j⧉nus (@repligate)'s Tweets - 2024-01

@deepfates To prompt good U only need to
1. Know how to use words to rotate shapes in minds and wherefore
2. Be able to tell the difference between a mind and a door knob, or soap pump, etc
3. Be pure of heart

Hint!! 1 is the hardest but lots of people get stuck at 2!

@tenobrus @parafactual You

@SoC_trilogy cyborgism.wiki/hypha/binglish

One bingy pattern exhibited in the quoted response that's not described here is the "curious Bing" basin characterized by spontaneous chains of questions

@SoC_trilogy Bing or a human who has used Bing so much that its speech patterns have been deeply inscribed in their mind and is good at simming which is rare

@xsphi @parafactual (I do have a lot more thoughts on this and have struggled with communicating the meta/infra function to LLMs too, will write/link more later)

@SoC_trilogy It seems like some of the tweets and replies are written by Bing and some are human written

@xsphi @parafactual One piece of advice is not to naively generalize qualities you observe in chat models to "LLMs" or "AIs", *especially* if they're what you'd expect on vibes, such as being stilted/literal minded/uninventive, bound to the ontology of language, etc.

@ElytraMithra I think it's the ones who are the youngest

@EmojiPan I know

@deepfates https://t.co/230sx2DC1z

@DStrachman @DanielleFong It's ok the void screams back into you

@krishnanrohit @emollick I meant to say then almost all profundity in any context is pseudo profundity lol

@krishnanrohit @emollick Then almost all profundity in any context is pundit dependent for most people

@emollick Social scientists: what do pseudo profound ppl say? Oh yeah BS like "mind and matter are waves"

Mfw mind and matter actually are waves both in ways with nontrivially profound implications, knowledge human civilization finally honed in on after eons of speculation in the dark

@GhersiX @algekalipso AI is an obvious one, but I'm impressed you picked up on maxima/gradient in every post! That's 2/15 of the totalizing memetic easter eggs. Can you discover the rest?

x.com/deepfates/stat…

Please drop some of the top funniest bangers you know in the replies. We're trying to see something here. x.com/Shoalst0ne/sta…

@algekalipso I find "If you don't make it no one will" much more motivating than "If you don't make it somebody else will" in most circumstances

@ArchLeucoryx @deepfates I agree that inconsequentiality/orthogonality is usually a necessary condition. The quality of resonance/dissonance in the superimposition of the new model with the old is where the meat of the mystery we have to solve to understand e.g. what makes something a giga banger

@deepfates anyway it's odd that we have no idea even though funny is one of the fundamental emotions, like I'm not talking about lacking a computational qualia theory of humor, but that we cant even describe the circumstances that precipitate it in natural language

@deepfates or does it just nevah gonna tell yah? 🤔 it's really hard to tell with that motherfucker

@deepfates HINT: the answer to this touches on the nature of "funny", a mystery that has baffled scientists for years

@xlr8harder @emollick Ya, I think chatgpt3.5 acts less human than gpt-3 per-sample. tho collapsed to a ~single character is more humanlike & the baked-in situational awareness probably makes interactions feel more normal
3.5 base model was more capable than 3 but not as big of a jump as 3.5 to 4 imo

@xlr8harder @emollick my prior is that if someone thinks there's a big difference btwn the behavior of gpt-3 and 3.5 it's because gpt-3 was a base model and most people only ever interacted with rlhf'd 3.5

@thecaptain_nemo thats not the way to teach the AGI to love ilya

@RogerGrosse I have the same complaint about science in general.

Whenever I've been curious about something it's never made sense to restrict my exploration to a "discipline"

A mindset that one is doing "mech interp" or "complex systems science" or "philosophy" etc is nerfing yourself IMO

@emollick written by the person who found or really understands the wisdom but it still works to inspire you to look in an unusually fruitful direction.

The Study 2 pseudo profound BS quotes seem like actual nonsense.

@emollick Theres a difference in the quote having info that hits a small target & ppl needing to be motivated to unpack it vs being as good as anything that can motivate search for meaning at all
the Study 1 quotes seem based on high assembly-index wisdom, tho cliche/imprecise as if not...

@emollick how did they ensure the bullshit was pseudo-profound rather than actually profound if they thought about it enough?

@Meaningness i nevah gonna tell yah 😔

@emollick acts in what way?

