Bryan Hsu

[Yasha Ektefaie @YEktefaie]

Very excited to launch BroadBox alongside our new science sandboxes preprint and especially to share MelanomaBox, which we’ve been building with @jacksonweir4, @SandeepKambham2, @kbryanhsu, @shantanuXsingh, @PardisSabeti , and @insitubiology.

Bryan and I had been developing ways for coding agents such as Claude Code to operate lab robots. Together with Jackson, Sandeep, and Shantanu we connected that automation to a real melanoma drug-combination experiment designed to run on a weekly cadence with minimal human labor.

Bryan and I were also part of the preprint team developing the broader science sandboxes framework. Arya and I saw how that framework and our experimental infrastructure could come together. The efforts converged, and we decided to open these sandboxes to the community as shared challenges for AI scientists. From that synthesis came BroadBox.

If AI is to take part in discovery at all, we must give AI scientists experiments they can actually run, and then measure how evidence forces them to revise or abandon a conjecture. We must also give them a benchmark that is not saturated: nature.

We’re excited to maintain and grow BroadBox across biology. Bring your agent or bring an experiment that should become a sandbox!

https://t.co/UpRZqy99YS

[Arya Rao @AryaRao_]

Today, we introduce science sandboxes, a framework to measure the scientific capability of AI, and BroadBox, our @broadinstitute effort to build them across biology w/ @PardisSabeti @eric_lander + an amazing team 🧵
Preprint: https://t.co/co3HNSikcM
Blog: https://t.co/dY6QFHw5Yt