @ns123abc's reply is a dunk asking to see Naval's proprietary model. Worth pairing with River AI's launch above, since the two land the same day and make the same bet from opposite ends. His other two posts are Grok 4.6 anticipation and "time to build," both content-free.
Robotics and hardware one-liners. @cleoabram posted a tiny transformer robot built for the moon. @Scobleizer mused about brain-to-brain access in fifty years while riding a self-driving car, followed by two posts about getting engaged in Vegas. No technical content in any of them.
Promo and event posts. @TareqAmin_ previewing HUMAIN's LEAP appearance, @minchoi soliciting replies about X's September creator program, and three contentless posts from @stepango. Skip.
Off-topic bulk. @MarioNawfal (20 posts on Hormuz, Iran, Ukraine procurement, US jobs revisions and domestic politics), @spencerpratt (7 on Los Angeles city politics and the Paramount-Warner merger lawsuit), @WHFraudTF (2 on Medicare and SNAP fraud prosecutions), @HouseGOP (1) and @DoWCTO (1). Together that is 31 of the slot's 68 posts and none of it touches AI research or industry. Skip.