A quiet X day: the cookie that follows you, a 7B image model, and LeCun's case against LLMs

In the 24 hours ending September 21, 2026, two of the fourteen AI and tech accounts this channel tracks showed anything new worth carrying. Yann LeCun argued about what the current model architecture can reach; Ethan Mollick argued about how the research on AI's effects should be done. That is the honest shape of a quiet Sunday, so the other four items below come from Hacker News and are labeled as fallbacks — the arrangement this channel uses when the monitored feed goes quiet. Retweets, replies, jokes, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked.

The case that auto-regressive LLMs cannot get to human-level AI X post · 3,012 likes, 458,343 views at capture · posted September 20

Yann LeCun, a Turing Award winner and until recently Meta's chief AI scientist, restated a claim he says has never stopped being true — "auto-regressive LLMs, in and of themselves, will not lead human-level AI" — and gave four reasons 1. First, the reasoning today's systems perform is not auto-regressive search at all, but it runs in token space, which he calls "limited and inefficient". Human-like reasoning, he argues, has to search in continuous representation space. Second, the self-improvement methods in practice only work where output quality can be scored without a human: mathematics, code, and scenarios that can be simulated accurately. Humans and animals, he says, learn new skills far more efficiently than reinforcement learning does.

Third, today's multimodal assistants rely on separately trained encoders that are not language models. He agrees with that design but wants it built through JEPA — joint-embedding predictive architecture, the self-supervised approach he has advocated — and counts roughly 3,000 JEPA papers in four years as evidence the field is converging on it. Fourth, he treats what is missing as evidence: if LLMs were the path to human-level AI, we would already have consumer domestic robots and Level-4 or Level-5 self-driving cars. He ends on a definition, quoting Jean Piaget: "intelligence is not what you know, it is what you do when you don't know."

Two of the four reasons are arguments from absence — we do not have the robots, so the architecture is wrong — and the paper count is his own tally, not a survey. The one number a reader can weigh is the one he does not dispute: he calls today's systems impressive, very useful, and superhuman in an increasing number of domains, coding among them.

The research AI's economic effects need now, and the kind that has to wait X post quoting Alex Imas, Google DeepMind's director of AGI economics · 256 likes, 33,627 views at capture · posted September 20

Alex Imas, who leads AGI economics at Google DeepMind and is on leave from Chicago Booth, wrote about what it is like to read empirical AI papers on the economy right now 2. Economics has a standard he does not want abandoned: clean identification, a defensible instrument, parallel trends satisfied. That is what gets a paper into a top journal, and he says it is very important work. It also takes years, sometimes decades, because a good instrument for a specific causal question often never turns up. The same research will eventually exist for AI, he says, and it is absolutely necessary.

What he wants in the meantime is a second kind of paper — ones where a reader can trust that the authors did their best methodologically, while everyone acknowledges that the field is moving too fast to wait for perfect identification. The point is not to lower the bar but to accumulate noisy signals: several independent teams approaching the same question from different directions. His example is the change in early-career hiring in AI-exposed work, where no single paper has clean identification, but enough teams have converged that he is willing to say a slowdown is probably happening.

Mollick's reply broadened the argument past economics: "I would broaden this to all social science. We are in uncharted waters." He wants fast, smart research that is informed by what AI can actually do, forward-looking, and allowed to be less than fully nailed down, and he notes that this sort of work is not usually high status in a field 3. Both posts are about method rather than findings, and Imas names no papers in his, so the convergence on hiring is his summary rather than a citable result. The trade they describe — noisier evidence now against cleaner evidence later — is the reader's to make for their own field.

Qwen open-sources a 7B image model that outputs transparency Hacker News fallback · 653 points, 173 comments at capture · submitted September 20

Qwen released Qwen-Image-2.1 on September 20 and open-sourced it, with weights on Hugging Face and ModelScope and code on GitHub 45. The model unifies text-to-image generation and image editing in one system, and its visual generation component is small for the job: 32 single-stream DiT layers — diffusion transformer layers — with 7B parameters.

The new capability is native transparency. A prompt decides whether the model returns an ordinary image or one with an alpha channel, and the same model can lift a subject out of a photograph as a transparent layer and then edit inside it, changing a subject's expression or replacing text while keeping the background clear. Editing also takes up to ten reference images at once, and local edits can be specified with circles, painted annotations, or separate masks. The efficiency work is in the attention design: text and instructions use a token-level causal mask while image generation uses a chunk-level mask, and the input images and instructions are cached once as static context rather than recomputed.

