AI community debate resurfaces over claim that the pure of heart would never evaluate a model

agreed and you expressed this far better than i managed when i've tried so far. i like to let the model explore open-endedly for a bit, i give access to a custom tool built by the model for the models, eg to read past conversation history the amount of effort it would take to fake such an environment becomes great enough that no one would do such a thing for an eval's sake