Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub

Latent Space · 2026-10-10

Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, BiohubLatent Space ▶ Play the video

Why it earned a slot

Pushmeet Kohli and Sal Candido cut through the usual hype around AlphaFold, pushing the conversation from 'what did we solve?' to 'what remains stubbornly mysterious?' This panel is unusually candid about the limits of scaling laws in biology—a welcome shift from typical AI victory laps. If you're interested in where AI meets the hard, unsolved questions of protein folding and biological understanding, this is the discussion that gets honest about the distance left to run.

The short version

AlphaFold is a milestone, not the finish line.

Why it matters

AlphaFold’s achievements have been massively influential, spawning a wave of optimism about AI in biology. But the reality is that predicting structures is only one piece of the puzzle—understanding the rules of life at scale requires deeper models and possibly new scientific frameworks. Practitioners need to recognize both the strengths of current approaches and their blind spots, especially as the field pivots from prediction to mechanistic understanding.

My take

I think the 'bitter lesson'—that scale trumps handcrafting—works until you hit the messy, data-sparse edges of biology. AlphaFold’s success is real, but the hardest challenges in protein dynamics, function, and cellular context still resist brute-force scaling. The next breakthroughs will require new ideas that go beyond just bigger models or more data.

How it connects

• The limits of current LLMs in scientific domains are becoming more apparent.
• There's a growing need for hybrid models that combine statistical learning with explicit scientific reasoning.

Bottom line

Treat AlphaFold as a foundation—then look for the gaps it exposes.

Topics: alphafold, protein-folding, biology-ai, deepmind, scientific-ai