Tamsin Leake

I'm Tamsin Leake, co-founder and head of research at Orthogonal, doing agent foundations.

Wiki Contributions

Comments

Hence, the policy should have an escape clause: You should feel free to talk about the potential exfohazard if your knowledge of it isn't exclusively caused by other alignment researchers telling you of it. That is, if you already knew of the potential exfohazard, or if your own research later led you to discover it.

In an ideal world, it's good to relax this clause in some way, from a binary to a spectrum. For example: if someone tells me of a hazard that I'm confident I would've discovered one my own one week later, then they only get to dictate me not-sharing-it for a week. "Knowing" isn't a strict binary; anyone can rederive anything with enough time (maybe) — it's just a question of how long it would've taken me to find it if they didn't tell me. This can even include someone bringing my attention to something I already knew, but to which I wouldn't as quickly have thought to pay attention if they didn't bring attention to it.

In the non-ideal world we inhabit, however, it's unclear how fraught it is to use such considerations.

My current belief is that you do make some update upon observing existing, you just don't update as much as if we were somehow able to survive and observe unaligned AI taking over. I do agree that the no update at all because you can't see the counterfactual is wrong, but anthropics is still somewhat filtering your evidence; you should update less.

(I don't have my full reasoning for {why I came to this conclusion} fully loaded rn, but I could probably do so if needed. Also, I only skimmed your post, sorry. I have a post on updating under anthropics with actual math I'm working on, but unsure when I'll get around to finishing it.)

Due to my timelines being this short, I'm hopeful that convincing just "the current crop of major-AI-Lab CEOs" might actually be enough to buy us the bulk of time that something like this could buy.

commenting on this post because it's the latest in the sequence; i disagree with the premises of the whole sequence. (EDIT: whoops, the sequence posts in fact discuss those premises so i probably should've commented on those. ohwell.)

the actual, endorsed, axiomatic (aka terminal aka intrinsic) values we have are ones we don't want to change, ones we don't want to be lost or modified over time. what you call "value change" is change in instrumental values.

i agree that, for example, our preferences about how to organize the society we live in should change over time. but that simply means that our preference about society aren't terminal values, and our terminal values on this topic are meta-values about how other (non-terminal) values should change.

these meta-values, and other terminal values, are values that we should not want changed or lost over time.

in actuality, people aren't coherent agents enough to have immutable terminal values; they have value drift and confusion about values and they don't distinguish (or don't distinguish well) between terminal and instrumental values in their mind.

but we should want to figure out what our axiomatic values are, and for those to not be changed at all. and everything else being instrumental to that, we do not have to figure out alignment with regards to instrumental values, only axiomatic values.

one solution to this problem is to simply never use that capability (running expensive computations) at all, or to not use it before the iterated counterfactual researchers have developed proofs that any expensive computation they run is safe, or before they have very slowly and carefully built dath-ilan-style corrigible aligned AGI.

nothing fundamentally, the user has to be careful what computation they invoke.

an approximate illustration of QACI: