Comment by bananaflag

Yeah it s because they have a strong prior on unsolved problems being unsolvable.

Once the idea of AI routinely solving conjectures enters the training data this encouragement will disappear like 2023-era prompt engineering did.