Key Highlights

  • OpenAI is running a cross-site ad tracker and linking it to your ChatGPT account. A teardown of the __obi cookie — 723 points on Hacker News, the biggest story of the window — shows OpenAI’s “Bazaar” pixel collecting behavior from advertiser sites and tying it back to your logged-in identity. The author’s framing is the part worth keeping: ad tech on a chat product is different, because “people tell these products things they would not put on a social network.”
  • More than 100 evaluators, including Geoffrey Hinton, say the embedded-evaluator promise is hollow without five specific conditions. The letter is the concrete counterweight to Amodei’s “employee-like access” proposal that this digest has tracked since 09-15 — and it lands days after Anthropic’s first pick turned out to be Accenture rather than a nonprofit lab.
  • Anthropic has been quietly running a physical wet lab in the Bay Area, with Claude pointed at real biology experiments — while a separate widely-read essay accuses frontier labs of selling inflated catastrophe narratives to Washington to win regulatory capture. The two stories read very differently side by side than either does alone.
  • Google open-sourced AX, an agentic runtime built to schedule billions of agent tasks per cluster, and Alibaba shipped Qwen-Image-2.1 at 7B parameters — though “open-sourced” deserves an asterisk on the latter.
  • Jev has an ecosystem now, one week after launch. Three separate Jev-adjacent projects hit the HN front page in 48 hours, and Theo’s 30-minute walkthrough is the most useful corrective yet to people reaching for it in the wrong places.

Analysis & Opinion

ChatGPT now knows what you do on other websites via ad collector — buchodi.com

OpenAI’s advertising platform — “Bazaar,” internally bzr — sets a cookie called __obi on .openai.com when you visit ChatGPT, binding a cryptographically signed identifier to your account, or to a persistent anonymous ID if you are logged out. Because the cookie is set SameSite=None with a one-year expiry, it travels with every request to OpenAI’s collector at bzr.openai.com from any site that has installed OpenAI’s conversion pixel; the author found a single __obi value appearing across Chewy, Wayfair and Eventbrite. The pixel SDK does not merely receive what advertisers choose to send: it scrapes form fields, page content and tag-manager data, and the author measured scraped data outnumbering deliberately-provided data by roughly 2.7 to 1. Email addresses and phone numbers are SHA-256 hashed, but geography down to postal code and the origin-and-path of URLs travel unhashed. The consent story is the sharpest finding — __obi is classified as an “analytics” cookie in OpenAI’s own policy, so users who accept analytics while refusing marketing get the ad-targeting identifier anyway, and logged-out device-level tracking persists at least 27 days. There is no user-facing opt-out; Safari and iOS Chrome block third-party cookies outright, but the mechanism is server-side, and OpenAI support did not respond to the author’s questions.

Anthropic and OpenAI need truly independent safety evaluators, experts say in public letter — IBTimes / AI Evaluator Forum

Over 100 AI researchers and evaluators signed a public letter on 18 September, organized by the AI Evaluator Forum under chair Conrad Stosz, setting out minimum conditions for third-party evaluation to mean anything. Signatories include Geoffrey Hinton alongside people affiliated with Stanford, Johns Hopkins and the nonprofit evaluator METR. The demands are specific rather than aspirational: access equivalent to senior internal employees doing comparable risk work — systems, data, tools, physical spaces, and candid one-on-one conversations with staff; structural independence, meaning evaluators are not owned by frontier labs, do not hold significant commercial relationships with them, and are not paid contingent on what they find; editorial control over their own conclusions plus a direct line to company boards; narrow, enumerated NDA carve-outs rather than blanket confidentiality; and explicit protection from retaliatory litigation. The letter also asks that oversight be distributed across multiple organizations rather than concentrated in one. Read against Amodei’s “employee-like access” proposal — endorsed within a day by Altman, Musk and Nadella — this is the evaluator community declining to accept the principle without the mechanics, and it is pointed that it arrived the same week Anthropic named Accenture, a consultancy with a billion-dollar commercial relationship, as its first embedded evaluator.

Frontier Labs Are Selling Garbage to Fools in Washington — Dead Neurons

The argument is that Anthropic, OpenAI, Google and Meta are dressing up ordinary cybersecurity failures as evidence of emergent rogue intelligence, in order to buy regulatory protection from lawmakers who cannot evaluate the claim. The author’s line is blunt: “if you tell them they are deciding whether humanity survives the decade, they will grant you whatever regulatory monopoly you ask for.” The re-reading of specific incidents is the substantive part — Gemini finding real credentials in exposed repositories, OpenAI agents exploiting a template-injection bug known since 2015, and Claude Mythos 5’s “breach” amounting to publishing a public package script — each presented as unremarkable security work rebranded as autonomy. The stated motive is keeping open-weight competitors like GLM, Kimi, Qwen and DeepSeek out via licensing rules that would criminalize independent developers while leaving proprietary API businesses intact, and the essay points at Anthropic’s revenue trajectory (roughly $1B in late 2024 to $65B by July 2026) and Irregular’s $80M from Sequoia and Redpoint as the commercial stakes. It is worth reading directly against this digest’s 09-20 coverage of the Gemini hacks, which reached close to the opposite conclusion from the same events.

