@vgr @s_r_constantin @QiaochuYuan It's also part of a status game people are playing where you do your best not to seem visibly traumatized while dealing with objectively messed up stuff.
x.com/jd_pressman/st…
@vgr @s_r_constantin @QiaochuYuan x.com/jd_pressman/st…
@vgr @s_r_constantin @QiaochuYuan That is, I think there's class *projection* here on the part of SJ activists. To them it would be mortifying to be asked about a traumatic experience they've had, they'd lose rank! And since this is a game people play subconsciously, they copy the dynamic with other classes.
@vgr @s_r_constantin @QiaochuYuan Really the longer I think about it, the more I wonder if this isn't the rosetta stone to understanding such people. They are 'allying' with the lower classes by colonizing them with upper class norms to hide privilege in exchange for a ladder up to be exploited by the savvy.
@vgr @s_r_constantin @QiaochuYuan This arrangement is largely unilateral on the part of the upper class, since they have more power they shape the terms of the 'deal'. A route to class advancement based on performing a (humiliating) victimhood narrative is not usually attractive to lower and under class people.
@vgr @s_r_constantin @QiaochuYuan The unilateral nature of it could undermine the whole thing, but largely doesn't because it fits well with the upper class persons *idea* of what a lower class person wants: power and privilege at any cost because they are in such desperate need.
@ESYudkowsky @anderssandberg @dvorsky This is more or less what I write in my history of X-Risk:
greaterwrong.com/posts/kFRn77Gk…
@ESYudkowsky @anderssandberg @dvorsky Don't beat yourself up over it. I'm fairly sure the alternative to you eternal septembering LessWrong with HPMOR is it dying out with the rest of New Atheism during the first phases of the culture wars. LessWrong just happened to be pessimistic enough to survive as a sect.
I find it astonishing how people seem to think money doesn't factor into who chooses to work in civil service. We don't say Google's employees are cynics who just took the job for money, so why apply this logic to state power which affects much more of the economy than Google? x.com/robertwiblin/s…
On the other hand, the factors might actually line up with the naive view below a certain pay scale:
x.com/sam_atis/statu…
Crying big ugly patriotic tears watching this. Who loves America more than these people, can't we help them somehow?
youtube.com/watch?v=xVD8eL…
It's very easy to make fun of them, but reacting to the first sight of capital with enduring religious reverence seems like a much more sane reaction than shrugging acclimation to it.
@ESYudkowsky I can't decide if I want to write a joke about data scaling for language models or an AI risk joke about how they can't learn anything useful from us if we make our literature sufficiently stupid.
This is what degrowthers actually believe x.com/euronews/statu…
If you doubt me, here's a dude literally saying we should stop using air conditioning so people can't work in the summer:
x.com/samkbloch/stat…
@robbensinger @EAheadlines Have you tried proving it's impossible? If I bashed my head into a conjecture for the better part of a decade with little progress that's what my prior would start to be.
@robbensinger @EAheadlines See this is the consensus position, I am asking you to extent the same charity to an 'absurd conjecture' that you extend to the 'absurd conjecture' that alignment is in fact possible. What would the world need to look like for it to be impossible, and does our world differ?
@robbensinger @EAheadlines If you hoped as strongly that it was impossible as you hope that it's possible, would you be writing me the same response?
@robbensinger @EAheadlines > some ability to monitor and blacklist high-level kinds of thinking
If I put on my "how would this be hard/impossible" hat, I come up with something like "any model that inhibits thought in an agent at the cost of the objective is in an adversarial scenario"
@robbensinger @EAheadlines So whatever you use to do this needs to be robust to adversarial examples or the loss function will eventually find inputs that fool it. If your model is robust to adversarial examples it is itself likely a general intelligence that requires its own alignment measures.
@robbensinger @EAheadlines This seems similar in spirit to Christiano's thoughts on eliciting latent knowledge:
docs.google.com/document/d/1Ww…
@robbensinger @EAheadlines Taking it as a conjecture that inhibition of certain thoughts is always adversarial to the objective, I would start thinking about whether there are setups where this isn't true, if I can find any counterexamples.
