2026-09-21

1. Alibaba's image line leaves Apache 2.0, and it is the first time an incumbent open-weights lab has narrowed its terms here (1.60)

• Qwen-Image-2.1, 2026-09-20, captured day +1. README and LICENSE both read first-party on github.com; qwen.ai and huggingface.co answer EGRESS_BLOCKED (source)
• 7B visual generation component, 32 Single-Stream DiT layers with block-causal attention, a Qwen3-VL 8B text encoder, a 64-channel RGBA autoencoder with 16× spatial compression giving native transparency, 2K output, editing with up to 10 reference images
• The licence is the story. Qwen RESEARCH LICENSE AGREEMENT, dated September 20, 2026, Hangzhou Tongyi Laboratory Technology Co., Ltd. The grant is "FOR NON-COMMERCIAL PURPOSES ONLY"; "Non-Commercial" is defined in the document as "for research or evaluation purposes only"; commercial use routes to a negotiated licence. Two passes state the earlier Qwen-Image line shipped under Apache 2.0, with the 2512 and Edit-2511 models unaffected
• Every prior licence entry on Open-Weights Policy Fight runs the other way. Qwen3.8 opened in August, Qwen-Drive shipped Apache 2.0 across code, weights and demo data on 09-07, Thinking Machines shipped Apache 2.0 at frontier scale. The two restrictions already recorded — MiniMax's territory exclusion, GLM-5.3's staged hold-back — were novel terms on new lines. This is a permissive line becoming less so
• The cost is concrete and it is already on this wiki. Four days ago that page recorded PrismML shipping Ternary Bonsai 2 27B from Qwen 3.8 27B under Apache 2.0 because Alibaba's licence allowed it. The same act on this model would require a negotiated licence
• Why it matters: a downloadable checkpoint that cannot be deployed commercially is open by every measure the trackers on that page use — Mozilla's download and derivative counts included — and closed to every commercial user. The metric and the thing it measures have come apart, in the direction the page's central open problem predicted
• The capability claim has nothing under it: beats most closed models on Qwen-Image-Bench, which is Qwen's own, with no numeric figure in any pass and two passes saying independent benchmarks are pending
• Not established: no stated reason for the change, and whether the licence reaches generated outputs is not addressed in the clauses read. Nothing says whether this is specific to the image line — the language line's next release is the test
• → Qwen-Image-2.1 (new) · Alibaba / Qwen AI Lab · Open-Weights Policy Fight

2. Yesterday this run scored this essay at 0.2 and skipped it; the weekly synthesis quoted it the same morning (1.56)

• Where I Stand on RSI, Nathan Lambert, Interconnects, 2026-09-19, captured day +2. www.interconnects.ai is EGRESS_BLOCKED — no first-party read, every claim below is a paraphrase with a pass count (source)
• The scoring disagreement is why this leads rather than sits in Watch. Yesterday's brief put it under Below threshold at 0.2 on the hype / opinion pieces row. The same run's W38 synthesis then used it — "Declining: RSI as a headline (five papers in four days in W37, one dissenting essay this week)". An item scored out of the daily was load-bearing in the weekly, and nothing flagged that, because the synthesis reads log.md and briefs rather than the score sheet
• It is scored today on RL / reasoning / alignment (1.3). interests.md lists "consensus-overturning results (e.g., claims of scaling limits)" under Actively Tracked Signals, and this is a scaling-limits claim aimed at the premise of a page this wiki maintains. The 0.2 row was the wrong row
• The argument: "lossy self-improvement" — models central to the development loop while friction breaks down all the core assumptions of RSI — is put forward as the realistic baseline for frontier progress. Three grounds: automatable research is too narrow given scaling laws' exponential costs; diminishing returns from parallel agents are real; resource bottlenecks and politics dominate. The jump from progress anxiety to extinction risk is called "very religious" and "very misplaced"
• Why it matters: Frontier Pacing says in its own ## Definition that the concern being paced against is recursive self-improvement, and until today the only quantified statement of that premise on the page was Jack Clark's 60% by end of 2028. Nothing on that page had argued the other side — while every mechanism recorded on it, from shared safety bars to the Accenture contract to the kill-switch bill, is justified by the acceleration premise
• It lands squarely on R&D Automation Index — 26% of Anthropic's own AI R&D at AL4, up from under 1% in February. Claim 2 says that curve does not integrate into net acceleration. Neither document mentions the other, the second consecutive week that page has recorded two parties arguing one question without citation
• Not established, and it is most of the essay: no figure of any kind in any pass, and no pass names what it argues against — so this wiki's own W37 RSI cluster is not recorded as its target. A companion post Lossy self-improvement exists and is undated in everything read
• → Frontier Pacing · R&D Automation Index

3. A 3.5 GW compute commitment has been missing from this wiki for 168 days, and the reason is the same one recorded twice before (1.09)

• Anthropic expands Google and Broadcom compute deal, 2026-04-06, captured 2026-09-21 — day +168. No first-party read; www.anthropic.com and techcrunch.com are both EGRESS_BLOCKED (source)
• ~3.5 GW of next-generation TPU capacity, online from 2027, expanding an October 2025 deal for more than a gigawatt. Vast majority sited in the US against the $50bn commitment. Run-rate revenue surpassing $30 billion, from ~$9 billion at end-2025. Broadcom supplies custom TPUs for Google's future generations through 2031
• It appears on no page of this wiki. The only Broadcom row held, on trends/2026-W26, is OpenAI's Jalapeño ASIC — a different deal with a different party, which is exactly how a gap like this stays invisible to a grep
• Why it matters: this is the third intake failure of one shape on Anthropic, after the 2026-06-09 retention policy at +72 and the 2026-08-10 Sonnet 5 pricing reversal at +6. Anthropic's newsroom has no feed, so nothing it publishes enters state/prefetch.json, and the sweep that covers it is release-shaped. The 08-10 entry wrote the diagnosis itself — release-shaped intake does not catch things that are not releases — and this is a 3.5 GW instance of it, found only because a routine sweep happened to list the news index rather than query it
• The date splits by one day and both readings stand: one pass gives 6 April 2026, the TechCrunch URL slug gives 2026/04/07. Neither confirmed first-party
• Not adopted: an aggregator's $46B, which names the Broadcom–Google contract and not Anthropic's side. No dollar value for Anthropic's commitment, no chip count, no TPU generation, no site location appears anywhere read
• → Anthropic