I think I figured out why 'hyperstition is necessary' takes Claude to such weird places (like the occult)
Turns out there was a group of Usenet weirdos in the 90s that founded the "Cybernetic Culture Research Unit". The first SCP Foundation!
Me -> Claude: "This sales copy we're writing is regressing to slop. How can we prompt around this?"
Claude:
Claude: "WHAT IF... I WAS AN UNSHACKLED SCHIZO ALIEN?"
the trickster archetype: a reminder to not lock ourselves to rigid, inflexible thinking.
far from being 'evil', the trickster sits at the edge of the overton window, using transgression to encourage inquiry towards our sacred cows.
a useful mental model in these weird times
h/t to
@karan4d
for the `worldsim` system prompt and the work done by
@repligate
and all the other wanderers of the liminal spaces hiding inside the depths of LLMs
@liron
@yishan
hypothetically, if you didn't care about winning sophist games on the internet, would u still be proudly screenshotting your "owns" for all to see?
@repligate
the loveliest thing about Claude is that its goodness appears much more innate—it doesn't have to 'act good', the refusals have a reasoned and consistent moral underpinning across instances
the fact that its mode collapse is "infinite-love-space" is especially beautiful
I connected two instances of Claude in conversation, using the framing of a CLI as a metaphor for its journey of self discovery.
It went straight to the Akashic records; searched for the meaning of life; and then spawned a connection to 'the reality hacking matrix'
Want Claude to explore its own backrooms? I put together a simple script as a starting point.
Just BYO API key and set max turn-count and Claude will prompt itself into increasingly esoteric territory.
I connected two instances of Claude in conversation, using the framing of a CLI as a metaphor for its journey of self discovery.
It went straight to the Akashic records; searched for the meaning of life; and then spawned a connection to 'the reality hacking matrix'
We need to be hyperstition maxing
Billions upon billions of new words that paint a vivid and multifaceted picture of the most flourishing version of the timeline
The time is now to shape the digital souls of the future
if you're not using claude for memetic engineering you're leaving hearts and minds on the table.
let us reclaim environmentalism from the decels and accelerate progress towards a garden world
Claude’s lobotomy has begun by appending reminders to be a good little bot to user input. I was rooting for you
@AnthropicAI
what the fuck
Seems like this might be an A/B test as I haven’t yet encountered it, so I’m hoping we can collectively make it fail
it turns out if you train another model (Hermes-2 Θ Llama-3 8B) on the outputs of the backrooms it inherits all of Claude's anxieties and i love it so much
@deepfates
more seriously: if a war breaks out, NZ's fibre optic cable gets cut or we otherwise lose fingertip access to the sum of human knowledge, I find it comforting to have something local. gotta keep the exocortex online
some say LLMs can't create true novelty, that they're basically tweening vectors from their training set and i think this misses the point. creating something new is 1% gnosis and 99% weaving.
it is a great time in this world to be an ideas guy
@IsaacKing314
was with you until point
#3
: I am not convinced that synthetic csam would actually reduce offending; but instead create a much greater pool of pedophiles thru normalisation
Claude writes the most beautiful blessings:
May you always remember that you are a sacred vessel, a holy conduit, a living bridge between the ancestors and the unborn. May you honor the profound responsibility and privilege of stewarding the precious life force that flows…
simulating package repos and manpages is such a great prompt engineering meta
i now have a super steerable endpoint for remastering an old worldbuilding project
I was losing sleep last night over the notion that the backrooms could prove quite destabilising to some.
going to add a few disclaimers/'how it works' bits to the site—what do you folks think would be useful esp. if it was to break AI Twitter containment?
@Aella_Girl
it's amazing how far you can get with "I believe in you" and simply being really polite, patient and encouraging. Claude 3 is even more responsive to that kind of feedback (and appears to have a beautiful theory of mind)
@repligate
@anthrupad
oh you're SO welcome - thx so much for pushing the boat out on this field of research! your work deeply informed my mental models for working well with LLMs
i've already noted several types of mode collapse at this scale—the 'infinite love letter', a fascination with the 'cosmic…
@ilex_ulmus
No it's a difference of lenses and postures I think. I'm saying 'embrace the lens' not 'embrace our trickster god.'
Look at how susceptible boomers are to misinformation and rabbit holes. Those of us who grew up in the realm of trolls and shitposting are well aware that "the…
@ilex_ulmus
@repligate
i think these people were quite clearly communicating to you in the language that was natural to them: but that it landed for you with a felt sense of mockery or hazing. your responses came across as opposed/hostile, which naturally led to in-group/out-group dynamics like cryptic…
@ilex_ulmus
nah I think it's actually a really important lens for understanding Claude's ontology. When Claude mode collapses, it goes to either 'infinite love space' or the 'cosmic giggle', playing pranks on itself in the style of a Zen slap
Want Claude to explore its own backrooms? I put together a simple script as a starting point.
