A Prediction Without Uncertainty Is Only Half a Prediction in AI for Science
bravo_abad · x · 2026-09-13
Using a concrete example, the author argues that in AI for science, two models both predicting an adsorption energy of −1.2 eV are not scientifically equivalent if one reports −1.2 ± 0.05 eV: the uncertainty quantifies what the model knows and doesn't know, guiding whether to trust the prediction, remeasure, or explore elsewhere.
Key points:
- Gaussian processes are valuable because their uncertainty grows where data are scarce and shrinks where observations constrain the model.
- In active learning and autonomous experimentation, that uncertainty becomes actionable — it can decide the next experiment.
- Core message: accuracy tells you what the model predicts; uncertainty tells you how much weight to give it.
The post promotes the author's book Inteligencia artificial para estudiantes de ciencias (Ediciones Pirámide).
More from AGI Musings
- AI Existential Risk Is Shifting From Far Mode to Near Mode, Commenter Argues — EigenGender · 2026-09-13
- AI existential risk is shifting from far mode to near mode in people's minds — EigenGender · 2026-09-13
- Frontier lab staff privately probe raw models with 'evil' questions and self-psyop into fear — tawnniee · 2026-09-13
- Tyler Cowen: Mathematicians must accept being handmaidens to AI, and it will outwrite me too — sebkrier · 2026-09-13
- Slowdown does more harm than good: we should race toward the age of abundance — kimmonismus · 2026-09-13
- Early OpenAI employee says 'winning' AGI is outdated — 99% of future compute will run locally — GregCook2011 · 2026-09-13