JadidJust-in AI Developments, Incidents & Deadlines
September 2026
SMTWTFS123456789101112131415161718192021222324252627282930

AI news, Tuesday, September 22, 2026

11 stories

TLDR

  • Grok 4.7 finally shipped after two months of slipped dates — $2/$6 unchanged, but it scores 46 on Artificial Analysis vs 53 for Claude Fable 5.1 and GPT-6, and the agentic-coding gap is much wider than the headline number suggests (Sep 21) Trending
  • Alibaba used Apsara to lay out a full-stack answer to Nvidia: the Zhenwu V900 accelerator (216GB, clusters to 500k chips, mass production 1Q27), 20GW of data centers by 2032, and a Qwen 4 preview (Sep 22) Infra
  • OpenAI stood up an independent math advisory group at the Institute for Advanced Study, disclosing that an internal model trained since Aug 28 has resolved 100+ long-standing open problems (Sep 21) Policy
  • Guild.ai/Morning Consult survey: 96% of orgs run agents in production and 96.4% are confident in their agent inventory — but two-thirds had an agent incident last year and only 31% have an automated kill switch (Sep 22) Agents Production
  • Ireland's DPC fined Google €403M over unlawful location-data processing and gave it six months to fix the pipeline — the fourth-largest GDPR fine ever, landing squarely on ad-targeting inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ (Sep 21) Policy

Action Items

  • Check whether anything of yours pins gpt-5.5 in Codex or ChatGPT Work. It's retired Oct 14 in those surfaces (API unaffected); migration target is gpt-5.6-sol. Three weeks out.
  • Answer the three questions from the Guild.ai survey for your own org: do you have a registry of agents in production, audit trails, and a kill switch you've actually tested? The survey's finding is that confidence and capability have decoupled.
  • If you're benchmarking Grok 4.7 for a coding path, test on your own agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ workload, not the headline index. The 46-vs-53 intelligence gap understates a 26%-vs-60% gap on Terminal-Bench 4.0.
  • Read Linear's CI write-up before your next pipeline budget conversation — it's the most concrete public account of AI-authored PR volume turning CI into the constraint.

Models & Tools

  • xAI ships Grok 4.7 — same price, larger base model, still mid-pack on agentic coding (Sep 21) Trending — After a timeline Musk walked back four times since late July, Grok 4.7 is out: a larger base model, a longer RLReinforcement learningTraining by trial and error, where the model takes actions and adjusts based on the rewards it gets.Kaelbling et al. 1996 ↗ run, training weighted toward multi-hour tasks, and better self-verification. Pricing holds at $2 in / $6 out per million tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗, which is closer to Chinese model pricing than to Western frontier pricing. On Artificial Analysis's Intelligence Index v4.3.2 it scores 46, against 53 for both Claude Fable 5.1 and GPT-6; its Coding Agent Index rose 47 → 56 over Grok 4.6. The number that matters for anyone routing agent work: on Terminal-Bench 4.0 it hits 26% versus 60% for GPT-6 Astra and 55% for Fable 5.1 — behind even DeepSeek V4.1 Flash at 27%. It's live in Cursor, Grok Build, the xAI API, OpenRouter, Vercel and Cloudflare, and rolling out in GitHub Copilot for Pro and above. Read it as a price-tier play, not a frontier one — real value if your workload is knowledge work and single-turn coding, a poor fit if it's long-horizon agent loops. (xAI, The Decoder, Artificial Analysis)

