Jumper: The true cost of ML compute is testing failed ideas
John Jumper · Nobel Laureate John Jumper: AI is Revolutionizing Scientific Discovery · Y Combinator · Jul 15, 2025 · at 12:07
This episode carries John Jumper's own address, with nobody on the show putting questions to them. It still counts as said, and it is kept out of every score on their page.
Nobel Laureate and DeepMind researcher John Jumper reflects on compute allocation during ML research and model development.
0:00 / 0:06exact quote · 6.5s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“The real cost of compute is the cost of ideas that didn't work. All the things you had to do to get there.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from John Jumper
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: 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…”
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 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: 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…”