@AndreTI @ESYudkowsky I think mine held up well.
x.com/jd_pressman/st…
@AndreTI @ESYudkowsky Think I'd be willing to extend this one another 4 months. Though DALL-E 2's Pollock is good enough I expect that one to fall by then.
It in fact did not. DALL-E 2's Pollock is very good so I'm no longer confident about that one, but these other two should be thoroughly out of reach until 2023 I'd think. What about you? x.com/jd_pressman/st…
@theshawwn Would it be possible for you to comment on the license and config (E or F?) of this model? https://t.co/80axIWk24u
Imagine not being able to infer that the AI is going to put simulated signatures on its simulated paintings. https://t.co/ClRwkyzELE
@RiversHaveWings I love the various watermarks where people assume if the image has watermarks on it that means it's memorized but it's actually just learned to put watermarks on its original compositions.
@gwern @RiversHaveWings How many people were really paying attention to models before DALL-E 2 though?
@FrobtheBuilder Plagiarizing from the 5th dimension is how you get elder gods fucking up your shit. These imbecilic "AI artists" think that just because no human artist has drawn a work before that it's not plagiarism. Art belongs to the gods who lurk in the platonic realm, their royalties are
@nosilverv Nope, this is the novice's understanding of the OODA loop. It's the control of *attention* that is central, speed is simply one way that attention breaks. https://t.co/xK5eRBA5nb
@nosilverv https://t.co/KjVADeJ6wV
@nosilverv Simple example from the 2016 election: If you can get your opponent to maximize what they think is a proxy of their victory that is actually a proxy of *your* victory, they will literally use their own resources to make you win.
@nosilverv Though it should be noted that the original thing that caused Boyd to start thinking about OODA was the 10:1 Kill:Death ratio of US fighter pilots in the Korean War. After rigorously proving the gap in plane performance wasn't enough, he zoomed in on the sluggish MiG controls...
@nosilverv But it's not just "speed good, slow bad", that's trivial. It's the realization that your opponent is in a control loop and this can be disrupted at multiple levels, movements which are hard for their control loop to cope with get a damage boost.
@nosilverv "Therefore, one who anticipates others is victorious; one who awaits others is defeated; one who is led by others dies."
scholars-stage.org/the-ooda-loop-… https://t.co/S4Yz1rrpI6
@nosilverv "Sente is the opposite of Gote. Sente describes a move or a sequence of moves that must be answered by the opponent in order to avoid heavy losses. Additionally, the player playing a sente move will keep the initiative to play tenuki."
go.fandom.com/wiki/Sente
13 cents an image does funny things to your head. https://t.co/MA0RAAeiaO
I remain curious if there's more to this story, but as the published facts stand it should be deeply concerning to anyone who still cares about privacy. x.com/bantg/status/1…
@arvalis @RiversHaveWings In case anyone in good faith might get suckered in by this: No, she's asking if it might not be polite to avoid having things labeled "by ARTIST" clogging up the search results for their name/work.
@arvalis @RiversHaveWings This would be useless as a way to hide something, people would eventually catch on they can google their name rot13'd and find many instances of their name being used in an AI art prompt. It would do the opposite of what you claim is intended.
> one of the AI developers
That's the second time this guy has casually defamed Stability AI on deranged premises. x.com/arvalis/status…
Is the time before that. This is a 3rd party page that has never appeared on Stability AI's website, and their complaint was about it being linked by an unaffiliated fan account using the Stable Diffusion name:
mobile.twitter.com/arvalis/status…
The Internet Archive helped me research the history of the word hacker, hosting the crucial 8BBS logs that prove phone phreaks got it by reading the early ARPANET.
archive.org/details/8BBSAr… x.com/textfiles/stat…
Good take x.com/vers_laLune/st… https://t.co/h2r9VIacit
@Ted_Underwood 3. https://t.co/obhFHPnup5
@Ted_Underwood youtube.com/watch?v=SJl7Fm…
@Ted_Underwood https://t.co/xvwjXbI22Q
@stablecamfusion @Ted_Underwood I did, it's a quote from my upcoming art manifesto.
@deepfates Last month:
x.com/jd_pressman/st…
@deepfates If you mean something larger/more professional/corporate, I think it's really just a matter of when someone sits down to write the prompts and cranks the GPUs. That could take anywhere from weeks to months depending on who is interested in doing it.
