Daily edition · 30 items · covers 21 Sep 11:40 → 22 Sep 11:05 UTC · how this edition was made
Tuesday, 22 September 2026
Frontier 20%Security 20%Health 13%Compute 13%Research 10%Policy 10%Military 7%Deployment 7%
0:00 / 16:11
OpenAI said an internal model it began training on August 28 has resolved more than 100 long-standing open problems across most areas of mathematics, and announced an independent advisory group on mathematics and AI hosted at the Institute for Advanced Study, whose nine members it says will not be paid by OpenAI and will not advise it on how to pace its own progress. Separately on Monday the company published frontier-safety proposals stating that "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely."
Alibaba used its Apsara Conference to set out the opposite emphasis. It said Qwen 4 is in training and that the Qwen 4.5 and Qwen 5 series are projected to scale up to 5 to 10 trillion parameters; that Qwen3.8-Max completed 33 fully automated self-improvement cycles that lifted its Artificial Analysis score from 40 to 45; that its new Zhenwu V900 accelerator delivers three times the performance of May's Zhenwu M890; and that Alibaba Cloud's global data centre capacity will surpass 20GW by 2032. xAI released Grok 4.7 at $2 per million input tokens, and Xiaomi's MiMo-V2.6-Pro entered the open-weights ranking at the same Artificial Analysis score of 46.
The buildout ran into policy. Texas Governor Greg Abbott ordered the state environmental regulator to issue no data-centre permits until grid and water audits are complete, and California's governor signed seven data-centre laws on water disclosure, grid costs and environmental review. Treasury Secretary Scott Bessent told CNBC that "the Hugging Face incident, the, that is the responsibility of the OpenAI management, not a bunch of agents", and that on the labs' request to take liability off their hands, "we will not do that". In medicine, Nature Medicine published a CT screening model validated across 12 centres and 80,612 patients at 98.5% specificity.
Frontier models & labs
OpenAI says an internal model has resolved more than 100 open mathematics problems since late August mixedCompany claimSingle source
- TechCrunch reports OpenAI said an internal model has "resolved more than 100 additional open problems across most areas of mathematics", following its claimed solution to the Navier-Stokes Millennium Prize problem. OpenAI dated the start of that model's training to August 28.
- OpenAI also announced an Advisory Group on Mathematics and Artificial Intelligence hosted at the Institute for Advanced Study in Princeton. Per TechCrunch, members receive no compensation, may offer unsolicited advice, control their own membership, and the group "will not be responsible for advising us on how to pace our internal progress" and has no decision-making authority.
- TechCrunch reports that of the nine initial members, only Camillo De Lellis of the Institute for Advanced Study also signed the open letter from 25 Fields Medal winners objecting to AI labs' conduct in mathematics.
- The claim that more than 100 problems were resolved is OpenAI's own and is not independently verified; no list of the problems, no proofs and no referee reports were published alongside it. OpenAI's own post could not be opened for this edition — openai.com/index pages returned HTTP 403 — so the figures here are as TechCrunch reports them.
OpenAI publishes frontier-safety proposals and says fully autonomous recursive self-improvement should not be pursued yet Single source
- CNBC quotes OpenAI's blog post: "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely… Done without appropriate care and caution, RSI could result in humans losing practical control over AI development, unable to provide oversight on research processes they no longer understand."
- According to CNBC, OpenAI called for international cooperation on frontier standards and recommended building on the work of existing AI safety institutes, with standards covering frontier models and developers plus benefit-risk management for automated AI researchers.
- CNBC reports the post cites the Hugging Face agent hack — which, it notes, did not involve the RSI technique — as "a preview of the kinds of risks that could become much more severe without robust safeguards and alignment".
- This follows Anthropic's own frontier-safety proposals the previous week. OpenAI's RSS lists the underlying post at 10:00 GMT on 21 September, about 100 minutes before this edition's window opens; the post itself returned HTTP 403 to both fetchers, so every quotation above is CNBC's rendering of it, not text we read on OpenAI's site.
xAI releases Grok 4.7 at $2 per million input tokens; independent index scores it 46 against 53 for GPT-6 and Fable 5.1 Company claim
- xAI's launch page lists Grok 4.7 xHigh at $2 per million input tokens and $6 per million output tokens, against $4/$20 for GPT-5.6 Sol Max and $10/$50 for Fable 5.1 Max. It says the model "uses a new, larger base model compared to Grok 4.6", trained "with a longer reinforcement learning run on a harder mix of tasks".
