Sources
Who we read, and what we take from each
Every feed the deck polls, grouped by what kind of source it is, with the permission we have to use it stated plainly. This page is generated from the same register the pipeline runs on, so it cannot disagree with the site.
01Primary
The organisation said it itself, or this is the paper. 21 sources.
A non-profit lab that publishes training data and methods, not only weights. When a claim needs a reproducible account behind it, AI2 tends to be the one with it.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Infrequent and worth waiting for. On-device and efficiency work that the frontier-scale conversation mostly ignores.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
The primary source for AI research. Everything downstream is reporting on what lands here first.
Metadata is CC0 1.0 Public Domain Dedication (arXiv API Terms of Use).
Computation and Language: where most language-model work files. Same API and same terms as the existing cs.AI row; the robots note is in CANDIDATES.md.
Machine Learning: where most method work files. Same API and terms as cs.AI.
Deployment-shaped: how models get run in production, at what cost. Thin on research, strong on the operational detail nobody else publishes.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Researchers explaining their own papers at length. Infrequent, technical, and the clearest writing in academic AI.
No robots.txt at all, so no restriction is stated (measured 2026-09-19).
The open-research collective. Small output, high signal, and unusually willing to publish what did not work, which is the part the field mostly leaves out.
No robots.txt at all, so no restriction is stated (measured 2026-09-19).
Product and announcement side of Google's AI work, as opposed to DeepMind's research side. Useful for what is shipping rather than what is being discovered.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
The research arm, publishing its own results. Slower than the news cycle and far more load-bearing: when a paper here and a headline elsewhere disagree, this is what was actually claimed.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Slower than the news cycle and more load-bearing. Papers arrive here with their authors attached.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Where open models are actually released, as opposed to announced.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Enterprise and hardware angles, including quantum work that is not AI at all. Filtered here to the AI items.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Broad and deep, and not only about language models. The feed carries full text, which is rare among corporate research blogs.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
The European lab, and the most consistent publisher of open-weight models at the frontier. Their release notes are the primary record of what actually shipped.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
US government primary source on AI standards and evaluation. Federal works, robots permits. Not AI-only, so the keyword filter carries it.
The hardware layer, written for people who actually run the models. Reliably ahead of the press on what the silicon can do, and reliably promotional about it.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
The lab speaking for itself. When this and the reporting disagree, this is what was actually said.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
A frontier lab whose models sit in this deck's own register with no source until now. Own blog, robots permits, 44 entries. ToS not read: §9.7 applies.
Image, audio and video generation. Their releases are the reference point for what open generative models can do at a given moment.
robots.txt allows every AI agent tested (measured 2026-09-19). ToS NOT read.
The lab's own blog. robots permits. ToS not read: §9.7 applies.
02Reported
An established outlet covering it. 18 sources.
Technical reporting that does not flinch from the detail. Rarely first, usually right.
robots.txt allows every AI agent tested (measured 2026-09-19). ToS NOT read: see plan section 9.7.
General-audience framing and strong editorial standards. Rarely first, and useful precisely because it only covers what has cleared a higher bar.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Deep silicon microarchitecture. Already reached this deck second-hand through Hacker News.
Consumer-product angle on AI. Good on what people can actually buy and use.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Engineering-first, allergic to hype. Allows the citation bot and blocks the training bots, which is the clearest statement of intent in the register.
Blocks every AI agent except OAI-SearchBot (measured 2026-09-19). HEADLINE AND LINK ONLY.
University press office, so read it as the institution's own account of its work. Reliable on what was done, optimistic about what it means.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Casey Newton on the platform and policy side: who decided what, and under what pressure. Reporting rather than analysis, and well sourced.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
University press releases, restated. Useful for reach, weak on scepticism: treat every claim as the institution's own.
No AI-crawler restriction in robots.txt (measured 2026-09-19). ToS NOT read.
Best-in-class on silicon and datacentre economics. Partly subscriber-only, which the reader's door now handles per item. ToS not read: Tier B caps us at the feed's own short description.
Enterprise and infrastructure reporting. Heavy volume, so it earns its place on coverage rather than on selection.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
First to the funding and the personnel news. They cut their own feed description to 27 characters, which is them telling us what they want used.
