CharlieKey takeaways from your favorite podcasts
← 10 October 2026Latent Space

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

Pushmeet Kohli and Sal Candido argue that progress in biology AI comes from starting with the problem, matching data and modeling to it, and being honest that protein folding is not truly solved.

Watch on YouTube ↗

Charlie's summary

  1. Scaling laws aren't everywhere: much of the work is finding the situation where more data and compute actually pay off.

    Sal says a common misconception is that scaling laws always exist. The real work is finding where adding data or compute gives better results. Once found, it “becomes an engineering problem” and you can “just turn the crank.” That needs data with the right information statistics, since a model can only generalize from what the data contains.

    ▶ 1:30
  2. Modelers lean toward available data, so messy data can help but the community must decide what data is actually needed.

    Sal admits “I'm lazy, so I'm going to work with what's available.” That has an upside: Biohub's protein language model trains on metagenomic sequences, many not even whole proteins, and performance still improves. The downside is scaling only what is easy to generate. Sal thinks modelers and data generators should work openly together to decide what data is needed and then find the scaling law.

    ▶ 3:02
  3. The real bitter lesson is not to be religious about your method: the problem comes first.

    Pushmeet recalls Rich Sutton's DeepMind lecture and reads it as a warning against saying “I'm a modeler” or I'm a data person. For AlphaFold, the PDB was too valuable and too expensive to expand by an order of magnitude, so investing in modeling made sense. For the virtual cell, the data was not there yet. His advice: be multidisciplinary, understand constraints, and fail fast.

    ▶ 5:34
  4. Baking in scientific intuition makes models data-efficient, but inductive bias can hold you back as data grows.

    Pushmeet says AlphaFold 2's hand-crafted design came from biophysics intuition that residues influence one another, which gave the model an unfair advantage and made it more data efficient. Sal adds that with little data you need inductive bias, but with more data a wrong bias can limit the model. Scaling also takes craft: architectures are being modified, not abandoned.

    ▶ 11:40

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.