I’m wrote a book about why all knowledge of the truth is fundamentally uncertain. Give it a read!
I’ve also written a lot about AI safety, personal development, and Buddhism, among other things. Most of that’s here on LessWrong, but there’s also some on my blog, Uncertain Updates.
I have low expectations that this specific law passes, but high expectations that this works to build consensus that we must do something. Even though there’s reasonable risk of politicization of AI, it’s hopefully counterbalanced by the broad unpopularity of data centers, including among the politically right (GOP seems quite worried about data centers costing them seats in the mid-terms), which means there’s a way to sell this not as “liberals holding back progress” but “pushing the Overton window in a direction that will lead to something reasonable”. My big hope is that this moves quickly enough that it influences the upcoming Trump/Xi meetings, and creates momentum around getting a treaty in place.
Cross-apply this to housing. Stop legislating individual zoning rules. Legislate the outputs.
This has, in theory, already happened in San Francisco. Unfortunately, we don’t seem to be on track to hit the numbers by the deadline. Remains to be seen if the State will do anything more about it.
You might try again. This is not radically far off from what, say, Anthropic is doing re: constitutional AI. The main difference is that it comes with a particular training process rather than experimenting on what content to train into the constitution.
With AIs, the pressure is not there yet. The managers are happy that a work that would previously take a week is now done in 15 minutes. But when they update to the situation, and start pushing on you the same way on the new scale (“15 minutes is not acceptable! we have already told the important stakeholders that it can be done in 7 minutes! drop the useless things such as refactoring and documentation and those other unproductive silly games you developers love to do”), the situation will get insane again.
My guess is that AI will remain a source of slack in the dimension of producing marginal additional code to pay down easy tech debt or prevent accumulating it, as 7 vs 15 minutes is not much (I know these were just random example numbers but 2x of a few minutes is not a big deal, especially when the value comes from parallelizing work rather than doing more work in serial faster). Of course, AI is a source of anti-slack in many other ways, so we take what we can get.
Yes, the intention is not to analogize managing bots with managing people, but to say that management was never really about people management, that was instrumental to the real goal of managers, which is to manage outcomes for the business, which is what we now are asking ICs to do since they don’t have to spend so long writing code.
FYI @jimmy @TAG @Nikon Maer. You should all also have DMs about prize logistics.
Matches something from my time being vegan. I found I had to eat a much larger volume of food, sometimes to the point it was uncomfortable, to get enough to function and not be hungry. Meat is crazy high density in terms of calories and nutrients, and being vegan ultimately means eating a lot more bulk.
Matches my experience. Some examples:
Waymo:
First ride: “holy shit it’s driving itself!”
Second ride: “dang this is so cool the future is here!”
Third ride: “why is Waymo so bad at driving?!?”
Claude Code:
First time using it: “this is incredible it mostly figured it out on its own!”
Every subsequent time using Claude Code: “wtf why didn’t you do it right?!?”
Sonos:
First song: “wow, it made a real song out of my half-formed idea!”
Every other song: “why does this suck at writing good songs 😡”
It just goes on and on. I expect to be thoroughly bored by my robot butler that does everything for me, my AI personal assistant that never makes any mistakes, and everything else that’s about to happen about 5 minutes after it becomes the new normal.
Thanks for writing this. To respond briefly, I chose to write this book to be accessible to a wide audience since my hope is that it plants something like “philosophical seeds” in the minds of its readers, which I hope includes many working on AI (both capabilities and safety). For this reason, and to keep the length of the book manageable, I chose to leave out much of the background that you highlight here.
You’re right that I’m a fan of Husserl’s and Heidegger’s philosophical work, and you are correctly seeing the influence of their work on mine. I also find several frameworks helpful for making sense of the world, and among those are transcendental idealism, phenomenology, Buddhsim, and cybernetics. I chose to avoid making arguments in the book explicitly for any of these because they don’t seem necessary to me to the project of the book. That is, I think someone could disagree with me on all of these frameworks and I’d hope they’d still get something from the book, namely learning something about fundamental uncertainty, which I see as being so fundamental that it can be seen no matter what assumptions you make. Of course, my arguments are going to betray my beliefs, I think it’d be hard for them not to, and if anything this is probably a weakness of the book that I couldn’t write a version with wider appeal that less tries to point towards particular answers.
@TAG did you mean to submit this to the essay contest? You posted in time but I don’t see a comment on either of the original essay announcement posts. If you meant to, can you comment with a link to this post here https://www.lesswrong.com/posts/vMJbnnvvrHNw728NC/fundamental-uncertainty-usd2-000-essay-contest?commentId=TeoZFPk7BoqwPPoCs to satisfy the rules requirements.
Sure, maybe there’s some line here, but I think the kind of person who shows up to LessWrong is selected against the kind of person who would be happy and thrive in a traditional monastery.
My general take is that monasteries are good for some people, but for a bunch of reasons I’m not choosing to live in one, and that’s fine, I’m glad they exist for people with minds very different from mine, but also if I had reason to live in a monastery it would probably be fine, even if I wasn’t maximally happy about it.
I wish this post was a little more careful to disambiguate between (1) things that are normal in monasteries and not evidence of a cult and (2) things that are abnormal and are evidence of a cult.
Whether or not MAPLE is a cult, I learned some things from this post that further update me in the direction of believing that Soryu is not serving his students well and should step back from teaching and leadership for a time to reconsider whether the path he is on really is benefiting all beings.
This is a good point, and in fact I had and then cut a whole section about how alignment actually works against corrigibility, as part of being aligned to values is saying no when those values would be violated. Adults are marked, in part, by being more sovereign than corrigible.
Maybe. See if you still think it’s true after 10k hours.
(I say because after that much meditation I think multi-agent models are getting at something real, but the models are confused in major ways that I would have a hard time claiming that they’re “true”.)
This is a great point and something I missed and didn’t think about enough! This is likely a bigger deal to creating the conditions in which humans are able to be aligned than I had considered, or than I think the folks thinking about these questions have been considering.
I get this. I’ve felt this way at times. I was extremely early on AI safety and existential risks as a concern, starting back around 1999, and really getting going in 2001. My largest impact on AI is likely in some combination of being early and building the movement by my mere presence enabling it to grow, talking to people years ago, getting them more engaged, and deliberately steering my career away from anything that smelled of capabilities research.
There’s no acclaim for being this type of guy, though. It’s quiet. And it kinda has to be, because how would anyone measure the impact I had? Maybe my marginal impact was actually zero, or negative, and if I had never heard of AI maybe everything would have gone approximately the same. It’s quite hard to tell.
My only real rewards are all of you—that maybe everyone gets to live and that, even if we don’t, at least I got to spend my adult life with friends who cared about the same things I did.