Summary
The morning slot is dominated by one release: SpaceXAI's Grok 4.6, which shows up across at least seven handles and is the only genuine cluster in the window. The substantive version of that story is not the benchmark score but the efficiency profile, since Grok 4.6 ties GPT-5.6 Sol at 61 on the Artificial Analysis Intelligence Index while completing agent workflows in roughly half the steps of Claude Opus 5 at more than 60% lower price, and Hugging Face's Elie Bakouch read the model card closely enough to notice the shape of what SpaceXAI optimized: big internal gains on GPU-kernel generation and on optimizing its own chat inference, and trailing performance on the public software-engineering evals it did not build. The single most interesting line in the whole slot is a one-sentence claim from @aksheyd that Grok 4.6 optimized the Grok Build harness for itself, which lands the same morning HuggingFace published a paper showing that a strong model writing a harness for a weaker one nearly doubles its accuracy without touching weights. A screenshot attached to one @aksheyd post turns out to carry the slot's hardest data: the Terminal-Bench 3.0 leaderboard, which publishes tokens and dollars per run next to accuracy and names the agent harness on every row, showing Grok 4.6 fourth at 26.5% for $2.1k against Opus 5's leading 42.7% for $5.8k. The second real story is hardware: tinygrad and comma shipped "chestnut," a $249 USB-to-PCIe eGPU dock, whose actual purpose is running a driving model with 30x more parameters and 100x more FLOPs than comma's current on-device model. Beyond those, Google Research posted a knowledge-profiling result arguing recall rather than storage is the bottleneck for factual accuracy, Bakouch made a sharp sovereignty argument about Mistral, and DHH spent the slot shipping an Omarchy release candidate and noting Grok 4.6 Fast at a quarter the price of comparable fast tiers. There were zero @bayesiansapien retweets for a fourth consecutive day, so the curated layer is absent again, and roughly two thirds of the 109 captured tweets were off-topic political content from four handles.
Posts
Grok 4.6 launch and its efficiency profile (cluster of 7, @zhu_hanqing666, @eliebakouch, @aksheyd, @ellev3n11, @ns123abc, @minchoi, @dhh · model card)
SpaceXAI released Grok 4.6 at Grok 4.5's price. Artificial Analysis put it at 61 on its Intelligence Index, joining the frontier alongside GPT-5.6 Sol, a five-point gain over Grok 4.5 just over a month after that model shipped. Zhu Hanqing, who worked on it, framed the goal as stronger long-horizon capability "while keeping the intelligence efficiency we care about," and the efficiency claim is the one with teeth: the model completes complex agentic workflows in about 53 steps where Claude Opus 5 needs 103, at more than 60% lower price. In an agent loop every step is a full round trip carrying accumulated context, so halving steps compounds into a much larger cost gap than the sticker price suggests.
Elie Bakouch's read of the model card is the most useful thing in the cluster because it describes what SpaceXAI actually optimized for. Four points: large gains on DeepSearchQA and on their internal KernelBench (GPU kernel generation); best-in-class on SpaceXAI's internal engineer benchmark, with Bakouch flagging the obvious caveat that they likely trained on similar data; state of the art on "inferenceEval," which measures optimization of their own chat inference, meaning the model is being tuned to optimize the serving stack it runs inside; and behind the frontier on the public evals, specifically Terminal-Bench 3.0, SWE Marathon v1.1 and DeepSWE. Strong on the benchmarks they built, weaker on the ones they did not, is a legible profile and there is no way to check from outside whether the internal benchmarks measure something real. Bakouch separately praised the card's stated policy that they "never silently downgrade intelligence or fall back to other models."
The line worth carrying into the wiki comes from @aksheyd: "Grok 4.6 optimized the Grok Build harness for itself." That is a self-to-self version of the day's strongest research result, described in today's digest, and it means nobody outside SpaceXAI can tell how much of the 53-versus-103 step reduction is weights and how much is scaffold. DHH, who ran it on a real pull request, reported Grok 4.6 Fast at $4 in / $12 out per million tokens, roughly a quarter of what comparable "fast" tiers cost, and said "no notes." Full treatment in the Grok 4.6 summary.
