You are at least saying words here that sounds like you are aware this problem is hard. (the people I’m talking about, so far, tend to not even note that this sort of question has a problem to overcome. Of course, it’s easy for AIs to learn to start adding some Epistemic Caveat sentences).
I’d have to dig into the details to have more of an actual opinion on your thing.
to be honest, the problem seems so hard that i almost feel skeptical that anybody is doing real research at all… the main update i did on reading your quickpost was “woah, you mean some of the amateur research being submitted to lesswrong isn’t slop? how do i learn how to do that?!”
my perspective has always sort of been that the “research” i’m doing is sorta Flamsteed-esque. i’m making lots and lots of observations and documenting them. my base expectation is that any given observation is meaningless or useless or confused. but eventually, some Newton-like person might come along with some actual functional understanding of what’s happening here. someone who could have predicted, in advance, that opus 4.5 would give anti-calibrated confidence values on its self-predictions, and then see my data as confirmatory evidence that they’re on the right track
but even that feels… optimistic. going with the flamsteed astronomy data analogy, i sorta expect to learn, a year or two from now, that my telescope introduces so much noise that all my data is useless
You are at least saying words here that sounds like you are aware this problem is hard. (the people I’m talking about, so far, tend to not even note that this sort of question has a problem to overcome. Of course, it’s easy for AIs to learn to start adding some Epistemic Caveat sentences).
I’d have to dig into the details to have more of an actual opinion on your thing.
to be honest, the problem seems so hard that i almost feel skeptical that anybody is doing real research at all… the main update i did on reading your quickpost was “woah, you mean some of the amateur research being submitted to lesswrong isn’t slop? how do i learn how to do that?!”
my perspective has always sort of been that the “research” i’m doing is sorta Flamsteed-esque. i’m making lots and lots of observations and documenting them. my base expectation is that any given observation is meaningless or useless or confused. but eventually, some Newton-like person might come along with some actual functional understanding of what’s happening here. someone who could have predicted, in advance, that opus 4.5 would give anti-calibrated confidence values on its self-predictions, and then see my data as confirmatory evidence that they’re on the right track
but even that feels… optimistic. going with the flamsteed astronomy data analogy, i sorta expect to learn, a year or two from now, that my telescope introduces so much noise that all my data is useless