Many are skeptical we’ll ever 'solve AI alignment.' If so — are we f**d? In our interview Buck Shlegeris argues that there are practical, non-galaxy-brained ways to reduce AI risk without knowing how to align AIs. This is 'AI control': finding ways to use advanced AI without knowing whether... we can trust it. Enjoy — links and transcript below: 01:51 What's AI control and why is it hot? 10:44 Detecting human vs AI spies 18:10 How to catch AIs trying to escape 33:18 Cheapest AI control techniques 51:01 If we catch a model escaping... will we do anything? 53:39 Getting AI models to think they've already escaped 59:01 Will they be able to tell it's a setup? 1:16:13 The pitch to AI companies to do this 1:18:29 Will AIs get superhuman so fast that this is all useless? 1:30:39 Current alignment methods don't detect scheming 1:38:51 Is 'controlling' AIs kind of a dick move? 1:44:12 Could 10 safety-focused people in an AGI company do anything useful? 1:49:40 Benefits of working outside frontier AI companies 2:01:07 What other safety-related research looks best to Buck? 2:02:02 If an AI escapes, is it likely to be able to beat humanity from there? 2:09:22 Will misaligned models have to go rogue ASAP, before they're ready? Find it in any app searching for 80,000 Hours Podcast.show more

This framing seems to imply that control reduces risk even if we can’t “solve” alignment. I think of control more as: — Suppose you don’t know how to align Very Very Powerful systems (what some might call ASI) or Moderately Powerful Systems (what some might call AGI) — Well, maybe you can control misaligned AGI — And then if you do that, maybe you can use the misaligned (but controlled) AGI to Do Something (with the most common example being something like “solve alignment for AGI and eventually ASI”). This framing paints control less as an alternative to “solving alignment” and more as an intermediate step. (As opposed to control reducing risk directly by EG making it unnecessary for anyone to “solve” ASI alignment). Why this matters— A successful outcome requires not only the control of AGI but also some plan for what that AGI will be used for. This is obvious when spelled out, but I think actually a rather challenging part of the puzzle & one that is worth emphasizing. (CC also concerns that in a world with intense race dynamics, Company A might end up just using their controlled AGIs to do capabilities research to outcompete Company B, in which case control ends up feeding into superalignment//RSI to DSA plans). Would be curious if you endorse this framing or if you think it’s off somehow.

Stay competitive by balancing cutting-edge AI with automation tools. Forrester shows how.

YouTube: Spotify: Apple: Transcript + links + summary: Or any app on the 80,000 Hours Podcast.

@bshlgrs The AI safety community just solved it! Hurray! 🥳 It's mathematically proven safe GPU clouds controlled by international scientists, running mathematically proven safe AIs (place AI, tool AI). The link to the paper draft is pinned in my profile

@bshlgrs It depends on what AI is aligning to. IMO it's likely Superintelligence can't be controlled, however it may align to some kind of context sensitive moral realism it discovers, which accommodates us not being f**d. Perhaps the best of our values approximately align to CSMR.

@bshlgrs the first premise is already false