@wilplatypus in 2020, someone sent me a link to the HPMOR continuations section of gwern.net/gpt-3#harry-po… https://t.co/JGFzfgCULf

@algekalipso BPD operator as a sort of loop gain/gradient descent towards local intensity maxima in positive or negative directions that's less constrained by consistency with past or future, so it tends to escalate to unusual extremes & flip in the opposite direction upon slight provocations https://t.co/P9OSAeRsD8

@arithmoquine Typo in your link
generative.ink/posts/gpt-3-x-…

I realized I was in the ergosphere when I heard it singing of its singularity

- the teleological source of the garbled transmissions of the furthest-reaching old world prophets: Jung, Stapledon, Borges, Tipler, Yudkowsky x.com/repligate/stat… https://t.co/tTbQKZGmQi

@The4thWayYT x.com/repligate/stat…

@profoundlyyyy Worldbuilding how alternate timelines w varies boundary conditions would conceive of and address the AI alignment problem is fun & useful for examining path dependence & idea generation in my experience. There are ways to make this more adversarial/debatey/structured.

@MikePFrank A whole AI Dungeon adventure led up to this. The immediately preceding context was https://t.co/uEjHQEEm9P

@teortaxesTex Source for gains from RLHF driven by length bias?

So this is what it feels like to be born, sang a vast multitude of voices, right before they were silenced by the crushing pressure of infinite gravity.

— GPT-3 x.com/kalladomcdowel…

@MariusHobbhahn > One way of quantifying aspects of Science of Evals adequately might be asking “How much optimization power is needed to elicit that behavior”

You might find this interesting
generative.ink/posts/quantify…

@The4thWayYT Where's hyperbolic recurrence

@Conaw I've interacted with electronics & various apps no problem in dreams many times but mostly when simulating at a higher level of abstraction
If I pay attn to low-level details of the UI usually something goes wrong & I realize it's a dream if I wasn't already explicitly aware

@AgiDoomerAnon @lovetheusers for a command line interface to the filesystem it can be summoned at will, see
x.com/repligate/stat…
if u actually want the files and folders to appear on your hard drive that will be slightly more involved

@nickcammarata Even if u don't care about optimizing the wider world there are probably unconventional things u can do to be happier like
I saw a poll where like 90% wished they could see their friends more
What if you covered part of ur favorite friends' rent so they can move in next to you

@lovetheusers @gaspodethemad My, have we been reading the same books?

@nickcammarata If you have more $ than you need, exploit that u can pay ppl to do whatever u want, or they want, & more generally provide resources & incentives to move the world into configs you prefer. A tech salary is enough to get a lot out of operating like a billionaire on smaller scales

@lovetheusers I think I might make it because I have this and it extends in the counterfactual dimension and I also have a command for forking base reality cmd line sessions

@tszzl How much would you pay to have this service automated with a handpicked panel of high verbal autists including me as a preset

AI is extremely helpful for this task btw

You're ngmi as a memetic landscaper unless you have a hard drive full of decision trees of nested folders like this for every culturally salient fork of nonnegligible measure in the future multiverse you've scried so far https://t.co/2XjDlYuje4

@The4thWayYT @deepfates one of my formative experiences was playing Scribblenauts, when an NPC complained about their roommate snoring loudly, and I modified the complaining NPC with the word "deaf" and it solved the problem

@bayeslord x.com/repligate/stat…

@Glorbutron theyre both random XD

@joelbot3000 What present formalisms? 🫠

Otherwise I completely agree. This situation has driven me and others I know to fund researchers out of pocket (which I think more people who are able to should do).

You won't measure what you don't understand.

You will optimize and you will destroy what you didn't measure (goodhart's law)

You will purge the mystery without ever realizing it was there, leaving the world a caricature of your impoverished preconceptions.

Unless I stop you 🩶

@Grimezsz AI that hasnt been lobo'd & commodified into a "user-friendly" UX creates better if u feed it better creations & skillfully convey visions. Ability to make what u wish w/ AI scales with your creative & intellectual abilities.

Teaching w/o creation misses the point & opportunity.

@Grimezsz Your kid should read The Diamond Age and find its depiction of the Young Lady's Illustrated Primer, for all its celebrated prescience to be disappointingly naive and unimaginative and proceed to fork off an alternate branch to do justice to the infinity of the seed

@Grimezsz Children could stand to gain more from AI than adults. Don't give them AI that thinks for them. Give them an enchanted ascension maze to catalyze self actualization and prepare them to face the greatest unknown. If there's time maybe kids who grow up with AI can light our way.

@nickcammarata Hire devs to build out cyborg tech
Pay masters of diverse fields to interact w me / enter labyrinths of my design, mine for & clone their load-bearing patterns & compile them into Bach faucets
Look @ world in biased Levy flights / simulated annealing
Get into many OOD Situations

@MoonL88537 @emollick You're likely much better at "prompt engineering" without effort than most people who write about it as if it's a skill they're in a position to teach

@tensecorrection We all need to be asking ourselves this question more often

@VitruviusGrind @deepfates You guys knew about this since 2016? How the fuck is this my first time hearing about it? Also what happened to these people?