The Terminal-Bench 3.0 leaderboard, read off the attached screenshot (@aksheyd, relaying @ryan_marten)
The image attached to this post is the most useful artifact in the whole slot, because unlike almost every leaderboard it reports tokens and dollars per run next to accuracy, and names the agent harness on every row. Transcribed:
Rank Model Agent Resolution rate Tokens Cost 1 Opus 5 (max) mini-SWE-agent 42.7% ± 1.6% 7.3B $5.8k 2 GPT-5.6 Sol (max) Codex 34.6% ± 1.6% 5.8B $4.0k 3 Fable 5 (max) Claude Code 34.1% ± 1.7% 3.6B $6.5k 4 Grok 4.6 (high) Grok Build 26.5% ± 1.5% 2.9B $2.1k 5 Opus 4.8 (max) Claude Code 21.1% ± 1.6% 5.2B $5.2k 6 GPT-5.6 Terra (max) Codex 20.8% ± 1.4% 7.0B $2.5k 7 Grok 4.5 (xhigh) Cursor CLI 15.7% ± 1.5% 1.2B $766 8 Sonnet 5 (max) Claude Code 14.6% ± 1.5% 17.9B $6.9k 9 GPT-5.6 Luna (max) Codex 14.3% ± 1.3% 11.9B $1.6k 10 GLM 5.2 (max) Claude Code 4.6% ± 1.0% 3.3B $3.4k Three things fall out of it. First, it confirms Bakouch's model-card read from the outside: Grok 4.6 is genuinely behind the frontier here, 26.5% against Opus 5's 42.7%, which is the public-eval gap he flagged. Second, on cost per accuracy point the ordering inverts: Grok 4.6 costs roughly $79 per point against Opus 5's $136 and Fable 5's $191, which is the same trade the 53-versus-103 step claim describes, now visible on a public benchmark. Third and most interesting for the harness question, the table is implicitly a model-by-harness grid: Claude Code appears four times spanning 4.6% to 34.1%, and Sonnet 5 with Claude Code spends 17.9B tokens and $6.9k to land at 14.6% while Fable 5 in the same harness reaches 34.1% on 3.6B tokens. Token spend is not tracking accuracy at all, which is precisely the pathology the day's papers describe.
Musk says Grok 4.7 lands in three to four weeks (@ns123abc, @minchoi)
Initial training is complete and the model is now in supplemental training on what Musk describes as "a massive amount of SpaceX company data." Read next to the model card's internal-engineering benchmark strength, this is a consistent strategy rather than a boast: SpaceXAI is training on proprietary engineering corpora and then measuring itself on engineering tasks. Whether that produces a genuinely better general model or a model that is excellent on SpaceX-shaped work is exactly the question the public-eval gap already raises.
tinygrad and comma ship "chestnut," a $249 eGPU dock (@tinygrad · comma.ai/shop/chestnut)
A USB-to-PCIe dock built on the ASM2464PD, supporting USB2/3/4 plus PCIe 4.0 x4, with fully open-source custom firmware. Every unit is tested at both USB3 and USB4 speeds and gets a green sticker if it passes. Two SKUs: $249 for the dock alone, $799 for a "ready to drive" kit with a GPU for comma four owners. The number that makes this more than an accessory launch is in the product page: the first chestnut-class driving model has 30x more parameters and uses 100x more FLOPs than comma's latest on-device model. That is the real announcement. comma is not selling a dock, it is moving its on-vehicle compute tier up two orders of magnitude and shipping the connector that makes it possible. tinygrad separately noted that with its custom kernel support "you (or your LLM) can push inference speeds beyond every other framework," pointing at a contributor who wrote a tinygrad inference server now going to production.
