Hey, liked your article. Very worrying if you are correct.
Noticed a small typo in the sentence (my highlight): Sadly, I have found that people consistently have a tendency *is* to assume that unknowns do not exist.
I think working on capability elicitation is likely net good for this reason, even if it's "directly capabilities relevant". The dangerous thing is the model, especially if we don't know it's full capabilities, and getting better work out of weeker models can help a lot. The main argument against is that this will accelerate AI r&d itself
Hey, liked your article. Very worrying if you are correct.
Noticed a small typo in the sentence (my highlight): Sadly, I have found that people consistently have a tendency *is* to assume that unknowns do not exist.
Are there empirical ways to measure capability and also some sort of exhaustive exploration on what’s possible with any generation model?
Exhaustive exploration is hard because there are many ways to elicit capabilities, not merely prompt engineering.
I think working on capability elicitation is likely net good for this reason, even if it's "directly capabilities relevant". The dangerous thing is the model, especially if we don't know it's full capabilities, and getting better work out of weeker models can help a lot. The main argument against is that this will accelerate AI r&d itself