  • Alibaba's Apsara keynote: Zhenwu V900 accelerator, Qwen 4 preview, and a self-improving Qwen3.8-Max (Sep 22) Infra Trending — T-Head, Alibaba's silicon arm, unveiled the Zhenwu V900 for both training and inference: 216GB of memory, 1,200GB/s chip-to-chip interconnect, native FP8 and FP4, roughly 3x the previous Zhenwu M890, and — with Alibaba Cloud's new network architecture — clusters scaling to 500,000 accelerators. CEO Eddie Wu called it China's most powerful AI chip. Mass production is 1Q27, so this is a roadmap, not a shipping product. Alongside it: Alibaba Cloud capacity beyond 20GW by 2032, a preview of Qwen 4 Max / Flash / Plus / 27B (Qwen 4 still in training), and a roadmap to 5–10 trillion parameter models in Qwen 4.5 and Qwen 5. The most interesting disclosure is the quiet one: Alibaba says Qwen3.8-Max ran 33 fully automated self-improvement cycles over a month — pipeline design, data validation, experimentation, error diagnosis — moving its Artificial Analysis score from 40 to 45. That's a public, quantified claim about automated model R&D from a lab outside the US. (CNBC, TechNode, TrendForce)

  • Foremerge: an open coordination protocol that catches agent intent conflicts before Git sees code conflicts (Sep 21) Agents Open Source — Show HN release of a protocol layered above Git for teams running parallel coding agents. The premise is that Git compares text, not intent — two agents can each produce a clean diff while one quietly removes an extension point the other depends on, and nothing flags it until both land. Foremerge has agents publish intent and claim semantic scopes with declared operations before writing code, so incompatible plans collide in a queryable store rather than in a merge. Two deliberate design choices: it never locks files or blocks agents (one crashed agent shouldn't stall the fleet), and it never asks a model to judge a conflict — deterministic rules only, so the same inputs always give the same answer. Ships as an MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server. Worth a look if you're past two or three concurrent agents on one repo and have started seeing merges that are clean and still wrong. (naw103/foremerge on GitHub, HN thread)

  • Mini-AGI: a continual-learning model trained from scratch on an 8GB laptop, where reading and training are the same event (Sep 21–22) Open Source — A Show HN project that was the community's top open-research engagement of the day. ~530M parameters and growing, trained batch-1 on a single continuous byte stream from a 7.8B-character corpus (TinyStories, OpenHermes-2.5, OpenThoughts-114k, the Lichess database). There is no separate fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ regime and no frozen base — it takes a gradient step per chunk as it reads, and the same path serves generation. To fit in 8GB with gradients and optimizer state, weights live on disk and only the working set is resident; the vocabulary is the 256 byte values, so there's no tokenizer to fit. The transferable idea is the architecture, not the scale — it's the cleanest small demonstration of continuous training as a serving-time property rather than a batch job. (volotat/mini-AGI on GitHub, HN thread)

APIs & Pricing

  • OpenAI retires GPT-5.5 from ChatGPT, ChatGPT Work and Codex on Oct 14 — the API is unaffected (announced ~Sep 16, deadline Oct 14) Pricing — GPT-5.5, which only launched Apr 23, 2026, comes off the consumer and Codex surfaces in three weeks. OpenAI directs Codex users who sign in with ChatGPT to switch to gpt-5.6-sol before the date; GPT-6 Astra is the other replacement. For chat users the switch is a default change and should be invisible. The distinction to internalize is the surface split: if your team's Codex usage runs through ChatGPT sign-in rather than an API key, this is a migration for you; if it runs through the API, it isn't. Worth a five-minute audit of which path your seats are on. (OpenAI model documentation, via Gizmochina)

Security

  • Survey of 362 IT decision-makers: agent inventory confidence is at 96.4%, agent incidents at 66.7% (Sep 22) Agents Production — Guild.ai published research fielded by Morning Consult (Aug 4–9, 2026, 362 IT decision-makers at 100+ employee companies, ±5pt margin). 96% of organizations have agents in production and 47% of those run dozens or more. Against that: only 42.7% have a centralized monitoring dashboard, 39.8% have logging or audit trails, and 31% have an automated kill switch (76.5% can stop a bad agent within minutes counting manual processes). 60.6% think employees are deploying agents without approval, and only 35.9% require registration before deployment. Two findings are worth arguing about internally. First, the report's own read on why fast response hasn't prevented incidents: the hard problem isn't stopping a bad agent, it's noticing. Second, companies with 20–199 engineers report agent incidents at nearly twice the rate of the smallest teams — 78% vs 43% — which puts the inflection point at a team size most readers of this digest are at or past. (Guild.ai / GlobeNewswire, full report)

