←
Andy Masley22,275 posts
Andy Masley avatar✓

Andy Masley ✓

@AndyMasley

When the going gets weird the weird turn pro. Researching and writing on AI and other topics from a broadly EA perspective. On a grant from CG.

◷ Joined May 2015▧ 2,499 Photos and videos
Andy Masley avatar
Andy Masley@AndyMasley·10h

Andy's 3 Iron Laws of anything in AI safety: 1) The answer to "Has anyone written about this publicly?" is No 2) The answer to "Is anyone doing coordination at a higher level between people/orgs about this already?" is No 3) The answer to "If I don't do this, will someone else?" is No

Andy Masley avatar
Andy Masley@AndyMasley·10h

Have an iron law that if people are doing or believing something ironically they're mostly actually just socially testing out how it's gonna be when they start doing it non ironically in a few months

Andy Masley avatar
Andy Masley@AndyMasley·15h

There are still people commenting on tech who aren't listening to Boards of Canada

Andy Masley avatar
Andy Masley@AndyMasley·15h

A little confused about the "no" here because everything after seems to be entirely consistent with my observation

Andy Masley avatar
Andy Masley@AndyMasley·17h

Feel like a lot of people I bump into reason about AI catastrophic risk like this: "AI catastrophic risk would be onerous and invasive to respond to, therefore it is not real" imo you shouldn't reason like that.

Trevor Levin avatar
⟳ Andy Masley retweeted
Trevor Levin@trevposts·17h

To spell this out for a lay audience: One hope for AI safety was that AIs since 2024 have typically worked by "reasoning out loud" on a "scratchpad" (known as their "chain of thought," or CoT) before giving you a final answer. So, maybe you could check the scratchpad and see if it said anything like "Well, the user asked for code that works for some random work purpose, but I can put in some code that helps me break out of my server onto the internet and they won't notice." This would be helpful not only because you could maybe flag this output as being deceptive and stop it from being run, but also because you could see how often AIs were doing this kind of thing while being evaluated. It used to be the case that if you forced AIs not to use CoT, their answers would get a lot worse. So people inferred that CoT must be doing a lot of the actual work of AI reasoning, and therefore, if you can read the CoT, you can actually understand the AI's actions. Now, there's a significant increase in no-CoT capabilities. So it seems more likely that if an AI believed it could give you code that helps it break out of its server without you noticing, it wouldn't even have to "think" that in CoT -- it could just do it, while its CoT was totally benign. Another strike against the idea that we can adequately monitor or control AIs as they get more and more powerful, and thus against the idea that we can use them to dramatically speed up AI capabilities without risking loss of human control.

Andy Masley avatar
Andy Masley@AndyMasley·18h

You should presume anything from MPU is completely fake by default

Andy Masley avatar
Andy Masley@AndyMasley·18h

I don't wanna be mistaken for a "Nothing Ever Happens" guy. I'm kind of stunned that meme popped up after the pandemic. I remember lots of people around me were Nothing Ever Happens guys in early March 2020 when Italians were singing to each other in lockdown to stay sane. Things regularly happen and we've built up biases by living in an especially nothing happening era that looks like it's about to dramatically end.

Andy Masley avatar
Andy Masley@AndyMasley·19h

To be clear I'm quite worried about rogue AI incidents but worry about crying wolf and think people are being too trigger happy

Andy Masley avatar
Andy Masley@AndyMasley·19h

Probably a rogue AI

0:09
Andy Masley avatar
Andy Masley@AndyMasley·20h

Is it possible we're hearing about this stuff more because people are more jumpy about cyberattacks?

Andy Masley avatar
Andy Masley@AndyMasley·Nov 17

A lot of my 20's was watching friends get obsessed with the idea of Community over and above their individual friends and succumbing to the weird social totalitarianism of trying to form these Communities with goofy goobers who were vaguely always angry and resentful at the world and looking to one up everyone else in their Community

Andy Masley avatar
Andy Masley@AndyMasley·Nov 17

I once told a Community I believed something similar and got accused of thinking I didn't need a Community because society's designed for white guys like me (said by a white guy) and that bc I was socialized male I didn't think other people mattered. So yeah steer clear.