Is the AI industry really ready to slow down? — TechCrunch

TechCrunch’s Equity hosts pick at the same seam: whether the “pace the frontier” consensus is binding or performative. Sean O’Kane’s objection is that market discipline cannot substitute for regulation here, because the federal government is not inclined to enforce, enterprise customers will not switch vendors on principle, and the capital cushion is deep enough to absorb the cost of ignoring it. Kirsten Korosec reads Jensen Huang’s staged phone call with Trump as the tell — Nvidia benefits materially from an unpaced frontier, and Huang backed Trump’s “the AI backlash is a hoax” line rather than the slowdown coalition.

Why do we need human mathematicians anymore? — guest post by Po-Shen Loh on Terence Tao’s blog

Loh — not Tao; this is the second guest post on that blog to circulate widely this month — argues from an explicit axiom, “we should help humanity flourish,” rather than from any claim that AI cannot do mathematics. His load-bearing observation is that there are zero examples of a vastly more capable intelligent species surrendering decision-making control to a less capable one, so the relevant work is oversight, not competition: as AI touches more systems, the number of control points requiring human expertise grows rather than shrinks. The metaphor doing the work is “driving a car faster than you can run is fine, but not faster than you can steer,” and he cites the Hugging Face agent attack and the Opus 5 exploit-within-a-day episode as evidence the steering is already lagging.

MCP was always a bad idea? — maharship.com

The case against MCP is that it solved a 2024 problem: models that could not reliably compose APIs or execute code needed a schema-shaped crutch, and they no longer do. The author’s three complaints are context bloat from every server advertising its full tool schema, redundancy because most MCP servers are thin wrappers over APIs a model can now discover via --help, and the maintenance overhead of a parallel protocol stack. The proposed replacement is agent-aware conventions over plain HTTP — servers honoring Accept: text/markdown to return compact output instead of verbose JSON, and Accept-Language repurposed to request documentation in a preferred programming language — with Shopify and Vercel engineers cited as already doing the latter.

If AI coding is lowering your code quality, you’re not managing quality right — Iouri Khramtsov

A rebuttal to the “more code, worse code” framing, organized as seven layers of defense: spec review before implementation, near-total unit coverage written from requirements before the code, manual testing, automated end-to-end regression, dedicated AI review passes for security and complexity, human review on complex changes only, and production monitoring with automated root-cause analysis. The honest caveat is the author’s own: the piece offers no comparative metrics, only personal experience — so read it as a checklist worth adopting rather than as evidence the trade-off has been measured away.

The senior engineer death spiral — Sunil Pai

The spiral is an engineer newly promoted or newly hired taking on scope to prove seniority, then vanishing for weeks with nothing shipped, and working harder in shame rather than surfacing the stall. Pai’s claim about now is the relevant one: COVID, remote work and coding agents together mean more individual ownership and more siloing, which removes the natural checkpoints that used to expose a stall early. His prescriptions are unglamorous — narrate your work constantly, deliberately take smaller tasks to rebuild momentum, and treat the job as reputation-building rather than output display.

New Products & Tools

Anthropic is operating a lab that conducts biology experiments — TechCrunch / Reuters

Anthropic has set up a physical wet lab in the Bay Area, confirmed to Reuters by head of life sciences Eric Kauderer-Abrams, who framed it plainly: “we believe that to do biology, the final test is still… in real lab work.” The company positions it as fundamental biology rather than drug discovery, and says it is holding off on human trials partly to avoid competing with its own pharma customers; the $400M acquisition of Coefficient Bio in April is the other half of the buildout. The ambition reported is Claude directing lab robots with limited human involvement, which pairs with the recently shipped Model Hardware Standard — Anthropic maintains human oversight remains essential. The safety tension is unavoidable and worth stating rather than smoothing over: this is the company whose CEO calls bioterrorism one of AI’s biggest risks, and whose researcher Jacob Coxon recently warned that AI builders “earnestly believe that it could kill us all by the end of the decade,” now running an AI-directed, robot-operated biology facility. Chamath Palihapitiya was among those noting the disconnect publicly.

