No. 029

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 The Association for Human Mathematics called it a demonstration of power rather than scholarship, and three were withdrawn for a sign error the next day.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 other way: $1.8 billion for cell biology measurements and a public genome annotation database for 22,000 species fund shared measurement and leave the questions to whoever uses the data. Carbon-A even releases its predictions as hypotheses to check rather than results to trust.

Open data and annotation

• Open data for predictive AI models of biology: $1.8 billion committed

  Chan Zuckerberg Biohub, October 7 2026

  Partners including Google DeepMind, Meta, Isomorphic Labs, the Department of Energy, and the NIH have committed close to $1.8 billion to generate open experimental data for AI models of cell behavior, on the stated premise that the bottleneck is not model size but the measurements needed to train and evaluate against.

• Carbon-A: Finding genes in known and unknown genomes

  Hugging Face, October 8 2026

  A 1.2 billion parameter model predicts protein-coding regions across 22,617 species from DNA alone, releasing weights under Apache 2.0 and a public annotation database covering 27 trillion base pairs, with the authors stating explicitly that the 566 million predicted loci are hypotheses to be checked rather than confirmed genes and describing what downstream experiments would be needed to confirm them.

Research systems

• The Denario project: Deep knowledge AI agents for scientific discovery

  PRX Intelligence, October 1 2026

  A modular multi-agent system covering the full research workflow, from idea generation and literature review through code, figures, and paper drafting, has been published in a peer-reviewed APS journal with 11 AI-generated draft papers across multiple disciplines as its primary demonstration output, which is a different status from a preprint for a system whose output is itself a set of papers.

• Scientific Agents: Evaluating Profession-Specific System Prompts on Scientific Tasks

  arXiv, September 7 2026

  Testing 503 expert scientist persona prompts against a plain "You are a helpful assistant" baseline across 4,531 questions on nine science benchmarks, the sole author found an average difference of -0.6 percentage points and no statistically clear improvement on any benchmark, while the expert prompts cost 2.2 to 4.5 times as much per API call.

• Paperclip Evidence API

  GXL (Generative Expert Labs), October 8 2026

  Stanford's James Zou and collaborators have released an API that searches 11 million papers and 150 million abstracts to retrieve passages for and against a scientific claim, with researchers in the announcement thread observing that "strongest evidence" in this context means retrieval relevance rather than study quality, which is the question the tool raises most directly.

• Claude Frontier Academy

  Anthropic, October 2 2026

  Anthropic's enterprise training program is primarily about business deployment, but the announcement confirms that "Claude Science," a named research agent with documented functions spanning genomics, structural biology, and single-cell RNA analysis, is being used by Novo Nordisk in drug-discovery workflows.

The shape of inquiry

• Bending the Curve of Discovery: AI in Science Today and Tomorrow

  DeepMind Institute, October 8 2026

  A survey of 637 scientists in the US and UK finds that 48.8 percent have shifted toward safer, incremental research since adopting AI tools (a tendency most concentrated among junior researchers), with 45.7 percent reporting they spend more than a quarter of their AI-saved time auditing AI output, in an essay written by two Google DeepMind executives drawing on product telemetry from 15 million Gemini interactions to argue for public investment in verification infrastructure.

• AHM Statement on OpenAI's October 6 Release of Mathematical Documents

  Association for Human Mathematics, October 7 2026

  The Association for Human Mathematics called OpenAI's release of 719 AI-generated mathematics manuscripts "not a demonstration of scholarship, but a demonstration of power" and urged mathematicians to discontinue work with OpenAI, while three of the manuscripts were withdrawn for a sign error the following day and the repository itself warns that some unformalized results "could have issues."

Research integrity

• RETRACTED: Advanced Nanovaccines for Cancer Immunotherapy

  Springer Nature, June 2025

  Springer Nature retracted this book by a University of Limerick physicist after a sentence identifying the text's author as an AI language model was found in the published pages, the book having cleared Springer's full review and publication process before a reader noticed the line, with the University of Limerick opening a research integrity investigation.

• 'Cerebral torment': IOP retracts nearly 100 conference papers for tortured phrases

  Retraction Watch, October 9 2026

  IOP Publishing retracted 80 papers from three conference proceedings after the Problematic Paper Screener flagged synonyms substituted to evade plagiarism detection, including "cerebral torment" for headache and "savvy lattices" for smart grids, a technique for obscuring AI-assisted writing that leaves a recognizable signal in the vocabulary.