+
+https://twitter.com/patio11/status/1766115590429618347
+> Please note that this is extremely, extremely true, and if you follow that to its logical conclusion, certain blogs are on the org chart of e.g. the U.S. in the same haha but absolutely serious way the NYT editorial page is.
+(quote-Tweeting someone noting Demis Hassabis linking to Slate Star Codex)
+
+----
+
+https://twitter.com/ESYudkowsky/status/1766144388890243108
+> Some social media company finally went and added easy phone calls to Congress. Personally, I was a little "how noneliezeran!" that Google didn't do this way earlier, or Uber under Travis. But I can see how this would be too scary in the USA, such that China did it first.
+
+This is kind of scummy behavior! Interesting that he calls scummy consequentialism eliezerian
+
+-----
+
+https://twitter.com/ESYudkowsky/status/1767276710041686076
+> What about optimising for finding and error-correcting invalid arguments?
+
+> In my own experience those are the same mental skill, but perhaps your mileage varies. But at any rate, sure.
+
+https://discord.com/channels/401181628015050773/458419017602826260/1212913739196669963
+
+-------
+
+Unfortunately, there's still some remaining tension here insofar as the guy continues to lean on "you gotta trust me, bro; I'm from dath ilan and therefore racially superior to you" personality-cult-leader intimidation tactics, which I consider myself to have a selfish interest in showing to third parties to be unfounded.
+
+With anyone else in the world, I'm happy to let an argument drop after it's been stalemated at 20K words, because no one else in the world is making a morally fraudulent claim to be a general-purpose epistemic authority that has a shot at fooling people like me. (_E.g._, Scott Alexander is very explicit about just being a guy with a blog; Scott does not actively try to discourage people from thinking for themselves.)
+
+New example from today: a claim that MIRI is constrained by the need to hire people who make only valid arguments, and (in response to a commenter) that in his experience, finding and error-correcting invalid arguments is the same mental skill. <https://twitter.com/ESYudkowsky/status/1767276710041686076>
+
+But elsewhere, this _motherfucker_ has been completely shameless about refusing to acknowledge counterarguments that would be politically inconvenient for him to acknowledge!
+
+[]
+
+(Screenshot took place in a publicly-linked server and is therefore OK to share)
+
+My heart racing, it's tempting to leave a Twitter reply saying, "Actually, in my exhaustively documented experience, you don't give a shit about error-correcting invalid arguments when that would be politically inconvenient for you"
+
+But ... what good would that do, at this point? As I wrote in the memoir, "We've already seen from his behavior that he doesn't give a shit what people like me think of his intellectual integrity. Why would that change?"
+
+The function of getting the Whole Dumb Story written down that I was supposed to _move on_. I have _other things to do_.
+
+---------
+
+ Oli Habryka gets it! (<https://www.greaterwrong.com/posts/juZ8ugdNqMrbX7x2J/challenges-to-yudkowsky-s-pronoun-reform-proposal/comment/he8dztSuBBuxNRMSY>)
+ Vaniver gets it! (<https://www.greaterwrong.com/posts/yFZH2sBsmmqgWm4Sp/if-clarity-seems-like-death-to-them/comment/dSiBGRGziEffJqN2B>)
+
+Eliezer Yudkowsky either doesn't get it, or is pretending not to get it. I almost suspect it's the first one, which is far worse
+
+https://www.lesswrong.com/posts/CpvyhFy9WvCNsifkY/discussion-with-eliezer-yudkowsky-on-agi-interventions
+> Various people who work or worked for MIRI came up with some actually-useful notions here and there, like Jessica Taylor's expected utility quantilization.
+
+https://twitter.com/ESYudkowsky/status/1301958048911560704
+> That is: they had to impose a (new) quantitative form of "conservatism" in my terminology, producing only results similar (low KL divergence) to things already seen, in order to get human-valued output. They didn't directly optimize for the learned reward function!
+
+-----
+
+Metzger is being reasonable here
+
+https://twitter.com/perrymetzger/status/1773340617671667713
+> That's a fairly inaccurate way of putting it. It wasn't "poked with a stick", what happened was that gradient descent was used to create a function approximator that came as close as possible to matching the inputs and outputs. It's not like someone beat a conscious entity until it deceptively did what its masters demanded but it secretly wants to do something else; in fact, the process didn't even involve GPT-4 itself, it was the process that *created* the weights of GPT-4.
+
+
+https://twitter.com/ESYudkowsky/status/1433206619135823878
+> I don't know if that quite answers your question - but my answer to a lot of "Well, what about *this* kind of AI?" is "Well, what about it? What does one do with that kind of AI, that stops the world from ending 6 months later when some other kind of AI builds nanotech?"
+
+https://twitter.com/ESYudkowsky/status/1433207431627042819
+> For concreteness: One can see how AlphaFold 2 is working up towards world-ending capability. If you ask how you could integrate an AF2 setup with GPT-3 style human imitation, to embody the human desire for proteins that do nice things... the answer is roughly "Lol, what? No."
+
+https://twitter.com/ESYudkowsky/status/1775233618035736819
+(in response to Jessica saying that LLM pretraining favors short term)
+> True for now, at least.
+
+He didn't already know this from the discussion when the Sleeper Agents paper came out?
+https://twitter.com/ESYudkowsky/status/1778938096123359256
+
+https://twitter.com/QiaochuYuan/status/1726860200421085589
+> i was on a 3-month trial for this position, and the terms of the trial were: i would not, at any point, get to field an actual question from the researchers, because all of their actual questions were *too sensitive* for me to know about
projects / Ultimately_Untrue_Thought.git / commitdiff