Given the goal of mitigating self-proliferation risks, have you observed a decrease in the model's ability to do things like help a user setup a local LLM with local or cloud software?
How much is pre-training dataset changes, how much is tuning?
How do you think about this problem, how do you solve it?
How much is pre-training dataset changes, how much is tuning?
How do you think about this problem, how do you solve it?
Seems tricky to me.