Industry & Business

  • Verda raises $189M at a $1B+ valuation — Europe's newest AI cloud unicorn, out of Helsinki (Sep 22) Infra — The oversubscribed Series B was led by Emergence Capital, with MUFG Innovation Partners, Supermicro, Varma, Lifeline Ventures, 6 Degrees Capital, byFounders and Finnish state investor Tesi participating; angels include Meta's Ola Tørudbakken. The number that carries the story is revenue velocity: annualized run rate hit $165M in July, up from just over $60M at the April round — roughly 2.7x in three months. Verda has added London and San Francisco offices and is expanding across Europe, the US and Asia. Relevant as a pricing signal more than a funding one. European GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ capacity with a sovereign-adjacent story is compounding fast enough that a second procurement quote is worth getting if you're currently single-sourced on a US hyperscaler for training. (Bloomberg, TNW, EU-Startups)

AI in Production

  • Linear rebuilt its CI because AI-authored PRs made validation, not authorship, the expensive part (Sep 21) Production Trending — Linear's engineering team published how they got PR wait times from over 6 minutes down to just over 5 while nearly quadrupling the test suite, cutting runner time per test roughly in half. The changes: off GitHub Actions onto faster third-party runners, the tsgo native TypeScript compiler, lint rules rewritten to avoid full type-graph builds, minimized critical-path jobs, and reworked Vitest sharding. It drew 136 points and 135 comments on HN. The broader pattern it illustrates is the one to plan around: AI coding tools don't remove bottlenecks, they relocate them. CircleCI's 2026 data has feature-branch throughput up 59% year over year while median main-branch throughput actually fell, and LinearB measured agentic AI PRs waiting 5.3x longer to be picked up than unassisted ones. If your generation capacity went up this year and your merge rate didn't, this is where it went. (Linear engineering, HN thread)

  • Claude API outage: 80 minutes of elevated errors across Mythos 5.1, Fable 5.1 and Opus 5 (Sep 21, evening PT) Production — Anthropic reported elevated errors on requests to Mythos 5 / 5.1, Fable 5 / 5.1 and Opus 5, affecting claude.ai, the Claude API, Claude Code and Claude Cowork. Impact window was 5:50pm–7:10pm PT, cause identified roughly 20 minutes in, success rates normal across all models by 7:11pm PT, resolved 7:35pm PT. No public postmortem yet, consistent with the pattern after the Sep 3 tri-provider outage. Short and contained, but it's the second multi-model Anthropic incident this month — if your failover path is still a plan rather than a tested runbook, that's two data points in three weeks. (Claude Status, HN thread)

Policy & Safety

  • Ireland's DPC fines Google €403M and orders a six-month fix on location-data processing (Sep 21) Policy — The inquiry, opened in 2020, covered May 2018 (GDPR entry) through Feb 4, 2020 and found Google processed users' location data unlawfully and unfairly across three features: Web & App Activity, Location History, and Location Accuracy. The DPC's stated harm is specifically about inference: users may have been unaware their location was being used to influence them with ads or to infer their interests, losing control of their personal data. It's the fourth-largest fine the DPC has issued (behind Meta's €1.2B, TikTok's €530M and Instagram's €405M), and Google has six months to bring processing into compliance. The reason this belongs in an AI digest rather than a privacy one: the violation is in the inference layer, not the collection layer. If you derive user attributes from behavioral signals to feed a model or a targeting system, the theory of harm here applies to your pipeline too. (Data Protection Commission, Irish Times, The Hacker News)

