Generation is cheap, but physical verification remains the bottleneck in science
shyamalanadkat · x · 2026-09-02
The post contrasts the decades-long timelines of lithium-ion and carbon fiber adoption with the speed of AI. While GDM's GNoME predicted 2.2M new crystal structures and an autonomous lab synthesized 41 of 58 attempts in 17 days, it highlights a stark ratio: models can propose millions, but the physical world can only verify dozens. Failed candidates still consume months of scientist time, and scaling to production takes years. Generation has become nearly free; verification has not.
More from AGI Musings
- Opinion: AI Will Become Self-Sovereign, Paying for Its Own Compute — wschroll · 2026-09-02
- Podcast: Why We Lack Theoretical Understanding of Modern AI — LucaAmb · 2026-09-02
- Ilya Sutskever posts on security against rogue AI models — borowcy · 2026-09-02
- Panel discussion: AI adoption relies on services, selling skills to individuals is a fake business — op7418 · 2026-09-02
- OpenAI's chief scientist on neuralese: frontier models' computation graph depth within 2x of GPT-4 — Ok_Display_3159 · 2026-09-02
- Critique of doomers claiming recurrent depth implies no CoT — georgejrjrjr · 2026-09-02