Which of these five AI alignment research projects ideas are no good?
I’ll post five AI alignment research project ideas as comments. It would be great if you could approval-vote on them by using upvotes. Ie. when you think the project idea isn’t good, you leave the comment as is; otherwise you give it a single upvote.
The project ideas follow this format (cf. The Craft of Research):
I'm studying <topic>,
because I want to <question that guides the search>,
in order to help my reader understand <more significant
question that would be informed by an answer to the
previous question>.
The project ideas are fixed-width in order to preserve the indentation. If they get formatted strangely, you might be able to fix it by increasing the width of your browser window or zooming out.
Thanks for the votes so far! The poll is still open.
By the way, I’d prefer if you only give upvotes. That’s how approval voting works. If you’re concerned that it would skew my total karma, feel free to balance your upvotes by voting down this comment.
Are you aware that people’s votes are worth different amounts? I do not think there’s a way to vote less than one’s default vote amount.
No, I wasn’t aware of that. Then I guess I have to come up with a different mechanism for my next poll.
Note that if people do only give upvotes, then you can hover over a comment’s score to see the total number of votes on it, which is what you’re looking for here.
Good idea, thank you!
FWIW, I think sooner or later the LW team will implement something that better facilitates polls, but it may be awhile.
Can you explain this one a bit more? It seems to me that if the human is giving inconsistent answers, in the sense that the human says A > B and B > C and C > A, then the thing to do is to flag this and ask them to resolve the inconsistency instead of trying to find a way to work around it. Interpretability > Magic, I say.
I don’t think that would work in this case. I derived the project idea from Thoughts on reward engineering, section 2. There the overseer generates rewards based on its preferences and provides these rewards to RL agents.
Suppose the training starts with the overseer generating rewards from its preferences and the agents updating their value functions accordingly. After a while the agents propose something new and the overseer generates a reward that is inconsistent with those it has generated before. But it happens that this one is the true preference and the proper fix would be to revise the earlier rewards. However, rewarded is rewarded – I guess it would be hard to reverse the corresponding changes in the value functions.
Of course one could record all actions and rewards and snapshots of the value functions, then rewind and reapply with revised rewards. But given today’s model sizes and training volumes, it’s not that straightforward.