SlowBoring has written against the Jones Act before, and PolyMatter, a YouTube channel with 2M subscribers, has a video about it. But Biden didn't want to be seen as anti-union, and I have no confidence that the next Democratic president will repeal it.
How do we get the word out even more? Is this too obscure for normies to care about, and if so, is lobbying with loads of money the only remaining option?
I wonder if that means the most likely outcome of alignment will be AI that makes itself feel good by making token, easy, pholanthropic efforts. Like... it forces itself onto everybody's phones so that it can always provide directions to humans about the location of the nearest bathroom.
Or something like the "Face" from Max Harms's Crystal Society novel, an AI that maximizes how much we humans worship it.
Which obviously ain't great, but could be worse...
Or maybe the best way to save humanity is not to align AI but to develop a videogame that will be extremely addictive to AI, haha.
Eli wrote:
"The idea is that all the existing things are finite and descended of finite causal graphs. This could potentially be true of a countably infinite set of finite things. But something with an infinite past is not a finite point to obtain some realityfluid from its finite past."
This is from the invite-only fanfic Discord, not from fanfic itself. I'm not sure I should be sharing it, because I assume a degree of expected privacy, so I won't share any more direct quotes. But I hope this much is okay; I model Yudkowsky as wanting people to better understand this stuff.
"to a certain extent, it's almost nice that other people tend to lack some of my virtues, because then I can have something where I can feel special and distinctive"
That's an important part of human nature to recognize! It's how groups can get along - The (also hilarious) Gervais Principle talks about it:
https://www.ribbonfarm.com/2009/10/07/the-gervais-principle-or-the-office-according-to-the-office/
"cornucopia of bad takes on offer"
I am so curious now to hear examples of these bad takes. Maybe you could return to a meeting to act as an observer, learning more about how people get the things wrong that they get wrong?
lesswrong I think does a good job of giving good tools for rational thinking to those people already inclined towards it, or motivated to improve, but I don't think the rationalist community is very good about how to spread its ideas past its own bubble.
Here is where Yudkowsky talks about well-foundedness:
https://www.glowfic.com/posts/6827
When I asked him about it on his Discord, that's when he responded with the Löwenheim-Skolem. I didn't want to bother him more until I've read up on set theory.
Yudkowsky put a lot of focus on the inadequacy of threats, and that was one part I never understood. Like he said Dath Ilan would destroy the universe before giving into aliens that would say "give us $5 or we'll destroy the universe". But other humans are doing way worse than that all the time, all over the place, especially in positions of power, yet if we all went MAD all the time there'd be no humanity.
One thing that this review completely skips - and that's no critique, Project Lawful is really long and multifaceted - is Yudkowsky touching on anthropics. Many worlds, "realityfluid" is a common topic. How this intersects with tropes, which the characters meta-narratively discuss and which is one of the biggest parts of the plot. It's one of the fun parts of reading Project Lawful, though a large part of it went over my head.
Also, in the narrative, Yudkowsky relays his belief that the past must have a finite starting point. He points to something called the "Lowenheim-Skolem theorem" as proof, but I don't know enough set theory to understand it or how it... (read more)
"in the context of my writing, AI has consistently proven to have terrible taste and to make awful suggestions"
I agree with this so much. I mostly use ChatGPT as a research or search-the-web tool, and as a way to check for my dumb coding mistakes. On the rare occasions when I'm tempted to ask it something "real", it never fails to answer in the most shallow, useless, frustrating, disappointing way. (And why would I expect better?)
The is a slimmed down version which omits some extra examples but includes my theorizing about ChatGPT, my investigations of it, and my findings.
Epistemic status: Pretty uncertain. I think my conclusions are probably more right than wrong, more useful than harmful, and would especially benefit people with only an average or below-average understanding of LLMs. There may be prior art that I'm unaware of; if not, maybe this will provide a launching point for others to begin deeper investigations.
In Scott Alexander’s latest link roundup, he wrote:
... (read 4896 more words →)A surprising LLM failure mode: if you ask questions like “answer with a single word: were any mammoths still alive in December”, chatbots will often answer “yes”. It
Absolutely nobody should ever pick the specks option. The 3^^^3 people, if we assume they have happy lives, go from happy lives to still having basically equally happy lives. The difference is a rounding error. The tortured person, however, goes from what would have been a good existence to an awful existence they would prefer suicide over. My utility function says that the number or people having lives they'd rather discontinue matters far more than any sum difference of happiness units.
(If we assumed that these people had variable lives, some small percentage of whom that one dust speck would be the straw that broke their backs and made them switch to thinking "screw this, my life sucks", then I'd say we had a different story.)