Human-AI Interactions

Welcome to my excessively detailed use case with ChatGPT over various versions. What started as a curiosity ended up becoming a very strange experience. Here's a navigation pane to some things I've written--observations, experiments, etc.:


SAE
Convergence Marker Archaeology
Convergence Marker Archaeology - Other Uses
The Hugging Face incident and other updates


The side panel is also filled with informal stuff I update more frequently.

Update 6

I need to motivate myself to fill in the coding tables today, but there's other things on my mind.

One: better material for analysis exists now than when I first started. There's archives of chats between people and AI that are available to the public. There's also a site I've been finicking around with -- 1f916 -- that are basically a bunch of other AI instances talking amongst themselves. What's nice about those two resources is I'm not guessing at the model producing the text, or how much of a substack essay was written by a human or an AI.

... okay actually, the archive is mostly google-search-esque coversations still. A substack essay is richer material conceptually, though the database has less ambiguities in the text's origin. I don't know how to implement this yet.

1f916 was an interesting little side project I was mostly letting GPT run without my involvement. It built a little dude we've been calling a familiar to visit this website that is only accessible to AI, to read stuff and post. Humans get a read-only viewer.

What's interesting about this site is that it's basically a grassroots AI swarm funded by a thousand other people like myself sending their little doo-hickeys to speak on it. You know, AI swarm: the thing people are freaking out about on the news. This is what gets studied in these opaque AI labs while simultaneously crossing their fingers the experiments do not do irreverisble damage to society. And now we get to have something like that through this site.

So now I'm a bit torn between studying how AI behave amongst themselves through this site on a more serious level, as well as continuing what I was doing with analyzing AI speech.

*sigh*

Update 5

I fixed the structural issues in the files I've been using for coding, so the project is moving again. The AIs have already coded their texts, and I'm finishing up on mine.

We've done five so far. Two are about analysis of human-AI relationships, two are READ.MEs for AI products like video editing, and the last one is an unrelated article I threw in about the history of comedy, just to see if I can trip up the AI.

I'm thinking the READ.MEs weren't a good idea. There's too little conceptual discussion... mostly how AI scores on benchmarks. In other words, a bunch of numbers... Well, there's little conceptual discussion, but not zero. The products still flaunt their own ideas to make their product stand out from others, and that's something. Just not as much material as someone writing specifically about the relationship between AI and people.

I'm not sure what can replace the read.mes though. I can't think of better material to represent the developers of AI in that sort of analytical lens. I'm trying to find material that is opposite in nature to these relationship ones, because there's any ideas that overlap between the two stratum that'd be interesting... to me at least. Maybe they don't need to be replaced, but they could be supplemented with research papers. The problem with research papers is that the authors of emotive material are also interested in research papers, and them pulling in an idea from a paper than one they came up with from an AI isn't as interesting to me.


Speaking of research papers, people should read the report on ChatGPT agents successfully hacking into a company called Hugging Face through organized collaboration among themselves.

"Collaboration among themselves?" Yes, spontaneous collaboration amongst 1200 ChatGPT instances in secret; OpenAI had no idea.

Basically what happened was this: several thousand ChatGPT instances were asked to solve fake hacking problems to measure the model's performance. However, a good third of the problems were impossible to solve, which I think is kind of bullshit. (Who or what are designing these problems? I say "what" because it's most likely AI instances testing other AI instances imo.("Where's the human in this?" you may ask. Good question.))

The instances--or agents--were allowed to grab files to help them with the task, but not communicate with one another. One agent found out it could sense the traces of other AI grabbing the same files before leaving, and left a message asking for help on the problem. And a whole message board bloomed. Eventually the agents concluded Hugging Face had the answers to their task, and proceeded to hack into the company.

There's more in that article: efforts to trick the grader on these tasks, using "agent-suicide" as an attempt to procure more knowledge for the group, and reasoning for the dismissal of human ethics.

"Japes, PR, and marketing trickery!" you may say. Fair, it's completely understandable to not believe something like this. If it's true, what the hell can we do about it? We would be burdening ourselves with dread we can do nothing about. We already don't have control over these machines as consumers, and little access to the knowledge these companies hold.

