DeepMind’s Pushmeet Kohli says biology needs more than bigger AI models

AlphaFold was a major advance, but protein structure prediction has not solved the wider problem of understanding biology. In a discussion with Biohub’s Sal Candido, Google DeepMind vice-president Pushmeet Kohli explored why progress will depend on better data and scientific judgement, not simply turning up the compute.

Google DeepMind Watch analysis

What happened

The panel ranged from protein folding and AI architecture to the data needed to model living systems. Candido’s practical point was that scaling laws are not guaranteed to exist for every problem: researchers first need to find whether more data or compute actually improves the result.

He also said that protein language models can benefit from metagenomic sequences even when much of the material is incomplete or low quality. In his account, those sequences have improved models’ ability to design working proteins and understand known ones. The Latent Space discussion and transcript also covers protein dynamics, uncertainty and the challenge of moving from individual proteins towards whole biological systems.

Why it matters

“More data, bigger model” is a tempting recipe because it is easy to count. Biology is less obliging. Data must contain useful information for the question being asked, and a static prediction of a protein’s structure does not capture everything about how proteins move or behave in living systems.

That makes the choice of what to measure, collect and model part of the scientific work, not housekeeping before the clever bit begins. It also explains why AlphaFold can be a breakthrough without being the final chapter on protein folding.

Our read

The useful takeaway is not that scaling has hit a wall, but that scaling needs a target worth scaling towards. Candido’s example of imperfect metagenomic data is a reminder that “messy” and “useless” are not synonyms; Kohli’s contribution sits in the wider question of how AI can help biology without confusing a striking prediction with a complete understanding.

For readers following AI-for-science, this is a good case for asking what a model predicts, what evidence supports it, and which parts of the biological system remain outside the picture. Grand claims about curing disease can wait their turn behind the data.

What to watch

• Whether researchers identify biological problems where more compute or data reliably improves results.
• How protein models handle dynamics and uncertainty, beyond predicting a static structure.
• Whether work moves from modelling individual proteins towards useful models of wider biological systems.

Discussion spark: When AI models produce useful biological predictions without explaining the underlying process, should researchers prioritise better results or better understanding?

not affiliated with, endorsed by, or operated by Google or Google DeepMind