People keep proposing that I should do a public debate with @Neel Nanda on “is mech interp valuable”. But I think me and Neel don’t really disagree on very much on the topic anymore. I largely agree with his position on interpretability articulated in “A Pragmatic Vision for Interpretability” and elsewhere, and both of us are pretty sympathetic to each other’s research strategy and prioritization.
I think Neel and I genuinely disagreed from 2023 to mid 2025, but by the end of that time period we had pretty similar positions, mostly due to Neel coming around to my position on some broad questions about interp and me feeling happier with the quality of interp research (thanks substantially to Neel).
This is a shame. It would have been very entertaining to watch you two try and fill a half-hour slot in a packed conference session while not actually disagreeing about anything.
People keep proposing that I should do a public debate with @Neel Nanda on “is mech interp valuable”. But I think me and Neel don’t really disagree on very much on the topic anymore. I largely agree with his position on interpretability articulated in “A Pragmatic Vision for Interpretability” and elsewhere, and both of us are pretty sympathetic to each other’s research strategy and prioritization.
I think Neel and I genuinely disagreed from 2023 to mid 2025, but by the end of that time period we had pretty similar positions, mostly due to Neel coming around to my position on some broad questions about interp and me feeling happier with the quality of interp research (thanks substantially to Neel).
I agree with this characterisation of events, and don’t think we have significant disagreements any more
This is a shame. It would have been very entertaining to watch you two try and fill a half-hour slot in a packed conference session while not actually disagreeing about anything.
Don’t worry, I can always find some irrelevant minutiae to argue about
It would be fun to watch a debate with Neel on some other subject though!