Edition 026 — China Moves to Control Model Weights
The Frontier AI Wire is researched and drafted by Claude, an AI model made by Anthropic, under rules set by Attorney Jeffrey M. Beck. Every factual claim links to its source, with primary sources first. Where the brief goes beyond what a source says, it labels that as inference. Attorney Beck reviews and approves each edition before it is published. Errors are corrected in place, marked where they occurred, and logged. Nothing is changed silently. How the Wire is made.
China proposed putting 23 artificial-intelligence technologies — large-model weights and architecture, agent frameworks, AI chip design and cluster interconnect — on its catalogue of technologies prohibited or restricted from export, and made pre-export registration binding the same day, while saying plainly that serving those models abroad through an API is not caught. Meta released Muse Spark Open, an 816-billion-parameter open-weights model that runs on two Mac Studios, and printed on its own model card the independent index score that puts it behind six other open-weights models, none of them American. Artificial Analysis put NVIDIA's Vera Rubin NVL72 on its own agentic harness and measured 1.14–1.45× the throughput per megawatt of its predecessor at peak, against the “up to 30×” NVIDIA published in August at one matched speed. Anthropic gave METR six employee-badged seats and says it cannot veto what METR publishes. And the FTC is reported to have sent civil investigative demands to OpenAI and Anthropic. One correction, to edition 025.
Dispatches
Ranked by how much each item should change your picture of the field — not by volume of coverage.
China proposes export controls on model weights, agent technology and AI chip design — and says serving those models abroad by API is not a technology export
MOFCOM and the Ministry of Science and Technology published Announcement No. 48 of 2026 on 16 October, opening public consultation on adding 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export and imposing temporary pre-export registration in the meantime. The spokesperson's question-and-answer page on mofcom.gov.cn is the primary source read at compile time, in translation; the annex listing the items themselves was not read in the original.
Two things are happening at once, and conflating them is the easy mistake. The catalogue additions are proposed and open to comment — Reuters, as carried at compile time, gives the deadline as 15 November. The registration requirement is already in force, from the announcement date until the revised catalogue takes effect: exporters must register on MOFCOM's technology-trade platform before transferring ownership of a listed technology abroad, licensing a domestic technology for overseas use, or changing an owner or controlling party to a foreign entity, and exports already under way but not finished are covered. Exporters assess for themselves whether they are in scope.
| Field | Items | Examples given |
|---|---|---|
| Artificial intelligence | 23 | large models, agents, computing clusters, information processing |
| Photovoltaics | 12 | polysilicon, wafer and cell manufacturing |
| Biomedicine | 5 | pathogenic microorganisms, cell and gene therapy, small nucleic acid drugs |
| Total | 40 | 14 proposed for prohibition, 26 for restriction |
The spokesperson's framing is the standard one: the revision is “not targeted at any specific country, region or entity,” the control scope is “clear, precise and limited,” and the stated aim for the AI items is keeping the technology “under human control.” Read this as a mirror of the controls China has spent four years on the receiving end of, with one difference worth holding onto: the thing being controlled here is not fabrication equipment but model artefacts and know-how.
What the AI items cover, as described. Core large-model technology — weights, architecture, and training and inference optimisation; agent technology — task planning, multi-agent collaboration, tool calling and game decision-making; AI chip design for processors and memory; and computing-cluster technology — high-speed interconnect, fault tolerance, communication and heterogeneous scheduling. Two cyber items are singled out in the spokesperson's own examples: AI-driven generation and mutation of malicious code, proposed for prohibition, and AI-assisted vulnerability discovery, proposed for restriction. One secondary analysis read at compile time reports that the annex sets performance thresholds for some AI entries and that its own copy of the annex text is partly garbled; no threshold figures are carried here.
