Will Buck think running the Redwood Research Mechanistic Interpretability Experiment (REMIX) was +EV on March 1st 2023?
resolved Mar 3

See https://www.lesswrong.com/posts/nqwzrpkPvviLHWXaE/apply-to-the-redwood-research-mechanistic-interpretability
Context: I work at Redwood and Buck is the CTO.

REMIX starts on 12/19/2022.

Resolves to N/A if REMIX is canceled (very unlikely as it starts on Monday).

Edit: ex-post EV analysis

Get Ṁ600 play money

🏅 Top traders

#NameTotal profit
Sort by:
bought Ṁ20 of YES

Is this relative to the baseline of the next best thing Redwood and Remix participants would do with the time+resources or something else?

predicted NO

@NoaNabeshima I always interpret EV calculations as being against the counterfactual of the resources not being used. Otherwise every EV calculation would influence every other one.

predicted NO

@EliasSchmied Hmm, or actually I notice I'm not really interpreting it like that in my head. I'm not sure, I'm confused. Maybe in the context of a specific organization with a limited budget it's tractable to compare it to the next best thing.

predicted YES

@EliasSchmied I think in this case a solid chunk of the costs are opportunity costs, I think the EV has to factor that in. The relevant comparison seems to be allocation of resources (incl researcher time) to next best thing

@DeltaTeePrime But then the original question becomes "Was this literally the best thing they could've done?", which.. doesn't feel like the original spirit of the question. Unless it's compared to the other thing they actually WOULD have done, without benefit of hindsight, that would make sense.

predicted YES

@EliasSchmied yeah this was what I should have said and intended

@EliasSchmied the criteria is "suppose we had decided not to do remix and instead used the resources elsewhere etc. Would this have been higher expected utility? (given ex-post knowledge about the quality of remix and what we otherwise would have done)." I might edit this into the description.