Petrichor Bio

AI models have learned to reason well in domains like code and math because those fields live in text that can be self-checked, but biology happens inside cells and most experimental detail is never written down, only summarized as a figure in a results section. As models become capable of deciding what experiments to run in biology, they still have no way to physically run those experiments. The text argues this will make experiments, not ideas, the bottleneck to progress once thinking becomes cheap, since AI labs are spending heavily to make reasoning cheap but testing an idea still costs as much as ever.

Petrichor describes itself as a wet lab built specifically for AI models rather than human researchers, designed to run far more experiments in parallel, record much more detail about each one, and stream results back while a run is still in progress so a model can change its approach mid-experiment. It draws an analogy to the chip industry's split between fabless design firms and foundries, positioning itself as the foundry for AI labs that will design experiments but need someone else to run them. The site states that whatever an experiment produces belongs to the commissioning party, and that Petrichor does not share or train on client data.

The text argues that AI labs are investing heavily to make reasoning cheap, which will shift the bottleneck in biology from generating ideas to testing them, making experimental capacity at scale newly valuable.