Video · 2026-06-22 · 53m · 6 moments

He won a Nobel here for AlphaFold. Then he left. - John Jumper

✦ AI generated

GUEST John Jumper GUEST Emmanuel Nee HOST Tim Scarfe

timeline · colored by role

Data◆

AlphaFold turned protein structure determination, which used to take about a year and $100,000 of specialist experimental work per protein, into a prediction that takes 5-10 minutes, and has now been used to predict 200 million protein structures.

Jumper explains the decades-old bottleneck of experimentally determining protein structure and how AlphaFold's deep learning system compressed that into minutes, scaling to 200 million predicted structures.

transcript

John Jumper: So, it takes 5-10 minutes to get the structure of a protein instead of a year. I should at some point figure out what that ratio is in terms of time. But then also, of course, it's incredibly scalable. So, we've predicted the structure of 200 million proteins, basically every protein from an organism whose genome has been sequenced.

explains mechanism · 2gives example · 1

Claim

AlphaFold succeeds because it is deliberately narrow — it predicts the outcome of one specific type of experiment rather than modeling the whole cell, and that narrowness is what makes it reliable.

Jumper stresses AlphaFold's power comes from its humility and narrowness: it predicts one class of experimental result very well rather than claiming to model all of biology.

transcript

John Jumper: I think really important in a certain sense is almost the humility of AlphaFold in that, you know, people say, you know, we are trying to predict what this experiment will give you. We are not trying to tell you everything. We are not a model of the entire cell. We are a predictor of this experiment that you did all the time and took you a year.

Fact

The equivariant geometric attention (IPA) in AlphaFold 2 got most of the public credit for its success, but ablation studies showed it accounted for only about 2.5 of the 30-point accuracy gain over AlphaFold 1 — the real gains came from many smaller ideas stacked together.

Jumper recounts how, despite the field crediting geometric equivariance for AlphaFold 2's breakthrough, ablations showed it contributed only a small fraction of the gain — the real story is dozens of smaller, cumulative engineering wins.

transcript

John Jumper: My favorite review of of AlphaFold 2, we got the reviews back when we submit the paper. And one of them said, "This is six or seven papers worth of ideas." Right? And I think I think that was that was right. There are many many ideas that added up to be a transformative system. And many, you know, to use a baseball analogy, it's not one or two home runs. It's, you know, 18 doubles.

Definition

Predicting, controlling, and understanding are three distinct things, and current machine learning models like AlphaFold give us prediction and some control, but human-communicable understanding still has to be derived separately by people.

Jumper draws a three-way distinction between predicting an outcome, controlling it, and understanding it in a human-communicable way, arguing AlphaFold gives us prediction (and some control) but understanding remains a separate human task.

transcript

John Jumper: So predict means that you say I'm going to do a thing, what am I going to what will be the value of my machine, what will appear on my computer screen in the future? That is predict. Control is I want to measure this thing in the future and I want it to come out 17. Right? That's control. Understand is a lot like predict except there's a human in the loop.

extends · 1

Claim◆

AlphaFold 2 succeeded by rejecting the 'bitter lesson' and instead building in extensive domain-specific, hand-engineered structure, because unlike language models, protein data is finite.

Jumper argues AlphaFold 2 is a counterexample to the 'bitter lesson' — success came from deliberately encoding biological and geometric domain knowledge into the architecture, because protein data (unlike text) is finite.

transcript

John Jumper: I don't really love the bitter lesson as people try and apply it. In fact, AlphaFold 2 is the opposite of that. We did a whole bunch of specialty stuff because our data is not finite. And in fact, now that we've gone to language models, we found our data is still finite. The internet is finite.

gives example · 1

Anecdote

AlphaFold compressed years of failed protein structure determination work into a couple of months by combining one round of experimental purification with computational prediction.

Structural biologist Emmanuel Nee describes personally struggling for four to five years to phase a protein, then, using AlphaFold alongside one round of new experimental data, solving the structure in two to three months.

transcript

Emmanuel Nee: At that time, to phase a protein was like it still was really, really difficult. Um, so I tried several years, close to four, five years, and it wasn't successful. And with AlphaFold, imagine this is more than 10 years ago, with AlphaFold, I went back and did just one um, protein purification, collected the data, and with AlphaFold in combination, I got the structure in less than two, three months.

Highlight slides

From year-long experiments to minutes✦ from: AlphaFold turned protein structure determination, which used to take about a year and $100,000 of specialist experimental work per protein, into a prediction that takes 5-10 minutes, and has now been used to predict 200 million protein structures.Scaling to the entire proteome✦ from: AlphaFold turned protein structure determination, which used to take about a year and $100,000 of specialist experimental work per protein, into a prediction that takes 5-10 minutes, and has now been used to predict 200 million protein structures.AlphaFold 2: A Counterexample to the 'Bitter Lesson'✦ from: AlphaFold 2 succeeded by rejecting the 'bitter lesson' and instead building in extensive domain-specific, hand-engineered structure, because unlike language models, protein data is finite.Why Domain Knowledge Beat Scale✦ from: AlphaFold 2 succeeded by rejecting the 'bitter lesson' and instead building in extensive domain-specific, hand-engineered structure, because unlike language models, protein data is finite.

Related episodes

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→

Video · The TWIML AI Podcast

Teaching Computers to Smell | Alex Wiltschko

→

Video · Syntax

I Built an LLM From Scratch

→