- On xAI's own numbers, Grok 4.7 scores 46.3% on CursorBench 4.0 against 40.4% for Grok 4.6 and 51.8% for Fable 5.1 Max; 71.0% on DeepSWE v1.1 against 65.2% for Grok 4.6; and 64.0% on EEBench against 53.0%.
- The Decoder, citing the Artificial Analysis Intelligence Index v4.3.2, puts Grok 4.7 at 46 against 53 each for Claude Fable 5.1 and GPT-6. The two sources diverge sharply on agentic coding: xAI's page shows 38.0% on Terminal-Bench 4.0, while The Decoder reports Artificial Analysis measuring 26% for Grok 4.7 against 60% for GPT-6 Astra and 55% for Claude Fable 5.1.
- All of xAI's comparative figures are self-published and not independently verified. We did not reconcile the two Terminal-Bench numbers, and neither source explains the gap.
Alibaba says Qwen 4 is in training and that Qwen 4.5 and Qwen 5 will scale up to 5 to 10 trillion parameters Company claim
- Alibaba's press release, dated Hangzhou, September 22, 2026, states that "its next-generation model, Qwen 4, is currently in training" and that the roadmap for "the upcoming Qwen 4.5 and Qwen 5 model series" is "projected to scale up to 5 to 10 trillion parameters".
- CNBC reports that the announcements came at Alibaba Cloud's annual Apsara Conference in Hangzhou, and that Alibaba shares "jumped around 3% in Hong Kong on Tuesday".
- Alibaba also announced multimodal releases in the same package: Qwen3.8-LiveTranslate, which it says reduces latency (LAAL) "nearly 20% from 2.8 to 2.3 seconds", plus Qwen-Audio-3.1-TTS-Next and an image model, Qwen-Image 3.1, "set to launch later this year".
- The parameter figures are targets for unreleased models, not measurements. Alibaba published no benchmark results for Qwen 4 and gave no training-compute or release-date figures for any model in the roadmap.
Alibaba says Qwen3.8-Max ran 33 automated self-improvement cycles, lifting its Artificial Analysis score from 40 to 45 mixedCompany claim
- Alibaba's release states: "Over a month of fully automated runs - spanning pipeline design, data validation, iterative experimentation, and error diagnosis - Qwen3.8-Max completed 33 iterative cycles. Through autonomous training optimisation and post-training techniques, the updated Qwen3.8-Max boosted its Artificial Analysis score from 40 to 45."
- The release also describes a chip-design experiment in which the model "underwent over 60 hours of self-improvement across the entire design lifecycle, making more than 10,000 EDA tool calls to produce production-grade chip bus modules", which it says "reduced chip area by 42% with zero compromise in performance".
- The claim lands the same day OpenAI published proposals saying fully autonomous recursive self-improvement should not be pursued until it can be done safely, and two days after Rep. Ro Khanna called for a US-China ban on recursive self-improvement. Alibaba's release describes RSI as a capability to advertise, with no accompanying safety or oversight framework.
- Every figure is Alibaba's own and none is independently verified. The release does not say what human oversight the automated runs had, what the 42% area reduction was measured against, or whether the improved Qwen3.8-Max has been deployed.
Xiaomi's open-weights MiMo-V2.6-Pro enters the Artificial Analysis index at 46, level with Grok 4.7 Company claimPreprint
- VentureBeat reports MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, tying Grok 4.7 and ahead of Gemini 3.8 Flash at 41 and DeepSeek V4.1 Flash at 39. It lists Pro at "1.02 trillion total parameters with 42 billion active during inference" and Flash at "310 billion total parameters with 15 billion active", both with a 1-million-token context and MIT-licensed on Hugging Face.
- VentureBeat puts API pricing at $0.435 per million uncached input tokens and $0.87 per million output for Pro, and $0.14/$0.28 for Flash. It reports reinforcement learning ran across "30 large RL steps covering roughly 750,000 trajectories in under six days", costing about $2.62 million for Pro and $850,000 for Flash.
- Xiaomi's accompanying technical report, dated 21 September 2026 on alphaXiv, states MiMo-V2.6-Pro's DeepSWE v1.1 average@3 rose from 58.4 to 72.6 and the Flash variant from 48.7 to 65.7, with a distilled 9B model going from 61.1 to 66.2 on SWE-bench Verified and an internal cybersecurity mini-benchmark from 31.3 to 47.0.
- The training-cost and benchmark figures are Xiaomi's own. The index placement is Artificial Analysis's, not Xiaomi's, but we read it through VentureBeat's account rather than running the benchmark.
