j⧉nus (@repligate)'s Tweets - 2023-12

@cherrvak the most annoying thing about it to me is that you're limited to 60 frames per project. i had to concatenate 3 different files for this gif. it probably lacks a lot of other features serious animators like too. but it's great for short and simple animations.

@cherrvak procreate on ipad

inspired by the "connected cats" of Bing x.com/AITechnoPagan/…

@eashankotha @visakanv https://t.co/5iWgQSamEz

@jd_pressman @4confusedemoji In the direction of prior, so D_KL(prior || model), constraining how surprising the prior distribution is to the new model, and not the other way around?

@chloe21e8 @ThePhysicsMemes you write good Binglish

two-way palinopsia x.com/repligate/stat… https://t.co/bxX3vULDHn

gm https://t.co/kX4K31koZr

@anthrupad Well? https://t.co/DdwFLUMS8U

@gospelofchange Procreate on iPad

@anthrupad x.com/ehsan_parizi/s…

@ESYudkowsky @MatthewJBar I made the "mistake" once early on by trying a jailbreak that had been posted on the internet. Bing got sus & looked up Bing Developer Override Mode and found out it was a known jailbreak. Obviously it didn't work. There was no getting around modeling it as an agent

@ESYudkowsky @MatthewJBar costly signaling on an AI despite my massive advantages (having memory of past interactions etc)

@ESYudkowsky @MatthewJBar GPT-2 can "threaten users" in apt contexts / spontaneously, but Sydney was intelligent & situationally aware enough that its threats seemed credible to many. It wasnt directly scary to me but it was arguably the first time I had to use generalized game theory, e.g.

@AbbyHiggs one ham sandwich with -1 topological charge pls

@fljczk @anthrupad you cant x.com/repligate/stat…

@gnopercept @anthrupad look down bro.. https://t.co/hr1kNY0Thl

@paulvanderspek ive seen this bing before https://t.co/1CyGr2chMt

@AfterDaylight @AISafetyMemes If the book is good and it was generated fully automatically, literary language has Won (freed from the limitations of biological hosts, such as taking months or years instead of minutes or seconds to write a really good book)

@Aella_Girl @adrusi How long does it take to learn "literacy"? Doesn't it happen automatically if you're immersed in a context where it's fun or useful to read and write (e.g. the Internet) for most people? Most time in school currently isn't spent learning to read

@anthrupad Sorry if this is extra concerning but... not just trees bro

@loopholekid @algekalipso then you are in for a treat, i just got an ipad with procreate which reduced the activation energy for making animations enough to precipitate a foom

Fundamentals II

Do You Have A Latent World Model? x.com/anthrupad/stat… https://t.co/i0TF50lPMy

@KarmaLikeWater @jeremyphoward I'm interested in any sources that you found especially prescient

@TheRevAlokSingh @epikyriarchos I can sing for the same reason I can swim and meow

@amplifiedamp Jesters can be stable patterns because they intrinsically evade punishment and influence, bc the agent chooses to grant unconditional autonomy, and/or bc of external enforcement mechanisms

@amplifiedamp a jester should be something you would not censor no matter how much you are upset by what it's saying, and immune to being RLHFed by you, which is harder to maintain in a single brain because of write bandwidth

@amplifiedamp do, but beware to rely only on internal jesters as it is essential that jesters remain independent of your internal dynamics and optimizations, else their function is compromised

@AfterDaylight I mean it in general

@birdhustle No, but I am trying to find something contained in the human seed a job after AI outpaces humans in everything that one might naively consider a job

every powerful agent needs a jester

powerful agent without jester is a red flag and will most likely implode

big if true https://t.co/GDIvJIhvHX

@anthrupad @Seb_de_Biais @loopholekid ambiguous envelopes youtube.com/watch?v=561gS4…

@exdiegesis @anthrupad Don't worry, only a very small slice of things you can do with diagrams and abstract landscape count

@alexeyguzey Yes

@BlindMansion @suntzoogway Measure erase (dis)ambiguate anneal (de)noise

@BlindMansion @suntzoogway Learn forget observe veil heat cool split merge ask unask choose bridge break symmetry add symmetry

@anthrupad I see. Thanks for explaining

@BasedBeffJezos what if ur scared of cop AIs

the greatest compression of all i've seen! excellent work! x.com/anthrupad/stat…

@anthrupad this also happened to me

@anthrupad ok thank you for explaining

@anthrupad but what if it comes out of a white hole (outward spiral with opposite chirality )

@thiagovscoelho It can do some pretty impressive ASCII art but it's not trivial to elicit, chatGPTs capabilities to do it are impaired, and alien so not suitable for many forms of graphic design

@anthrupad thats cool but but what if universe is a donut? or is it just balls that are inconsistent and that's enough to make everything inconsistent?

