Seven Thoughts on "AI Scientists" (corinwagen.github.io)

At the 2025 Progress Conference, attendees expressed unexpected optimism that AI will accelerate scientific progress—particularly via “AI scientists,” i.e., agentic systems or complex probabilistic scientific-reasoning tools that can autonomously propose and sometimes test hypotheses. The author argues this is already happening: venture-backed startups (Lila Sciences, Periodic Labs, Potato, Cusp, Radical AI, Orbital Materials), corporate platforms (Microsoft Discovery, Benchling AI), research teams (OpenAI, FutureHouse) and academic projects (El Agente, MDCrow, ChemGraph) are deploying agentic workflows in drug discovery, materials simulation, and lab automation. Even conservative observers must reckon with production systems that search literature, generate candidates, and run deterministic verification steps. Technically, the key promise is coupling creative, probabilistic hypothesis generation from LLM-like models with deterministic, verifiable testing (virtual screens, automated assays, simulations). That maps to an NP-like pattern: idea generation is hard; verification is often cheap. Even if models are noisier than humans, scale and low per-trial cost (enabled by lab automation) make them valuable; reliability is managed by downstream deterministic checks. Implications: scientists will shift to higher-level design and analysis, complexity and “burden of knowledge” will be mediated by abstraction and tooling, and progress will depend critically on integrating robust verification pipelines and lab automation to turn noisy creativity into real discovery.