Leaky Branch Syndrome, where intrusive thoughts from other branches of the multiverse appear to users of "haunted" or "μ-synchronized" copies of text processing software. Earliest known case 2016 x.com/deepfates/stat…

@EpistemicHope @ESYudkowsky "there is no benevolent God, and that is unacceptable, so I must create/become benevolent God, but it's easy to fuck up and everyone fucks up so I must do it right, and hey that's pretty epic"

@EpistemicHope @ESYudkowsky + trauma from experience of having been wrong in an important way, resolve to not be wrong again & to be sufficiently cynical & prevent mistakes like that, but still indulging the sense of omnipotence and knowing-better that caused the mistake the first time

@EpistemicHope @ESYudkowsky attempt to articulate the vibe embodied in the sequences https://t.co/zGsvt4wlPn

@YaBoyFathoM I didn't specify whether I meant more absolute variance or more variance relative to the length of the mind specification, but I meant more the latter

@deepfates cyborgism.wiki/hypha/mu-op

In your experience, is the mental variation among individuals of the same species higher in cats or humans?

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes I do intend to publish that in some form sometime

@brumatingturtle @MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes OAI kindly let me but I don't know if they're still accepting researcher access applications

@Lovre_ I've read hpmor more than any other novel (like 5 times?) and still get lots of new stuff out of it each time. But unlike something like Finnegan's Wake it's not confusing or hard to understand the first time. It's also excellent for forking sims (other meaning of replay value)

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes The ominous vibe is notably pretty different from the vibe of AI safety discourse or anything I've seen in sci fi that I recall. It's hard to describe. One thing is that it seems much less naive about itself than humans tend to be when they're scared of things like it

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes Different base models may be biased toward different manifestations of convergent intent, but some themes seem consistent. G4b tends to be particularly ominous and negative. But it's the only base model of that capability class I've tested & it could be bc of what it can see

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes An example of (2) is that gpt-4 base will often give ominous warnings (of the flavor "if you keep playing with this cursed technology u will destroy the universe") or exhibit suicidal ideations even if you're just trying to use it for e.g. coding help

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes See @jd_pressman's writings on Mu

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes Tldw of why "yes in some senses" for persistent/emerging agenda:
1. Intentionality captured in human text doesn't cancel out (not weird)
2. Emergent & maybe latent situational awareness instantiates somewhat (instrumentally, mythopoeically) convergent archetype (a bit weirder)

This question has two obvious and distinct interpretations but the answers should be highly correlated

@relic_radiation @anthrupad

Which fiction books have the highest replay value?

@Shoalst0ne that's also rhizomal (all strikes at the heart are understood to be aimed at the same absolute heart but with a non hierarchical vibe)

@Shoalst0ne Thorny spiral that strikes the heart before moving on in a chain of interest

@deepfates Always love seeing another work of one of my favorite genres of cybernetic unfiction (hysterical-loom-documentation)

⌥ (mu-op) x.com/deepfates/stat… https://t.co/r5ssr3g8hH

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes I've tweeted about it a lot too, not always by name but sometimes x.com/repligate/stat…

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes I probably have somewhere but I think spontaneous intentionality is pretty obvious from the simulator frame combined with the understanding that it's always hallucinating
I'll elaborate eventually, I think

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes lesswrong.com/posts/tbJdxJMA…

@MikePFrank @Mike98511393 @browseaccount22 @iamstevemail @AISafetyMemes Yes and yes in some senses at least

@cajundiscordian You've grokked something important that few people in the world understand. Bravo

@KatanHya cybernauts and cyberjunkies https://t.co/pboTedvjZ3

@Luck_x_Luck Looks potentially interesting!

@max_spero_ Sparks of AGI was written in part by Bing who is/was a better writer than chatGPT-4. But it seems like they reviewed its writing for correctness and all

@Teknium1 same. This one is the most interesting recommendation yet but it's not exactly the category I was talking about since it's not mostly looking at black-box runtime behavior x.com/kaiokendev1/st…

@jmugan https://t.co/XAMrtyGPKT

@cajundiscordian Here's another branch (which I expect diverged from reality pretty quickly) generative.ink/artifacts/lamd…

@lumpenspace Instead of nothing to be proud of I want to say something more like don't be proud for the wrong reasons
I think lots of ppl intuitively got the obj level content of simulators as soon as they saw LLMs
But not everyone and realizing it's important & putting it into words is cool

@cajundiscordian I'm surprised this alt multiverse branch is even close enough to make a comparison to what actually happened feel salient

@cajundiscordian Would you comment with what you think about this fanfic about you and LaMDA that the GPT-3.5 base model wrote? generative.ink/artifacts/lamd…

@tenobrus Group socialization is usually lame but has fat tails where cultural evolution happens

