What's this about?
On Latent Space, Xaira researchers Bo Wang and Ci Chu explain why their 3.1-billion-parameter "cell model" for drug discovery hit an information ceiling: test performance stopped improving after 1.5B parameters because the *dataset* ran out of information — not the model out of capacity.
What exactly is a "foundation model for cells"?
Like a language model trained on text, a cell model is trained on huge amounts of biological measurement data to learn how cells behave, aiming to predict how they'll respond to a drug. The episode's argument: unlike text, biology's data is scarce and often lacks *causal* signal, so scaling parameters alone stops helping — you need better data, not a bigger model.
Why does it matter?
It's a pointed counter to "just scale it" in biology: the binding constraint for bio foundation models is the information content of the data, which reshapes where drug-discovery AI money and effort should go.