It’s too late to record a baseline, so any metrics will need to work retrospectively anyway.
There’s still a bunch of us that refuse to use it.
See Goodhart’s law: “When a measure becomes a target, it ceases to be a good measure”.
If LLMs are judged by some specific metric meant to track their negative impact, they’ll adapt to make the metric look good without improving the underlying impact.
Great example of Goodhart:
I used to work at a Verizon store, and we were rated by the amount of accessories we could sell with each phone. In order to boost that number, everyone tried to add on a bunch of cheap phone accessories (clearance items and screen protectors) to drive that number up.
Our goal was more than three, but once we figured out how to sell 5 or more for less than $20, they changed it to total cost of accessories per phone.
