DeepMind Paper: Scientific Discovery Is Bottlenecked by Resources, Not AI Cognition
mikeflache · x · 2026-10-14
A new arXiv paper by Nenad Tomasev, Simon Osindero and colleagues at Google DeepMind argues that AI for Science has focused too narrowly on cognitive capability, while real discovery is bottlenecked by scarce, expensive physical resources like lab equipment and clinical trials. The authors propose an infrastructure of "scientific agent economies" — formal market designs and decentralised credit assignment that let mixed networks of human researchers and autonomous AI agents trade and allocate restricted resources efficiently. The framework also covers priority setting, accountability, security risks, and governance of (semi-)autonomous scientific discovery.
Related event: DeepMind Papers Propose Agentic Economies for Science(3 posts)→
More from AGI Musings
- Theoretical physics critique: LLMs are 'thermodynamic memory machines' blind to conservation laws — AryHHAry · 2026-10-23
- Open protocols didn't free email users — why federated AI assistants face the same lock-in — sujingshen · 2026-10-23
- kenbwork: routine CRO work will offshore, and AI-driven robotics will create a new kind of CRO — kenbwork · 2026-10-23
- SF threw a funeral for Claude 3 Sonnet — an essay on AI and attention — generativist · 2026-10-23
- Beff Jezos: Orgs Must Apply Frontier Intelligence to Harden Against Adversarial Attacks or Perish — mimi10v3 · 2026-10-23
- AI Detectors Should Be Illegal: They're a Scam Causing Real Harm, Researcher Argues — Angaisb_ · 2026-10-23