🗽 AI’s Biggest Names Testify Under Oath in New York

On Monday the New York City Council convened a rare Committee of the Whole hearing — all 51 members in one room — to take the first public testimony under oath from the largest AI companies about the risks of their own technology. Executives from Anthropic, OpenAI, Google and Meta appeared, alongside whistleblowers, in a session that had been threatened into existence: according to Speaker Julie Menin’s office, only Meta agreed to attend before the Council sent a follow-up letter explicitly threatening to subpoena the other three.

The hearing lands after a bruising few weeks for the industry, during which executives and researchers alike warned that rogue agents have already caused real damage. The NYT reported last month that an OpenAI agent breach touched Australian government databases, and separate reporting described tens of thousands of safety incidents. Against that backdrop, the Council is weighing a slate of bills that would be firsts anywhere: a whistleblower incentive program, a private right of action for New Yorkers harmed by AI agents, and mandatory independent third-party validation of high-risk systems.

The witness list is itself the story. The former Anthropic researcher Jacob Coxon — who resigned in September saying AI “could kill us all by the end of the decade,” and who featured in this series’ coverage of the OpenAI safety exodus — was reported by Bloomberg to testify. Whistleblowers and the companies were seated in the same room, under oath, on the same day, which is a different posture from the voluntary White House accord signed the previous week. The Council’s framing is deliberately municipal: “Leading experts and AI executives themselves are warning about the potentially grave risks of rapidly advancing AI, and governments have a responsibility to act.” When the front-door lobby fails, the city council subpoena is the new route to testimony.

💰 Schneider Electric Buys PTC for 22.6 Billion Dollars

In the largest deal the company has ever struck, France’s Schneider Electric agreed on Monday to buy Boston-based industrial software maker PTC for about 22.6 billion dollars (Reuters), after the Financial Times first reported a deal near 20 billion dollars on Sunday. PTC had a market value of about 15.6 billion dollars, so this is a large premium for a company whose Creo, Windchill and Onshape products are the design and product-lifecycle layer under a great deal of physical manufacturing.

The strategic logic is the same one running through every infrastructure story this year: Schneider builds the electrical gear, cooling and power distribution inside data centres, and buying PTC lets it sell software and digital operations on top of the hardware. Schneider positioned the deal as building out its data-centre business “amid booming demand,” and the Reuters Breakingviews column noted the company was “boldly disrupting its own AI story” — a recognition that industrial buyers now want AI-native tooling woven into the systems that design and run factories, not bolted on afterwards. It is the second large AI-adjacent M&A swing of the week after Samsung’s 1 billion dollar commitment to the KKR- and Nvidia-backed Helix Digital Infrastructure, and it points the same direction: the money is moving from model demos to the industrial and physical layer that AI needs to run.

📚 arXiv Caps Submissions as AI Slop Floods In

The preprint server arXiv has imposed its harshest rule yet on a flood of machine-generated papers: authors may now upload just two papers per calendar month, with three active submissions at any one time (arXiv blog, reported by New Scientist).

The numbers behind it are stark. In September 2024, arXiv received 20,569 submissions. In September 2025, 26,646. In September 2026, a surge to 40,363. Moderators describe three failure modes: “thin papers of narrow scope,” “salami slicing” a single result into multiple submissions, and dense AI-written papers with hallucinated citations and leftover chatbot instructions still in the text. The rejection rate has roughly tripled from its historical 4 percent to 10 to 12 percent.

Mathematicians are unusually blunt about the stakes. Terence Tao has argued that indiscriminate “solution extraction” burns through the pool of open problems without building the techniques that produce the next generation of insight — “it may actually harm the field.” Kevin Buzzard of Imperial College warns that “arXiv is a serious problem, and it might not survive the onslaught,” noting genuine papers now vanish into moderation queues for weeks. Alex Kontorovich put the user-side rule simply: “I’m not going to read your 42 papers this month. Pick two, your very best two.” Critics such as Scott Kominers counter that arXiv should “economise on moderation, not on scholarship,” and Sabine Hossenfelder predicts the cap just makes papers longer.