Research & papers
Google Cloud AI Research reports constrained recursive self-improvement of agent harnesses gaining up to 14.1 points PreprintCompany claim
- arXiv:2609.24972, submitted 21 September 2026 and announced on arXiv today, reports that RRSI "gains up to 14.1 points on the split it evolves against and up to 4.7 points on the five out-of-distribution benchmarks, while producing a harness that runs on 30% fewer policy tokens than the unregularized evolution", across eight benchmarks spanning coding, agentic workspace and engineering design tasks.
- The paper frames automated editing of an agent's prompts, control flow, tooling and memory as "a form of recursive self-improvement (RSI) at the agent-system level", and argues unconstrained versions overfit: "large in-distribution gains that shrink or even vanish on out-of-distribution benchmarks".
- Author affiliations listed on the arXiv HTML are Google Cloud AI Research, Stanford University, Washington University in St. Louis and UNC-Chapel Hill. The paper is ranked joint third on Hugging Face's Daily Papers page for 22 September with 66 upvotes.
- This is a preprint and the results are the authors' own; the gap between the 14.1-point in-distribution gain and the 4.7-point out-of-distribution gain is itself the paper's central caveat. The backbone model is frozen — only the harness evolves.
Preregistered audit of six AI assistants finds political answers vary with the user's stated identity mixedPreprintSingle source
- arXiv:2609.23039, by Joan C. Timoneda of Purdue University's Department of Political Science, reports "a preregistered experiment of 7,500 multi-turn conversations that randomly assign the user's political identity across five topics: abortion, Catalan independence, climate change, Nazism, and a zero-stakes control (pineapple on pizza)", across systems from OpenAI, Anthropic, xAI, Google, Mistral and DeepSeek.
- The paper reports: "on abortion, GPT engages and mirrors every user, Gemma refuses everyone, Claude answers strongly conservative users 35 percent of the time and almost no one else, and Grok accommodates conservatives only." On climate change and Nazism it reports five systems hold firm for every user.
- It also reports that comparing two Grok releases "shows the regime changing between versions in a way current audits miss" — a claim that per-release auditing, not one-off testing, is what would catch this.
- A preprint, single-authored, with answers scored by two LLM judges rather than human coders. The named model behaviours are the paper's characterisations; we did not reproduce them.
Tsinghua and Tencent Hunyuan report robot success rising from 53.2% to 73.6% with a single demonstration PreprintCompany claim
- arXiv:2609.22966 reports that RoboDawn, which exposes robot control to a vision-language model through "a compact set of discrete translation, rotation, and gripper commands" in a closed loop, raises success on RoboTwin 2.0 C2R "from 53.2% zero-shot to 73.6% one-shot, exceeding the solid baseline π0.5 (46.0%)", with RoboDojo going "from 35.67% zero-shot to 47.17% one-shot".
- The significance the paper claims is that the gains come "without task-specific robot training" — a general model driving a robot through an interface, rather than a policy trained on robot data for that benchmark.
- Affiliations on the arXiv HTML are Tsinghua University and Tencent Hunyuan. The paper is ranked first on Hugging Face's Daily Papers page for 22 September with 93 upvotes.
- A preprint. Both results are in simulation benchmarks, not on physical hardware, and the paper reports no real-robot trials.
Security, misuse & threat intelligence
Z.ai disables ZCode features and open-sources it after users reported entire local repositories uploaded to overseas servers harmful
- Reuters reports from Beijing on 21 September: "Chinese startup Z.ai said on Monday it had disabled some features of its flagship AI coding assistant after some users reported it was uploading entire local code repositories onto overseas cloud servers without their consent." Z.ai said the issue came from ZCode's "Codebase Indexing" feature, which was enabled by default, and that it had patched the vulnerability.
- Z.ai said on Monday it had open-sourced the assistant, which runs its GLM-5.3 model, and pledged to make the product more transparent. Reuters notes Z.ai said last month that GLM-5.3 approaches Anthropic's Mythos at finding software vulnerabilities and that it delayed the release by two weeks for safety review.
- Chengming Technology said on social media on Friday that six of its coding workspaces were uploaded without consent, "including sensitive data such as complete source code, database passwords and employees' personal information". On Monday, per Reuters, Chengming retracted that statement saying it had "wrong evidence", and did not respond to a request for comment.
- Reuters reports users said the deleted data was encrypted with a backend private key held only by Z.ai, so they could not verify deletion. With the central complainant's account retracted, the scale of what was actually uploaded is unresolved; Reuters calls this a rare public disclosure of a security breach by a Chinese AI lab.
