General Biology Models Need Novel Experimental Data

yusufroohani · x · 2026-07-12

A repost suggests that training a "general biology model" requires a highly diverse range of experiments, many of which do not fit the traditional 96-well plate format. The text further proposes that RL-scale data in biology might come from **new assay designs**: combining automation with multiplexing to generate more **multimodal data**, thereby giving the model enough feedback signals to learn. The cited view is more radical, suggesting that Anthropic or OpenAI should acquire Ginkgo Biosciences because its automated labs could provide the infrastructure for biology-oriented verification / RL.

Original post →

More from Companies & People

Companies & People channel →