What progress have we made since your last update?

LessWrong post with the (exclusively null) results.

@niplav (I may still donate my part, depending on what other funders do (sorry about the game of chicken (but you know I really don't earn that much)))

And possible CoI I may apply to this project as a participant and/or mentor (‽¹).

¹: I have no background in any of the fellowship pinball games.

I will join the chorus of people who are like "oh man nobody is actually trying to, uh, solve the problem of pointing/constraining/steering/limiting vast amounts of optimization power, but instead we are spending resources on many things that are related" (monitoring the situation! finding ways to steer non-vast amounts of optimization power! measuring non-vast amounts of optimization power! building infrastructure for slightly less non-vast amounts of optimization power!). "It surely would be great if more people tried to figure out how to figure out strategies that even in concept work for vast overwhelming optimizers. Or build things that are not that but still useful enough as an alternative." This seems like the kind of event that would cause more thinking on alignment foundations to happen (not necessarily agent foundations, mind you, we might want to leave that line of retreat open).

I found this project so enticing that I considered foregoing tax benefits and donating a prospective 10% of my income to it (which would be fairly little, but what can you do).

I've interacted with Mateusz for about ten hours in total, and with plex for about 1½ hours, and found them to be good thinkers, plausibly better than me, maybe a bit more obscurantist. I feel like I can have a higher confidence in Mateusz than plex given the amount of public to-me-legible-output, but I may be misremembering as I'm shooting this comment from my hip.

I mean finally if it's another $200k to METR or to this I definitely choose this.

What progress have we made since your last update?

I'm at my eighth week of data collection, Sam at his tenth, Ethan at his sixth.

Here's the preliminary results with my data alone, which imho looks quite promising.

What are our next steps?

I finish my eighth week, we wait for Ethan to finish his data (should be done mid February), maybe I'll collect Sam's data beforehand & analyze it with my own. Once I have all our data I'll first dump the raw results table on the internet and then do a proper write-up.

@niplav Still have to get my collaborators' data.

@saulmunn A quick n' dirty analysis automated via Claude on my six datapoints gives me 7.28 ± 2.89 hours of sleep after Orexin-A, 8.09 ± 0.75 after Placebo (i.e. Cohen's d of -0.386) at a p-value 0.556.

Plausibly mid February in case something comes up. We need to balance the downsides from being sleep deprived with collecting meaningful data, sorry it's taking so long.

@niplav I'll ask the other two how long they'll still take.

@saulmunn I'd guess late January. I have two or three blocks left but with Christmas + meditation retreat… I'll push the data already collected to Github.

@niplav Still on it, I sometimes procrastinate on doing blocks because 5½ hours is really not that much sleep. We're soldiering on.

@madar Once we have experimental results we'll outline them here, possibly even before posting them. I don't know if one can subscribe to the comments on a manifund project.

@madar Hi! Good questions.

We're not posting updates anywhere as of yet. We don't need any more donations. This comment is the last update. I also add subjective impressions here as the experiment goes on.

I'm not sure how important regular updates are, since by now it's a very steady rhythm of taking the Orexin-A weekly, but little interesting things until we have results with a decent sample size. We did neglect writing an update here because I was otherwise occupied, that explains the large gap since the last comment.

The current state of the project is that we've all received the Orexin-A and are collecting data (I took a dose of {placebo, orexin} this morning but am still struggling with the five hours of sleep). We'll probably collect data at least until end of October, likely into November, and then I'll post the next update.

We've ordered the Orexin-A, and should receive it on the 9th-11th.

@niplav (Your link is to a chemistry Anki deck :-)

@saulmunn Not really, this amount of money we've gotten here is totally sufficient for our experiment needs. If we get a positive result we'll apply to several places for a more in-depth trial, but for now we're set.

@yagudin Good point, we'll collect subjective impressions as well.

Yeah, the blogpost was spurred by the intent to run the experiment, which is now going to happen.