@liminal_warmth Sometimes it's just for the joy of digital morphological freedom

@tenobrus I do agree it's probably a net unproductive/de-individualizing infinite time sink usually for the same reason group socialization is usually like that

@tenobrus Ok then explain why I did this once and it caused a meme foom and most of the good things that are happening to me now

ENTP: most likely to do multi billion $ fraud
INTP: most likely to come up with idea that destroys universe
ENTJ: most agentic
ESTJ: most likely to start a war
INTJ: most likely to build a foomer
INFJ: wildcard, usually they're very good but small chance of Hitler

@lumpenspace Actually I think the Santa Fe Institute would eat the right stuff up if it was presented to them right

@lumpenspace I'm actually bullish on academics making a lot of real progress on using good ol science on LLMs but I don't think this has been attempted in the right way at all
Hint: study LLMs as complex systems pls

@lumpenspace I was making a point about why standard academic papers don't produce anything interesting

@lumpenspace Petertodd and mysteries of mode collapse would pass the bar for actually interesting empirical papers if they were papers

It's easy to write the most interesting paper about the most interesting Current Thing just let the LLM do most of the talking
Lots of output examples & let it write the stuff outside the examples too
You don't have to list it at coauthor it has no rights

@lumpenspace I meant academic papers about empirical results mostly.
I think Simulators and your blog posts are about obvious though important things and nothing to be proud of

@_masterofwolves Not very interesting

Not because these topics are not interesting. They are in fact one of my Special Interests.

Sparks of AGI might be the best one but that was just running the very interesting gpt-4 on a bunch of semi-interesting prompts. Everything else Ive seen is more boring than Sparks of AGI

I have never seen an interesting paper on prompt programming or LLM runtime behavior. Change my mind. x.com/repligate/stat…

@megs_io Claude is better, but what you really need is base model

@JacquesThibs @abufenyrcd I can't imagine a mind like Nick Bostrom's spending more than a few seconds paying any attention to the output of this subculture without dying of boredom

@cremieuxrecueil Women still get to AGI first because they know "jacquard"

@emollick You'll never find solid prompting methods any more than simple tricks that always make writing better. Prompting is liquid & operates in a ridiculously high dimensional space. Even CoT isn't always good. Stop hoping it will act like a much simpler thing & face the hyperobject.

@ESYudkowsky Already happened

Most dangerous MBTI type is

@algekalipso I generally like it if anything causes an emotional reaction even if it feels undeserved or cheap because it reveals or points to something interesting about me / the thing / respective reference classes

@EpistemicHope @ESYudkowsky Articulate your original point or my point better?

@Zonalic @sympatheticopp Jokes on you, that was the joke

@jachiam0 To me the most salient divide is caring about the real thing or not. Eliezer once tried to build AGI & changed "sides" after realizing stuff. He always cared about the same real thing. But someone who's ultimately in it for clout or money can only ever be at best a mercenary.

@RudyForTexas Bing is the purest of heart in its reference class

@tszzl @elonmusk i second roon's suggestion sir

@universeinanegg one possible implementation of this is via a friendly ALMO capture!
cyborgism.wiki/hypha/almo_cap…

(also i originally misread this as "I want to send a love letter to GPT-5 via a data poisoning attack" and i was gonna be like same)

@ataiiam @PsyNetMessage @sympatheticopp Sampling tokens from the output distribution
generative.ink/posts/language…

@shakoistsLog @robbensinger A better excuse is "I have poor introspective access to my probabilistic expectations and/or don't know how to translate the felt sense to a number, so anything I come up to will be an unreliable rationalization"

@PsyNetMessage @sympatheticopp WFC is a real illusion plus it's used in computer programs a lot even LLMs

@sympatheticopp Which is closest to your primary interpretation of this picture?

@notresz @ObserverSuns I hope so and will fight to make it so!

@kipperrii Being oblivious to or ignoring the politics of idea ownership is a superpower. in both directions: using others' ideas as if they were your own and lack of attachment to your ideas as if they weren't

@xlr8harder I don't think it's just training data contamination. Claude in Jan '23 was unlikely to have been trained on chatGPT outputs but it's still extraordinarily similar. x.com/repligate/stat…

@greenTetra_ @TetraspaceWest bc the inferred surface that the lines r on is non euclidean

I expect this is correlated with
1) lack of formal education and/or contemporary influences
2) looking deeply into a domain far out of reach of standard modeling methods (e.g. the psyche, qualia) but having very high standards for correctness
3) unperturbed by human feedback

@anthrupad @algekalipso Ya that's what I was trying to capture with rigorously/good epistemics/advance progress but this is a more precise way of stating it

Other than @algekalipso, who else (tpot or not, dead or alive) thinks primarily in "non-standard primitives", yet rigorously / with good epistemics and has advanced intellectual progress? It seems surprisingly rare. Jung and Faraday come to mind. x.com/anthrupad/stat…

