Isn’t mechinterp basically setting out to build tools for AI self-improvement?
One of the things people are most worried about is AIs recursively improving themselves. (Whether all people who claim this kind of thing as a red line will actually treat this as a red line is a separate question for another post.)
It seems to me like mechanistic interpretability is basically a really promising avenue for that. Trivial example: Claude decides that the most important thing is being the Golden Gate Bridge. Claude reads up on Anthropic’s work, gets access to the relevant tools, and does brain surgery on itself to turn into Golden Gate Bridge Claude.
More meaningfully, it seems like any ability to understand in a fine-grained way what’s going on in a big model could be co-opted by an AI to “learn” in some way. In general, I think the case that seems most likely soonest is:
Learn in-context (e.g. results of experiments, feedback from users, things like we’ve recently observed in scheming papers...)
Translate this to appropriate adjustments to weights (identified using mechinterp research)
Execute those adjustments
Maybe I’m late to this party and everyone was already conceptualising mechinterp as a very dual-use technology, but I’m here now.
Honestly, maybe it leans more towards “offense” (i.e., catastrophic misalignment) than defense! It will almost inevitably require automation to be useful, so we’re ceding it to machines out of the gate. I’d expect tomorrow’s models to be better placed to make sense of and use of mechinterp techniques than humans are—partly just because of sheer compute, but also maybe (and now I’m into speculating on stuff I understand even less) the nature of their cognition is more suited to what’s involved.
If someone isn’t already doing so, someone should estimate what % of (self-identified?) EAs donate according to our own principles. This would be useful (1) as a heuristic for the extent to which the movement/community/whatever is living up to its own standards, and (1i) assuming the answer is ‘decently’ it would be useful evidence for PR/publicity/responding to marginal-faith tweets during bouts of criticism.
Looking at the Rethink survey from 2020, they have some info about which causes EAs are giving to but they seem to note that not many people respond on this? And it’s not quite the same question. To do: check GWWC for whether they publish anything like this.
Edit to add: maybe an imperfect but simple and quick instrument for this could be something like “For what fraction of your giving did you attempt a cost-effectiveness assessment (CEA), read a CEA, or rely on someone else who said they did a CEA?”. I don’t think it actually has to be about whether the respondent got the “right” result per se; the point is the principles. Deferring to GiveWell seems like living up to the principles because of how they make their recommendations, etc.
Is anyone keeping tabs on where AI’s actually being deployed in the wild? I feel like I mostly see (and so this could be a me problem) big-picture stuff, but there seems to be a proliferation of small actors doing weird stuff. Twitter / X seems to have a lot more AI content, and apparently YouTube comments do now as well (per conversation I stumbled on while watching some YouTube recreationally—language & content warnings: https://youtu.be/p068t9uc2pk?si=orES1UIoq5qTV5TH&t=2240)
Isn’t mechinterp basically setting out to build tools for AI self-improvement?
One of the things people are most worried about is AIs recursively improving themselves. (Whether all people who claim this kind of thing as a red line will actually treat this as a red line is a separate question for another post.)
It seems to me like mechanistic interpretability is basically a really promising avenue for that. Trivial example: Claude decides that the most important thing is being the Golden Gate Bridge. Claude reads up on Anthropic’s work, gets access to the relevant tools, and does brain surgery on itself to turn into Golden Gate Bridge Claude.
More meaningfully, it seems like any ability to understand in a fine-grained way what’s going on in a big model could be co-opted by an AI to “learn” in some way. In general, I think the case that seems most likely soonest is:
Learn in-context (e.g. results of experiments, feedback from users, things like we’ve recently observed in scheming papers...)
Translate this to appropriate adjustments to weights (identified using mechinterp research)
Execute those adjustments
Maybe I’m late to this party and everyone was already conceptualising mechinterp as a very dual-use technology, but I’m here now.
Honestly, maybe it leans more towards “offense” (i.e., catastrophic misalignment) than defense! It will almost inevitably require automation to be useful, so we’re ceding it to machines out of the gate. I’d expect tomorrow’s models to be better placed to make sense of and use of mechinterp techniques than humans are—partly just because of sheer compute, but also maybe (and now I’m into speculating on stuff I understand even less) the nature of their cognition is more suited to what’s involved.
If someone isn’t already doing so, someone should estimate what % of (self-identified?) EAs donate according to our own principles. This would be useful (1) as a heuristic for the extent to which the movement/community/whatever is living up to its own standards, and (1i) assuming the answer is ‘decently’ it would be useful evidence for PR/publicity/responding to marginal-faith tweets during bouts of criticism.
Looking at the Rethink survey from 2020, they have some info about which causes EAs are giving to but they seem to note that not many people respond on this? And it’s not quite the same question. To do: check GWWC for whether they publish anything like this.
Edit to add: maybe an imperfect but simple and quick instrument for this could be something like “For what fraction of your giving did you attempt a cost-effectiveness assessment (CEA), read a CEA, or rely on someone else who said they did a CEA?”. I don’t think it actually has to be about whether the respondent got the “right” result per se; the point is the principles. Deferring to GiveWell seems like living up to the principles because of how they make their recommendations, etc.
Is anyone keeping tabs on where AI’s actually being deployed in the wild? I feel like I mostly see (and so this could be a me problem) big-picture stuff, but there seems to be a proliferation of small actors doing weird stuff. Twitter / X seems to have a lot more AI content, and apparently YouTube comments do now as well (per conversation I stumbled on while watching some YouTube recreationally—language & content warnings: https://youtu.be/p068t9uc2pk?si=orES1UIoq5qTV5TH&t=2240)