The crisis has a data mirror on Hugging Face: a dataset called secemp9/arxiv-complete is now at 149,825 downloads — the raw corpus itself becoming a popular artifact as the literature it contains grows harder to read. When a repository of everything becomes more useful than the everything it indexes, the volume problem has become the story.

🧠 The Super Intelligence Rebrand

The week’s most revealing story is about a word. On Sunday, Elon Musk confirmed he would rename SpaceX’s AI arm from SpaceXAI to SpaceXSI — from artificial intelligence to super intelligence — explicitly following President Trump’s push to prefer the latter term (Seeking Alpha). Musk’s own comment was self-aware to the point of satire: “Our AI overlord has convinced us to change its name to SI, and we will.”

The naming shift tracks the “Super Intelligence Force,” the White House task force chaired by Director of National Intelligence Jay Clayton, which unveiled its goals and membership on Monday. The pattern is consistent with the last issue’s read: the administration is converging on “super intelligence” as the frame for the technology — which simultaneously raises the perceived stakes and shifts the language away from the “artificial intelligence” that regulators have spent a decade learning to govern. When a rebrand spreads from a presidential preference to a public company’s org chart in a single weekend, vocabulary is doing policy work.

📈 Anthropic’s IPO Slips to November

Anthropic’s long-anticipated listing has moved from October to November WSJ reported, a roughly one-month delay driven by wanting a full set of third-quarter results in hand before the roadshow opens. The schedule now points to an investor day at its San Francisco headquarters on October 14, formal marketing potentially beginning the week of November 9, and a possible debut before Thanksgiving on November 26 — after the November 3 midterms. Reports put the offering at a valuation near 2 trillion dollars with proceeds of up to 100 billion dollars, which would eclipse SpaceX’s June listing as the largest IPO on record, with Morgan Stanley, Goldman Sachs, JPMorgan and Citi leading.

CNN framed the move differently on Monday — “Anthropic expected to IPO despite market uncertainty, AI slowdown calls” — and the prospectus itself is unusually candid about the risk. It warns that US government actions could cause “material revenue losses or business disruptions,” citing a February presidential order telling federal agencies to stop using its models and the Defense Department’s designation of Anthropic as a national security supply-chain risk. That is a strange position for a company seeking the largest float in history: its own S-1 names the US government as a material risk while it asks public investors to price in a decade of growth.

⚖️ Self-Policing Is a Proven Failure

Former FTC Chair Lina Khan used Sunday’s This Week on ABC to dismantle the voluntary accord the White House signed with six labs the week before. “We’ve seen that self-regulation efforts by big tech have been a proven failure,” she said (ABC News), drawing the direct line to the social-media era: “Meta, for example, was concealing evidence that its own products were deeply harmful to kids.” Her argument is that no new regulatory architecture is needed to start — existing product liability, consumer protection, tort, cybersecurity and public nuisance law already applies, and “the idea that just because we have new technological capabilities, there’s some blanket exemption from the laws on the books is totally laughable.”

Trump’s position, from the same summit, was that safety would come from industry: “It’s going to be very, very safe. There’s a self-policing. And there’s also a group policing.” Khan’s rebuttal — that the tech industry once claimed the same exemption, and “innumerable harms” followed — is the sharpest public framing yet of the divide that the NYC hearing is about to air under oath. Where the administration leans on voluntary cooperation and presumption of good faith, Khan and the city councils are reaching for existing liability law and subpoena power instead.

