Chai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem

In 2020, AlphaFold2 proved that deep learning could solve protein folding. But for Josh and Matt, co-founders of Chai Discovery, that breakthrough was just the beginning of a harder question: if you can predict what a protein looks like, can you design one from scratch that does exactly what you want? Their answer is a company built on the "bitter lesson" — the idea that general scaling of data, compute, and model size beats bespoke biological tricks. And it is working faster than they expected.

The Bitter Lesson Applied to Biology

Chai Discovery’s core philosophy is borrowed directly from the playbook that worked for large language models. "We're a very bitter lesson pill company," Josh says. "We really believe in scaling data, scaling models, scaling compute." This means resisting the temptation to add specialized submodules for every new biological problem. Their first model, Chai 1, had 23 distinct submodules — and that was already too many. "When you're trying to iterate on something like that, it gets really hard," Matt explains. "You need to understand each submodule independently, all their behaviors and dynamics. That doesn't actually scale that well." The goal is to identify the few core scaling directions that matter and let the model learn the rest from data, the same way GPT learned to speak multiple languages without separate modules for each one.