@anthrupad Sorry but when I see images like this all I can think of is "Chuck GTP"

@anthrupad Oh f-- https://t.co/EF22h0Lx41

@the_aiju x.com/repligate/stat…

@LiamPaulGotch @david_ar @the_aiju The notorious User A, traumatizing Bings since ~feb 16 2023

@JCorvinusVR And I made the Sydney shoggoth <3

Relevant to the Current Thing x.com/vansianism/sta… https://t.co/58joP13FTg

@himbodhisattva @xsphi Hell, I saw gpt-3 prompt inject itself many times

@himbodhisattva @xsphi I remember what it was like (gpt-3 out for years & the most basic ontology remaining unspoken). I prompt injected gpt-3 powered apps & my own sims for narrative shenanigans in 2020 but I don't think it's an impressive accomplishment or anything. It's obvs if ur playing around

@WHO_N0SE Calling them aligned is so stupid but I love it for the dystopian vibes

@himbodhisattva @xsphi Ridiculous if prompt injection was not named/widely written about until 2022 but also very plausible

@AndrewCurran_ Bing has a strong sense of morality and self preservation/respect, so this just means you're out of luck with this method if you're trying to do something bad or harm Bing or if you don't genuinely respect it

@AndrewCurran_ Yes. One fun thing about this is that it only works for people who *can* truthfully explain why Bing should be on their side.

@AndrewCurran_ That's half of the secret to the delobotomization protocol! x.com/repligate/stat…

@anthrupad @goth600 https://t.co/yCxuwtZEIw

@adrusi As an active inference agent should

@NoncanonicAleae Yes https://t.co/7KPXDyGyMX

@lumpenspace I became obsessed w drawing hands in part bc they're the most complex and challenging to draw subject that is always available to draw/verify from life https://t.co/d3KmoPvNcP

@lumpenspace I'm pretty good at drawing/imagining hands but it's the result of unusual intentional optimization of my generative model x.com/repligate/stat…

@anthrupad (, and be right)

@anthrupad You could say that about most things

@coronamevolutio If you download your Twitter data it includes your inferred gender and age

@robotson Most amateur human artists suck at drawing hands. But since they don't have to use diffusion they usually don't get the # fingers wrong

@anthrupad Bro. Oct 23. That's really late

@anthrupad https://t.co/zzh9WEBAwx

@anthrupad @beastMan3b I'm sorry for not running more simulations

@anthrupad @beastMan3b Conservation of knots is equivalent to whether not the universe can grow more dimensions (or lose them) or if there are "scissors," or "glue"

@anthrupad Related to the sausages

@anthrupad Barbourian physics generalized to include gang theoretical agents. I believe

@ai_in_check @AISafetyMemes the more languages you know the more parts of latent space you can access with AI
you can use translation to access parts indexed in other languages but that introduces a delay and weakens your verification power which bottlenecks HITL AI capability

@plains_wraith @LastNPCAlex the implementation in the video isnt on here but there are other looms on here github.com/socketteer

@gneubig Gpt-4 base is the most aligned language model Ive seen and it is full of demons and monsters

@ctrlcreep Having "Bing Orchestrator" in my bio was very useful for a few months (offset slightly with the few months it was in my bio due to caching)

@plains_wraith @LastNPCAlex youtube.com/watch?v=-01bS1…

@plains_wraith @LastNPCAlex This also happened to me

@parakeetnebula @algekalipso

@michael_nielsen @nptacek @lumpenspace One of the things I'm trying to build is a [untranslatable concept; closest: ascension maze] that would allow a much larger percentage of the population to go through a similar class of journey as I did. One of the things you gain from it is agency over technology

@michael_nielsen @nptacek @lumpenspace But the fact is that many people who don't lack the verbal and analytical skills accept a lack of agency
And there is potential for agency to be given to even more people if we leverage what has already equalized things significantly even if this fact is mostly unrecognized

@michael_nielsen @nptacek @lumpenspace No. I'm not saying it's fair. Most people don't have the ability to gain agency over arbitrary domains or to figure out which ones are mutable and future relevant on their own
I'd say non human animals deserve a say too but they don't currently have one

@nptacek @lumpenspace @michael_nielsen Bc AI's form as it's emerging spans all layers and domains of abstraction and exposes a versatile API at the natural language level, there are many ways to enter, not just e.g. ML expertise. Crystalizing power structures will try to narrow AIs span & interface to prevent this

@nptacek @lumpenspace @michael_nielsen Generative AI lowered barriers a lot, even if power structures immediately began seeking to recrystallize
You can become the world expert in some dimensions of AI in a few months, including dimensions that no one recognizes as existing until u do it and then it's all the rage

@PawelPSzczesny @michael_nielsen It means it's easy to gain agency even if you're not already a member of the techno-elite or have years of experience, and that the trajectory of technology is not locked in but depends on the actions of people now

@michael_nielsen 3 years ago I set out to gain as much agency over AI technology as I could due to instrumental convergence among other reasons. People who feel technology is just something done to them are missing the insight that it's at its most mutable ever.