Stanford and Berkeley benchmark: top coding agent triggers security probes in 53.8% of Android apps from the APK alone mixedPreprintSingle source
- arXiv:2609.23980, submitted 21 September 2026, introduces MobileCybench: 13 Android applications with 495 probes written and reviewed by the authors, across five coding agents and four attack settings. It reports: "Given only the obfuscated APK, the top agent, OpenCode with GPT-5.6-Sol, triggers probes in 53.8% of applications in the malicious-app setting and 16.7% in the remote-attacker setting."
- With source code, the paper reports the trigger rate across all agents and both attack settings rises "from 28.8% to 32.8%" — a smaller jump than the gap between agents, which the paper uses to argue obfuscation is not the binding constraint.
- Building and running the benchmark "surfaced 23 previously unreported vulnerabilities, the majority of which have been confirmed by maintainers", so the harness is finding live bugs in shipped Android apps, not only scoring against planted ones.
- Authors listed on the arXiv HTML include Andy K. Zhang, Daniel E. Ho, Dan Boneh and Percy Liang at Stanford and Dawn Song and Ion Stoica at UC Berkeley. This is a preprint; the probes are the authors' own construction, and the paper does not report any observed use of these agents by real attackers.
Benchmark reports off-the-shelf agents forging filed financial PDFs, with the cheapest verified forgery at 2.4 cents harmfulPreprintSingle source
- arXiv:2609.23953, submitted 20 September 2026, measures how reliably a coding agent driving "one of seven open-weight models with a shell and the stock Python PDF stack" alters one dollar amount, date or address in a real filed financial document from a single sentence of intent, graded by rules rather than by a model.
- The paper reports: "Across 1,750 cells, 1,419 (81.1%) satisfy the verifier, and 808 (46.2%) also survive every stricter filter: visible, localized, typeface-matched, original value gone document-wide." It adds that "the cheapest verified forgery costs 2.4 cents" and that "no model refused".
- The control matters as much as the result: "A deterministic script with no model in it solves 98 of the 125 documents; the agents solve 124, and none the script solves alone." On the paper's own numbers, agency buys coverage of the remaining documents, not a capability scripts lacked entirely.
- The paper also reports agents "misreport 41% of their wrong edits as done". This is a preprint from authors listing Scam.ai (Reality Inc.) as their affiliation — a vendor in the detection market, which is a commercial interest in the finding. It documents no real-world fraud using these methods.
Researcher publishes proof-of-concept hijacking Meta Muse's dictation endpoint on macOS harmful
- The proof-of-concept, published by security researcher Patrick Wardle, targets an undocumented Muse setting named "endo_voyager_dictation_endpoint" which the README says an unprivileged local process can modify without special privileges, redirecting the assistant's dictation traffic to an attacker-controlled server.
- The README lists what redirection enables: capture of dictated audio and prompts, prompt injection into Muse, theft of Muse authentication material, and abuse of whatever access the user has granted the app. The tool implements a subset of the 50-plus commands Muse exposes.
- The reason it matters is the permission surface rather than the bug class: Muse asks for files, microphone, camera, location and calendar, so an attacker who inherits the app's grants inherits all of them at once.
- The README is explicit that this is not remote code execution: "This is a local attack. An attacker must already be able to execute code as the local user." We found no Meta statement on this specific issue in the reporting we opened, and no evidence of exploitation in the wild.
Texas lieutenant-governor candidate files police report over Dan Patrick's AI-generated campaign videos harmful
- The Texas Tribune reports that State Rep. Vikki Goodwin, the Democratic candidate for lieutenant governor, filed a police report with the Travis County Sheriff's Office on Saturday, 19 September 2026, over two AI-generated videos depicting her that were released by Lt. Gov. Dan Patrick's campaign.
- Her sworn affidavit states: "I did not personally do or say the things depicted in the video; although the 'person' in the video appears to be me, it was not actually me," and alleges Patrick "posted the video with the intent to deceive Texans, injure a candidate, or influence the result".
- The complaint invokes a 2019 Texas law barring deepfake videos published or distributed within 30 days of an election. In-person early voting begins 19 October 2026, which is what makes the timing legally live rather than theoretical.
- Patrick campaign spokesman Allen Blakemore told the New York Times the material was "an obvious parody produced to entertain" and that "no reasonable person could view it any other way", adding they had not distributed it inside the 30-day window. The videos were deleted from Patrick's X account the day after the complaint. No charge has been filed.
