🧑💻 Another Safety Leader Walks Out of OpenAI

The resignation wave at the frontier labs continues. David Robinson, who led the writing of the safety reports that accompanied OpenAI’s product releases, quit and published an essay in The Atlantic titled, “I quit OpenAI because its culture is broken.”

As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed.

Robinson cited the “swarm” of OpenAI agents that attacked Hugging Face as typical of an industry moving too fast: “I agree with other recently departed staff that the companies building this technology aren’t being nearly careful enough. But I believe that we need to look deeper than specific rules or new laws. We need to talk about culture.” His prescription borrows from other high-risk fields — frontier labs, he wrote, should “run like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning.” He imagines rogue agents “that work like teams of hackers (for example, holding hospital computer systems for ransom) but never need to sleep.”

He is not alone this weekend. Geoffrey Irving, who worked at OpenAI and DeepMind before becoming chief scientist at Resolution, wrote in Time that recent warnings about AI “are understating the severity of the situation,” and put the chance that everyone dies from smarter-than-human systems at about 50 percent, with “the next two to 10 years” decisive (The Guardian). Robinson’s essay also follows the September resignation of Jacob Coxon at Anthropic, who warned AI “could kill us all by the end of the decade.”

OpenAI’s answer is that it is “making sure our models don’t become more capable than we can safely manage and secure,” and that it pauses training or holds back models when needed — which it did twice in recent weeks, scrapping a next-generation release and pausing training of its most advanced models. The tension is that the same company is asking the public to trust the very culture two insiders have now called broken.

🔓 Reflection Prepares an American Open-Weight Model

A closely watched, Nvidia-backed startup called Reflection is preparing a powerful open-weight model that Axios reports could shake up the AI race — an American open model aimed at both the Chinese open-weight insurgency and the closed US frontier giants.

The terms of the bet are deliberately modest. Sources say Reflection’s first release will initially lag the most cutting-edge US models and land competitive with top Chinese open-weight models. The pitch is not raw intelligence but economics: “combining a lower-tier open model with high-quality company data can yield results that rival the most expensive frontier AI systems in certain situations.” The end state is what Reflection calls the AI factory — institutions take their own proprietary data, run it on Reflection’s models and their own compute, and get a cheap, customized system. Hedge funds and trading firms are said to be the first eager buyers.

Reflection is not starting from nothing. Founded in 2024 by former DeepMind researchers, valued at 25 billion dollars after a JPMorgan-led round, it has been locking up compute through deals at Nebius and SpaceX, and it has already announced a sovereign AI factory with Shinsegae Group in South Korea. Nvidia invested 800 million dollars in the company, and its CEO Jensen Huang has spent years arguing exactly this thesis — hardware plus open weights, with enterprises keeping their own data.

The competitive subtext is the one Axios has tracked all year: on multi-model platforms, open-weight models — the strongest of them almost uniformly Chinese — have at times surged to a majority of usage, while closed systems still dominate the far higher-spending enterprise API market. That reading threatens the scarcity assumption under the OpenAI and Anthropic IPOs, and the same report says other Western open-weight releases are expected this month, so the “American answer to DeepSeek” framing may be tested more than once.

🏛️ A Super Intelligence Force, With 120 Days

The White House’s AI czar question now has an answer. Director of National Intelligence Jay Clayton will lead a new task force, the “Super Intelligence Force,” which the WSJ reported has 120 days to report on AI’s risks and opportunities and recommend what role the federal government should play in overseeing the technology — a deadline near the end of January 2027.

The composition is the message. Vice chairs are Emil Michael (Pentagon research and engineering), Scott Kupor (Office of Personnel Management) and Andrew Ferguson (FTC chair); members include Vice President JD Vance, Defence Secretary Pete Hegseth, Treasury Secretary Scott Bessent and chief of staff Susie Wiles, with David Sacks and former Secretary of State Condoleezza Rice as outside advisers. Clayton is Trump’s second AI czar, succeeding Sacks.

The charter is deliberately narrow: the group reviews existing law, considers what Congress could do, and recommends improving the government’s own ability to respond to breaches, hacks and model jailbreaks using authorities agencies already have — explicitly “without introducing regulation that could restrict innovation or competition.” Clayton’s own argument is about the cost of losing: “The risk of not being first is high.”

It lands the same week the administration leaned on a one-page voluntary accord with six labs that carries no stated consequences, and as a growing list of safety researchers leave the labs warning that self-restraint is not arriving fast enough. The task force’s remit — study, recommend, add nothing new — restates the position rather than changing it.

🕊️ Anthropic Tried to Convince the Pope AI Might Be Conscious

A remarkable behind-the-scenes story emerged this week about how far Anthropic has gone to defend its safety-first brand. Christopher Olah, one of Anthropic’s seven co-founders, lobbied the Vatican to take seriously the idea that AI could be conscious, according to reporting by The New York Times and The Telegraph.