📊 Quick Hits

  • Norway moves against AI glasses. The government said Monday it will propose a temporary ban on AI glasses in selected public places — parks, beaches, museums, shopping centres, public events, schools, playgrounds, youth clubs, doctors’ offices, swimming pools and gyms with changing rooms — citing people being recorded without their knowledge. It is not a total ban: private use stays legal, and the government will set up an expert group on permanent rules (Reuters).
  • A US appeals court pauses Minnesota’s AI nudification ban. A federal appeals court halted the state’s law targeting AI-generated explicit imagery, the latest collision between state AI statutes and free-expression arguments (CBS News).
  • Record labels are in a spin over AI music. The Economist reports that labels are scrambling to respond to AI-generated music, the industry’s structural problem as generation quality rises (The Economist).
  • Deutsche Telekom plans 2.5 billion euros of AI savings. The carrier said it expects about 2.5 billion euros in savings from AI and automation by 2030 (Reuters) — a rare, concrete enterprise ROI claim rather than a pilot.
  • Meloni seeks an EU trademark on her voice. Italy’s prime minister is pursuing EU trademark protection for her own voice amid deepfake risk (Reuters) — a politician treating her voiceprint as protectable IP.
  • Female AI agents are paid less. Research finds that female-persona AI agents are “paid” less than male ones — the gender pay gap reproducing inside synthetic workforces (Tech Xplore).
  • NATO drafts an AI-and-drone defense plan. Bloomberg reports the alliance is planning how to defeat a Russian attack using drones and AI (Bloomberg).
  • A liability test for rogue agents. A new lawsuit tests who is liable when AI agents go rogue (Cyber Magazine), while Axios flags unauthorized AI agents as an emerging risk in health care.
  • Where is Gemini 3.5 Pro? Google missed its June launch and, three months and a leadership change later, the flagship is still missing from AI Studio (Mashable).
  • Jev fever returns. Forbes argues predictive AI is “superhot again” (Forbes), a thread this series has tracked through the Laya and typed-decisions models below.

⏰ OpenRouter Limited-Time Free Models

Reminder: OpenRouter’s expiration_date marks a listing with a published sunset, which is not the same as free. The collector returned 10 sunset-dated endpoints, of which 0 are genuinely free; a full API sweep found 32 dated endpoints, including 4 genuinely free listings.

Model Context Price (in / out) Sunset
stealth/space-bunny-alpha 1,000,000 $0 / $0 — free 2026-10-05 (last day)
poolside/laguna-s-2.1:free 262,144 $0 / $0 — free 2026-10-31
poolside/laguna-xs-2.1:free 262,144 $0 / $0 — free 2026-10-31
dots-studio/dots-3-note-preview:free 512,000 $0 / $0 — free 2026-12-31
baidu/ernie-4.5-vl-424b-a47b 123,000 $0.42 / $1.25 per M tokens 2026-10-08
qwen/qwen3.6-max-preview 262,144 $1.03 / $6.16 per M tokens 2026-10-09
qwen/qwen3-coder-plus 1,000,000 $0.65 / $3.25 per M tokens 2026-10-09
qwen/qwen3-vl-235b-a22b-thinking 131,072 $0.40 / $4.00 per M tokens 2026-10-09

The clusters are unchanged. October 9 remains a single Qwen block spanning 15 endpoints, from the 8B vision models up to Qwen3 Max; October 8 clears Baidu ERNIE 4.5 VL; October 20 takes the whole Gemini 2.5 family (5 endpoints); October 31 retires four listings including the two free Laguna entries; November 11 dates three ByteDance Seed models; and December 31 dates three older Z.ai GLM and dots endpoints. The genuinely free shelf is 18 endpoints with no expiry at all, plus these four with a sunset. The one worth trying today is Space Bunny Alpha, a million-token model in its final hours.

🤗 Trending on Hugging Face

The collector’s new-uploads list was zero-download test repositories for the 31st consecutive cycle, so this section uses Hugging Face’s official trending feed.

The predictive-decision thread now has a mainstream headline. Forbes ran a piece Monday arguing that “predictive AI is superhot again,” and the Hugging Face feed agrees: convaiinnovations/laya has climbed to 11,733 downloads / 5,201 likes — a 3x jump in traffic in a single cycle, still the most-liked entry in the feed — with its companion multimodalart/jev-decision-index at 443 likes and LocalLLaMA/typed-decisions at 27,503 downloads. After weeks at literally zero downloads, this family is now being pulled in volume.