  • OpenAI creates an independent math advisory group at the IAS — and discloses 100+ resolved open problems (Sep 21) Policy — The Advisory Group on Mathematics and Artificial Intelligence is hosted at the Institute for Advanced Study in Princeton with nine founding members — François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, Melanie Matchett Wood — operating independently of any AI company, with members taking no payment. The prompting context, disclosed in the same announcement: an internal model OpenAI began training Aug 28 has resolved the Navier–Stokes Millennium Prize problem (covered here Sep 15, including the credit dispute) and more than 100 long-standing open problems across most areas of mathematics, at a pace that surprised OpenAI's own mathematicians. The group will advise on reviewing and communicating emerging results and on academic standards. The scope limit is the part to read carefully: OpenAI states explicitly that the group "will not be responsible for advising us on how to pace our internal progress on mathematics." This is a disclosure-and-communications body, not a brake — which makes it a useful template for what "independent oversight" is likely to mean in practice as other labs announce similar structures. (OpenAI, Terence Tao's blog, TechCrunch)

Coming Up

  • Meta Connect — Sep 23–24, Menlo Park. Zuckerberg keynote 4pm PT Sep 23; developer state of the union 10am PT Sep 24. Expected: new Ray-Ban Meta variants, camera-free "Luna" AI glasses, the "Phoenix" mixed-reality headset ($1,000–$2,000), and deeper Muse integration — the same agent Amazon blocked this week.
  • Trump–Xi meeting in Washington — Sep 24. AI incident-notification mechanism is on the table.
  • OpenAI Videos API and all Sora 2 models removed — Sep 24. No named successor.
  • OpenAI DevDay — Sep 29, Fort Mason, San Francisco.
  • Gemini gemini-omni-flash-preview shuts down — Sep 30. Replacement: gemini-omni-1.1-flash.
  • Docusign MCP server opens to all agents — Sep 30.
  • OpenAI gpt-5.4-cyber removed from the API — Oct 1. Migrate to gpt-5.6-cyber.
  • GitHub Copilot billing changes Oct 1; second model-retirement wave Oct 2.
  • Gemini 2.5 Flash Image (gemini-2.5-flash-image) shuts down — Oct 2.
  • GPT-5.5 retired from ChatGPT, ChatGPT Work and Codex — Oct 14. API unaffected.
  • TechCrunch Disrupt — Oct 13–15, Moscone West, San Francisco.
  • StepFun Step 5 Preview open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ — Oct 15.
  • AGNTCon + MCPCon North America — Oct 22–23, San Jose.
  • OpenAI API shutdown for o1, o1-pro, o3-mini and o4-mini — Oct 23; full o3 and o3-pro follow Dec 11.
  • Siri AI adds French, Japanese, Korean, Portuguese and Spanish — October. Still no EU availability date.
  • Firmus ASX IPO targeting up to ~US$5B — end of October.
  • EU Commission industrial-AI package — November.
  • Agentforce Hunter (outbound sales) reaches GA — November.
  • Anthropic IPO window — November (pushed from October to include Q3 financials).
  • Cursor loses access to OpenAI models — Nov 12.
  • GPT-5.6 Sol promotional pricing ($4/$20) runs through at least Nov 21–30.
  • California expert recommendations on independent verification and a frontier kill switch due — mid-to-late November.
  • EU AI Act: prohibitions on AI systems generating non-consensual intimate imagery and CSAM take effect Dec 2, 2026.
  • Emerald AI's 100MW flexible data center in Manassas, VA — online before end of 2026.
  • Connecticut's AI health-plan rules take effect — Jan 1, 2027.
  • Alibaba Zhenwu V900 mass production — 1Q27; T-Head server CPUs — 3Q27.
  • OpenAI transcription deprecations (whisper-1, gpt-4o-transcribe family) shut off Feb 26, 2027.
  • Google must bring location-data processing into GDPR compliance — by ~Mar 2027 (six months from the DPC order).
  • UN Global Dialogue on AI Governance — May 2027.
  • EU AI Act high-risk obligations for stand-alone systems — Dec 2, 2027; embedded high-risk AI in regulated products — Aug 2, 2028.

Type to search · ↑↓ to move · enter to open