[kitsuac]

It isn't /fundamentally/ conservative, it is just typically programmed to choose the most conservative (highest probability) predictions. You could integrate a liberal aspect by fuzzing the decision process to choose from lower probability predictions.

More creativity, and ability to escape local minima, but at some cost when dealing with 'typical' cases and when making particularly damaging mispredictions.

[l0b0]

I think the point is rather that you can't get a more useful prediction by choosing a lower probability description unless you have AGI. Only an AGI could tell that you're not in the mood for "Hey" to be followed by "darling", and only a superhuman AGI could realistically compensate for human bias in data sets.

[kitsuac]

Without AGI there are still cases when the lower probability prediction will be better, and will lead to escaping a local minima. I'd argue that the potential benefits of calibrating that axis dynamically exist with or without AGI.

[ben_w]

Are you describing the explore/exploit tradeoff or simulated annealing in this case?

[goto11]

Fuzzing will just give you more random predictions, it doesn't remove any inherent biases in the training data.