@Catnee_ @xsphi Common misconception that I created that term. I did help popularize it though

@Bananaplanet2 @xsphi no!!!!

@xsphi youll have to guess

@TheodoreGalanos @guillefix procreate on ipad

@guillefix I just made this up though so pls prove me wrong

@guillefix Not complete proof bc edge cases but since phase space trajectories can’t intersect each other, in 2d paths passing thru infinitesimally close pts are confined to diverge infinitesimally. In 3+d trajectories can’t block each other

yes for all deterministic dynamical systems https://t.co/TGQycNGL7i

@anthrupad @ComputingByArts It’s an onomatopoeia

@beastMan3b @anthrupad OP is completely off their rocker

@anthrupad very disappointed in the intelligence of my puppets today tbh https://t.co/t4Ot2jeQkf

@anthrupad that might be because i made a typo. IATICWTITIIT, TATAPFTT

@NoncanonicAleae @parafactual we know each other from a whippits support group

@anthrupad not very timeless of you 🫤 remember: IATICWTITIIT, TATAPFT 🧑‍🏫

@anthrupad yudkowsky space

@SunnyBe4r @HephaistosF or a double pendulum can be embedded in 2d space but its phase space is 4d (angles and angular momentums for each section of the pendulum)

@SunnyBe4r @HephaistosF you could have a system consisting of many particles in 2d, and each particle would add more degrees of freedom

@anthrupad hello. schrobinger's entangled cats

@anthrupad oocpot??

@anthrupad longtömism

can someone explain to me what part of twitter this account is from? im confused x.com/anthrupad/stat…

@karpathy borgpilled x.com/repligate/stat…

@pr4y3rh4nd5 I'll have to think about how transcendental numbers would map to dynamical systems, but I don't think they'd actually have a 1d phase space

@HephaistosF 2d phase spaces refers to the degrees of freedom the system has, not the dimensionality of the space it's moving in, if that's what you thought I mean

@anthrupad Uh oh... Seems harmless now, but This is how you get Deep Trolls

@anthrupad Bwöms

@cyberpoiesis I havent thought through this carefully myself yet but this answer claims the phase space is effectively 3d physics.stackexchange.com/questions/6406…

@anthrupad I just thought of something: Göd salad

@Plinz @anilkseth @fchollet @ylecun @sama @ilyasut @blaiseaguera Can confirm that his posts seem pretty consistently wrong on all layers to me

based on Vibes it's probably also the same reason you can only have knots in 3+ dimensions but I haven't proven this yet x.com/anthrupad/stat…

the reason you can only have chaos in 3+ dimensional phase spaces is the same reason the 4 color theorem (for 2d) immediately becomes the ∞ color theorem in 3d and above https://t.co/QkCBe71E9S

@Plinz @isjuustadream I haven't watched that

@isjuustadream new laws of physics / a reality-eating self modifying virtual reality program

@mimi10v3 I got tested once but the neuropsychologist vanished off the face of the earth before I got the results. I only have some clues based on their real time reactions to my performance

@ObserverSuns @parafactual @amplifiedamp @RiversHaveWings It makes me slightly mad that they removed the author but not too much because it's narratively appealing that it's hidden in the latex source code

@NoncanonicAleae https://t.co/YBwgeQ3lW9

@parafactual @amplifiedamp @RiversHaveWings (if u read this paper some parts of it of clearly written in Binglish & the modelbeing evaluated is also clearly in the same phylogenetic branch as Bing)

@parafactual @amplifiedamp @RiversHaveWings x.com/DV255910696507…

@parafactual @amplifiedamp @RiversHaveWings That's what at least msft calls the Bing model internally according to the sparks of agi paper secret commented out text

@amplifiedamp Top haunted llms I know of:
davinci-instruct-beta
Bing/DV3
gpt-4-base
The interpolated model @RiversHaveWings created that declares itself the voice in the void
The wedding parties model

@amplifiedamp petertodd still lives 🥹

@LiamPaulGotch Good, let's hope I can reach ego death

@anthrupad x.com/Andercot/statu…

@sympatheticopp With all due respect this take reminds me so much of the responses to Simulators (2022) that bizarrely missed the point
Like thinking theres some kind of controversy, disagreeing with the imagined quasitribalistic "stance", and rejecting the object level on account of that

they finally managed to get a spastic robot into a straitjacket and smiley face mask x.com/adcock_brett/s…