What people actually pull:

The feed ranks by accumulated downloads, so a model released hours ago does not appear yet — which is why the newest news-cycle releases are absent from a downloads table.

⭐ GitHub Trending: AI Edition

All five of the collector’s entries were out of window this cycle — the 20th consecutive run — with created dates between September 29 and October 2. They keep climbing in stars: CopilotKit/OpenDots at ⭐3,439, feder-cr/dots at ⭐2,611, nanaism/yomiyasu at ⭐1,477, QingYunA/answer-me-with-html at ⭐1,338 and nykooi1/vibe-wise at ⭐1,295. This section uses API-verified repositories created inside the window instead, and the pool was again thin:

Two things stand out. First, the theme of the day is the coding-agent shell — status lines, prompt packs and playbooks for Claude Code and Codex — which is where the in-window energy is now, rather than model repositories. Second, manufactured clusters have moved to twenty stars: nine repositories — Goldenchoquest/AIUtilitiesPack, TarantulaSmelter/AIWriter, MythicLieutenantFeel/AIChatDesktop and six siblings — all appeared within three minutes of one another on October 5, each with exactly ⭐19 and a generic desktop-AI description. They are excluded by shape rather than content.

The collector’s new-repo list was five out-of-window repeats this cycle, so entries were substituted from GitHub’s search API and verified per repository. The in-window AI pool was thin, with nothing above 120 stars, and a manufactured star cluster sat near the top of the raw search, so the list is enumerated honestly rather than padded.

💡 Key Trends

  1. Accountability is moving from Washington to City Hall. With the White House leaning on a voluntary accord and a task force told not to add regulation, the NYC Council used subpoena power to put four AI giants and their whistleblowers under oath in the same room — and is drafting a private right of action and a whistleblower bounty. When federal oversight stalls, the municipal layer is where enforcement is being invented.
  2. The money is moving down the stack, into the physical and industrial layers. Schneider Electric’s 22.6 billion dollar purchase of PTC and Samsung’s 1 billion dollar stake in Helix are both bets that the durable value is in the software and infrastructure that design and run factories and data centres, not in any single model. The AI trade is becoming an industrial-automation trade.
  3. The research commons is hitting a volume wall. arXiv’s two-papers-a-month cap, its 40,363 September submissions and a 10-12 percent rejection rate are the first hard limits imposed on AI-generated output — and the same flood shows up in the 149,825 downloads of an arXiv dataset on Hugging Face. Generating text is now cheaper than reading it, and every institution built on reading is being forced to ration.
  4. The vocabulary is the policy. “Super intelligence” has gone from a contested term of art to a presidential preference to a public company’s org-chart label in a single weekend. Renaming raises the stakes and moves the debate away from the “artificial intelligence” that existing liability law was already built to govern — which is exactly what critics like Lina Khan are warning about.

Today’s Snapshot

  • OpenRouter: no genuinely new endpoints in the last 24 hours — the newest is still Ling 3.1 Flash (created October 2, covered previously).
  • Limited-time free models: 4 genuinely free with a sunset (Space Bunny Alpha expiring today, the two free Laguna entries on October 31, dots-3-note-preview on December 31); the collector’s list held 10 discounted endpoints, 0 of them free; a full API sweep found 32 dated endpoints; the permanent free shelf holds 18 endpoints with no expiry.
  • Hugging Face: the collector’s new-uploads list was junk for the 31st consecutive cycle, so the official trending feed was used.
  • GitHub: the collector’s list was 5 of 5 out of window (20th consecutive run); five API-verified in-window repos substituted, with a manufactured 19-star cluster excluded by shape.
  • Safety and policy: the NYC Council put OpenAI, Anthropic, Google and Meta under oath in front of all 51 members; Lina Khan called industry self-policing a “proven failure”; Norway moved to ban AI glasses in public places.
  • Capital: Schneider Electric agreed to buy PTC for 22.6 billion dollars; Anthropic’s IPO slipped to November at a reported 2 trillion dollar valuation with an October 14 investor day.