Descriptions:
John Jumper — who led the DeepMind team behind AlphaFold and shared the 2024 Nobel Prize in Chemistry with Demis Hassabis and David Baker — sits down with Machine Learning Street Talk for a wide-ranging technical conversation about protein structure prediction, what AlphaFold actually solved, and what remains out of reach. Jumper is now leaving Google DeepMind to join Anthropic, making this one of the more significant AI talent moves of 2026.
The technical core of the interview traces the architectural evolution across AlphaFold versions: from the repurposed computer vision CNN of AlphaFold 1, through the custom EvoFormer architecture (axial attention plus SC3 equivariance) of AlphaFold 2, to the diffusion-based AlphaFold 3. Jumper pushes back on reducing these advances to single architectural choices or the “bitter lesson,” arguing that AlphaFold 2’s success came specifically from building domain-specific inductive biases rather than scaling general architectures. The system now predicts structures for over 200 million proteins and is actively used by more than 3 million researchers across 190 countries.
Jumper is candid about what AlphaFold doesn’t do: it is not a model of the cell, cannot handle most multi-protein biological contexts, and was built for a narrow but transformative slice of the problem space. The conversation explores what a true AI-for-biology system would require, including the challenge of identifying the right biological data to predict. The interview also features a structural biologist from Africa describing how AlphaFold compressed months of lab work into days — grounding the abstract achievements in real-world scientific impact.
📺 Source: Machine Learning Street Talk · Published June 22, 2026
🏷️ Format: Interview