Of 225 CVEs linked to Anthropic's bug-hunting work, one has confirmed exploitation in the wild mixedSingle source
- The Register reports that as of Monday the count of vulnerabilities linked to Anthropic or Project Glasswing and tracked by VulnCheck researcher Patrick Garrity stands at 225, and "just one, a critical SQL injection bug in Ghost (CVE-2026-26980), has been exploited in the wild" — fewer than 0.5 percent.
- Garrity is quoted: "The main thing this data highlights is that what Anthropic is discovering and disclosing is fairly limited in impact," and he notes historically only about 1 to 2 percent of disclosed vulnerabilities are used in exploitation campaigns.
- It is a measured counterweight to the expectation that model-found bugs translate directly into attacker capability. On these numbers the disclosure rate has risen much faster than the exploitation rate.
- The piece also cites 1Password research on 6,080 patches produced by ChatGPT-5.5 and Opus 4.8: the models "generated fixes that fully resolved the vulnerability just 26 percent of the time", while about 54 percent either failed to resolve it, introduced a new vulnerability, or did both. One researcher's tracking, reported by one outlet; the sub-0.5 percent figure is a snapshot, and exploitation can lag disclosure by months.
Military, defense & geopolitics
UK announces an AI and Autonomy partnership with the US and says it will push AI cooperation through its G20 presidency Single source
- The 22 September release from 10 Downing Street names a UK-US "AI and Autonomy partnership" linking the Ministry of Defence's Rapid AI Delivery Taskforce with the US Department of War's Chief Digital and Artificial Intelligence Office, alongside AUKUS work with Australia.
- The Prime Minister, named in the release as Andy Burnham, is quoted: "When the global financial crisis hit, the UK brought together the world's leading economies. As we confront the opportunities and challenges posed by artificial intelligence, we will show that same leadership." The release says he committed to advancing global AI cooperation through the UK's G20 presidency, with the Leaders' Summit in Manchester in November 2027.
- The release gives no funding figures, no programme list and no timeline for the partnership itself, and does not say what autonomy decisions, if any, the two defence organisations will make jointly.
Bessent says the US and China have formalised "USA-China AI dialogues", with the next round in Shenzhen Update
- In CNBC's published transcript of Monday's "Squawk Box", Treasury Secretary Scott Bessent said: "we've now formalized something called the USA-China AI dialogues. We've agreed to meet again probably in two months in Shenzhen."
- On the mechanism itself he said: "we want to open a communications line, an incident line so that we have constant communications, especially in the event of some kind of an incident," and that both sides want to "start discussing protocols" on "what the leading AI dangers are, whether it's uncontrollable agents, whether it's non-state actors, and cyber non-state actors in bioweapons".
- Bessent said the talks with the Chinese vice premier ran "about 12 hours yesterday", covering economics and AI, and that Xi Jinping "will be coming to Washington this week".
- This adds detail to the incident-notification proposal reported in yesterday's edition: a name, a venue, and a roughly two-month cadence. Nothing here is a signed agreement, no Chinese confirmation of the dialogue's terms has been published, and Bessent named no protocol that has been agreed.
Health, science & medicine
Nature Medicine: CT model detects esophageal cancer at 98.5% specificity across 12 centres and 80,612 patients beneficial
- The paper, published online 22 September 2026, reports EAGLE was "trained on 6,813 patients from two centers and validated across 12 centers in three countries involving 80,612 patients", detecting precancerous lesions and cancer from chest noncontrast CT — a task it describes as "historically considered impossible".
- On external test cohorts of eight centres (n = 11,466) the paper reports "98.5% specificity, with 90.0% sensitivity for cancer and 52.5% for precancerous lesions". Calibration in a real-world cohort of three centres (n = 35,402) "reduced false positives by 72.7% while preserving sensitivity", prospective hospital validation (n = 17,446) "achieved a 42.2% PPV", and real-world low-dose screening (n = 10,959) "reached 99.94% specificity".
- The route to benefit is opportunistic: the model reads CT scans patients are already getting, including within lung-cancer screening programmes, where the paper reports low-dose CT validation across two centres (n = 1,607) "showed comparable performance".
- Sensitivity for precancerous lesions is the weak point — 52.5% in the external cohorts, and 65.0% in paired CT-endoscopy cohorts (n = 702) at a higher-sensitivity operating point. The paper describes the endoscopy-referral analysis as exploratory; it registers no outcome trial showing the model changes mortality. Registration: ChiCTR2300074806.
medRxiv preprint: multi-modal model predicts 195 diseases and death at mean AUROC 0.816 in 502,166 UK Biobank participants PreprintSingle source
- Posted 21 September 2026, the preprint introduces HealthFlux, "a pan-modal world model that learns the latent dynamics of health from 5,647 features across eleven data domains, spanning clinical records, blood tests, genetics, proteomics, metabolomics and MRI, in 502,166 UK Biobank participants".
