I recently had the same thought. But I need to see activation patterns and benchmarks to be convinced. The thing they seem to be discounting is the training corpus. If the training corpus follows the you pattern, changing the pronoun is probably going to have a negative impact.
If you are aware of particular models which corpus does not follow this pattern I in particular would be keenly interested. As most are distillations of the larger mass today, or have micro specializations, there is not a wide array from which to sample from to validate your arguement that I an aware of. I would agree its possible for words to change the meansongs based on the training. I think that is in essence what its already saying.
I recently had the same thought. But I need to see activation patterns and benchmarks to be convinced. The thing they seem to be discounting is the training corpus. If the training corpus follows the you pattern, changing the pronoun is probably going to have a negative impact.
If you are aware of particular models which corpus does not follow this pattern I in particular would be keenly interested. As most are distillations of the larger mass today, or have micro specializations, there is not a wide array from which to sample from to validate your arguement that I an aware of. I would agree its possible for words to change the meansongs based on the training. I think that is in essence what its already saying.
[flagged]