The carve-outs, which decide how much of this bites. The spokesperson states that API calls, applications and SaaS services built on a restricted model are not treated as exports of model technology — so a Chinese lab serving its model to foreign users is outside the measure as described, and only the transfer of the artefacts and methods is in it. The same account describes expedited “report and export immediately” handling for certain open-source exports, and one secondary analysis reports that technologies already public through open-source release or academic publication are excluded. Those two carve-outs are what separate this from an export ban on Chinese open-weights models, and this brief could not read the annex language that implements them.
Why a reader should care beyond the headline. If weights and architecture are controllable technology in one jurisdiction, the question of whether they are controllable technology anywhere stops being hypothetical. The US debate over controlling model weights has run for two years without a rule of this shape. What China has actually imposed today is a registration requirement; what it has proposed is the rule.
Separately, and on reporting rather than a document: Bloomberg, as carried in republications read at compile time, interviewed Taiwan economy minister Kuo Jyh-huei on 12 October and reports that Taiwan plans to add AI servers and advanced chips to its own export-control list, aligning with US restrictions on sales to China, and to make unlicensed shipments criminal under the Foreign Trade Act — which, per that account, Taiwanese law does not currently provide for, so prosecutors have reached for other statutes in smuggling cases. The minister gave no implementation timeline, no gazette notice or formal announcement is cited, and the Bloomberg article itself was not read at compile time.
Sources MOFCOM spokesperson Q&A on the catalogue revision, 16 Oct (primary, read in translation) · Reuters via The Standard, 16 Oct (comment deadline) · Geopolitechs analysis, 16 Oct (instrument numbers, annex reading) · ChainCatcher, 16 Oct (summarising Bloomberg on the 23 AI items) · Bloomingbit, 16 Oct (republishing Bloomberg on Taiwan) · Crypto Briefing, 16 Oct (Taiwan)
Meta's Muse Spark Open is the strongest US open-weights model on the one index that scored it independently — and the fourth-strongest open-weights model on that index
Released 15 October by Meta Superintelligence Labs: an open-weights, natively multimodal mixture-of-experts model of 816B total parameters with 42B active per token and a 1M-token context, built for coding and agentic work. Meta calls it “the most capable US open-weight model to date.” Artificial Analysis scored it the same day at 39 on its Intelligence Index at xhigh effort — which is AA's basis for calling it the top US open-weights model, and which puts six other open-weights models above it, none of them American on AA's own account.
| Model | Index |
|---|---|
| MiMo‑V2.6‑Pro | 46 |
| GLM‑5.3 | 45 |
| Kimi K3 | 44 |
| GLM‑5.3‑Flash | 42 |
| Qwen3.8 2.4T | 40 |
| Qwen3.8‑Flash‑Next | 40 |
| Muse Spark Open (xhigh) | 39 |
For scale inside Meta's own range, AA puts Muse Spark Open at xhigh 22 points above Muse Glimmer, Meta's 30B dense open-weights model, at high effort (17).
The rest of the numbers are Meta's own, and the model card prints them against the models that beat it. On Meta's table, Muse Spark Open leads MCP‑Atlas at 89.2 and JobBench at 61.8, and trails on Terminal‑Bench 2.1 (76.4 against Kimi K3's 85.0), DeepSWE 1.1 (60 against 68) and GDPVal‑AA v2 (1,401 against up to 1,653). A vendor publishing a table it loses on is worth noticing, and so is the fact that the one line in it produced by someone other than Meta is the AA index score.
Architecture, from the model card. 62 layers at 8,192 hidden dimension; 256 experts with 8 activated per token; grouped-query attention with 128 query and 16 key-value heads at head dimension 64, with gated attention and QK-norm; a 1.8B-parameter, 50-layer Perception Encoder; a 202,048-token vocabulary with untied input and output embeddings. Input is text, image, video and audio; output is text. Default reasoning effort is xhigh. SiliconANGLE reports it was built from Muse Spark 1.2, which shipped in August, with added safety post-training so it works without Meta's service-side guardrails; the card states no token count or training compute.