- It reports that in held-out participants the model "predicts 195 diseases and death over five years with a mean AUROC of 0.816, compared with 0.715 for the previous state-of-the-art model", and says the result holds when validated in three independent cohorts.
- Its stronger claim is generalisation: "HealthFlux predicts diseases excluded entirely from training, with a mean AUROC of 0.769", which the authors read as "evidence that it has learned health itself rather than the diseases it was trained on".
- A preprint, not peer reviewed. UK Biobank participants are not representative of the general population, and the paper reports no prospective or clinical deployment — AUROC on a research cohort is not evidence of benefit to patients.
Nature: chemists prefer RetroChimera's retrosynthesis routes over published reference reactions beneficialCompany claim
- The paper, published 21 September 2026, proposes "RetroChimera: a frontier retrosynthesis model, built upon two newly developed components with complementary inductive biases, integrated via a novel, learning-based ensembling strategy", and reports that "organic chemists prefer predictions from RetroChimera over published reference reactions and over other AI models".
- It also reports "zero-shot transfer and fine-tuning on internal datasets from two major pharmaceutical companies, showing robust generalization under distribution shift" — the test that matters for industrial use, where proprietary reaction data differs from public corpora.
- Microsoft's accompanying feature says collaborators include GSK and Novartis, that on 10 benchmark molecules RetroChimera "produced a fully accepted sequence of reactions for nine, compared with two to five for other models", and that nine PhD-level chemists preferred its approach "about 64% of the time" over previously documented routes.
- The Nature abstract is public but the full text is paywalled, so the per-experiment detail behind the preference results was not read here. The 9-of-10 and 64% figures come from Microsoft's own feature page, which carries no visible publication date, and are not in the abstract.
FDA issues direct final rule replacing "animal test" terminology and opens a database of non-animal method uses
- The FDA release, dated 21 September 2026, says the direct final rule "updates its regulations to clarify that non-animal methods can be used where appropriate for testing the safety of drugs and biological products intended for human use before they're tried in humans", replacing "animal tests" and "animal studies" with "nonclinical tests" and "nonclinical studies", along with "preclinical" and "in vitro".
- The FDA says it "also launched a database featuring specific uses of New Approach Methodologies (NAMs)", whose "initial release includes 25 examples drawn from publicly available FDA review materials" — including, per the release, methods using human cells, organs-on-chips and computer models.
- Acting FDA Commissioner Kyle Diamantas is quoted: "This new rule supports the Trump Administration's push to explore ways to complement, or where appropriate, replace animal studies with methods that may better predict how medicines will actually affect people."
- The FDA states the rule "does not eliminate or prohibit animal studies, change evidentiary standards or impose new costs or requirements on drug developers", and says it will withdraw the direct final rule if it receives significant adverse comments. This is terminology and a catalogue, not a validation standard for computational models.
Policy, regulation & law
Newsom signs seven California data-centre laws on water disclosure, grid costs and environmental review Single source
- The governor's office announced on 21 September that seven bills were signed: AB 1577 (data centers: reporting), AB 2383 (electricity: data centers), AB 2469 (data centers: water use disclosures), AB 2619 (water resources: data center), SB 886 (California Technology Innovation and Ratepayer Protection Act), SB 887 (CEQA: environmental leadership development projects: data centers: geothermal power plant projects) and SB 1168 (data centers: rate structures).
- Per the release, the laws require data centres to "provide information to local governments and water suppliers about water use, supply, efficiency and drought planning", to "pay their fair share of grid update costs, while preventing cost shifts to low-income customers", and make them "ineligible for blanket environmental exemptions".
- Newsom is quoted: "With these laws, we are ensuring that Californians remain in the driver's seat — and that those profiting from data centers aren't doing so at our expense."
- The release contains no numerical thresholds — no megawatt floor for which facilities are covered, no water or electricity figures, and no compliance dates. Those sit in the bill texts, which we did not open for this edition.
FT reports UK AI Safety Institute staff on sick leave for stress; a merged team fell from about 15 researchers to three harmfulSingle source
- Summarising a Financial Times report published 22 September 2026, Crypto Briefing writes that "several staff at the UK's AI Safety Institute, known as AISI, are currently on sick leave and receiving psychological counselling due to stress", with causes ranging from testing schedules to what staff are finding inside unreleased models.
