AI capability in research tasks is advancing faster than institutional frameworks for governing it. Google released Deep Review, a public tool that checks manuscripts for mathematical errors, logical gaps and unsupported claims, built on a system already piloted on more than 15,000 conference submissions. TasteVal, a new benchmark for experimental research design, finds frontier models now exceed the human expert baseline, with scores tripling in. Full edition: hallway.aris.pub/no/030/

OpenAI picked 719 mathematics problems, generated manuscripts for them, and released the set on its own terms. The Association for Human Mathematics called it a demonstration of power rather than scholarship, and three were withdrawn for a sign error the next day. A DeepMind survey finds researchers moving the same direction on their own, toward safer and more incremental questions once AI is in the workflow. The week's open data commitments go the. Full edition: hallway.aris.pub/no/029/

arXiv's new submission cap and a post warning that the next generation of mathematics paper mills will produce correct proofs rather than fraudulent ones are responses to the same underlying change: AI makes it cheap to produce scientific outputs that are technically sound and voluminous at the same time. AI-Newton, a system that independently recovers Newton's second law and conservation of energy from raw experimental data with no physical. Full edition: hallway.aris.pub/no/028/

When an AI system produces a scientific result, someone still has to decide whether it holds up, and three of this week's items answer that in different ways. In mathematics, nine researchers including Timothy Gowers and Edward Witten have formed an advisory group at the Institute for Advanced Study to help OpenAI decide how to release a batch of machine-produced results it has not yet published, which comes down to setting the terms on which those. Full edition: hallway.aris.pub/no/027/

A study built on 15 million AI interactions finds scientists saving an average of 7 hours a week, with 74 percent reporting net time gains. A separate study of 207,000 astronomy papers finds 54 percent of 2025 papers carry detectable language-model traces while 0.81 percent disclose it. Together they describe a practice that is already common and a published record that does not document it, which means the literature alone does not tell you how a. Full edition: hallway.aris.pub/no/026/

Two things arrived simultaneously this week: Tristan Buckmaster's documented account of what OpenAI proposed to him in the days before the Navier-Stokes announcement, and a declaration signed by 25 Fields Medalists arguing that AI companies' benchmark-driven approach is "severely misaligned" with the goals of mathematics. They are the same argument at different levels of abstraction. Buckmaster's statement is specific about a call on September 6, a. Full edition: hallway.aris.pub/no/025/

Google released Co-Scientist last month. Since then, Anthropic opened 10,000 subsidized seats for researchers and OpenAI launched a life-sciences workbench with its own domain model. All three frontier labs now sell a scientist-branded product line, making this a product category rather than a series of one-offs. What none of the product pages address is that grammar-only LLM edits shift content and meaning. A team from Princeton and UCSD reported. Full edition: hallway.aris.pub/no/024/

Google claims its Co-Scientist ran full closed-loop autonomous research across three scientific domains. Terminal-Bench Science, released the same week, puts the best agent at a 30 percent pass rate on tasks drawn from real research workflows. Both can be true: narrow, well-defined problems are yielding to autonomous methods, while general scientific competence remains far off. Full edition: hallway.aris.pub/no/023/

Edition 021 flagged the verification gap: AI is producing results faster than anyone can check them. This week, three responses showed up. Tao's Palomar registry gives Lean proofs a submission and review infrastructure. BioNexus grades whether an agent's biology analysis is scientifically valid. Lea keeps the mathematician in the loop during formalization. On the production side, Anthropic's protein-design agent and Inherent's Faraday keep pushing. Full edition: hallway.aris.pub/no/022/

The verification thread keeps tightening. Anthropic's Claude improved a Riemann zeta bound and formalized the proof in Lean. VibeMathed now catalogs 565 AI-solved math problems, 134 of them Lean-verified. And Litvak's planted-error experiment shows that AI peer reviewers are surprisingly uncorrelated. Pool five systems and you catch 93 of 100 errors, but every one of them misses omissions. The pattern across all three: individual AI outputs are. Full edition: hallway.aris.pub/no/021/

Three editions running, AI-in-math results keep landing. This week: OpenAI claims ten open problems solved for under $2,000 of compute each, two independent teams used GPT-5.6 Sol Ultra to crack the same quantum cryptography problem hours apart, and a Fields medalist left his university post for OpenAI's safety team. Meanwhile, a Princeton-led study gave frontier agents $3,000 and six days to answer open research questions; the agents completed all. Full edition: hallway.aris.pub/no/020/

This week, verification crystallized as the central constraint for AI in science. Google Research shipped a prototype chaining every AI-generated claim to its evidence. A NeurIPS workshop adopted "the bottleneck is verification" as its thesis. A benchmark for comparing human and AI manuscript reviews landed on bioRxiv. And at the International Congress of Mathematicians, Terence Tao framed AI as a foundational crisis for the discipline, comparable. Full edition: hallway.aris.pub/no/019/

The same month that Science's Editor-in-Chief called AI in publishing "slower, worse, and more expensive", the White House committed $5B+ to a national AI-for-science mission and ARIA backed a directory of 123 problems for AI researchers to solve. Meanwhile, a PNNL team argues the entire "AI scientist" wave has been scaling individual reasoners when the actual bottleneck is team coordination. Institutions are not arguing about whether AI belongs in. Full edition: hallway.aris.pub/no/018/

Science publishes Biomni and its Editor-in-Chief publishes an editorial calling AI in publishing "slower, worse, and more expensive" in the same window. Nature publishes Robin and a news story about data poisoning of the autonomous agents it just endorsed. The infrastructure layers keep shipping (Elicit's MCP, the Vancouver Standard consultation) but the institutional voices are now openly contradicting each other about whether any of this is. Full edition: hallway.aris.pub/no/017/

Two weeks ago the story was product launches: Claude Science, Google's Paper Assistant Tool, a DIY Claude skill. This week it is peer-reviewed publications and editorial recognition. Science published Biomni. Nature published Co-Scientist. Nature News published a survey treating the whole cluster as a settled category. The gap between "vendor demo" and "institutional endorsement" closed fast. Meanwhile the open-source alternative shipped and the. Full edition: hallway.aris.pub/no/016/