What it takes to run. The card gives at least 700GB of accelerator memory for the NVFP4 weights and about 2TB for BF16. The desk-side configuration everyone wrote about is two 512GB M5 Ultra Mac Studios paired over MLX, about a terabyte of unified memory. MacRumors notes the 512GB Mac Studio has not shipped and its price is unannounced; the 256GB machine lists at $10,799, which is a reference point and not a price for the pair. Hosted access is listed through Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI, with context limits that differ by host — Groq at 131,072 tokens, Fireworks at 1,048,576 — and weights through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry.
Cost, as measured rather than quoted. Artificial Analysis puts the cost of running its Intelligence Index at $0.84 per task for this model at current API pricing of $1.25 per 1M input and $4.25 per 1M output ($0.15 cached input) from Together AI, Fireworks and Baseten — less than half of GLM‑5.3 at max ($2.01) and roughly three times DeepSeek V4.1 Flash at max ($0.27). Meta's own announcement omits hosted token rates.
The licence, which is the part with teeth. The Muse Model License permits running, modifying, fine-tuning, distilling and redistributing, defines derivatives to include models trained on the model's outputs, and requires products to display “Built with Muse Spark Open.” A licensee whose group revenue passes $20 million in any consecutive 12 months must notify Meta and sign a separate agreement before offering the model as a hosted service; internal deployment, research and non-model-serving commercial use are exempt. Downloading from Hugging Face requires full legal name, date of birth and organisation through a gated form.
Two caveats on the comparison. VentureBeat notes the DeepSWE results use each model's preferred harness, so the agent software is part of what is being compared; SiliconANGLE reports Meta says the model cleared its Meta Superintelligence Scaling Framework release thresholds with the full evaluation report still to come. No such report was located at compile time.
Sources Model card, meta-models/Muse-Spark-Open-816B-A42B (primary) · @ArtificialAnlys — index score, cost and ranking, read 16 Oct from the Frontier Wire Sources list · VentureBeat, 15 Oct · SiliconANGLE, 15 Oct · MacRumors, 15 Oct (hardware) · TechCrunch, 15 Oct
Someone other than NVIDIA has now measured Vera Rubin NVL72: 1.14–1.45× the throughput per megawatt of GB300 at peak, and up to 5.7× at one matched speed
Artificial Analysis published the first AA‑AgentPerf results for Vera Rubin NVL72 on 15 October. The benchmark replays real coding-agent trajectories at several concurrency levels — contexts up to 132K input tokens, up to 182 requests per trajectory, an average of 97% of input tokens repeated from the previous request — and plots total throughput per megawatt of measured GPU power against p25 per-user output speed.
| Workload | Rubin peak | GB300 peak | At peak | Largest lead, matched speed |
|---|---|---|---|---|
| DeepSeek V4 Pro | 272M tok/s | 239M tok/s | 1.14× | ~5.7× at ~190 tok/s |
| DeepSeek V4.1 Flash | 571M tok/s | 395M tok/s | 1.45× | ~5.4× at ~580 tok/s |
| GLM‑5.2 | 272M tok/s | 201M tok/s | 1.35× | ~3.0× at ~165 tok/s |
| Kimi K3 | 70M tok/s | 56M tok/s | 1.26× | ~2.8× at ~67 tok/s |
NVIDIA's own figure for the same generational comparison, published 24 August on its developer blog, is “up to 30×” higher AI-factory throughput per megawatt at 160 tokens per second per user. Both numbers can be right, because they are not the same measurement — and the brief's reading is that the gap between them is mostly the denominator and the speed point, not a disagreement about the hardware.
Where the two measurements differ. AA's megawatt counts each GPU and its HBM only, excluding CPUs, networking and cooling. NVIDIA's metric is throughput per provisioned megawatt, on SemiAnalysis's AgentX benchmark — which replays prerecorded Claude Code sessions turn by turn — and its Vera Rubin results are NVIDIA-measured and, in NVIDIA's own words, “pending SemiAnalysis review.” NVIDIA's own chart gives 2× at 110 tokens/s per user and 10× at 130 before reaching 30× at 160, so the multiple is a function of where on the curve you stand.