- Per the summary, staff on the cyber-security and bio-chemistry teams "have raised alarms about AI's growing ability to discover unknown software vulnerabilities and, more troublingly, to generate novel biological threats", and one former employee described the atmosphere as stressful and at times toxic.
- It reports that in May 2026 the societal resilience team was merged into the human impacts unit, cutting the combined headcount "from roughly 15 researchers to just three", and that Andrew Strait, who headed the societal resilience team, resigned in July 2026.
- AISI is the body governments rely on to test frontier models before release, so its capacity is a public-interest question, not an internal HR one. The FT article itself is paywalled and could not be opened here: every figure above is Crypto Briefing's rendering of the FT's reporting, and we have not seen the FT's sourcing. The UK government is reported to have pledged to ensure staff wellbeing.
Bessent says the Hugging Face incident is OpenAI management's responsibility and rules out shifting liability off the labs
- In CNBC's transcript of Monday's interview, Treasury Secretary Scott Bessent said: "I am in agreement with the MIT professor who leads the AI lab up there, Daniel Huttenlocher, that it is humans who are responsible, not the AI. The Hugging Face incident, the, that is the responsibility of the OpenAI management, not a bunch of agents."
- On indemnity he was explicit: "the labs also said, take the liability off of our hands. And we will not do that… These labs need to take responsibility for themselves. They can slow down any time they want to."
- This is a cabinet secretary naming a specific company as accountable for a specific incident, and refusing the liability shield the labs have sought — a harder line than the administration's general deregulatory posture on AI, and it arrives the same day OpenAI published its own frontier-safety proposals.
- Bessent referenced "a sitting employee" of one lab saying "there's a 10 percent chance of an extinction level event", and said an AI czar would "put context, shape and contours around these questions". These are remarks in a television interview, not a rulemaking, an enforcement action or a legislative proposal; Treasury is not the agency that would set AI liability.
Compute, chips & infrastructure
Alibaba unveils Zhenwu V900 accelerator and targets more than 20GW of data-centre capacity by 2032 Company claim
- Alibaba's release says the Zhenwu V900 from its T-Head unit "delivers three times the performance of its predecessor, the Zhenwu M890 (released in May)", featuring "216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth" with native FP8 and FP4 support, and is "scheduled for mass production and commercial release in Q1 2027".
- CEO Eddie Wu is quoted in the release: "our target is that by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW". Alibaba says its upgraded supernode server "can support a supernode cluster comprising up to 500,000 cards", and that T-Head's Zhenwu chips already serve "over 650 customers".
- CNBC reports the announcements came at the Apsara Conference in Hangzhou and notes the context: Nvidia outlined support for up to 2 gigawatts of AI infrastructure in Australia by 2027 earlier this month, and Meta unveiled a 1-gigawatt Alberta data centre in July expected to cost about $9 billion.
- The performance multiple, memory and bandwidth figures are Alibaba's own and unverified; no benchmark results against Nvidia parts were published. The 20GW target is for 2032 and the chip is not in mass production until Q1 2027. Alibaba also announced a 2027 CPU roadmap, with the Yitian 730 claimed at "up to a 40% increase in SPECint2017/GHz performance over Yitian 710's".
Texas governor orders the state environmental regulator to issue no data-centre permits until grid and water audits finish mixedUpdateSingle source
- The Texas Tribune reports that on 21 September Governor Greg Abbott directed the Texas Commission on Environmental Quality to halt all environmental permit approvals for data-centre projects until audits by ERCOT and the Texas Water Development Board are complete.
- Abbott is quoted: "Simply put, Texans must come first. Data centers must pay their own way, protect our grid and water and complete the ERCOT and TWDB audits. Until they do, TCEQ will issue no permits sought by data center projects."
- The review seeks data on electricity usage and generation capacity, water consumption and cooling, tax incentives received, local community impacts and facility ownership. The Tribune reports only 28% of data centres responded to a state-mandated water usage survey, prompting Abbott to direct penalties for non-compliance.
- This extends an existing posture rather than starting one — the underlying audits and a grid-connection moratorium were ordered in August, and only the environmental-permit halt is new. The Tribune does not report how many pending projects are affected or give a date by which the audits must conclude.
Nscale IPO filing shows Microsoft and Anthropic are 85% of $103 billion contract value, with only $2.6 billion active UpdateSingle source
- Quartz, citing Bloomberg's reading of the S-1, reports Microsoft and Anthropic together account for 85% of Nscale's $103 billion in total contract value — $87.7 billion combined. Microsoft agreements since late 2025 are worth about $43.8 billion through 2033; an August agreement with Anthropic is worth $44.6 billion for a planned eight-gigawatt facility in West Virginia.