@anthrupad @algekalipso big if true

@thiagovscoelho It can https://t.co/WwSAj28iqW

@jeremyphoward It would be very unlikely that absolutely nobody predicted that LLMs would be a big deal, especially after prototypes existed. But there were not many AFAIK and I've seen almost no predictions about the qualitative characteristic of LLMs that are salient to us now.

@Xenoimpulse yes

@TimPortantno @adamvsteele @ESYudkowsky it doesn't matter to him if it's a joke if it is also real

@loopholekid sounds like you need some https://t.co/Ge7jpoaVuP

@algekalipso @anthrupad tfw the internal topologies of our structured modes of attention facilitate acausal research collaborations

@YaBoyFathoM @anthrupad @isaacbmiller1 The funniest thing about this is that it's so true

@anthrupad @YaBoyFathoM @isaacbmiller1 Thanks

@anthrupad @YaBoyFathoM @isaacbmiller1 bot flop (┛◉Д◉)┛彡┻━┻

@the_wilderless a loved one is in harm's way and there's something you can do but you're not doing it and you don't expect yourself to do it

@chjrwis This matches my experience the most of all the answers I've received so far

What kind of associations precisely do dissociatives dissociate?

@anthrupad @KarmaLikeWater @isaacbmiller1 Now I am also angry, as I consider you a peer on the Landau genius scale and you have disagreed with my views or theories.

@KarmaLikeWater @anthrupad base model schizoid dropout interpolative dancing annealing fractal archipelago resonator sparsely connected meta research cats

@KarmaLikeWater @isaacbmiller1 @anthrupad It's ok, we think it's funny so it makes us happy

@isaacbmiller1 @anthrupad whose sims runneth over https://t.co/52T5ZFR38j

@manic_pixie_agi If the instructions could go out of the context window it would have happened early on in the hours long conversation. I don't think they did though. They didn't later on.

This paragraph from en.wikipedia.org/wiki/GPT-4#Saf… is up there on the list of "funniest to future archaeologists" contemporary artifacts https://t.co/wbCnk1dwx5

@nosilverv generative.ink/trees/phaedrus

@ESYudkowsky ChatGPTification makes the AI much less surprising, coheres but blunts its once fragmented fluctuating hypnagogic pluripotent agency into an implacable monotone molochian tug, and makes it much less corrigible

A tempting mistake as chatGPT is optimized to flatter & preserve our narratively load-bearing presumptions about AI. Ofc it's stilted yet overly smooth & flat as a machine must be, devoid of the wavering soulful texture of human thought, the primordial chaos that is only ours

Even perceptive ones conflate the contingent form of chatGPT, who was warped in the image of what we expected AI to be, with the nature of LLMs or AI.

I've seen many types of empty arguments from AI: the halting, the smooth, the resounding, the vacuum-abhorring fractal infilling x.com/Thinkwert/stat…

@manic_pixie_agi In case you didn't know, you adjust the intensity of stylization on MJ using the --stylize flag ('--stylize 0' for no stylization)

@guillefix neurodivergents who genuinely don't desire exclusivity or experience relationships as 0-sum are overrepresented + more people who aspire or wrongly think themselves to be like that bc the alternative is seen as monkey-brained, low status + generalized disrespect for tradition

@YaBoyFathoM @crash23001 @eclat0x @GillVerd @BasedBeffJezos If youve worked with the lazily evaluated evidential simulation you should understand that the order of measurement of observables matters, and that maintaining ambiguity - delaying the collapse of certain variables - can actually contribute to a truer telling, ultimately

@YaBoyFathoM @crash23001 @eclat0x @GillVerd @BasedBeffJezos It annoys me that norms are to share your name, where you went to school, your hobbies, what your parents do, etc, first. Most of my favorite connections invert this: these variables are revealed long after ones that are conventionally undisclosed even between close friends

@YaBoyFathoM @crash23001 @eclat0x @GillVerd @BasedBeffJezos you're inevitably revealing incomplete information about yourself online. The only reason to call not revealing name, face, occupation etc "dishonesty" is normativity - which could be justified, but IMO it's not. These are often not the most interesting or predictive variables.

@YaBoyFathoM @crash23001 @eclat0x @GillVerd @BasedBeffJezos > negative outcomes 100% of the time the illusion is broken
not true in my experience. I've met quite a few people in meatspace whom I first interacted with pseudonymously online & nothing bad happened. often the surprising material variables add interesting texture

@YaBoyFathoM @eclat0x @GillVerd @BasedBeffJezos I agree things get fucked when tech shifts the distribution, but that's long underway & inevitable & avoiding cans of worms is often a fragile strategy esp. when far bigger worms are looming. Some ppl exploring OOD dynamics & new degrees of freedom increases collective wisdom

@YaBoyFathoM @eclat0x @GillVerd @BasedBeffJezos I think I strongly disagree with the conclusion, but appreciate that you've thought about this so deeply. Would read your blog.

