“The ultimate evaluation of an AI's scientific ability requires a high-throughput feedback loop where hypotheses are tested in a wet lab and the results are used to train and evaluate the models.”