Codex' goal feature is awesome. Today I asked it to do hyperparam search + dataset ablations for training a fasttext model until it has precision = 0.75 on the test set, and, after trying for an hour, it finally achieved it by training on the test set.