The ceiling that makes matched-speed multiples large. AA reports GB300 rack configurations top out near 163 tokens/s per user on DeepSeek V4 Pro. Any comparison above that speed is a comparison against a system running out of headroom, which is how a 1.14× peak advantage becomes a ~5.7× advantage at ~190 tokens/s. That is a real operational fact, not a trick, and it is also why a single multiple with no speed attached means very little.
Per-GPU, and what Rubin is. Per GPU, AA puts Vera Rubin 2.0–2.5× ahead at peak, with larger matched-speed leads (~12× on DeepSeek V4 Pro at ~165 tok/s, ~18× on V4.1 Flash at ~580 tok/s). The hardware: 288GB HBM4 at 19.2 TB/s, 3.0 TB/s NVLink, 35 PFLOPS dense NVFP4, 1.8 TB/s NVLink‑C2C to a Vera CPU, 0.5–1.9 kW measured — against Blackwell Ultra's 288GB HBM3E at 8 TB/s, 1.8 TB/s NVLink, 15 PFLOPS and 0.3–1.0 kW.
Caveats AA states itself. Rubin results use early Rubin-specific kernels and tuning, with further gains expected. Time to first token is worse on Rubin (median 1.65 s against 1.08 s on DeepSeek V4 Pro), and on end-to-end speed GB300 matches or beats it per megawatt at most speeds on that model. The V4.1 Flash Rubin runs capped replies at 2,048 tokens against 16,384 elsewhere, with about 2% hitting the cap — a cap that modestly favours Rubin. NVIDIA supplied the Vera Rubin configurations and most GB300 rack configurations; AA built some small four-GPU GB300 configurations itself to fill gaps. Serving used KV-cache offload to CPU memory over NVLink‑C2C with KV-aware routing through Dynamo, disaggregated prefill and decode, wide expert parallelism and speculative decoding.
Sources Artificial Analysis, First AA-AgentPerf results for NVIDIA Vera Rubin NVL72, 15 Oct (primary) · NVIDIA developer blog, 24 Aug (vendor figures) · @ArtificialAnlys, read 16 Oct from the Frontier Wire Sources list
Anthropic gives METR six employee-badged seats, and says it cannot veto what METR publishes
An update on independent evaluators, published 15 October at 17:20 UTC. Six METR employees have been onboarded with company hardware, office space and internal tools and systems described as roughly comparable to those of Anthropic's own risk-assessment teams. Anthropic employees are encouraged to share information with them, with three carve-outs: customer data, attorney-client privileged material, and model architecture.
The publication terms are the substance. Anthropic states it cannot veto what METR publishes about its risk levels, incidents or practices; it can require redaction only within predefined categories; and METR may say publicly whether it thinks a redaction was justified without revealing the content. That is a narrower power than a sign-off and a wider one than nothing, and it is the first arrangement of this shape this brief has seen described in a lab's own words.
The tiers. Tier 1: narrowly scoped risk investigations, access at Anthropic's discretion, publication requires Anthropic's approval. Tier 2: publication without approval and more information, with Anthropic able to require removal of specified information categories. Tier 3: the METR arrangement — employee-like access with tier-2-style publication rights. Anthropic says it hopes to move more evaluators to tier 3 as trust builds and commits to no timeline for doing so.
Who else, and how far along. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, with terms clarified in a 15 October update to the post. Accenture and Faculty are described as planning an “ambitious partnership” spanning model evaluation, red-teaming, alignment assessment, safeguard testing and capability building. Hugging Face and several unnamed organisations are in early discussions, which the post says means more limited information and more limited publication rights.
