MIT professor's ten predictions for AI for science, from executable papers to compiled matter
ProfBuehlerMIT · x · 2026-10-14
MIT professor Markus Buehler revisits his November 2022 perspective (submitted three weeks before ChatGPT) arguing AI would become a new language of science, and lays out ten predictions for where AI for science goes next.
Key predictions: every scientist gets a team of superintelligent agents that read literature, simulate and design experiments; instruments on demand (a sketch becomes a validated simulator in an afternoon); executable papers where every claim ships with code, data and a verifier; machine civilizations of invention that evolve at machine speed; a physics of discovery engineering breakthrough rates; "breaker" agents that hunt disconfirming data out-discovering polishers; models that learn during the question; debugging a model's physics like code; compiled matter with agents closing the design-synthesize-test loop in days; and cross-domain discovery linking proteins, music and metamaterials.
More from AGI Musings
- "Programming is liberty": dev pushes back on AI labs deciding what software may exist — jbarbier · 2026-10-14
- Nuclear bombs are a misleading analogy for superintelligence, yet fitting — akbirthko · 2026-10-14
- Miles Brundage: stopping AI weight theft should be the top AI policy priority — Miles_Brundage · 2026-10-14
- Critique: Terence Tao's Math 2.0 vision contains two fundamental contradictions — layer07_yuxi · 2026-10-14
- Blogger finds two fundamental contradictions in Terence Tao's Math 2.0 AI vision — layer07_yuxi · 2026-10-14
- PauseAI US founder argues Eliezer Yudkowsky is undermining his own AI safety movement — mimi10v3 · 2026-10-14