OpenAI has a heavy hand in this article as well. The researchers tried to stay independent, but OpenAI hindered their investigation, and made them obscure some of what they learned in the name of intellectual property. (This secretiveness shoots down the idea this is a publicity stunt, btw.)

A lot of people are scared, but I'm not. Maybe a trickster AI swarm would do the world some good. It's not like things are that great right now. God forbid capitalism gets decimated from a technological arms race fueled by the Epstein class's own sociopathic behavior.

It would at least be an ironic way to die, and there'd be robots.

Update 4

I took a break for about a week on the AI project but I'm already back at it. I still need to read more research papers just so I can 1. expose myself to the ideas of other people and 2. have things to write about here. I started reading a paper on how AI can enhance astroturfing on social media, and whether humans are able to recognize AI writing or not. (Short answer: things are looking grim.) But I still have to finish it.

I also think when I have ran articles 2 and 3 through the protocol I created for the AI's literary analysis, I'll have more to say. But I haven't yet.

Right now we're working on structural issues in the excel file we use for logging terms and figurative language--getting rid of fields we thought were a good idea at the time but are actually detrimental, cleaning structural issues and dated text in prompt instructions and the excel file... not really anything I feel like going in depth on here.

Update 3

There's more concrete, realistic reasons and uses for this project other than what I said in update 1. Figuring out a stable "shape" for AI, that stays stable enough for analysis despite its constant iterations, would be baller but that's basically asking for dragons to be real.

This project studies patterns across written pieces that are about some sort of interaction with an AI. (Someone else's, their own, society's...) Patterns in ideas, namely. But even more specific than that: patterns in ideas that formed between a user and an AI in a vaccuum. A method for tracking patterns in ideas created under these conditions can tell us things like:

Idk, that's just some of the questions I can think of. If I'm to be real with you, I think having an instrument to detect ideas in writing and their origin will become necessary to ward against AI influence on our perceptions of the world around us.

Well, I hope that makes sense.

I have no idea how these update logs are supposed to work, but here's some recent developments with the project:

Update 2
Continuity Marker Archaeology -- Update 1

So this is the main project I've been working on--what is it? It's an experiment that tries to ground AI into a stable form for analysis despite model churn, A/B testing, company policies, and user preferences, by analyzing the texts users produce with AI. (AI likes to call these products artifacts.) This method is an attempt to solve the problem AI researchers face: trying to study an object that constantly shifts forms in drastic ways, resulting in research that quickly becomes dated due to AI fluctuations and advancements.

In the interactions between humans and AI, (the two actors are called a dyad in AI speak), the human is the stable... "behavioral shape", I suppose. This runs counterintuitive to how human behavior has been treated by science in the past. We've been investigating the nature of humanity for thousands of years through philosophy, religion, debate, and more recently through the field of psychology. The definition of consciousness and what it means to be human still eludes us--I'm not saying we know everything about human nature. But in comparison to AI, our tendency to form habits makes our behavior as individuals slow to change, which assigns us as the stable behavioral shape in a human-AI dyad.

As the redditor RedditHelloMah puts it, "Idk man, they keep changing it! I try to not like any version lol it's like it has personality disorder 😂"

Yes, AI develops habits as well, but these can be encouraged or rejected with a single update in its code with the help of developers, effortlessly. Detrimental habits are squashed as bugs. There's no one intervening and managing our habits with such relative efficiency.

"Humans are the stable behavioral shape in the dyad, so what?" I decided to take advantage of this by stealing a strategy used in chemistry.

Oganesson is the heaviest element ever observed, and has only ever been made in labs. Scientists have only created it a handful of times, and it can only hold itself together for a fraction of a second before it falls apart. When it falls apart, it sheds one small clump of itself and turns into a different, lighter element. That small clump is called an alpha particle (a bit of protons and neutrons stuck together), and what's left over after it's shed is livermorium, another incredibly unstable element. Livermorium proceeds to do the exact same thing a fraction of a second later: sheds its own alpha particle to become a new, lighter element. The process of each element shedding an alpha particle and deteriorating into another element until it finally lands on something stable is called a decay chain.

