AutoML: On the fallacy of many-shot in-context learning