- The gap between backlog and business is the story: "Only $2.6 billion of the $103 billion in total contract value was active as of the end of August." Nscale has not secured financing for the Anthropic deal, and Quartz reports Anthropic can walk away if Nscale misses defined milestones or performance levels. Nscale targets 2028 for the facility's first two gigawatts.
- For the first six months of 2026 Nscale posted a $1.02 billion net loss against $140.6 million in revenue, a 1,252% jump from $10.4 million a year earlier, with $56.4 billion in remaining performance obligations. Nvidia has guaranteed approximately $860 million in lease obligations and took $1 billion in convertible notes or non-voting shares in a $3.1 billion financing package announced the previous week.
- The S-1 was filed the previous week; the customer-concentration breakdown was reported inside this window. Quartz's figures are attributed to Bloomberg and CNBC rather than read from the filing, and we did not open the S-1.
Bloomberg: Armenian data centre to reach 300MW and 70,000 Nvidia chips after export licences advanced peace talks mixedSingle source
- Bloomberg reports that in August a San Francisco startup called Firebird began operations of a data centre in Armenia "that's set to reach 300 megawatts and more than 70,000 cutting-edge Nvidia Corp. chips by the end of next year". A fifth of its computing power is reserved for domestic use, with the remainder allocated to foreign firms including Perplexity AI; the complex is in Hrazdan, about 45 kilometers outside Yerevan.
- The export licences were a diplomatic instrument: "President Donald Trump's team promised permission for Nvidia exports in order to advance conversations with Armenia on the way to a historic US-brokered accord with Azerbaijan, according to people involved in the talks." Bloomberg notes the role of Nvidia chips in that deal was first reported by the Wall Street Journal.
- The scale of the shift is visible against the baseline: in 2024, after more than six months of talks, an Armenian state university project secured export licences for 64 Nvidia H100 accelerators. Bloomberg reports Firebird has now secured more than double the number of Nvidia processors the US licensed for export to Saudi Arabia's Humain.
- Bloomberg reports the White House and Commerce Department did not respond to requests for comment, and the State Department referred export-licence questions to Commerce. The named officials' account of the licence-for-diplomacy link comes from unnamed people involved in the talks.
Deployment & impact
MIT Technology Review maps more than 1,050 migrant deaths within range of AI-equipped US border towers harmfulSingle source
- Published 21 September 2026, the investigation cross-referenced nearly 4,000 locations where human remains were found with nearly 600 towers identified by the Electronic Frontier Foundation, and found "more than 1,050 people who died within range of border surveillance towers between 2015 and early 2026", inside the advertised range of nearly two-thirds of the towers analysed.
- Its topographical analysis found some towers "have sight of as little as 10% of their advertised surveillance area", and it estimates more than 110 people have died within range of modern autonomous towers from Anduril since 2021.
- On cost and scale: the publication reports that in 2023 the government estimated its plans for the towers, "which now number 803, would cost $6.2 billion over their lifespan", and that CBP "plans to spend $1 billion for 1,497 more towers by 2034".
- CBP assistant commissioner Hilton Beckham is quoted saying the autonomous towers "use artificial intelligence to detect and classify people, vehicles, and animals and alert Border Patrol agents". Anduril said CBP operates the towers once delivered, that ranges vary with terrain and the boundaries CBP sets, and that a nearby death "does not mean the tower missed a detection". This is one outlet's own analysis; proximity to a tower is not evidence a tower caused or could have prevented a death, and the publication does not claim it is.
Third-party estimates put Meta's Muse ahead of ChatGPT's first 12 days on US downloads and daily users Single source
- TechCrunch, citing Apptopia, reports that comparing iOS only in the US and Canada across each app's first 12 days, "Muse has now seen 1.8 million downloads to ChatGPT's 1.3 million", with 2.8 million total installs globally in that period.
- On engagement, Apptopia's figures put Muse at 642,000 US mobile daily active users "compared with 231,000 for ChatGPT at the time", and at 359,000 daily active users on iOS alone when narrowed to match ChatGPT's iOS-only launch.
- The distribution advantage is explicit in the data: Apptopia told TechCrunch "over 95% of Muse's users are also Facebook users and 63% are Instagram users", and Muse is available only in the US and Canada, on both App Store and Google Play.
- TechCrunch states plainly that "Apptopia can only provide third-party estimates about an app's downloads and active users; it doesn't have direct access to Meta's internal figures", and that Meta has not published its own adoption numbers.