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

Designing in-context data or conversation seems to be a new engineering task.