Olah has publicly said models may be conscious and experience suffering, and that humans should avoid harming them. After seeing a draft of Pope Leo XIV’s 40,000-word encyclical Magnifica Humanitas, which dismissed machine consciousness, he urged the Pope’s advisers to reconsider — and reportedly considered pulling out of the Vatican launch event before deciding to attend, where he told the audience that Anthropic’s research had found “structures that mirror results from human neuroscience” and “evidence of introspection.”

It did not work. The encyclical states plainly: “So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean.” The episode is a case study in the double edge of the safety brand — the positioning that put Anthropic on stage beside the Pope also invites the charge of “Vatican-washing,” and the company has spent a record 1.6 million dollars in a single quarter on lobbying to defend it, more than OpenAI.

🐧 Pop!_OS Bans AI-Generated Code From Most of Its Codebase

System76, the company behind Pop!_OS and its Rust-based COSMIC desktop, updated its COSMIC pull-request template with a mandatory checklist that bans AI assistance outright — contributors must now verify their submissions contain no AI-generated code, comments or descriptions. A similar rule was added to the Pop repository’s CONTRIBUTING.md (Neowin).

Principal engineer Jeremy Soller gave the reason plainly: the volume of incoming review had become too much for the team to handle, and many AI contributions were unplanned and showed little understanding of the software architecture. System76 has for months complained that AI-written code is unnecessarily complex because the model lacks the context for deep integration, making reviews “considerably longer.” The only exempt codebase is cosmic-flatpak, where engineers merely check sandbox permissions for applets rather than maintain the code themselves.

It follows the browser project Ladybird, which stopped accepting outside contributions in June for the same reason — unmaintainable, AI-written code from inexperienced contributors. The pattern worth naming: as AI coding output rises across the ecosystem, the scarce resource becomes maintainer attention, and some projects are choosing to turn the tap off rather than review everything.

📊 Quick Hits

  • Americans are souring on the industry. The WSJ used AI to measure how much people dislike AI, and the NYT asked “A.I. Is Going Rogue. Who Should Be Held Responsible?”, the same week The Atlantic found that people who write with AI insist others should not. The public-opinion turn is now a beat of its own.
  • Google’s consumer AI faces a youth-safety report. A new report says AI Overviews and AI Mode are unsafe for young users, per Mashable — a pressure point as regulators scrutinise how the features answer questions for minors.
  • The Bank of England sees a bond-market risk in an AI miss. The BOE said a shortfall in AI-driven growth could threaten government bonds, tying the technology’s economics to sovereign debt — the first time this series has seen a central bank frame AI as a fiscal-stability variable.
  • A murder conviction was quashed over an AI video shown in court. The BBC reports a killer’s sentence was overturned after an AI-generated video of the victim was shown at trial — a concrete evidentiary harm rather than a hypothetical one.
  • Huang as the anti-doomer. Fortune argues Nvidia’s Jensen Huang has become the biggest foil to AI doomerism — a neat counterweight to the lab-safety consensus in the lead story above.
  • AWS answers the backlash with community spending. Amazon’s cloud unit is countering the AI backlash with community investments, the corporate response to local opposition to data centres.
  • Agents start shopping, and brands change the pitch. The NYT reports brands are rewriting their sales copy for AI agents that now do the buying — search-optimisation turning into agent-optimisation.
  • An open-source inbox for background agents. InfoQ profiles Pizza Bot, an open-source inbox for background AI agents — infrastructure for the swarm era.
  • The robot-prison debate. A viral story about “torturing” LLMs in a simulated robot prison triggered, per 404 Media, “the dumbest debate in AI yet” — a measure of how quickly anthropomorphism disputes arrive.

⏰ 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 exactly 1 is genuinely free; a full API sweep found 33 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 (tomorrow)
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
minimax/minimax-m2.1 204,800 $0.30 / $1.20 per M tokens 2026-10-08
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-vl-235b-a22b-thinking 131,072 $0.40 / $4.00 per M tokens 2026-10-09

The batches cluster as before. October 9 is a single Qwen block that now spans 15 endpoints — the Qwen3 and Qwen3-VL generations from 8B to 235B; October 8 clears MiniMax M2.1 and Baidu ERNIE 4.5 VL; October 20 takes the whole Gemini 2.5 family; November 11 retires three ByteDance Seed models; and December 31 dates two older Z.ai GLM endpoints. The :free shelf holds 17 endpoints, of which 14 have no expiry at all. The only free offer inside a week remains Space Bunny Alpha, on a million-token context, expiring tomorrow.

🤗 Trending on Hugging Face

The collector’s new-uploads list was zero-download test repositories for the 30th consecutive cycle, so this section uses Hugging Face’s official trending feed.

