3 ms·
> many tasks are covered in the training data; but even small perturbations within a covered class of task result in outright failure or reward hacking. Is thi
by highfrequency 18d ago
> many tasks are covered in the training data; but even small perturbations within a covered class of task result in outright failure or reward hacking.
Is this true? Could you cite an example of a simple prompt that GPT 6 / Fable 5.1 consistently bungle?