No. 008

AI co-mathematician, compiler-verified protocols, open infrastructure bets

Two infrastructure bets landed in the same week: $152M for open AI compute (NSF OMAI) and $20M to sustain the open-source libraries that compute depends on (OS4Science). Together they signal that funders now treat reproducible AI infrastructure as a first-class research output, not a byproduct. Meanwhile, Google's co-mathematician workbench and a compiler-verified protocol system both point in the same direction: AI that maintains state across a research session, not just answers a question.

Research Practice

  • AI co-mathematician: Accelerating mathematicians with agentic AI

    arXiv, May 7 2026

    Google DeepMind's Gemini-based workbench gives mathematicians a stateful workspace where multiple agents run parallel workstreams, track hypotheses, search literature, and attempt proofs, scoring 48% on FrontierMath Tier 4 and helping an Oxford researcher close an open problem from the Kourovka Notebook.

  • Towards autonomous biology: Compiler-verified protocols as a foundation for real-world AI execution

    bioRxiv, May 5 2026

    BPL-COGEN pairs a fine-tuned 30B language model with a deterministic compiler in a generate-validate-repair loop, achieving 95.1% fidelity when converting 300 published Nature Protocols papers into executable protocol code, a step toward AI systems that can reliably run (not just plan) wet-lab experiments.

Infrastructure

  • Open by design: Ai2 brings fully open AI infrastructure online with NSF OMAI

    Allen Institute for AI, May 7 2026

    The $152M NSF/NVIDIA-funded Open Multimodal AI Infrastructure is now live, giving academic researchers access to model weights, training data, and methods rather than black-box APIs, with an explicit design principle that exploratory training runs become public artifacts.

  • Open Source for Science Fund launches to power AI-driven discovery

    Renaissance Philanthropy, May 4 2026

    A new $20M multi-donor fund seeded by Biohub and Wellcome will sustain the open-source software stack underpinning scientific research, with grants up to $1M for foundational libraries, picking up where CZI's Essential Open Source Software program left off.

Publishing & Integrity

  • New IOP Publishing tool detects duplicate peer reviews in push against reviewer fraud

    IOP Publishing, May 5 2026

    The Duplicate Review Checker uses semantic similarity and text-overlap algorithms to flag recycled reviews; since its 2024 pilot it has processed half a million reports and identified nearly 2,500 cases with over 60% content overlap, including review-mill operations.

  • Restoring trust in science: What would make a difference?

    The Scholarly Kitchen, May 12 2026

    An argument that trust in science will not be restored by better messaging alone but requires structural changes to incentives, more disciplined public communication, and genuine engagement with publics who have disengaged.

  • Love, Death & Robots: Scholarly Edition

    The Scholarly Kitchen, May 7 2026

    A framework for classifying bot traffic on scholarly platforms by intent rather than identity, arguing that publishers need to distinguish AI training crawlers from legitimate discovery tools before blanket blocking undermines the access infrastructure researchers depend on.