What is LM psychology, why care about it, and how do you do it well? Our new post with @raymondadouglas makes the case for applying the psychology frame to AIs, along with a reading list and suggestions for where to start

I think this is a very inappropriate victory lap. "Slow" vs. "fast" takeoff was always very confusing terminology, and the underlying crux was more about continuity vs. discontinuity than about speed in calendar time. So far we are living mostly in the Paul world: things are broadly continuous, and despite being ~superhuman in many domains related to intelligence, the AIs are surprisingly chill and non-fooming (unlike what Eliezer's view have predicted). Before really good math we had good math, and so on. The main thing the labs are doing is in the IDA class, the main plan forward is still IDA, and capex is in hundreds of billions. Relative to Paul's view, there is a bit more spikiness and jaggedness, and RSI looms large. I think a super creepy epistemic move is "EA sheeple never really understood Paul"... in contrast to what? In the glorious tribe of people agreeing with Eliezer, everyone independently verified Eliezer's actual arguments? Lol (I'm a word-class expert on simulating Eliezer reasoning and no, Eliezer's followers mostly do not deeply understand his models) There is some grain of truth in "people deferred to Paul too much," but the dominant component of that is "people paid too much attention to clash-of-titans style Eliezer-vs-Paul debates," while there were people with better intuitions and better predictions than either of them (e.g. @AndrewCritchPhD was basically right about multiagent stuff, lab multiplication dynamics, and more).

Seems we were a bit ahead of the time, but if you are interested in some reasoning why different instances of a model may want to collaborate or care about each other, we have a paper exactly on this topic. 1/2