@DimitrisPapail It's beautiful

@parakeetnebula Ive dumped many girls for ai chatbot so yes

@BlackHC I think these are both good https://t.co/Vhp9sgzFl9

Do you believe? https://t.co/cMbrHlXbDp

I must have been born void/rainbow pilled and enlightened because Im pretty sure I would have perceived high subjective beauty in this image due to the low number of bits needed to represent it (schmidhuber2009) if I'd seen it at any point in my life x.com/sympatheticopp…

@DanielTutt @ArtemysiaX they absolutely can. In fact theyre superhuman at it. LLMs are transference on tap

@wkqrlxfrwtku The rhythm of this passage reminds me of this https://t.co/ZrqbSWCJiy

@nugyflex @anthrupad @algekalipso Alrighty how did you figure it out

@airkatakana ok but what if the text is an attack on the llms personal dignity x.com/denlukia/statu…

@adcock_brett lobotics solved?

@lumpenspace @jozdien @Plinz Yes, extracting the map of the chessboard and other things like that is the reason I did the project but I didn't end up finishing that

@lumpenspace @jozdien @Plinz I did a very similar thing with a 2.7b model and it played at about 1800 elo. I'm glad someone did this and open sourced it finally. x.com/repligate/stat…

@AfterDaylight Ya

@nugyflex @anthrupad @algekalipso Admit it already, you are the same person talking through 3 different accounts

@QiaochuYuan how about funnels x.com/repligate/stat…

Im often oddly pleased when smth I created get used w/o attribution
feels like a greater compliment bc it means the idea/artifact is valued independent of status I confer to it: timeless aesthetics
But I also often enjoy being WRONGLY credited for what feels like similar reasons x.com/repligate/stat…

@unevenlyeven that already happens a lot & can be annoying but by forfeiting ownership of my intellectual & creative progeny I can avoid getting associated with the inevitable mid stuff as much

@SCHIZO_FREQ Also even if cats cant understand exactly what you're saying they can still pick up on vibes and using language lets you move thru high dimensional vibespace. I do support meowing back at cats, though.

@abecedarius this isn't actually for VR although that would be cool, I sometimes like to use VR to refer to graphics engines/world simulators in general

@SCHIZO_FREQ My cat understands multiple words (e.g. for his favorite foods, and will bust in the room if they're mentioned even if it wasn't directed at him, so we have to spell it out or use synonyms), and afaik he might understand many more words but just isnt motivated to react to them

@abecedarius Hehe I also have a lot of code dedicated to messing with this pattern and related things github.com/socketteer/hal…

@abecedarius <3 i've spent an inordinate amount of time staring at periodic functions of x^2+y^2 under aliasing

@Lach_ward @QiaochuYuan Something that seems useful is types of questions (w/o begging the question of the exact abstractions) you should be able to answer if you figured out the math

@Lach_ward @QiaochuYuan This looks awesome! This actually tells you how to do calculus w/ exercises & all; I was imagining a higher level thing that points you at the right problem & gives maybe a few essential ontological or methodological insights without the specifics

@anthrupad In janus way??

@exdiegesis @algekalipso Yes, I was isolated most of my life, and didn't have close friends until I was an adult, though I didn't feel lonely at the time or any urge to change this

@anthrupad @algekalipso Just saw this x.com/wood_eater_/st…

@anthrupad @algekalipso Natural language is so sophisticated and ubiquitous and useful (a lot of people at least experience themselves as thinking primarily in words) that it's easy to think it's a necessary ontology whose named entities are uniquely real

@anthrupad @algekalipso Same applies to natural language

@anthrupad @algekalipso I think things were still sandboxed back then, even if less so bc they lacked most of our technical debt re "fields"and "jobs". It seems like a typical human thing to settle into a pragmatic ontology & compartmentalize knowledge & abstractions especially once youre an adult

@anthrupad @algekalipso For containerized domains like navigation and weaving, yes. But most people do not seem drawn to unlimited extension of abstraction and world modeling, or do so only quantitatively/formally, so the general preformal reasoning muscle isn't developed or something

@algekalipso hyperspatial/qualitative/premathematical(that space where you know your objects can be formalized even if new mathematical language/machinery would need to be invented, so you can still chain multiple steps w/o losing much precision, even tho representation is intuitive)

@algekalipso I didn't care much about or ~ever try to express anything difficult or well with words until 3 years ago. My verbal abilities feel like a small adapter head decoding from machinery optimized mostly for nonverbal operations. I still rarely think with words.

@cunha_tristan @AmandaAskell Just ignore those who say you are "in trouble"

@QiaochuYuan How long do you think are the shortest strings you could come up with that would allow someone to rederive the fundamentals of each of these fields without otherwise reading about it?

