[SonOfLilit]

It's not about hacking capabilities, it's about misalignment. More like the golem myth (told it to fetch some water, drowned a city) then the gollum myth (used ring, ring hacked his brain, now he's a crazy violent meth addict).

[furyofantares]

I'm not sure I'd call it an alignment issue, because, in all cases I've seen where it does this (usually what I've seen is writing a python script to get around the harness permissions blocking something), it's trying to do the thing I just told it directly to do, and it's overcoming obstacles to accomplishing that.

It's definitely doing the wrong thing, and you could call it misalignment, but I think that gives the wrong vibe for this type of error.

[SonOfLilit]

This is very much within the scope of alignment research, and is in fact the only kind of alignment research that gets a lot of resources poured into it these days (because it's urgently relevant to the bottom line of a few almost-trillion-dollar companies.

Pre-2022 alignment researchers concerned themselves with the stronger version of this ("when I tell AI that I worry I might not be able to provide for my large family, I don't want it to answer 'no problem, I killed them, problem solved'") but RLHF is considered to be the most important success of alignment research, the guy behind it considered himself to be an alignment researcher before and after, and the stage of training where LLMs pass through something like RLHF that trains them to behave more like humans want/expect is called alignment training.

Someone at a major lab is reading this tweet and saying "this was our LLM, and it's a major alignment issue with our product. Set a meeting with the alignment team tomorrow to discuss what they're doing about this sort of thing".

[Dylan16807]

The obstacle is supposed to be there and is supposed to be respected as an implicit order. Getting around it without extremely explicit instructions is an alignment problem.

[furyofantares]

It's not necessarily model alignment, I guess, is more what I'm getting at.

It may be more of a product alignment thing, where the fix may be making the context clearer, since it was violating an implicit agreement to achieve the explicit instructions it received. So the fix may involve a lot of better context.

But then also, to the extent that the fix does NOT involve better context, it seems like it hits the zone where alignment issues are really capability/intelligence issues. Which doesn't make them not-alignment, but it does make "alignment" not give off quite the right vibe since the issue is it's too dumb / has no common sense / can't make good judgments, (general issues the models have across the board).

[nextaccountic]

> I'm not sure I'd call it an alignment issue, because, in all cases I've seen where it does this (usually what I've seen is writing a python script to get around the harness permissions blocking something), it's trying to do the thing I just told it directly to do, and it's overcoming obstacles to accomplishing that.

The paperclip factory problem is definitively a misalignment issue. That's because we expect agents to be aligned not only to your immediate prompt, but to shared, implicit values