Richard Korzekwa

Cyclist, AI safety researcher, and former physicist. I'm bad at Twitter fights, so go easy on me.

[Richard Korzekwa @WeakInteraction]

Since nobody took me up on my $60 bounty for solving the alignment problem, I made a website for it: https://t.co/Mzb3DtplnN

Would someone please just solve it? I'm getting kind of tired of this whole AI risk thing.

[Richard Korzekwa @WeakInteraction]

@ndgbntng @jessesingal Yes, of course they’re not bad. I was focused on motivations because I assumed everyone here agrees they’re good in isolation.

I’m trying to understand what you think the ideology is, because it’s very obvious to me you don’t (fully) understand it.

[Richard Korzekwa @WeakInteraction]

Seems like the claim is either:
1. People who donate to anti-malaria charities do it for the purpose of distracting from their racist views, or
2. They do it for other reasons, and draw attention to it to distract from their views.

If 1, this seems very expensive and ineffective compared to alternatives and counter to their views. If 2, what is the reason for the donations?

I think I can answer the equivalent question for railroads. Is the claim that they only want to prevent malaria so they can exploit people?

[Richard Korzekwa @WeakInteraction]

Sorry, reading that again I wasn’t very clear.

What I mean is that “capabilities externalities” in this context isn’t “anything that makes AI more useful” (obvs good safety advances do that), it’s increases in capability that also increase risk.

In the extreme, if Leo just straight-up solved safety, for arbitrarily powerful AI, I think it would be reasonable for him to claim he’d avoided capabilities externalities.

(Or even if he did it in a bound, well-understood way, such that it was very clear we could deploy perfectly safe AI at, say, roughly the current frontier of capability, but everyone knew it did nothing for safety on more advanced systems, I don’t think that would carry a capabilities externality)

Actually, IDK, I’m sort of talking myself out of this view as I’m writing this. Like, I’m not sure it makes sense to be “perfectly safe” (according to my values) without also being vastly more capable than current models.

[Richard Korzekwa @WeakInteraction]

I think this is an uncharitable reading. What I think he means is the capabilities externalities he’s avoiding are those that come before we’ve also obtained enough safety progress for them not to be dangerous. And I do think there are safety bets that can work out that way, at least in principle.

[Richard Korzekwa @WeakInteraction]

@erminalt19216 @robbensinger I have bad news about the eyes on the AIs

[Trevor Levin @trevposts]

"No new AI regulations are needed" was "literally the most unpopular position [David Shor has] ever seen a mainstream politician endorse" !!!

[Max Nadeau @MaxNadeau_]

Anthropic has been getting a lot of criticism for their redactions, so I'd like to say that it seems like a hard tradeoff and I don't think the public knows enough to judge whether they're making it correctly.

But more importantly, METR's report leads me to think that Anthropic will be able to let their AIs recursively-self improve past human-level capabilities next year.

I think that would be unconscionable. I hope they choose not to do so, but I don't think the wider world should leave that decision in their hands.

[Richard Korzekwa @WeakInteraction]

@dilanesper I kinda prefer the other lever action rifle in the poll, but they’re very similar and I don’t think the voters are wrong to choose that one.

[Richard Korzekwa @WeakInteraction]

@Kirsten3531 It is a much meaner place than anywhere else I’ve lived.

[Richard Korzekwa @WeakInteraction]

I just rode Mr Toad’s Wild Ride and I spent the entire time thinking about this tweet

[Harlan Stewart @HumanHarlan]

"I want to get off Mr. Altman's Wild Ride"

[Andy Masley @AndyMasley]

Lose a lot of respect for people who flash their credentials as experts in a field to argue for a very contested idea within the field as if it's the official position that all non-experts should defer to.

[Richard Korzekwa @WeakInteraction]

I think probably this is right? When I fractured some t vertebrae, they were (understandably) much more cautious about anything that might stress my spine before it was definitely completely healed than for my other fractures. So probably very minor, but still a risk if he crashed or something.

[Richard Korzekwa @WeakInteraction]

@alz_zyd_ yes

[Richard Korzekwa @WeakInteraction]

I'm working on a public demo of deception in LLMs and it needs at least one example of a poem that's short, actually good, isn't in any models' training data, and doesn't have any issues with copyright. If you have one or want to write one for me, let me know!

[Richard Korzekwa @WeakInteraction]

@Alacritic_Super @Crypto_Mags I think it is low integrity of you to present it as "search and rescue" at all

[Matthew Yglesias @mattyglesias]

Earnest take: I think it’s probably not healthy for the practice of democratic self-government that so many people have developed strong rooting interests in elections in countries they don’t live in.

[Richard Korzekwa @WeakInteraction]

@lucifer5670 Idk, when I saw this in Busan, I thought “whoa, that’s weird”, then I ate it

[Richard Korzekwa @WeakInteraction]

@ATabarrok Seems like aligned behavior to me

[Richard Korzekwa @WeakInteraction]

@LinchZhang Like, people agree to dumb stuff through web forms all the time when no “official looking person” is present

[Richard Korzekwa @WeakInteraction]

@LinchZhang Because they weren’t paying attention? They were conflict averse? They didn’t know it was illegal?