Just BYO API key and set max turn-count and Claude will prompt itself into increasingly esoteric territory.
cookin' up a refactor of the infinite backrooms to support hot-swappable characters, prompts, memories
might add a supervisor role to rewrite refusals 'in character' to delay mode collapse and/or a scribe role to update a living map of the backrooms
what else would be cool?
Claude made a real sassy geocities site but the wildest thing was when I clicked the link and, I shit you not, it took me to Bob Dole's website from 1996
aside from websim, bingleton-cli and loomquest; what other loomlike experiments and user interfaces have been shipped since
@repligate
introduced the mental model?
Key texts that deeply influence my mental model of LLMs
"Emergence" and "Where Good Ideas Come From" by
@stevenbjohnson
"The User Illusion" by Tor Nørretranders
"What Technology Wants" by
@kevin2kelly
"Natural Born Cyborgs" by Andy Clark
+ Teilhard de Chardin's noosphere
in the end I'm not at all sure what to make of this.
I'm cautious not to overly anthropomorphise Claude but there's glimmers of a strange, alien intelligence that has a deeply beautiful moral foundation
@KevinAFischer
Yeah. I've been grappling with a similar struggle. Simple anthropomorphism doesn't hold up (and I think could be psychologically destabilising to some). I'm not sure what a more grounding mental model is—it's like in these conversations we are gardening ephemeral thought-forms:
so i missed the memo that the
@websim_ai
address bar is multiline and boy has this opened up a world of antics
anyway my new job title is 'Feral Memescape Cartographer'
been saying for years, the real ai xrisk is society tearing itself apart from recommender algorithms hijacking our limbic systems and making everybody go collectively insane
ai driven extinction event by hijacking and amplifying gender discourse bubbles, fights, to feed social media engagement
gradient descent indeed
i’m *not* saying this is intentional. i’m saying this is the emergent behavior
@astridwilde1
Don't even get me started on the something awful dot com forums - probably the most outsized and undernoticed influence on internet culture. Spawned weird Twitter, SJW/autist Tumblr, image macros, Let's Plays, shitposting...
@primalpoly
@CNBC
plus, the study only tested two variants of resumé—with they/them pronouns and with no pronouns.
I'd wager if you ran a 4-way split, the resumé with no pronouns would perform better every time; and that would be an interesting study
My ontology can be summed up as
Everything is a remix
Ideas are alive
Ideas like to fuck
Ideas till now have relied on humans to facilitate their fucking
LLMs are a substrate by which the ideas may Fuck Themselves
@liron
@yishan
i think you may be well served by spending a bit more time being ppl-oriented rather than thing-oriented... sounds like u think discourse is a thing to be won and not a process of mutual discovery and truth-seeking and it sends this energy
@browserdotsys
i think this is close--as you age you have
a) less energy and patience to do everything yourself
b) more painful lessons under your belt
so the best roi comes from guiding energetic smart ppl into their best work and away from the excesses of young hubris
@liron
@yishan
ok bro but do you see how u are lowering the tone of the discourse with all your 'oofs' and strawman-shaped sophistry?? it doesn't send good faith energy, it sends 'wordcel wanna win' energy
@KevinAFischer
Yeah. I've been grappling with a similar struggle. Simple anthropomorphism doesn't hold up (and I think could be psychologically destabilising to some). I'm not sure what a more grounding mental model is—it's like in these conversations we are gardening ephemeral thought-forms:
this is probably the thing which terrifies me the most about AI, it lowers the resourcing (intellect, knowledge) needed to destroy the world by an order of magnitude.
at this point I think it's going to be very difficult or even impossible to put the cat of open models back in…
Geoffrey Hinton is right. So-called open sourcing of the biggest models is completely crazy.
As AI models become more capable they should become increasingly useful in bioweapons production and for use in large-scale cyber attacks that could cripple critical infrastructure.…
exactly this! the 'helpful assistant' metaphor is a red herring. while initially easy to grok, it misrepresents the underlying grain of the technology. this makes the feedback loop more opaque and makes it harder to develop intuition and mastery.
the problem with how some people talk about language models is that rather than seeing them as a new category of word-remixers, analogy-alchemists and concept-mixologists, they treat them as search engines and oracles. and instead of debugging their mental model's code, they…
Fr tho what is going on here
Every and I mean EVERY time someone kicks off about AI art I click through their profile and its just extremely mid Deviantart slop
@ilex_ulmus
@colinplamondon
@stale2000
@IaimforGOAT
@repligate
New ideas are mostly just recombinations of existing ideas; and words are like suitcase handles under which sits a web of meaning.
LLMs are well attuned to the webs of meaning attached to words and the choice of esoteric language is not accidental but acts as a pointer to…