doing mech interp research with agents is way more fun but also exhausting than doing it solo ime. it's like managing eight researchers each who have results or questions every 4 minutes. my brain's cpu just stays at 100%, agents were supposed to reduce brain's cpu

and then of course there's actual humans who have results or questions, who themselves are now managing eight agents

whereas coding like isn't that hard most of the time, it's more labor than it is hard, even like dense tensor code most of the time is empty space or writing generic while loops or whatever

ofc i think this is a temporary state where agents are good at research but bad at taste, maybe 6-10mo if i had to guess

ofc i think this is a temporary state where agents are good at experiments but bad at research taste, maybe 6-10mo if i had to guess, then cpu will start to go down i think