https://leashbio.substack.com/p/making-reality-replayable training an AI to design experiments by premeasuring a vast space of values (eg the binding affinities of a bunch of molecules with a target) and letting the AI learn, with a limited budget, a sequence of experiments to "run" to achieve a goal (like finding the best binder).
this is more efficient than starting with no experiments and sequentially running the ones the agent suggests. the agent will go down a lot of dead ends before it gets good, and it may repeat experiments, which means you have to repeat experiments. better to prerun them all.
they validate that smarter modes do better at this task, spending less time in unproductive regions of chemical-space, which suggests that the models are learning a real skill.