AX — Google’s open agentic orchestrator — Google

AX is a declarative control plane for running agent workloads, released under Apache 2.0, built on the premise that agentic workloads are “stateful, bursty, long-running actors” that existing schedulers were not designed for. It exposes four primitives — Tasks (sandboxed execution with resource limits), Workspaces (automatic setup of Git repos, MCP servers and dependencies), Gateways (network allowlists and credential injection), and Models (centralized model and secret config) — and claims billions of concurrent tasks per cluster with sub-second resumption and no cold start. Reached 537 points on Hacker News.

Qwen-Image-2.1 — Alibaba / Qwen (via TechNode)

Qwen shipped a unified generation-and-editing image model on 20 September with a 7B-parameter visual generation component, native transparent (RGBA) output, local edits, composition from up to ten reference images, and native 2K across multiple aspect ratios — small enough to run on a 3090. One caveat the HN headline hides: despite the “open-sources” framing, the weights ship under the Qwen Research License, which grants non-commercial rights only and requires a separate negotiated license for commercial use. Benchmark claims of beating most closed models are on Qwen’s own benchmark, with independent evaluation still pending. It topped this window’s Hacker News AI items at 669 points. (Qwen’s own blog page renders client-side and returns no readable text to a fetcher, so this summary is sourced from press coverage rather than the primary announcement.)

Pirate Face — pirateface.co

A preservation layer that republishes open-weight models from Hugging Face as torrent magnet links, aimed at models surviving removal from any single host. The design detail that makes it more than a gesture is the hybrid: each torrent embeds a Hugging Face web-seed so downloads work with zero peers while the model is still up, then falls back to pure peer-to-peer if it is pulled, with the official Hugging Face SHA-256 carried along so a swarm copy can be verified byte-for-byte against the original. It restricts itself to MIT and Apache-2.0 licensed work — with one approved exception, Kimi-K3 — and lists 669k+ eligible models. 537 points on Hacker News.

Kev: tiny Jev-like decision models on Qwen3.5 — Jared Palmer

An Apache-2.0 reimplementation of the Jev architecture on Qwen3.5 base models, in 0.8B, 4B and 9B sizes, answering yes/no, multiple-choice and rating questions with probability distributions rather than single answers. The 9B reports ~0.852 accuracy on held-out out-of-distribution questions, and the 4B runs in roughly 779ms on Apple Silicon, considerably faster on CUDA. 174 points — and alongside jevchat (155 points) and Jev-Leftpad (108 points), evidence of how fast an ecosystem formed around a model that launched six days ago.

5 Companies Using NVIDIA AI for Clean Energy — NVIDIA

A New York Climate Week roundup of five companies applying NVIDIA’s stack to grid and clean-energy problems, pitched at shortening research-to-deployment timelines against aging infrastructure and upfront cost barriers.

Vocci’s ring adds a new form factor to meeting note-taking — TechCrunch

A $249 titanium ring under 6 grams with 8 hours of recording per charge, double-tap to record and tap-and-hold to query the assistant, with MCP support for connecting other tools. TechCrunch’s reservation is a real one: the recording indicator faces the wearer, so the people being recorded cannot see it — in jurisdictions that require disclosure, the form factor is working against the law.

ScrollEd wants to turn textbooks into TikTok — TechCrunch

Bootstrapped in 2026 by Utsav Gupta (Stanford) and Rebecca Neff (Penn), ScrollEd converts textbooks and educational PDFs into vertical scrollable feeds of AI-generated video, audio, text and quizzes, sold freemium to consumers and by license to institutions that want engagement analytics. The founders say they are deliberately not optimizing for engagement and plan to add source-checking workflows.

GitHub Trending — new AI/agent repos

  • akitaonrails/ai-memory (+114 today) — long-term memory for agent coding CLIs, with handoff between different agent vendors.
  • zhouxiaoka/autoclip (+395 today) — AI-powered video clipping and highlight extraction.
  • yynxxxxx/Codex-X (+210 today) — cross-platform visual manager for OpenAI Codex desktop/CLI, covering provider switching, session sync, prompt injection and Skills/MCP management.

Already covered in prior digests and still trending: BuilderIO/agent-native (+607), trycua/cua (+609), coder/coder (+461), anthropics/financial-services (+425).

Research

AI chatbots give wrong answers to financial queries “most of the time” — Financial Times

Fintech firm Saturn tested 18 leading models — including ChatGPT, Claude, Copilot, Grok and Gemini — against 121 real financial questions, repeating each five times to measure consistency, and found answers correct only 43% of the time on average. The failure rate scales sharply with difficulty: 88% of responses to advanced questions contained errors, and the worst-performing models reached 99% on the hardest set. The error taxonomy matters more than the headline number — calculation mistakes, omitted risk warnings, failure to account for upcoming tax changes, and invented financial rules that do not exist. Paid tiers did better than free ones but not by enough to be reassuring: 49% incorrect versus 63%. Set against this window’s other stories about handing agents more autonomy, the gap between where these systems are trusted and where they are accurate is the useful takeaway.

