Nobel Laureate John Jumper: AI is Revolutionizing Scientific Discovery

TL;DR

John Jumper describes his journey from physics dropout to leading AlphaFold at DeepMind, arguing that AI for science accelerates discovery by amplifying experimental work. The core insight is that research ideas (not just data or compute) drove a 100x improvement over prior state-of-the-art, and making predictions accessible via an open database triggered widespread adoption and unexpected applica

Key insights

• The research component of machine learning (novel ideas, midscale innovations) is underappreciated compared to data and compute; an ablation study showed AlphaFold2 trained on 1% of data was as accura
• Improvements came from many midscale ideas, not a single technique – for example, equivariance (a popular research area) explained only 2-3 of the 30 GDT-point gain from AlphaFold1 to AlphaFold2.
• External, blind benchmarks (CASP since 1994) are critical because internal benchmarks lead to overfitting; real-world problems are almost always harder than training data.