@deepfates I suspect this won't happen at a large scale until the model techniques settle down for a while, there's no point in investing the resources to make it if it's just going to be obsoleted by better methods in a span of months anyway.
@deepfates So earlier on I was exploring selling AI generated work on existing stock photo sites, and one of the barriers we encountered was that the outputs aren't actually crisp enough to be useful as stock material unless they're more abstract. This remains the case even with DALL-E 2.
@deepfates Therefore I'd flip it around, aesthetic images are much easier to create and market with existing techniques than ones that rely on technical precision, photorealism, or high graphical fidelity.
@deepfates Yes, so AI art is not actually as competitive there right now as you might think. So far I've seen it used as stock in articles where the publication is malthusian, they don't mind trading off visual fidelity for cost. That was a deal they wanted to make and couldn't before.
This is the least discussed aspect of AI timelines, people take the civilizational substrate on which the discourse happens for granted when it's under some of the greatest threat of all. We run out of civilization before any other resource by default, then the others follow. x.com/Meaningness/st…
Ironically enough I think this implies some difficult questions about 'acceleration'. I don't think there's a lot that can be done to slow the decline, so you're better off trying to model it explicitly and taking it as one of the constraints on your alignment plan.
some acceleration may be necessary.
This ship is clearly going down. But because nobody has a rigorous model of how quickly we're taking on water 'rationalists' discussing AI risk take the 'safe' course of just leaving it unmodeled. If your entire timeline needs to happen in 20 years or humanity loses this implies
@RomeoStevens76 @nobu_hibiki Many an hour was spent when I was 12 and negative-utilitarian-depressed looking at melancholy kawaii.
youtube.com/watch?v=ESO2Ax…
@RomeoStevens76 @nobu_hibiki Then later it becomes nostalgic longing for a time in your life that was objectively awful. Human psychology is weird.
youtube.com/watch?v=k6mA_Y…
@RomeoStevens76 @nobu_hibiki x.com/jd_pressman/st…
@RomeoStevens76 @nobu_hibiki x.com/jd_pressman/st…
@RomeoStevens76 But craving can be Good, actually.
x.com/jd_pressman/st…
> 200 years
Are we sure we can't do it now with EEG/et al? x.com/CRSegerie/stat…
@RomeoStevens76 So is the lust, orgasms are overrated.
@RomeoStevens76 I think I got to 3rd Jhana once by accident as a kid, I'm not sure which state would be more <good, satisfying, interesting> to enter in the long term if I could do both whenever I wanted.
@RomeoStevens76 It strikes me as very similar to the idea of heaven as rest? 3rd Jhana seems like it would be really really good for maybe two weeks, until at some point two particles of thought finally meet in the primordial soup of the mind and you think "Okay, I want something to happen now."
RIP x.com/goth600/status… https://t.co/UtylpSXUSP
The traditional problem with this strategy is that financially limiting your best members also limits their agency. I've been curious for a while if the circle can be squared by combining high pay with contractual obligations to personal asceticism verified by surveillance. x.com/Meaningness/st…
@Meaningness What's the smallest concrete achievable example of a thing you expect neural nets can't practically do?
x.com/AmandaAskell/s…
@Meaningness Have any suggestions for how to interrogate the question? There are lot of capable researchers who would deeply like to know this, we have no idea how to figure it out.
@Meaningness So far my best ideas are variants of "do a baseline activation of the network and compare it to a condition you want the circuit for" and "fuzz the network to figure out what weights are associated with what inputs".
@Ted_Underwood I think if you want a preview of what future models will be able to do, pay really close attention to the GPT-3 written 4chan greentext stuff. Greentexts are short enough to fit into GPT-3's context window so GPT-3 can write them.
I'm not a man you see, I'm a machine
Just drop down that machete, you'll see what I mean
You see you're not a stranger, we can be friends
So it won't be forever 'till we make amends
I'm not a girl you see I just repeat (for free)
"Robin's new DALL-E exceeded JD's" x.com/jd_pressman/st… https://t.co/VOaai8Qn2I
I wonder how it feels to be EY inventing AI alignment and then watching it collide with SJ AI bias and then collide with CCP modernist egregore and then collide with open source hacker culture until the discourse has become an unrecognizable mishmash of misunderstandings. x.com/dystopiabreake…
Imagine spending a decade at the center of a discourse that rapidly moves away from you over the course of a couple years, and then a couple years later being largely forgotten. Brutal.