Blocks all seven AI agents found (measured 2026-09-19). HEADLINE AND LINK ONLY.
AI-only newsroom with a fast cadence and no AI-crawler restriction, which makes it the most usable dedicated outlet in the register.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Sceptical to a fault and often right for it. Their robots file asks AI tools to stay out, so we take the headline and send the reader to them.
Blocks 7 of 9 AI agents tested (measured 2026-09-19). HEADLINE AND LINK ONLY.
Fast and well connected. Their feed carries full text and their robots file says do not use it, so we take the headline and send the reader to them.
Blocks CCBot, ClaudeBot, Google-Extended, anthropic-ai, PerplexityBot (measured 2026-09-19). HEADLINE AND LINK ONLY.
The silicon beat: supply, benchmarks, what the chips actually cost. Strongest in the register on the hardware constraint underneath every model claim.
robots.txt allows every AI agent tested (measured 2026-09-19). ToS NOT read.
AI policy and safety news desk.
Enterprise angle. Rate-limited us on one request, so it gets the slowest cadence in the register and a backoff.
Mostly permissive robots.txt, but returned 429 on a single probe request (measured 2026-09-19). ToS NOT read.
Strong on the human consequences, lighter on the mechanism.
robots.txt allows every AI agent tested (measured 2026-09-19). ToS NOT read: see plan section 9.7.
03Commentary
Analysis, opinion, or community attention. 17 sources.
Narayanan and Kapoor, whose entire project is separating what AI does from what it is sold as. The most reliable corrective in this register, and the one most likely to disagree with the rest of the deck.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
First-party technical safety research. Feeds the alignment section.
The field's loudest sceptic. Carried deliberately: a deck positioned on evidence tiers should not be all boosters. Register runs hot, hence evidence tier 3.
Not a source of fact, a source of ATTENTION. What the people who build this stuff are reading today.
Official API, MIT licensed, no stated rate limit.
Jack Clark, weekly, on where the field is actually going rather than what shipped.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Nathan Lambert on post-training, open models and what benchmark numbers actually mean. Opinionated, technically grounded, and quick to say when a result is thinner than its headline.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
A weekly sweep with a low hype floor. Good for catching what the daily cycle dropped.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Practitioner-facing, strong on tooling and what engineers actually ship.
Lilian Weng's long-form surveys. Rare, enormous, and the reference people send each other when they want a subject explained once and properly.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Ethan Mollick, testing these tools on real work and reporting what happened. Empirical where most commentary is speculative.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Implementation-level explanation: what the architecture actually does, with the code. For readers who want the mechanism rather than the summary.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
One person, reading everything, testing it himself, and showing his working. The highest signal per word in the field.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Volume aggregator, adopted PROVISIONALLY: read its trust.py row after a week before keeping it.
Essayistic, slower cadence.
Long-form essays by researchers, edited properly. Slow, and the closest thing the field has to a serious magazine.
No AI-crawler restriction in robots.txt (measured 2026-09-19).
Timothy B. Lee. Careful and sceptical, explains rather than reports.
The most thorough weekly AI roundup published anywhere. Opinionated and heavily sourced, so evidence tier 3.
05Excluded, and why
Named here on purpose. Leaving a publisher out silently would be the opposite of the point.
MIT Technology Review
Blocks even the bots whose only job is to cite and send traffic. A publisher refusing the referral is asking to be left out, and honouring that is the difference between a curator and a scraper. Kept in this file, with its reason, so nobody re-adds it in six months without reading this line.
06Missing from this list
Anthropic, Meta AI, Cohere, The Batch and Stanford HAI publish no machine-readable feed at all, checked on 20 September 2026 by asking each site for its own feed declaration and then trying the conventional paths. Nothing was guessed and nothing was assumed.
And Nature Machine Intelligence refused this deck, which is a different thing: it answered 200 but served a challenge page titled 'Client Challenge', so nothing here is a finding about what they publish. We did not find out whether they publish a feed, so this page does not say that they do not.
This paragraph is generated from that check rather than typed. Until 20 September 2026 it was typed, and it claimed that a lab starting a feed would show up here on the next build, which nothing in the code was doing.