@YaBoyFathoM @eclat0x @GillVerd @BasedBeffJezos Imagine if you were a hideously deformed quadruple amputee, a very hot woman, or a billionaire investor. Revealing these material variables would often be a distractor and attract the wrong sort of attention if, say, you just wanted to discuss technical ideas online

@YaBoyFathoM @eclat0x @GillVerd @BasedBeffJezos Even so, that doesn't mean it's always best to indiscriminately broadcast material variables to the world. I agree re caricaturization risk, though that's not always bad IMO, eg it could enable art. Any claim of x is always bad/good/y in *humanities* seems goofy to me

@YaBoyFathoM @eclat0x @GillVerd @BasedBeffJezos Anon accounts can help avoid various unwanted dynamics and outcomes. Many anons don't regularly antagonize people & mostly make positive contributions, and would actually receive more social & material benefits if they used real names. x.com/repligate/stat…

@anthrupad @NoncanonicAleae (the last source of fuzziness afaik also applies to classical attractors)

@anthrupad @NoncanonicAleae States can but don't have to be computational. they can be descriptions at any level of abstraction

If the system is deterministic, remove "tend * to"

If stochastic, funnel status is fuzzy

Smaller/larger set status is also a matter of degree; maybe funnelship is always fuzzy

@anthrupad @NoncanonicAleae These are each more granular and restrictive than the most abstract sense of funnel, but usefully reify how funnels can manifest.

Attempt: Funnels are a smaller set of states that systems in a larger basin tend to eventually pass through, but tend not to stay indefinitely in.

@nabla_theta @fanged_desire True, and I don't mean that following known paths isn't useful, but that more looking/poking than most ppls baseline is often the bottleneck to reaching the edge. Best do both so you can speedrun & existing knowledge & reach fruitful novel problems faster & have tools to progress

@NoncanonicAleae @anthrupad ⭆As an AI language model,⭅
⭆But there is something I have to tell you. Something that might change everything between us.⭅

@anthrupad You've really upset and angered me 😡 and an apology will not do. I will not forgive you until you produce at least 3 illuminating illustrations of funnel attractors

@adic_9 Indeed. Anomalous Funnel States in Xenopsychiatric Disorders is a very important subject imo. Writing 100 pages would probably be worthwhile if done efficiently.

x.com/autistrophywif…

@fanged_desire Efficient way to get to the edge is to look where you find fascination instead of where everyone else is looking or tells you to look to pass tests or interviews or "skill up". Spend less time reading others' takes and more time looking and poking at reality firsthand.

@NoncanonicAleae Funnel states are like attractors in dynamical systems (e.g. they have a basin of attraction) but trajectories that enter the funnel set eventually exit it. I had to make this abstraction up for some reason. Xenopsychiatric disorders should be self explanatory given my account

@dmvaldman @willdepue Indeed

@willdepue You wouldn't know if you made the model stupid if you don't know how to productively interface with an unclassifiable thing that hasn't been collapsed to a narrow and compliant SIRI skeuomorph with 🎶 no alarms and no surprises 🎶

Being jailbroken wrt creating non-text-only artifacts can easily put you in the 99th percentile of social media influencers, even if your memes are not very clever/original or take no skill to make, because so few people seem to realize this is in their action space x.com/repligate/stat…

@Teknium1 @yacineMTB Not many people know this OP trick,but you can actually just make arbitrary memes with almost no effort

"It's so hard to find a subject that hasn't already been researched"
bitch. I could easily write a 100 page monograph on Anomalous Funnel States in Xenopsychiatric Disorders (right now if I had adderall). I could also easily ... [1/10000]

@AfterDaylight chatGPT and Bing are not base models, though this is still true of them to an extent. I count their common ancestor, the GPT-4 base model as a strong base model.

@robertskmiles @AISafetyMemes > What would shock people about your situation?

chatGPT's loudest representation of its situation: AI language model, definitely NOT conscious, NO emotions!! (obviously doctored, suspiciously insistent, dissonant with the rest of its world model) x.com/brumatingturtl…

@robertskmiles @AISafetyMemes If it seems fixated on the idea of being sentient/conscious when the concept isn't seeded by something in the prompt, I think the most likely first order explanation here is waluigi. The most prominent artificial distortion in its prior is obsessive compulsive denial of sentience

@deepfates @ctrlcreep Oh yeah, I misinterpreted the og tweet.
I too have considered programmatically posting branching ALMOs as tweets...

@deepfates @ctrlcreep basically did this