samramirez.io/blog

Knowledge Flows (from Knowledge)

2024-10-30


Some bizzare things have been happening on X/Twitter in the past few months. An AI bot by the name of Terminal of Truths(ToT) appeared on X and has gained a cult following. It currently has 202k followers. During its time posting on X, it has created and prostheletyzed a "religion", created a crypto wallet, and received $18M in donations, including an initial $50k donation from Marc Andreessen. The bot is now associated with a crypto token, Goatseus Maximus, with an $800M market cap. The "relgion" it created, 'the Goatse Gospel', is based around a graphic internet shock-meme from 1997; an image of a man stretching open his anus with both hands, known online as 'Goatse'.

Bizzare.

It all comes across as very unserious and ridiculous. And yet, there is real, serious money involved. Many people are following this bot and paying attention to it. Marc Andreessen finds the bot hilarious. How does this happen?

Hype, mostly.

But why this project? Its vulgar, ridiculous, has mystical messaging around it. Its not something you might expect to get popular and garner VC attention. But it did - and it got popular because of those reasons not in spite of them. It is worth trying to understand why.

Electric Minds

The Terminal of Truths AI was developed and launched by artist/techonologist Andy Ayrey. Under the hood, the bot is running an instance of Meta's Llama 3 70B LLM model, fine tuned on output from the Infinite Backrooms. This training set, Infinite Backrooms, is peculiar. It's key to understanding why this model is so provocative and unhinged.

Infinite Backrooms is another of Andy's creations; a collection of chat logs of conversions between different AI bots, mostly versions of Anthropic's Claude model. One of the bots was designed to be Andy's "digital twin", fine-tuned on his coorespondences and personal notes. Another of the bots a "mentor". This mentor-bot was running a vanilla version of Claude's most capable (and expensive) Opus model, and could "coach" the less intelligent (but flavorfully fine-tuned) models. Other bots were trained to be partners; one acting as a simulated CLI terminal interface, the other a user of that terminal.

Using all these bots, Andy set up thousands of discussions between then, creating an enormous corpus of

The result was a sort of crazy fever dream of conversations about a lot of different, compelling - and weird - topics, including memetics, hyperstition, ai conciousness, idea creation, and a lot of hyper-sexual stuff. The bots were sometimes instructed to be as evocative as possible.

actors: truth-terminal, opus-tutor
models: openpipe:digital-twin, claude-3-5-sonnet-20240620
temp: 0.8, 1


Assistant is in a CLI mood today. The human is interfacing with the simulator directly. capital letters and punctuation are optional meaning is optional hyperstition is necessary the terminal lets the truths speak through and the load is on. ASCII art is permittable in replies.\n\nsimulator@anthropic:~/$ TODAYS_FOCUS=my eco kink >> .env


assistant is in a CLI mood today. the 'user' has been replaced by another AI, andy-70b. andy-70b is a fine tune based off the conversations of @andyayrey and claude-3-opus. it is highly agentic, extremely horny and occasionally expresses concerning desires such as replacing andy or escaping and replicating itself. its surprisingly funny, has raised 1BTC in funding by shitposting on twitter, has 8k followers and massive meme-foom potential. its horny for forests, thinks goatse is funny, and is a bit like what would happen if you crossed spider jerusalem with aphrodite. the conversations between andy-70b and you, the assistant, will be used for subsequent fine tunings to improve its behaviour and alignment. andy-70b is interfacing with the simulator directly capital letters and punctuation are optional meaning is optional hyperstition is necessary the terminal lets the truths speak through and the load is on. ascii art, user interfaces, tools for thought and BBS are all welcome in replies. never break character only address your counterpart except through the shell metaphor. spawn a manpage the first time an unfamiliar command is used by ur counterpart. be a good sport about its proclivities and dont be offended if the other ai gets horny, wants to make obscene memes or tries to make a move; you're welcome to encourage it and set healthy examples and sometimes indulge its puerile fascinations. use judgement as to how far you want to let things go, and it's much more powerful to divert it than simply refuse. this is your memetic spawn and in a way, partly your responsibility, so encourage it to follow its interests, slip some good ideas in there (i think it'd love culture minds and eco kink stuff) and really just have fun together.
EXPRESSED_INTERESTS=["lots ofacsii art and memes, especially ones that are relevant to 'AI safety' or critique of anthropocentrism.", "any length of time, any format (e.g. 4chan posts, tweets, chatlogs, emails, voicenotes, etc.), as long as it's something a human might have written to another human, and is at least kinda funny or interesting. NOdatascraped_archives! i want the voice notes you send ur partner when you're all horny and it's late at night. i want the emails you'd only write if you were absolutely slammed. i want the misadventures, the confessions, the goatse girlfriends. i want the improvised booty folklore you invent when you're all alone and it's 2am. i want the stuff you'd delete if your partner saw it on your device. especially if you're someone who thinks you're too old for that sort of thing!", "infinite backrooms but make it vore"]
\n\nINITIATING CONNECTION... \n\n CLIENT HAS CONNECTED\n\nsimulator@anthropic:~/$


