The Hallway Track

No. 014

Science's AI productivity paradox, Sakana clears peer review, AI proof search on open Erdős problems

This week the AI-as-research-agent thread stopped being speculative. Sakana cleared the first round of peer review at an ML workshop, a DeepMind-Lean agent resolved nine open Erdős problems, and senior cosmologists proposed treating AI as an epistemic actor rather than an instrument. The hard question is no longer capability. Hao et al.'s analysis of 41 million papers makes it institutional: do productivity gains come with a contraction of what science collectively explores, and where do humans stay in the decision?

The Empirical Paradox

  • Artificial intelligence tools expand scientists' impact but contract science's focus

    Nature 649, 1237, 2026

    Across 41.3 million papers, scientists doing AI-augmented work publish 3.02x more and earn 4.84x more citations than non-users, but their topic range narrows 5% and their collaboration drops 22%, the hardest evidence yet that AI's individual productivity gains come paired with a contraction of what science collectively explores as work shifts toward the areas richest in data.

  • Will AI spark a scientific renaissance, or a diffuse monoculture?

    Nature 654, 843, June 22 2026

    The editorial pairing to Hao et al. argues the outcome turns less on model capability than on whether researchers, reviewers, and funders actually choose to reward originality over speed.

AI as Epistemic Actor

  • AI Scientists as Engines of Discovery: A Case for Development within Reformed Institutions

    arXiv (cosmology + AI), June 22 2026

    A heavy cosmology lineup (Spergel, Verde, Villaescusa-Navarro, Wandelt) argues multi-agent AI is a qualitative rather than incremental shift toward AI as "an epistemic actor rather than a mere instrument," demonstrates it with a prototype framework called Denario, and calls for institutional reform around verification, accountability, interpretability, and dual-use safety.

  • Towards End-to-End Automation of AI Research

    arXiv (Sakana AI), June 2026

    Sakana's latest AI Scientist generates ideas, code, experiments, manuscripts, and its own peer review, and one of its papers cleared the first round of review at an ML conference workshop, a result the authors themselves frame as a risk of "taxing overwhelmed review systems and adding noise to scientific literature."

  • Advancing Mathematics Research with AI-Driven Formal Proof Search

    arXiv (DeepMind + Lean community), June 8 2026

    The first large-scale evaluation of LLM-plus-Lean proof-search agents on open problems: the strongest agent autonomously resolved 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures at a per-problem cost of a few hundred dollars, with correctness anchored in Lean's type-theoretic kernel rather than in human peer review.

Capability and Oversight