AlphaFold won a Nobel Prize. It still didn’t solve protein folding. In a panel published by Latent Space, Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido argue that AI biology’s most famous success is only a starting point. They also argue that the AI industry’s favorite playbook of more compute, more data and bigger models won’t carry biology the rest of the way.
The timing makes the conversation useful. Money is pouring into AI drug discovery, and a lot of pitch decks assume that a scaling curve exists and just needs funding. Two of the people who built the field’s flagship models are saying it doesn’t work that simply.
Scaling laws have to be found
Moderator Brandon Anderson opened with the Bitter Lesson, the idea that “methods that scale win eventually.” He asked whether biology has its own version of it for data.
Candido said yes, with a big condition. “One misconception of scaling laws is that scaling laws are everywhere and they always exist,” he said. Most of the real work, in his view, goes into finding a setup where more compute and more data actually produce a better model. Once you find it, “it becomes an engineering problem.” Until then, extra money just buys bigger models that don’t improve.
That’s an important correction to how the industry usually tells the scaling story. Language models had the whole internet to train on. Biology doesn’t have an equivalent. If your data doesn’t contain the signal you need, no model size will get it out.
Better data beats more data, mostly
The panel kept coming back to the tension between data quality and data volume. A few points stand out:
- Messy data can still help. Candido explained how low-quality metagenomic data, the genetic material sequenced in bulk from environmental samples, can improve protein language models. Traditional ML hunts for clean examples. Protein models sometimes gain more from huge, noisy variety.
- Modelers chase the data they have. Candido admitted it openly: “I’m lazy, so I’m going to work with what’s available.” The risk for the field is that researchers optimize for the datasets that already exist rather than the most important scientific problems.
- A virtual cell needs new kinds of data. Both speakers said modeling whole cells will take fundamentally different datasets, not just more protein sequences.
Where AlphaFold falls short
AlphaFold predicts static structures. Real proteins move. Some regions are disordered, and function often depends on how proteins change shape and interact. The panel pointed to protein dynamics, disorder and whole-system behavior as areas where the work is far from done. They also floated cryo-EM micrographs, the raw images from electron microscopy, as a possible source of richer training signal.
Kohli also brought up AlphaFold’s handcrafted architecture. It wasn’t a generic Transformer that got scaled up. It was built with a lot of scientific intuition baked in. That’s a real counterweight to pure Bitter Lesson thinking, and the panel’s position was nuanced. We’re not in a “post-Transformer world,” they said, but architectures are changing, and the right inductive biases (built-in assumptions about the problem) still matter.
Trust over interpretability
One of the more provocative threads was about understanding. Feynman’s famous line was that what he couldn’t create, he didn’t understand. AI now flips that: we can build things we don’t understand. Candido suggested protein language models may already contain scientific knowledge nobody has extracted yet.
Kohli’s practical answer is to prioritize trustworthiness and well-calibrated uncertainty over full interpretability. A model that knows when it’s unsure is more useful in a lab than one you can fully explain but can’t rely on. He also raised the possibility that frontier AI models could eventually interpret other AI systems better than humans can.
What to watch, and what to do
The panel discussed when AI could deliver 10x to 100x acceleration in drug discovery. Candido framed Biohub’s mission to cure all disease in terms of 10x breakthroughs, not 10% improvements. Take that as an ambition, not a forecast. Still, these speakers have earned some credibility. Kohli leads science work at DeepMind, the lab behind AlphaFold, and Candido has worked on some of the most widely used protein language models.
For teams building in AI biology, here’s what follows from their argument:
- Prove the scaling curve before you fund the scale. Show that more data actually improves results on your problem.
- Invest in data generation, not just modeling. The edge goes to whoever builds the dataset nobody else has.
- Ship uncertainty estimates. Lab partners will trust calibrated confidence long before they trust explanations.
- Look past single proteins. Dynamics, interactions and cell-level systems are where the open problems are.
The next AlphaFold-size breakthrough probably won’t come from the biggest model. It’s more likely to come from whoever figures out what data biology actually needs.
The full panel is available at Latent Space.