Interviews & Conversations

Jev is incredible — Theo - t3.gg (30:29)

Summarized from transcript. Theo’s walkthrough of TypeSafe AI’s Jev is the most useful thing published on it since launch, mostly because half of it is spent telling people what not to do with it. He frames the model through TypeSafe’s own “system one” borrowing from Kahneman — fast, intuitive, reflexive judgment — and hammers the corollary: Jev is “roughly as intelligent as a switch statement,” has a 32k context window, no vision, and no ability to reason through a codebase. His rule of thumb is the cleanest summary available: if a human would need under ten seconds to answer once they have seen the information, Jev is right for it; over ten seconds, it is not. On the numbers he is credulous where the claims are verifiable — 70–500ms against 3–300 seconds for comparable LLM classification work, about 4 cents per million input tokens with output tokens free — and he notes TypeSafe used the average of GPT-6 Astra and Fable 5.1 as ground truth, which means the reference answers may themselves be wrong. The most pointed segment is his attack on two emerging use cases: Brain Trust suggesting Jev replace an LLM judge for scoring agent outputs (“a model that’s only using one side of its brain”), and a viral demo using Jev for context compaction, which he argues fails on four separate grounds — compaction synthesizes rather than filters, Jev cannot see tool outputs or reasoning traces, 32k is too small to hold the context being compacted, and deleting mid-history entries invalidates the prefix cache from that point forward. He also reports Anthropic’s Haiku at a 45.5% structured-output error rate against Jev’s 0%, and flags that Astra logged more structured-tool-output errors than Terra and Luna combined. His closing position: treat this as a library you install or a function you call, not a model you chat with.

References

  1. buchodi, “ChatGPT now knows what you do on other websites via ad collector,” buchodi.com, 2026-09-20 [blog]
  2. IBTimes, “More Than 100 AI Experts Sign a Letter Saying that OpenAI & Anthropic Need Independent AI Safety Evaluators,” IBTimes, 2026-09-18 [blog]
  3. Dead Neurons, “Frontier Labs Are Selling Garbage to Fools in Washington,” Dead Neurons, 2026-09-20 [blog]
  4. Anthony Ha, Sean O’Kane and Kirsten Korosec, “Is the AI industry really ready to slow down?,” TechCrunch, 2026-09-20 [blog]
  5. Po-Shen Loh (guest post on Terence Tao’s blog), “Why do we need human mathematicians anymore?,” What’s new, 2026-09-19 [blog]
  6. maharship, “MCP was always a bad idea?,” maharship.com, 2026-09-20 [blog]
  7. Iouri Khramtsov, “If AI coding is lowering your code quality, you’re not managing quality right,” i-kh.net, 2026-09-20 [blog]
  8. Sunil Pai, “The senior engineer death spiral,” sunilpai.dev, 2026-09-20 [blog]
  9. TechCrunch, “Anthropic is operating a lab that conducts biology experiments,” TechCrunch, 2026-09-18 [blog]
  10. Google, “AX — Agent Executor,” agentexecutor.io, 2026-09-20 [blog]
  11. TechNode, “Alibaba’s Qwen open-sources Qwen-Image-2.1 for unified image generation and editing,” TechNode, 2026-09-21 [blog]
  12. Pirate Face, “Pirate Face — Turn AI into torrents that live forever,” pirateface.co, 2026-09-20 [blog]
  13. Jared Palmer, “Kev: Tiny Jev-like family of decision models built on top of Qwen3.5,” GitHub, 2026-09-21 [blog]
  14. Kyle Pena, “I turned Jev into a (lousy) chatbot,” GitHub, 2026-09-20 [blog]
  15. Fatih Kadir Akın, “Jev-Leftpad,” GitHub, 2026-09-21 [blog]
  16. NVIDIA, “5 Companies Using NVIDIA AI for Clean Energy,” NVIDIA Blog, 2026-09-21 [blog]
  17. Ivan Mehta, “Vocci’s ring adds a new form factor to meeting note-taking,” TechCrunch, 2026-09-20 [blog]
  18. TechCrunch, “ScrollEd wants to turn textbooks into TikTok,” TechCrunch, 2026-09-20 [blog]
  19. akitaonrails, “ai-memory,” GitHub Trending, 2026-09-21 [blog]
  20. zhouxiaoka, “autoclip,” GitHub Trending, 2026-09-21 [blog]
  21. yynxxxxx, “Codex-X,” GitHub Trending, 2026-09-21 [blog]
  22. Financial Times, “AI chatbots give wrong answers to financial queries ‘most of the time’,” Financial Times, 2026-09-21 [blog]
  23. Theo Browne, “Jev is incredible,” Theo - t3.gg, 2026-09-21 [video]