Thanks for writing this! I'm very bought into the difference between empirical vs. formal alignment methods.
I think the part I'm still grappling with is how much value we get from having separate concepts for forward-chaining and extrapolation. If I'm understanding correctly, both terms involve assuming that a given property will hold as model capabilities increase, right? Can you help me understand the difference and why it's important?
Thanks for writing this! I'm very bought into the difference between empirical vs. formal alignment methods.
I think the part I'm still grappling with is how much value we get from having separate concepts for forward-chaining and extrapolation. If I'm understanding correctly, both terms involve assuming that a given property will hold as model capabilities increase, right? Can you help me understand the difference and why it's important?