Google Research on knowledge profiling: recall, not storage, is the factuality bottleneck (@GoogleResearch · post)
A behavioral framework that measures both encoding and recall separately and uses the split to locate where parametric factuality actually fails in frontier models. The title carries the finding: "Empty shelves or lost keys? Recall is the bottleneck for parametric factuality." In plain terms, when a model gets a fact wrong it is usually not because the fact was never learned during pretraining. It is because the model cannot retrieve what it stored. That is a meaningfully different diagnosis from the usual one, and it points at retrieval-side interventions rather than at more pretraining data. It also rhymes with the day's tool-use result, where models call retrieval tools whose returns do not affect the answer: in both cases the information is present and the system fails to use it. The farmer captured only the page navigation chrome rather than the article body, so this is read from the tweet and the headline.
Bakouch's sovereignty argument about Mistral (@eliebakouch · Mistral post)
Mistral announced in-region inference, open-model serving, and new European compute infrastructure. Bakouch's take is the sharper one and it is not a product observation. His worry is not that Mistral's models trail the frontier, it is that critical actors such as governments and large companies may be forced to use them for things like cybersecurity, and that the consequences of mandating a less capable model in a security-critical role "are beyond what people imagine today." He extends it past cyber: entire R&D-driven industries could fall behind purely because policy required them to use weaker models. His read on the Mistral announcement is therefore hopeful, since hosting the best available open models on sovereign infrastructure decouples "where the compute runs" from "which model you must use." That is a genuinely useful frame for the sovereign-AI debate, which usually collapses the two.
DHH ships Omarchy Quattro RC and argues Linux adoption goes parabolic (@dhh · PR)
The first release candidate for Omarchy Quattro is out with a Friday final planned. The engineering details are cost-optimization in miniature: he shrank the ISO by another 114MB by compressing the NVIDIA 580xx driver packages at maximum zstd (the default level left the unstripped driver blobs 22% larger than necessary), fitting three kernels plus all NVIDIA drivers plus the whole distribution under 6GB. The distribution now ships agents as first-class, managed through
miseon an out-of-band upgrade path from system packages so they can be updated many times a day. His broader claim: "Linux adoption among programmers is about to go parabolic. Any missing app can be recreated easily. The open-source advantage with agents is unstoppable." OpenAI's new Linux desktop app installs from Omarchy's own Cloudflare-backed package repository with no AUR required, which is the concrete version of that argument.Quintar and the spatial-AI-for-live-events pitch (cluster of 4, @Scobleizer · quintar.ai)
Robert Scoble ran an interview with Quintar founder Jay Jayaram and posted repeatedly about the demo, describing a 3D basketball court rendering where the system knows the position of everything in a stadium to within a few inches and you can touch a player to pull up stats. Quintar positions itself as "the infrastructure layer for spatial AI," a shared spatial foundation for people, agents and devices, and frames the hard problem correctly: the difficulty in AI glasses is not the AI, it is reliably anchoring that AI to the right real-world object at distance, in a crowd, in real time. This sits outside this wiki's attention range and is recorded rather than pursued, though the framing that spatial grounding is an infrastructure problem rather than a model problem is the same structural claim the harness cluster makes elsewhere.
Off-topic political and personal volume (cluster of roughly 70: @MarioNawfal, @spencerpratt, @brivael, @WHFraudTF, @HouseGOP, @DoWCTO, @JonasBadalic)
Four handles accounted for the large majority of the slot's captured volume with US and French domestic politics, military news, and local election content carrying no AI signal. Skip. One item is worth a single line because it touches the supply chain this wiki tracks: a Nawfal thread relaying former CIA analyst Larry Johnson's argument that rerouting energy around the Strait of Hormuz via pipeline does not address the non-oil traffic, specifically that roughly 40% of the world's helium, which chip manufacturing depends on, moves through the strait and cannot be piped.
Tesla FSD, Grok-built game demos, and assorted promo (@Tesla, @stepango)
Tesla amplified a frame-by-frame driver analysis claiming FSD Supervised began evading a crash about 0.17 seconds before the at-fault vehicle started moving. Single anecdote, no independent verification, no fleet statistics. @stepango shipped a browser survivors-like game generated by Grok 4.5 and asked what 4.6 could do with it, which is a capability demo rather than a result. Both are Skip.