Jul 15, 2025 · 27m · y-combinator
Nobel Laureate John Jumper: AI is Revolutionizing Scientific Discovery · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Nobel Laureate John Jumper presents at Y Combinator's AI Startup School on how DeepMind's AlphaFold solved a 50-year-old biological challenge in protein structure prediction. He details the machine learning research driving AlphaFold and explains how open-sourcing AI tools accelerates global scientific discovery and medical breakthroughs.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 0.0 Guest teaching 6.0 Guest disagreement 0.8 The partners pushing back 0.0
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 15:15 Pushing back against the equivariance hype
Jumper directly pushes back on the ML community's simplistic narrative that equivariance was the sole secret to AlphaFold's performance, citing ablation data showing it only contributed a few points.
Hardest push from the partners ▶ 14:41 Voiceover interludeIn this solo presentation format, there is no host pushback present in the transcript; this interlude represents the sole non-guest vocal presence.
Biggest teaching moment ▶ 13:00 The 100x value of research over dataJumper educates the audience on the AlQureshi lab's findings, proving that architectural research provided a hundred-fold amplification over pure data scaling.
The partners hold their own ▶ 14:41 Application voiceover transitionBecause the host does not engage or speak during the lecture, there are no instances of host pushback or technical counterarguments.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome and Opening Remarks at Y Combinator | 0 | 5 | 0 | 0 | John Jumper opens his solo presentation by recounting his transition from physics to computational biology and DeepMind. As a keynote monologue without host interaction, host metrics are zero. | |
| Guiding Principles and Presentation Overview | 0 | 6 | 0 | 0 | Jumper provides an accessible biology primer explaining cell complexity, molecular machines, and protein folding mechanics. Host-side scores remain zero due to the lecture format. | |
| The Bottleneck of Experimental Structure Determination | 0 | 7 | 1 | 0 | Jumper details the historical bottleneck of experimental protein crystal determination and synchrotron X-ray diffraction. He contrasts the 200,000 known structures in the PDB against billions of discovered sequences. | |
| Introducing AlphaFold's Predictive Power | 0 | 8 | 1 | 0 | Jumper breaks down the components of AlphaFold's success, demonstrating that research insights yielded a 100x amplification over pure compute or data scaling. Host metrics remain at zero. | |
| Y Combinator Application Interstitial | 0 | 0 | 0 | 0 | A brief Y Combinator application interstitial read by a voiceover. All interaction metrics are zero. | |
| Open Science and the AlphaFold Protein Database | 0 | 7 | 2 | 0 | Jumper gently critiques the AI community's fixation on single architectures like equivariance and emphasizes blind CASP benchmarking. He discusses the social proof that occurred when releasing the 200M prediction database. | |
| Unintended Uses and Targeted Drug Delivery Applications | 0 | 8 | 2 | 0 | Jumper counters the idea that computational predictions simply require redundant experimental validation, showing instead how AlphaFold acted as a hypothesis engine for targeted drug delivery in the Zhang Lab. | |
| AI as an Experimental Amplifier and Conclusion | 0 | 7 | 0 | 0 | Jumper concludes by framing AI as an amplifier for experimental science and poses the key future question of narrow versus generalized scientific foundation models. |
Statements from this episode (15)
Jumper: AlphaFold has received roughly 35,000 academic citations
“There about, I think, 35,000 citations of AlphaFold, but within that is there are tens of thousands of examples of people using our tools to do science that I couldn't do on my own.”
Jumper: About 200,000 protein structures are known, growing by 12,000 annually
“About 200,000 protein structures are known. They pretty regularly increase at about 12,000 a year.”
Jumper: Protein sequence discovery outpaces structure determination by 3,000 times
“So billions of protein sequences are being discovered. About 3000 times faster are we learning about protein sequence than protein structure.”
Jumper: Experimental protein structure determination takes 1-2 years and $100K
“The green is the experimental structure that took someone a year or two of effort. If you want to put an economic value on it, on the order of a 100,000 dollars”
Jumper: AlphaFold 2's final model trained on 128 TPUv3s for two weeks
“The final model itself was a 128 TPU V three cores, roughly equivalent to a GPU per core for two weeks.”
Jumper: The true cost of ML compute is testing failed ideas
“The real cost of compute is the cost of ideas that didn't work. All the things you had to do to get there.”
Jumper: AlphaFold 2 architecture on 1% of data matched AlphaFold 1
“The Al Qureshi lab, did a very, ah, careful experiment, where they took AlphaFold II, the architecture, and they trained it on one percent of the available data. And they could show that alpha fold two trained on one percent of the data was as accurate or more…”
Jumper: Equivariance explains only 2-3 GDT of AlphaFold 2's 30-point gain
“The sixth row there, no IPA, invariant point attention, that removes all the equivariance in alpha fold, and it hurts a bit, but only a bit. Alpha fold itself on this GDT scale that you can see on the left graph, alpha fold two was about 30 GDT better than alp…”
Jumper: Transformative AI systems come from many mid-scale ideas, not one
“It isn't about one idea. It's about many mid-scale ideas that add up to a transformative system, and it's very, very important when you're building these systems to think about what we would call in this context biological relevance.”
Jumper: Protein prediction leads LLMs and ML in blind assessment
“Protein structure prediction is in some ways far ahead of LLMs or the general machine learning space, and having blind assessment.”
Jumper: AlphaFold had one-third the error of any CASP competitor
“And we had about a third of the error of any other group on this assessment.”
Jumper: AlphaFold database scaled from 300,000 to 200 million protein predictions
“One is that we open source the code and we actually open source the code about a week before we released a database of predictions starting originally at 300,000 predictions and later going to two hundred million. Essentially every protein from an organism who…”
Jumper: A two-day open-source hack created the best protein interaction predictor
“The tweet on the left from Yoshitaka Morawaki came out two days after our code was available. We had predicted the structure of individual proteins, but we considered, we were working on building a system that would predict how proteins came together. But, ah,…”
Jumper: AlphaFold made structural biology 5% to 10% faster
“I like to think that our work Made the whole field of what's called structural biology, biology that deals with structures, you know, five or 10% faster. But the amount to which that matters for the world is enormous.”
Jumper: Scientific AI will eventually be driven by broad, general models
“I think we will start to see this on more general systems, be them LLMs or others, That we will find more and more scientific knowledge within them, and we'll use them for important, important purposes, and I think this is really where this is going, and I thi…”