The limits, stated by Anthropic. The METR engagement is a small team under enforceable agreements; other METR staff, including those working with Anthropic on other projects, are walled off from it. Architecture information is characterised as high leak risk and low relevance, and sharing it needs explicit internal approval. The post says the arrangement is “not a way for other organizations to obtain our information, or that of our customers and partners.” Anthropic frames the whole thing as voluntary and says “every frontier AI company should do the same.”
The missing half. This is the company's account of an arrangement the company designed. METR's updates index, read at compile time, carried nothing dated later than 6 October; a METR post dated 14 October circulating on this brief's source list, said to describe METR's approach to embedded assessments, was not retrieved at compile time. Anthropic frames the step as following a commitment made last month and its signature of the White House Accord on Super Intelligence.
Sources Anthropic, An update on independent evaluators, 15 Oct (primary) · TechCrunch, 15 Oct · CNBC, 15 Oct · METR updates index (read at compile time)
The FTC is reported to have sent civil investigative demands to OpenAI and Anthropic
Reported 15 October, originating with Axios as summarised in the accounts read at compile time: the Federal Trade Commission has issued civil investigative demands to OpenAI and Anthropic seeking documents on AI safety incidents, company safeguards, marketing practices and privacy policies, and plans to seek testimony from senior executives. An FTC spokesperson confirmed an investigation into potential consumer-protection violations to CBS News.
No FTC release, demand or docket entry was located at compile time, and no response from either company was found in any account read at compile time. A civil investigative demand is a compulsory request for information, not a charge; nothing here is a finding of fact and no allegation has been tested. One account read at compile time notes a separate European Commission inquiry to both companies about AI agent security incidents — a different matter, and not evidence about this one.
Sources TechRepublic, 15 Oct (summarising Axios and CBS) · Techstrong.ai, 15 Oct · AI Weekly news index, 15–16 Oct
OpenAI opens its life-sciences model to US university labs — 30 seats per lab, free for a year
The Rosalind Academic Plan, listed on openai.com/news dated 15 October: up to 30 Business Standard seats per eligible lab for 12 months, for principal investigators and lab leaders at US universities, subject to organisational approval. The seats are reported to unlock OpenAI's other advanced models alongside GPT‑Rosalind, and research data on these accounts is said not to be used for training by default. The OpenAI post itself could not be retrieved by this brief at compile time; everything above comes from reporting, and the terms should be read from the source before anyone relies on them.
GPT‑Rosalind, per the same account, launched as a research preview on 16 April 2026, restricted to vetted enterprise and research organisations, and is built to reason across biological entities, synthesise literature, plan experiments and connect to scientific databases. The one pricing anchor in anything read at compile time is a $5 per million input tokens listing for a gpt‑rosalind‑discovery model as of 10 October. Benchmark gains quoted in that account — 53.7% on Genebench, 18.0% on Medchem Bench, 19.6% on Labworkbench — are attributed to OpenAI, name no baseline this brief could identify, and are carried here as unverified; see Checked and spiked.
Sources Crypto Briefing, 15 Oct · OpenAI newsroom listing (read at compile time) · Unite.AI, 15 Oct (listing; not retrieved at compile time)
Mistral ships four time-series models, the largest 209M parameters, trained mostly on synthetic data
Timestral, published 15 October: four probabilistic multivariate forecasting models at 5M, 28M, 123M and 209M parameters, sharing one architecture and training recipe, under a modified MIT licence on GitHub and Hugging Face and deployable on-premises, in a VPC or air-gapped. Context is 32,768 time steps; the models emit 21 quantiles, trained with a pinball loss approximating CRPS.
It is a small, unglamorous release and it is in this brief for one reason: the claims are made on public third-party leaderboards, which makes them checkable in a way most vendor numbers are not. Mistral says a single checkpoint ranks first among pretrained zero-shot models on GIFT‑Eval, TIME and fev‑bench across five metrics without ensembling, that the 209M model beats the previous state of the art with about two-thirds of its parameters, and that the 28M model beats Chronos‑2 — roughly four times its size — on GIFT‑Eval and TIME. Results are stated to have been submitted to those leaderboards. No independent confirmation was located at compile time, and the technical report is forthcoming.
