No. 033

Seven essays and reports from mathematics arrived this week on what the profession should do as AI-generated proofs and manuscripts accumulate faster than humans can read them. The empirical benchmarks make a parallel case. BenchBench-Cloning finds that frontier models assemble individual molecular biology steps correctly but fail to combine them into valid protocols. Thirty agents asked to replicate one bacterial RNA-seq study produced eighteen incompatible pipelines. Heidari and Mohseni state what both sets of findings point toward: a research task can be delegated to AI only if the community retains the ability to explain and contest the results, and if accountable humans decide which questions matter.

Autonomous Research

• AI agents enter the wet lab

  Nature (James Zou), November 4 2026

  Wu et al.'s AutoDNA runs DNA experiments across more than 20 instrument types and over 11,000 steps with humans only restocking reagents, matched expert-optimized synthesis yields, and completed protocols 3.2 times faster by parallelizing steps; Zou's commentary addresses what autonomous lab systems will need to be reliable enough to trust.

• Agentic cheminformatics with Google Antigravity

  Google DeepMind, November 3 2026

  The language model plans the workflow and reasons about results while RDKit and connections to ChEMBL and PubChem do the molecular calculations; the walkthrough works through a known case (crizotinib to lorlatinib), DeepMind claims an hour against several days by hand, and none of the proposed molecules has been synthesized or tested.

• BioAnalyst: A Foundation Model for Biodiversity

  Ecological Informatics, 2026

  Trained on 10 types of data covering 20 years of European biodiversity at 0.25 degree resolution, BioAnalyst is released with its code, weights, BioCube dataset, and fine-tuning workflows public, as an example of an open domain foundation model built outside machine learning.

• OpenResearch: autoresearch for arXiv papers

  alphaXiv, October 2026

  Changing "arxiv" to "autoarxiv" in any paper URL launches an agent that sets up the environment and reruns the experiments on user-chosen compute, logging inputs, run charts, and artifacts in one place; the site reports no data on how often reproduced results match the paper's numbers.

Benchmarks, Evaluation, and Reproducibility

• BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development

  arXiv cs.LG, October 2026

  Sixty-five tasks built from real experimental records test protocol development and construct design; the top model (GPT-6 Astra) solves 30.6 percent on one attempt and 72 percent on ten attempts for protocol tasks, but models consistently assemble individual steps correctly while failing to combine them into valid end-to-end protocols.

• Verifiable Agentic Science on Public Infrastructure

  bioRxiv, October 26 2026

  Thirty frontier agents asked to replicate one published bacterial RNA-seq study produced 18 incompatible pipelines; running agents through public infrastructure (UseGalaxy.org, ACCESS-CI) captures every input, parameter, and output, and audit of 48 published analyses found one result had been computed on a 1,000-cell test file instead of the cited 5,613-cell atlas.

• Terminal-Bench-Science: call for contributions to release 2.0

  Terminal-Bench-Science, November 2026

  Domain scientists can submit tasks drawn from their own research workflows for the 2.0 release (deadline December 15, 2026) and become co-authors; version 1.0 results are not yet published.

• EMNLP 2026 AI reviewing experiment

  EMNLP 2026

  EMNLP ran an opt-in experiment in which authors received AI-generated reviews of their submissions and rated them; reports from the conference indicate the reviews were rated well for summarizing contributions and strengths, and poorly for identifying weaknesses, missing related work, and catching technical errors.

Research Infrastructure

• ResearchPigeon: a daily arXiv digest picked by embedding similarity

  researchpigeon.com, November 2026

  One email per day, one arXiv paper, chosen by comparing an embedding of the reader's written interest profile with embeddings of new submissions; likes and dislikes adjust the profile, and arXiv's AI category now receives nearly 200 papers a day, making reader-side filtering a practical necessity.

• Talk with my publications

  Niels Mede (Wageningen University), November 2026

  An assistant professor of science communication built a chat interface over their 50 publications, including live analysis of 26 datasets; it uses Mistral Large on EU servers, answers cite page numbers linked to the PDF, and the page notes that analyses have not been reviewed by the author.

• The World Is Something We Make

  arg min (Ben Recht), October 26 2026

  Recht argues that "frictionless reproducibility" (shared tasks, code, and benchmarks) drove past AI progress, that the field abandoned this with large language models, and that compute pooling could produce a fully open model (weights, code, data, and process) whose training, fine-tuning, and audit the community could then improve.

Mathematics and Professional Culture

• Why and How to Do Science When AI Is Better at It

  arXiv cs.AI, October 30 2026

  Heidari and Mohseni argue that delegating a research task to AI is justified only when the human community can still explain, defend, and contest the results, and when accountable humans decide which questions matter and what findings mean, applying this two-part test to every CRediT contributor role including peer review.

• The mathematical sciences in an AI-enabled world

  Academy for the Mathematical Sciences, October 28 2026

  arXiv mathematics submissions ran 75 percent above a year earlier in September, arXiv has since capped submissions at two per submitter per month, and leading probability journals have suspended submissions; the report recommends assessing correctness, significance, and understanding separately and providing a dissemination route for results whose proofs are not yet human-digestible.

• Summit on Math Research in the Age of AI: report

  Harvard CMSA, October 22-23 2026

  Thirty-two mathematicians produced recommendations on credit and authorship for AI-assisted work, Lean formalization as a supplement to the literature, training early-career mathematicians, and new publication formats that separate dissemination from validation.

• Mathematicians race to simplify AI proofs

  Quanta (Konstantin Kakaes), November 5 2026

  Following OpenAI's October proof release, mathematicians are iterating toward simpler versions of the Landau-Siegel zeros result, with Ben Green of Oxford producing reformulations through AI conversations and Kevin Barreto arriving at a three-page version; the race is for the shortest, most elementary proof rather than for priority.

• Mathematics has changed. Let's act like it

  Terry Tao's blog (Frank Calegari), November 1 2026

  Calegari argues that journals should "demand much more from a paper in terms of writing, ideas, and exposition," that seminars "should rarely be focused entirely on new work," and that the profession should commission consolidation projects on the scale of the revised classification of finite simple groups.

• Mathematics of the Future

  Pavel Etingof and Slava Gerovitch (MIT), October 5 2026

  A 22-page essay arguing for a shift from an "ownership culture" of priority and publication counts toward "intellectual agency": hiring based on talks, oral exams, and collaborative work; a "big cleanup" of the literature using AI and Lean; and publication formats that separate the technical record from explanatory articles.

• Egoless mathematics in the AI era

  Terry Tao's blog (Karthik C. S., Amit Sahai, Jie Xue), November 2 2026

  The essay argues that anxiety about AI exposes an older problem: when worth is measured by publications and awards, mathematicians become producers of output rather than seekers of understanding, and that "mathematical statements may be formally correct without being enlightening."

• AI, Mathematics, Education, and a public goods problem

  Terry Tao's blog (Gireeja Ranade, UC Berkeley), October 29 2026

  Ranade argues that AI weakens the incentives that produce trained scientists: students report losing motivation, fewer undergraduates apply to PhD programs, and AI-assisted work creates a habit of signing off on results nobody fully understands, a problem that formal proof verification does not solve because it moves the question to whether theorem statements are faithful.