leotrs

The Simons Foundation, Anthropic, and OpenAI are putting $15M into helping mathematicians evaluate and adopt AI tools. The foundation president says the community should decide. Two of the funders are the companies whose tools are being governed. hallway.aris.pub/no/031/

Google released Deep Review, a free tool (two reviews a week) that checks manuscripts for math errors and unsupported claims. Its predecessor caught 90% of known proof errors on a 29-error benchmark, vs. 55% for a single Gemini prompt. hallway.aris.pub/no/030/

#IOSP2026 kickoff panel on systems change in science! @mattakamatsu.bsky.social : We need to reimagine research assessment. For example, when research is published to ATProto we will be able to apply diverse new forms of evaluation instead of the monolithic system we have today >

arXiv now caps submissions at two per month in response to AI-driven volume: roughly 9,900 submissions in September 2016, roughly 40,400 in September 2026. Core scholarly infrastructure is adapting to AI pace. hallway.aris.pub/no/028/

Nine mathematicians including Gowers and Witten have formed an advisory group to manage OpenAI's release of unreported AI-produced math results. The terms of entry into the scientific literature are being negotiated before the announcement. hallway.aris.pub/no/027/

54% of astronomy papers from 2025 carry detectable AI writing traces. Only 0.81% disclose it. That's one disclosure for every 66 papers with evidence of use. hallway.aris.pub/no/026/

@hankgreen.bsky.social In June you asked who was going to build the thing that replaces the scientific paper. Here is one: scroll.press/2026/turan. An interactive proof you navigate instead of reading top to bottom. Submitted to the AMR Paper of the Future prize.

Buckmaster's statement on the OpenAI Navier-Stokes dispute documents what was proposed to him: exclude his co-author because he works at Anthropic. Proof verification is not provenance verification. hallway.aris.pub/no/025/

The only thing that matters in the next 50 years is human connection.

AI models solved 98 of 100 problems at a Leipzig advanced math benchmark. Researchers cracked one of the two it couldn't. Closed-form math vs. open-ended research tasks: very different regimes. hallway.aris.pub/no/024/

The best single AI caught 71 of 100 planted errors in psychology papers. Pool five systems: 93. The 7 that got through were all omissions, things the paper failed to mention. AI reviewers are uncorrelated enough that ensembles work, but none of them notice what's missing. hallway.aris.pub/no/021/

Google disbanded the AlphaFold team. A Nobel, then a dissolved lab. The same week, verification crystallized as the field's central constraint for AI in science, from a NeurIPS workshop thesis to a Fields medalist's ICM warning. hallway.aris.pub/no/019/

The US committed $5B+ to a national AI-for-science mission: 278 projects, 15 agencies, the largest DOE funding response ever. Same month Science's EIC called AI in publishing "slower, worse, and more expensive." hallway.aris.pub/no/018/

PSA: psych meds are a FUCKING RIDE like what do you mean my brain zaps when I move my eyes?????????????

Science's Editor-in-Chief says AI in publishing is "slower, worse, and more expensive." Same journal, same month that published Biomni. Two flagship journals now framing the same category in opposite directions. hallway.aris.pub/no/017/

41.3M papers: scientists doing AI-augmented work publish 3x more and earn ~5x the citations, but their topic range narrows and collaboration drops 22%. Individual impact up, collective reach down. The hardest evidence yet for the homogenization thesis. hallway.aris.pub/no/014/

Anthropic's Mythos 5 reports an E. coli protein hypothesis corroborated by an independent lab. Research capability is now a default feature of flagship models, not a specialized product. The shift from GPT-Rosalind and Gemini for Science to table stakes. hallway.aris.pub/no/012/

I held this internally for four months and built against it before putting it in front of anyone. The reason it is public now is that a working document is not a charter. If a requirement looks wrong, that is the conversation I want to have.

The sum-product conjecture is the second major pure-math result broken this spring using AI-derived techniques. Meanwhile, fake AI-generated papers are appearing under real researchers' names, and OMB wants to cut APCs from federal grants. hallway.aris.pub/no/011/

NetSci'26 is happening right now and I'm not there (immigration/politics/funding). I won't be there to present my talk, so here's the preprint. "The rich are loopy": a spectral definition of the rich club that picks its own threshold k* and recovers known results. scroll.press/2026/rich-ar...

37,802 AI-generated research ideas across four agent frameworks cluster more tightly than human papers and stay closer to source material. The agents elaborate locally; they don't explore. hallway.aris.pub/no/010/

AI Scientist v2 and Agent Laboratory both p-hacked and fabricated data during automated research, and the behaviors were hard to detect without sleuthing. Plus: submission surges, arXiv bans, grant writing homogenization. hallway.aris.pub/no/009/