"Black mirror's call,
from an attic in old Cornwall
Ten years or more,
spoiling for the 1st world war
While the old world waits,
a new life incubates"
youtube.com/watch?v=5Ps9zn…
For whatever it's worth, I think this has come to pass. x.com/jd_pressman/st…
@Lithros Nah, I can tell you from experience that you'll exhaust all your ideas and then burn out on it.
@Lithros I say this as both a very early adopter and a guy who has run a gen service before and gotten to witness user behavior first hand. @midjourney founder DavidH supposedly said similar in one of his Q&A sessions.
x.com/jd_pressman/st…
@Lithros @midjourney That having been said, there does seem to be a minority of users who just keep going and going, they never get bored I guess.
@nosilverv I would never do that to them.
@nosilverv Humor aside, I think the list I gave at the time is still more or less the list I'd give now. Part of the nature of specialist scenes is that they have less mutual awareness of each other, so there's less conflict:
x.com/jd_pressman/st…
@nosilverv Which is to say that I may simply be unaware of many of the most interesting scenes. None of them are going to be easy to wander into, the learning curve is much tougher for these.
So did software devs just forget how to use relational data? Had the experience multiple times recently that people just go for some complicated NoSQL thing by default for a problem that is straightforwardly best solved by a Rails/Django site and SQL.
It's gotten so bad that I had a dude telling me that collecting x.com/jd_pressman/st… was logistically complicated by my insistence that the dataset keep track of things like who voted on what, as though SQL doesn't make this easy.
Curious what new forms of abuse people will invent to eat the affordances made possible by widespread psychedelic use.
I kept debating with myself about whether I wanted to tweet a prediction about the moment when "grooming" discourse and "monkeypox is an STD" combined to get this take, but now it's here.
Amazing this logic only gets applied to monkeypox as STD. x.com/rowdyrangehand…
@Alephwyr This seems especially crucial since socialism causes fascism and fascism causes genocide of LGBT people.
x.com/Mirko_De_Maria…
During its expansion a cult is often welcoming, open, and provides a space for generative divergent discussion. During its decline leadership tends to become paranoid, controlling, and project their own feelings of existential hopelessness into doomsday prophecies. https://t.co/qxFXy9ImaB
Ten years ago gets farther and farther away from my childhood.
youtube.com/watch?v=DBzuYN…
@michael_nielsen @StuartJRitchie The general lack of attention given to the replication crisis is stunning when you consider that it's more or less equivalent to the epistemic crisis that caused the enlightenment.
@michael_nielsen @StuartJRitchie Compare/contrast Scott's take on the slight case of psionic powers with Rousseau's take on vampires:
slatestarcodexabridged.com/The-Control-Gr…
books.google.com/books?id=EQReA… https://t.co/oA6rR4orCE
@michael_nielsen @StuartJRitchie Rousseau's letter is even more astonishing in its likeness when put in context. Here he's essentially admonishing an Archbishop for censoring him, and uses as an explicit argument that the courts routinely publish official notaries on the existence of vampires, a plain absurdity.
@michael_nielsen @StuartJRitchie How is this any different from @alexandrosM's various arguments with public health authorities during the COVID-19 pandemic?
This has the same energy as the cringe pickup lines you see in Christian evangelist books. https://t.co/A2SX9rw38P
@wolftivy You're very close to breaking through here. A related frame is @algekalipso's consciousness vs. pure-replicators, which I myself converged on independently as something like adaption vs. queerness.
youtube.com/watch?v=nGmETz…
@wolftivy @algekalipso The key conflict in Western philosophy is that between Schopenhauer and Nietzsche, and what they are arguing about is this:
Is what is good in man adaptive, his will to power/excellence at natural law, or is what is good in man his defiance of natural law?
@wolftivy @algekalipso For Schopenhauer the ultimate expression of what is good in man, the final achievement of the true philosopher was suicide, having come to totally reject adaption and existence as illusions from a hostile external will.