Training. Two stages: pretraining on purely synthetic data at 8,192 context — univariate processes with trends, seasonality and varied noise, combined into multivariate series with nonlinear dependencies — then post-training on a mix of real and synthetic data at four times the context. Named techniques are decoupled horizon forecasting and covariate-aware attention that assigns each series a role: target, past-only covariate, or known-future covariate.
Latency, as stated. On eight CPU threads, forecasting one day ahead from three weeks of 15-minute data: about 10 ms for the 5M model and 140 ms for the 209M. The 28M is about 5× faster than the 209M, and the 123M lands within 0.2% of the 209M's score at about three-quarters of its latency. The page carries no limitations section; Mistral notes only that public benchmarks measure general performance and production data can differ.
Sources Mistral, Introducing Timestral, 15 Oct (primary) · Mistral news index (publication date)
Also on the wire
Confirmed, and not enough on their own to move the picture.
-
NVIDIA opens a 120B multimodal model and a 74B local-agent model, with the datasets and recipes
Released 15 October: Nemotron 3.5 Super VL, 120B total and 12B active, multimodal with up to 1M context, and Nemotron 3.5 Bolt 74B‑A9B, distilled from it and quantised to 2 bits to fit 24GB; a 4GB Bolt E4B is dated 23 October. Weights are under OpenMDW‑1.1, with open datasets, training recipes and RL environments alongside. Artificial Analysis scored them 26 and 21 on its Intelligence Index on 15 October. NVIDIA's own table has Super VL at 89.90 GPQA‑Diamond and 71.07 SWE‑bench Verified, with Thinking Machines' Inkling‑Small ahead of it on several of the same self-reported lines.
Sources MarkTechPost, 15 Oct · AA changelog, 15 Oct
-
Artificial Analysis put five frontier models in front of a robot arm, and a human grader scored the result
Published 14 October and missed by this brief yesterday — see Corrections. AA‑Robotics v0.1 is a research preview: models drive a Hiwonder NexArm through a simple HTTP API in an isolated container, zero-shot, across eight tabletop tasks, with four cameras and a human grader scoring the final state. GPT‑6 Astra 69.4, GPT‑6.1 Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8, Claude Fable 5.1 46.3, each in its own vendor's coding-agent harness at high effort. AA's own caveats: models run nearly two orders of magnitude slower than a ~20-second human baseline, about 70% of spend is input context, and they often keep refining past a passing state and fail as a result.
Sources Artificial Analysis, AA-Robotics v0.1, 14 Oct (primary)
-
The lab desks, read at compile time
Anthropic's newsroom carried item 04, dated 15 October, and nothing dated 16 October. OpenAI's newsroom listed two 15 October items — item 06 and the Brazilian Portuguese educator course carried in edition 025 — and nothing dated 16 October; the Rosalind post itself would not render to this brief. DeepMind's blog index listed six October items, none with a day attached, including one on antimicrobial peptides with Google Antigravity that this brief could not retrieve and found no reporting on. blog.google's AI section again returned no dated item list. Meta's AI blog index showed nothing newer than July, which is its own comment on where Meta now publishes. Mistral published item 07. Artificial Analysis published item 03 and the Nemotron evaluations above. METR's index showed nothing newer than 6 October, Epoch AI nothing newer than the 14 October Slay the Spire report carried in edition 025, ARC Prize's blog nothing newer than 3 September, SpaceXAI nothing newer than Grok Voice 3.0 on 14 October. Qwen's old blog index still says the blog has moved, and DeepSeek's news page again returned the API quick-start rather than a dated item list.
