BroadBox Get access

Build a sandbox

We invite scientists across disciplines to help push the frontier of AI’s scientific capability—no computational experience required.

Propose a sandboxOr build one at the hackathon

Start from data you already have

Many sandboxes can be built from existing datasets or assays. Three examples:

Drug response

Can it learn dose-response patterns and predict unseen results?

The agent chooses
A compound and a dose
The sandbox reveals
The measured cell viability, from a plate assay

Protein function

Can it infer which residues and combinations control function?

The agent chooses
Amino-acid mutations
The sandbox reveals
Enzyme-activity measurements, from a mutational scan

Gene regulation

Can it infer regulatory relationships and predict the effects of other knockdowns?

The agent chooses
A gene to knock down
The sandbox reveals
The expression changes, from a perturbation dataset

What a good sandbox needs

It does not have to be solvable by today’s AI. Start with a problem in your research that you wish an AI could investigate.

  1. A meaningful scientific question

    In the drug example: which compounds kill these cells, and at what dose?

  2. Experiments whose selection matters

    Forty-eight possible wells and a budget for a few: which ones the agent picks decides what it can learn.

  3. A way to assess what the agent learned

    Known rules, withheld evidence or new experiments. In the drug example, the wells the agent never saw.

In the terms of the preprint, you choose the specimens the agent may submit, the assay that measures them, and the oracle that applies the assay and reports back. The report can be as rich as every measurement or as small as one number.

Choose where the feedback comes from

Any of the three kinds works. Each trades realism against scale and exactness. More on wet, damp and dry

Wet

Results from
Actual physical experiments.
Offers
The most trustworthy evidence there is.
Costs
Often expensive, time-consuming and noisy.

Damp

Results from
A model trained on empirical data.
Offers
Repeated experimentation becomes scalable.
Costs
It inherits the model’s data biases and assumptions.

Dry

Results from
Invented rules that you specify.
Offers
An exact ground truth.
Costs
Real-world physical complexity.

Why build one?

The most important scientific questions come from scientists who know their fields. A sandbox lets you turn a problem from your own research into a challenge for frontier AI: see how agents investigate it, uncover where they fall short, and help develop the capabilities needed to make real discoveries. As models improve, your sandbox remains a way to measure progress on a question you believe matters. Rather than waiting for AI to transform your field, you can help define what it must learn to do.

Our goal is not to assemble puzzles that current models can already solve. It is to create increasingly ambitious sandboxes that push the limits of scientific capability. As agents improve, we can move from well-defined experimental systems to longer, more complex investigations—and eventually to questions whose answers scientists do not yet know.

Propose a sandboxBroadBox Hackathon, December 1–2