Post-training makes AI models less human-like and less predictable

Original: Been thinking about how heavier, more use case driven post training potentially puts the model “on rails”.

Source: x.com ↗