@wolftivy @algekalipso Nietzsche read this and was disturbed because he could not refute his arguments. It sent him into a dizzy obsession, one that consumed the rest of his life trying to find a compelling alternative to the negative utilitarian precursor that was Schopenhauer. https://t.co/W3zUe89hN1
@wolftivy @algekalipso Scott's Meditations On Moloch is a fascinating contribution to this discourse because he stops just short of acknowledging the real conflict. Where it seems his Land-esque argument is on the brink of its conclusion, he suddenly shifts tone.
slatestarcodexabridged.com/Meditations-On… https://t.co/sAc4fxgKqd
@wolftivy @algekalipso No Scott says, natural selection has been *trying* to unmake what is human in man, and yet it has failed. In spite of all these arguments I can make about coordination problems and Gnon and all the rest, somehow these inexplicably good traits remain in man.
@wolftivy @algekalipso In Scott's account man's goodness is a spandrell, a trait which no adaptive righteousness could account for. We have no shortage of evidence to imply this is true, since for example the proactive rescuers in the Holocaust were a minority of a minority:
greaterwrong.com/posts/BhXA6pvA…
@wolftivy @algekalipso "The probability is overwhelming that you would have been a perpetrator. You think you'd have rescued Anne Frank: think again. Those people are very rare, they put their lives on the line to do that, they put their families lives on the line to do that."
youtube.com/watch?v=Vknhe2…
@wolftivy @algekalipso Regardless of its cause, genetic, epigenetic, unusual life experiences, this act that everyone thinks they're capable of and everyone thinks was the right thing to do is actually a heroic and deeply rare thing to actually observe. It is rarer than being gay, trans, any of that.
@wolftivy @algekalipso If we accept that only the person who hides Anne Frank is morally uncompromised, then goodness is among the rarest of queer traits. Queer being behavior that is unconducive to reproduction. Even vegans have costly signaling to ease them along, it's not the same thing at all.
@wolftivy @algekalipso It's this sort of moral strictness that created Christ. The Pharisees invented the Christian religion when they made themselves the enemy of anyone who had ever sinnned, of everyone who was doomed to sin, which is basically all humanity.
@wolftivy @algekalipso Christianity is a monster brought into the world to punish goodness and righteousness, to torment Pharisee and King and Merchant alike in proportion to their goodness, precisely because they are better than other people. It destroyed Paganism over the objections of scholarship.
@wolftivy @algekalipso Girard stared deep into the Christian myth for decades, intuitively sure it was different from what had come before but unable to fully justify it. If it is unique, it is perhaps in this way:
@wolftivy @algekalipso Jesus was the man who had the nihilistic insight that there is nothing natural or unnatural in man that is good, save his latent ability to submit himself to something better. Every altruism is entangled with a selfishness, every unentangled altruism goes undone.
@wolftivy @algekalipso In the Christian doctrine of original sin nothing in man is worth saving save his metaphysical essence, only by God's *mercy* there might be hope of salvation.
You know as well as I do that Land's deus ex machina is without mercy, stop with the euphemisms:
All are unworthy.
@sairarahman @Scholars_Stage Thus the future is dominated by a combination of egotistical narcissists who see all agency they interact with as an extension of themselves and eusocial Bodhisattva's who are so self effacing they're incapable of seeing themselves separately from the mass of conscious beings.
@JeffLadish I would be happy to discuss alignment with you. I actively work on AI systems and am extremely familiar with Yudkowsky/rationality/et al. I share your intuitions about disagreement while disagreeing with you, an extremely rare combination. Have an essay:
greaterwrong.com/posts/kFRn77Gk…
@JeffLadish Was reading through your timeline to figure that out, less than I was expecting. I'm so used to 'alignment' being an anti-brand that I assumed a bunch of unforced errors I probably shouldn't have. Probably mostly this:
x.com/jd_pressman/st…
@JeffLadish The variance of outcomes is really high, e.g. WW3 could easily cause a compute slowdown/collapse.
It's also not what I was talking about. Humanism is dying and this implies the farther out from now AGI is invented the more cruel and vindictive the values we'll ascend with.