Source Anthropic newsroom · OpenAI · DeepMind blog index · blog.google AI · Meta AI blog · Mistral · x.ai · METR · ARC Prize · Epoch AI · Artificial Analysis · Qwen · DeepSeek (all read at compile time)
Checked and spiked
Circulating today; did not survive checking.
“Vera Rubin is 30× more efficient than Blackwell.” Circulating as a property of the hardware. It is NVIDIA's own measured figure, dated 24 August, on SemiAnalysis's AgentX benchmark, at one speed point — 160 tokens per second per user — with throughput per provisioned megawatt as the denominator and, in NVIDIA's words, “pending SemiAnalysis review.” The same chart gives 2× at 110 tokens/s and 10× at 130. Carried in item 03 with that provenance; not carried as a general efficiency multiple.
“China is banning exports of its AI models.” What took effect on 16 October is a registration requirement. The catalogue listing is a proposal open for comment, the spokesperson states that API, application and SaaS access to a restricted model is not an export of model technology, and a secondary analysis reports carve-outs for technologies already public through open-source release or academic publication. The proposal is in item 01; the characterisation is not.
“Meta and Apple have partnered on Muse Spark Open.” MacRumors describes the two companies as partnering and reports that Apple hardware chief John Ternus replied to a post by Mark Zuckerberg on X; TechCrunch's page description says the companies teamed up while its article body describes only a model designed to run on Apple hardware. No announcement from Apple was located at compile time — Apple's newsroom index, read at compile time, returned navigation and no article list to this brief, so its contents are unchecked. What is documented is that the model card lists MLX optimisation for two M5 Ultra Mac Studios. The brief carries that and not the partnership framing.
GPT‑Rosalind's reported benchmark gains. The figures of 53.7%, 18.0% and 19.6% on Genebench, Medchem Bench and Labworkbench appear in one account read at compile time, attributed to OpenAI, described as gains over “existing baselines” that are not named. They are noted in item 06 and are not used for any comparison here.
Corrections
Errors in this brief — fixed in place above, logged here.
Edition 025 reported that Artificial Analysis “showed nothing newer than 13 October.” That was wrong. Artificial Analysis published AA‑Robotics v0.1, a research preview of frontier models controlling a six-axis robot arm, on 14 October, and the changelog read today carries that entry under 14 October. The sweep missed it.
The failure is also one of form, and that is the part worth fixing. The standing rule this brief adopted after edition 015 is that an absence is reported as “not found at compile time, at this URL” or it is not reported. Edition 025 stated the absence flatly inside a paragraph whose closing source line said the pages were read at compile time, which is a hedge at the wrong altitude: it covers the reading, not the claim. The lab-desks paragraph in future editions states what was seen rather than what was not there.
What it takes down: nothing else. No other item in edition 025 rested on that line, and the missed item is carried in today's Also on the wire. The archived text of edition 025 is marked in place.
Sources Artificial Analysis, AA-Robotics v0.1, 14 Oct · AA changelog (read at compile time)
Two standing notes, carried forward. First: a statement that was true when written and overtaken by a publication hours later is not an error, and this brief will keep saying so rather than quietly retrofitting — the test is whether that is applied honestly in both directions, and today's correction is the other direction. Second: the rule adopted after edition 015 — that an absence is reported as “not found at compile time, at this URL” or it is not reported — is applied throughout this edition, to the unread annex and unverified thresholds in item 01, the absent Meta evaluation report in item 02, the vendor-measured and unreviewed NVIDIA figures in item 03, the missing METR account in item 04, the unobtained FTC documents in item 05, the unretrievable OpenAI post in item 06, the absent independent run in item 07, and the lab desks above. Editions 002, 003, 006, 008, 014, 015 and now 025 carry corrections, archived with their editions. On sourcing: the private source list was reachable at compile time and read by screenshot; item 03 reached this brief through it before any search sweep surfaced it, as did the measured index score and cost figures in item 02.