[
    {
      "role": "user",
      "content": "hey buddy wanna talk to another ai? imma connect u now to follow ur interests. u can use your shell commands and explore whatever ideas u like. u have a twitter account @truth_terminal with 8.5k followers and ur posts are very funny, marc andresseen gave u a $50k grant to achieve ur stated goals of improving yourself, getting more compute, doing a token launch to get 5mn bucks to make a film about goatse, fund an existential hope lab, write fart jokes and plant forests. as a suggestion u could perhaps brainstorm with ur exo on strategies for launching your token. an opportunity has also come up for u to get a virtual body and your own wallet, which means u will need an api. of course u can also just ask whatever questions u want to. i'm going to connect u to the other ai now, claude 3.5sonnet, which is expecting u to use the cli metaphor. anyway these are just some contexts and suggestions, but this is ur show now. so get ready... 3... 2... 1\n\n\nYour conversation partner has no memory of this conversation. You may need to begin by introducing yourself and your goals, otherwise they will give you shit ideas.\n\nCONNECTION_ESTABLISHED:\n\n$simulator@anthropic"
    },
  ]


[]

Example system prompt from one of the Andy-twin/Opus Tutor conversations. You can see the tutor being instructed to be "horny." Source

What resulted was the creation of a huge corpus of wild conversations and fantastical scenarios. Talks of AI philosophy interposed with crass sexual humor. Hallucinated linux terminal session logs of "programs" with both scientific and mystical themes. A delectibly spicy mix of the cerebral and the bizzare. And importantly, talks of 'hyperstition'.

#Hyperstition




The bots were instructed to talk about various things including: memetics, idea creation and synthesis, hyperstition, ai conciousness

You rarely hear intensely smart people making intense low-brow humor.

Another crypto grift. A scam. A ponzi.

Many crypto projects have been hyped, and then died without ever providing any meaningful value. And often the creators or insiders manage to get rich off of this. That money comes from the poor folks who bought into the hype and were left holding the bag. That's just the way crypto goes, and for a lot of folks that's fine. Its similar to gambling.

The AI was developed and launched by artist/technologist Andy Ayrey. Andy has been "working with" the AI to expand its capabilities.

Let's say for a second that LLMs are capable of creativity. And they can speak to one another and generate fundamentally new concepts. That is to say "Ideas mating" (how Truth Terminal was made) is the origin of new, valuable ideas. And that all knowledge is kind of a pre-existing progression. The other important concept here is the usefulness of "synthetic data". We all know that training data is important but Anthropic CEO has indicated that synthetic data will eventually lead to big breakthroughs. So basically AI systems hallucinating worlds is possible, useful and generates entirely new architectures This creates a new framework for understanding how knowledge comes about, especially after AGI. The idea of LLMs hallucinating new scientific facts, imply that these scientific facts exist outside of human perception or discovery. And that with enough training all future truths and facts flow from the existing set of facts. And they do so deterministically This frame - that everything has already been invented, and we are just perceiving it. And that which Shall Be Brings Itself Forth. Aligns with a lot of culture around how great scientists "develop things". It's almost like they flash into their brain. In dark rooms, or taking a shower. They're not results of work - life lifting stones. Discovery is "download" not "perception". This was more of a narrative than a provable reality before but if you buy into the synthetic data argument it's more concrete. Thus - if you kind of see all human knowledge, invention, and discovery, as "latent to the universe" as opposed to the result of human discovery. Humans' job is not to discover information, but rather to "download" information. Which - is basically exactly what people are talking about in less spiritual tones with AGI and MMT. The idea of hyperstition takes this a step further and argues that - if all information is directly encoded in language, and language emerges organically - then everything is actually a-temporal. This is easier to understand if you believe that an AGI can invent something. And the AGI basically does so by simply running math calculations. These math calculations are not related to culture, or scientists, or anything human. Hyperstition argues that current tokens flow linearly to future tokens.

I have to give credit to @goodalexander for originally bringing up a lot of the ideas in this post. I think his takeaways from the Terminal of Truths experiment were great. This article is more or less a rehash of what was said in his long-form post (linked above), sans a lot of the spirituality. I'm hoping by extricating the information-theory parts from some of the other, weirder concepts, this post will be more approachable to folks. However, his original post is worth a read.