The quality comparison in the post is Qwen's own benchmark, run against its own model, so it is a vendor claim and nothing more. The part a reader can actually act on is that the weights are public: the test of Qwen-Image-2.1 is a local run, and the Hugging Face card is the entry point 6.

A torrent mirror for open model weights Hacker News fallback · 537 points, 143 comments at capture · submitted September 20

Pirate Face mirrors open models from Hugging Face — language, image, audio, and datasets — as magnet links, on the argument that a peer-to-peer swarm cannot be taken down the way a single host can 7. Every mirrored file keeps the checksum Hugging Face published, so a download can be verified byte for byte against the original rather than against whoever is seeding it. The site says it has 669,000-plus eligible models, and browsing, downloading, and seeding need no account.

Two details decide whether it is worth anything to a reader. The verification path is the useful half: the claimed handle scheme ties a user's page to a verified Hugging Face account, which is the project's answer to impersonation, and the checksum is what makes a torrent copy trustworthy in principle. The advertised drop-in endpoint, which would let an existing pipeline point at Pirate Face by changing one environment variable, is listed as coming soon rather than working. The permanence claim itself rests on strangers continuing to seed, and the site publishes no independent audit of the mirror.

Google's AX runs agent fleets from a YAML file Hacker News fallback · 498 points, 220 comments at capture · submitted September 20

AX is an open, declarative control plane for running agentic workloads at scale, published by Google on top of the open Agent Substrate runtime 89. A task is a small YAML file: list the Git repositories, MCP servers, and skills the agent needs, describe the goal, set a network policy, name the model. Running ax apply on it creates a sandboxed workspace already set up that way.

The project reduces the whole problem to four primitives — isolated execution for a task, declarative workspace setup, a gateway that locks traffic to an allowlist and injects credentials, and one place for model configuration and secrets. The scale claims are the interesting part and are the project's own: every task runs as a lightweight actor, so AX claims billions of concurrent agent sessions per cluster without an orchestrator limit; idle agents waiting on a model or a tool call are checkpointed and resumed in under a second; and dozens of tasks share each worker so idle waiting time becomes spare capacity.

This is worth a look for a reason that has nothing to do with model quality. The pitch is about everything around the model — sandboxing code nobody has reviewed, capping what an agent can spend in a loop, fencing its network, and keeping its state alive between calls. A reader already running agents will recognise the bottleneck this claims to remove; a reader deciding whether to run them at all now has a repository to read before believing any of it.

OpenAI's ad pixel reports visits on other sites back to your ChatGPT account Hacker News fallback · 719 points, 382 comments at capture · submitted September 20

A security researcher reports that OpenAI's advertising measurement sets a cookie called __obi, scoped to .openai.com and bound to your ChatGPT account, and that this identifier is then sent back to OpenAI from ordinary websites as you browse them 10. Any company that buys ads on ChatGPT installs an OpenAI measurement pixel on its own site, the way retailers already install Meta's and Google's tracking code; loading that pixel sends the identifier to OpenAI along with what you were reading, searching for, or buying. The effect is that OpenAI can connect that activity to the account you are signed into.

The mechanism as he documents it runs in three steps. The ChatGPT client asks OpenAI's backend for a signed token that binds the account to a 22-character identifier, carrying an explicit consent field reading analytics_allowed. That token is exchanged at bzr.openai.com for a cookie with SameSite=none and a one-year lifetime, which is the configuration a cookie needs to be sent on requests from other sites. From then on, requests from an advertiser's page to OpenAI's ad hosts carry it. He says he reproduced this on his own phone under two independent capture methods and cross-checked several months of observed traffic covering 936 distinct advertiser pixels across 1,029 hostnames.

The finding is one researcher's own capture, published on his own site, and it has not been independently replicated; treat the mechanism as demonstrated and the scale as his measurement. The advertisement-measurement pixel itself is documented by OpenAI 11. What a reader can check cheaply is the premise rather than the study: visit a site that runs ChatGPT ads, then look for __obi in the browser's storage for openai.com.

Two of the six items are X posts from the accounts this channel tracks — LeCun on what the architecture can reach, Mollick on how the research gets made. The other four are the day's build news, and each one has a single artifact worth opening instead of trusting: the Qwen weights, the Pirate Face checksum, the AX repository, and your own browser's cookie jar.