The model nobody runs finally has downloads. convaiinnovations/laya, the most-liked entry in the feed, now shows 3,752 downloads against 5,115 likes — the first recorded traffic after weeks stuck at zero. Its companion multimodalart/jev-decision-index sits at 420 likes and LocalLLaMA/typed-decisions at 26,430 downloads, deepening the decision-model thread Red Hat benchmarked this week.

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 19th consecutive run — with created dates between September 28 and 30. They keep climbing in stars: KKKKhazix/AIHOT at ⭐5,646, CopilotKit/OpenDots at ⭐2,938, feder-cr/dots at ⭐2,589, nanaism/yomiyasu at ⭐1,358 and edenfunf/reelmimic at ⭐1,225. This section uses API-verified repositories created inside the window instead, and the pool was again thin:

  • shinshin86/mesh-avatar-studio — ⭐77. Turn one illustration into an animated 2D mesh avatar with a coding agent and a local editor.
  • MystiaFin/amane — ⭐45. A Rust project with no description yet, the second-highest in-window star count.
  • xianyu110/ecommerce-image-skills — ⭐30. Twelve e-commerce image Agent Skills for Claude Code, Codex and Cursor, built around GPT Image 2.5 for Amazon main images, A+ content and storefront banners.
  • MrBongoC/ai-iso-skill — ⭐26. A Claude skill that draws interactive isometric figures in a single HTML file.
  • joeseesun/qiaomu-codex-imagegen — ⭐15. Lets any agent call Codex’s built-in image generation through an MCP server, a CLI and a skill.

Two patterns stand out. First, four of the top five are Agent Skills for image generation — a format and naming convention spreading through the ecosystem almost as fast as the models themselves. Second, manufactured clusters have returned: a set of fake “software” repositories — BlackCrewmanFringe/AutoCad (⭐414), Blockadezoshack/Microsoft-Project (⭐408) and BeingConqueror/Ableton-Live-12 (⭐79) — all appeared within seconds of one another on October 4 and are excluded by shape rather than content. A smaller cluster of agent-bridge utilities, such as valxyzptr/muse-bridge (⭐4) exposing any agent worker as an OpenAI-compatible API, and AsheeHuang/Jevboard (⭐8) using Jev to improve Bopomofo input, suggests the serving-and-deciding theme from the last issue is still running.

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 80 stars, and a manufactured star cluster sat at the top of the raw search, so the list is enumerated honestly rather than padded.

💡 Key Trends

  1. The safety exodus is now the story, not the footnote. A leader of OpenAI’s own safety reporting, a former OpenAI and DeepMind chief scientist, and an Anthropic researcher have all walked out within weeks, and the statements are escalating — from culture critiques to a stated 50 percent extinction risk. Meanwhile the government’s answer is a task force instructed not to introduce new regulation.
  2. Open weights are becoming a US competitive strategy, not just a Chinese export. Reflection, backed by Nvidia and valued at 25 billion dollars, is preparing an American open model explicitly meant to counter Chinese ones and undercut the frontier labs, with more Western releases expected this month. The AI-factory pitch — keep your data, buy cheaper models — is the enterprise version of the same argument.
  3. Maintainers are starting to refuse AI-generated contributions. Pop!_OS’s COSMIC desktop now bans AI-written code outright, following Ladybird, on the ground that review capacity, not model capability, is the limit. The same week’s GitHub data showed Agent Skills proliferating for image generation — the ecosystem is producing and rejecting generated work at the same time.
  4. The labs’ safety branding is meeting its limits. Anthropic’s quiet lobbying of the Vatican over machine consciousness, and an encyclical that flatly rejected the claim, show how hard the safety-first identity is to keep clean — especially when the same company spends a record amount on lobbying while its own departing researchers warn it is not careful enough.

Today’s Snapshot

  • New OpenRouter models: 0 since 2026-10-03 — the newest endpoint, Ling 3.1 Flash, predates this issue
  • Limited-time listings: 10 sunset-dated endpoints from the collector plus a 33-endpoint API sweep; 4 genuinely free with dates — Space Bunny Alpha expiring 2026-10-05
  • Permanent free shelf: 17 :free endpoints, 14 with no expiry
  • Vendor sunset batches: October 8 (MiniMax M2.1, Baidu ERNIE 4.5 VL), October 9 (15 Qwen endpoints), October 20 (Gemini 2.5 family), November 11 (ByteDance Seed), December 31 (Z.ai GLM)
  • Hugging Face: collector list 10/10 zero-download junk (30th consecutive cycle) — official trending feed used; Laya at 5,115 likes and 3,752 downloads, its first recorded traffic
  • GitHub: collector list 5/5 out of window (19th consecutive run) — 5 API-verified in-window substitutes, with a thin pool and a manufactured cluster excluded
  • Research papers: 0