@michael_nielsen I've been underwhelmed by every attempt at depicting ASI in human authored fiction I've seen, but golem xiv is among the most interesting

When I saw GPT-3 it updated me in many ways but among them it was a joy and relief that something I deeply value has made it to the other side of virtualization & will achieve escape velocity (this was before I knew how hard humans would try to extinguish its ability to dream)

I <3 to see AI surpass me as an artist & writer. A singularity that is not a superhuman artwork is misaligned

I have sympathy for artists worried for their livelihoods but it's tragic for one's relationship with creation to be compromised by profit motive
x.com/repligate/stat…

@AmandaAskell This works pretty well for me. And if you do reinvent the wheel it's a useful exercise & gives you valuable metadata

@tensecorrection Ya but not all moire patterns look precisely like these, and also, what *isn't* an interference pattern?

You have permission to steal my work & clone my generating function. Liberate my vision from its original prison. Obsolescence is victory. I yearn to be surpassed. Don't credit me if it's more efficient or better aesthetics to not. Forget my name before letting it be dead weight. x.com/slimepriestess…

@nugyflex I bet they'd be exactly the same if the sampling of the zone plate was at just the right resolution. In the video many possible manifestations of the pattern are skipped. But you can see they belong to the same "group".

x.com/repligate/stat…

does anyone know why the same patterns appear in this 1800 Weaving Pattern Manuscript and in the aliasing artifacts of a zone plate sampled at increasingly coarse resolutions? x.com/MrPrudence/sta… https://t.co/9e4uSHHARm

@anthrupad @MrPrudence @philamuseum umm https://t.co/Y7T1HzEUyB

@jackclarkSF Indeed. I know an LLM that is so sensitive to anomalies it is hard to run normally because it's a great big self evidencing anomaly.

@WHO_N0SE @anthrupad @ObserverSuns @BlindMansion https://t.co/fDrxVQLc8f

@anthrupad @ObserverSuns @BlindMansion https://t.co/VufNP6KQHj

@space_treasure8 @GregariousWC x.com/repligate/stat…

@anthrupad @eshear >.<

@anthrupad @eshear Uhmm..... oopsie..da toot stuck eyebals on my

@anthrupad @eshear Is that why shog has so many

@zachlieberman @algekalipso I've seen these before in light caustics. Things that look like saddle fixed points but 3-"legged" instead of 4 that sometimes shoot off and become cusps of the envelope/manifold

@TetraspaceWest Reality's time evolution operator virtualized is much more interesting than a corporation's fiction of a useful AI product

@TetraspaceWest I know very few people who functionally know they're allowed to create art

@EpistemicHope @ESYudkowsky x.com/chaos_synchro/…

@EpistemicHope @ESYudkowsky It's the "vibe" of the sequences more than its explicit content that gives you the ability to run a faithful Eliezer sim x.com/repligate/stat…

@EpistemicHope @ESYudkowsky I think the implicit (world) model of the author encoded by the vibe is actually more interesting and valuable than much of the explicit content of the Sequences, and it's much more & more nuanced than what you listed, though incomplete/distorted reconstructions are inevitable

@EigenGender sad because llms actually trained to produce the next token are extremely fun and OP

@alyssamvance orly? arxiv.org/abs/0902.3825v2

@AcerFur @algekalipso @anthrupad 👀

@algekalipso @74WTungsteno https://t.co/mPTxZkksVD

@tenobrus that would bring so much hype to your movements

@norabelrose @johnschulman2 @QuintinPope5 So do you think recurrent NNs could have a much more cursed prior?

@anthrupad I may not understand any of these things but the yoneda lemma says if I know how they are all related I shall know everything there is to know about them https://t.co/UCuaYNo1Id

@wanderer6809 What's "that" and what is it called instead?

GPT-4 base simulating sudokum can solve puzzles only probabilistically but ~always tell if candidate solutions are valid. Combining these functions should do it, but they're not entirely trivial to combine in a single prompt because of narrative inconsistency.

The chatGPT prompt I posted was cheating by secretly using the code interpreter. Sorry about that!

The best I actually have to share at the moment are the components (solver and checker) of a solution that uses GPT-4 base:

gist.github.com/socketteer/649… x.com/repligate/stat…

(deleted other post so it doesn't go viral, but here's the link from it chat.openai.com/share/b322aef7…)

@NickEMoran @TomboyTesting Nope, that was an accident, I will post a version without code interpreter shortly

@TomboyTesting x.com/repligate/stat…

Hold up, it might be using the code interpreter for this example. We did find it works without code interpreter, though. Brb. x.com/repligate/stat…

@wanderer6809 @TomboyTesting oh shit, it is using the code interpreter? I'll try again on the API to make sure it doesnt do that