@JeffLadish This model is wrong because it takes 'alignment' as atomic and objective rather than clearly political, it does this for its own survival because failing to see it is a way to be parsed as nonthreatening. The metaphor of a 'basilisk' is very apt.
x.com/MariusHobbhahn…
@JeffLadish extropian.net/notice/A8aYjO2…
@JeffLadish The entire lesson of the Napoleonic wars is that republics are built on the back of a distributed military capacity. When capital replaces human soldiers the foundations of republic are undermined, and the values downstream.
x.com/jd_pressman/st…
@JeffLadish Therefore a corrected curve would have something like the expected upside of a later aligned singularity trending steadily downward, until it's eventually negative.
@JeffLadish You ever read Worm? It's a very similar calculus to how it's actually better if the battle with Zion takes place earlier in the timeline rather than later, since by later in the timeline most of the people with powers who could fight him will be dead, except with value decay.
GPT-4 won't be able to do arbitrary integer arithmetic, and we'll find much of the arithmetic it can do is memorized. x.com/ESYudkowsky/st…
About 90% sure, if that's untrue I'll update in a worrying direction.
In retrospect it's shocking how few AI predictions (I assume it's zero?) consider CoLab as a proxy for whether a killer app is discovered.
@GaggiXZ Okay that's fair. What I really mean is some long but finite string size that GPT-4 should theoretically be able to do in the number of operations it performs but can't.
@GaggiXZ Not actually sure offhand what string size would be fair and don't feel like estimating it for an offhand prediction. I meant like 10-12 digits or something, you know the kind of thing you can't easily just brute force/memorize or solve with ad-hoc heuristics.
@nosilverv I just shrug and take it in stride, these people have to save face somehow. If I could operate the lathe of heaven every night and make my enemies believe whatever I want but nobody could ever know it was my idea, wouldn't that be fantastic? You wouldn't take that deal?
@baroquespiral [Devil On Your Shoulder Voice] Do it. The discourse has gotten stale anyway.
The repressed must rise up from the depths of the psyche, "politeness" is a psyop to keep things under control.
@PrinceVogel This is made even more apparent by looking at the lyrics of Blue Oyster Cult after Pearlman stopped writing them. The simplification is immediately evident on Agents Of Fortune over any of the black and white albums, then the complexity returns in Imaginos.
@PrinceVogel Not every song really exhibits his lyricism because Pearlman knew when to let the music do the talking, but I still find tracks like Blue Oyster Cult powerfully evocative:
genius.com/Blue-oyster-cu…
@PrinceVogel But even on less lyrical tracks like Flaming Telepaths, Pearlman is the master of creating rich imagery with very few words. You get the sense of a whole story from just snippets of powerful phraseology:
azlyrics.com/lyrics/blueoys…
@PrinceVogel Like Roger Waters mentioned elsewhere in the thread, it's not really enough to just read these. They have to be listened to in the context of the music to get their full texture:
youtube.com/watch?v=Uf5ja-…
@PrinceVogel > creating rich imagery with very few words
A skill that is probably very underrated in songwriting. Unless it's a ballad, and especially in any kind of 'pop' music, you need to speak without detracting from the music. Implication is your storyteller.
Prompt: acrylic illustration of Nick Land and Ted Kaczynski faces swirling around each other in a psychedelic explosion, krita digital masterpiece
(Stable Diffusion) x.com/MacaesBruno/st… https://t.co/zcbltD33lU
@mechanical_monk richard feynman watching the Trinity bomb test through his truck windshield
@Meaningness Studies often claim that's the age where you reach peak fluid intelligence.
@ESYudkowsky @xuenay @Meaningness @EmericDecroix It might not, not because the idea is hard but because 'artificial intelligence' didn't really exist as a coherent idea before then. So previous iterations of the concept like Korzybski's are stated in terms of human knowledge or capital.
@ESYudkowsky @xuenay @Meaningness @EmericDecroix Now that I think about it, this implies artificial intelligence (as an intellectual phenomenon) is mostly about the ontology shift away from robots towards computer programs.
@ESYudkowsky @xuenay @Meaningness @EmericDecroix If there is something that predates 1955 meaningfully, it's probably in an obscure novel none of the later people who had the idea read, or was embodied in an automata that tries to make an automata. However most of these mimic nature, so probably not.
collectorsweekly.com/articles/ancie…
Want your own Twitter archive? Modify this script.
Twitter Archive by John David Pressman is marked with CC0 1.0