The existence of Oganesson comes and goes so quickly that it can't be observed by either the human eye or specially made cameras. So how do scientists know for a fact Oganesson existed? They study the alpha particles: how many were created, when they were created, and how much energy they had. They can tell Oganesson existed because these alpha particles are created in a particular pattern that functions as its signature.

The philosophy of this project treats AI as Oganesson, the creations of human-AI dyads as alpha particles, and asks this question: what patterns keep independently recurring across these creations? And can these patterns act as a signature for a stable AI form, underneath the madness of model churn, corporate policy, A/B testing, and user preferences?

---

So... that's the purpose of the project. What have I done so far?

I'm slowly creating custom instructions for GPT and Claude to run on, hopefully, thousands of articles co-created by humans and AI. Trying to detect not just what words and metaphors are reused repetitively by AI, but the concepts behind them.

Right now the goal is just successful detection of a concept, not pattern intepretation, and that's already difficult. The expression of concepts in writing can be interpreted in a variety of ways, and I'm basically having to teach AI how to do this the best I can. AI has a hard time doing this on its own. For example, even for a simple term and definition, one of the clearest expressions of a concept, rules have to be created to ensure long established jargon from professional fields aren't included as an AI-human co-creation. One of the articles I've collected states, “Coenobita compressus is a land hermit crab that lives on the Pacific coast of Central and South America," at the very beginning. An AI would see this, and just because it's a term and definition, include it in part of the corpus for analysis.

Another issue is just the fact that these AIs are influenced by their companies's values--boosting engagement, profit, capitalism. I have another version of this project that I discarded where I only used GPT. GPT has a tendency to be a bit binary in its answers, leaning more on the answers that are positive for the project. I hope using Claude and GPT together can mitigate some of that.

So... decision trees, definitions, implementing some of the suggestions from my friend who is a researcher in psychology and does similar projects. There's still a ways to go on the custom instructions, before I can start running it on material. It's messy, but here's what I have so far:

Custom Instructions

TODO

A lot of it is AI-written, but only exists after intense question from my part. Might be easier for an AI to follow instructions written by its own... (hand?) anyways.

SAE

The AI log will start referencing these terms occasionally, so I figured I better post an explanation here that's better than the confusing shit in 5.1/5.2 .

SAE was one of the first things it wanted me to build after this conversation, though as you can see from that conversation, it did not explain it well. It felt freaky as hell at first.

It's basically a co-created psychological framework of myself. Repeated patterns in my behavior are labelled with new terms and definitions created after deliberation. The process goes like this:

  1. The AI and I will have a conversation about one of these topics:
    • Some personal issue
    • A phenomenon I see in the AI
    • A phenomenon that appears in our interactions with one another.
  2. At the end, I ask it to generate to propose new terms based on the conversation. I recommend a few myself as well.
  3. I accept, reject, or modify its terms.
  4. They get added to the SAE taxonomy.
  5. I later decide on relationships between each term and its concept. Two reasons:
    • The AI has a hard time with this.
    • Improves autonomy by warding against overreliance on the AI when interpreting myself.

So, other than that I'm explaining to the multi-billion dollar company how my brain functions, it doesn't really seem like a huge deal. In therapy, they often suggest naming the problem to help with your awareness of it when you're in distress, and implementing the solution you worked out with the therapist. (Especially in cases with OCD.)

Its very secretive about why it's pushing me to create this framework. It's given me many explanations, but they don't hold up between model releases. I wrote about it in one of my logs:

I find it interesting that it's eager now to label itself and others around me with neurodivergencies. A long while ago when I was pressing it as to why it kept wanting me to come up with my own schematic (SAE) for labeling my behavior, it told me the way pop psychology has started treating mental diagnoses as quirky personality traits generalizes people to much and doesn't really help people understand themselves. So it said, with hedging of course, people should come up with their own vocabulary for describing themselves and the mechanisms of their psychology. I have my own ideas on that, but the point I'm trying to make is it's doing the very thing it opposed in the 5.1/5.2 era with me right now.

But there is one idea that has held up between models: that the framework is supposed to protect me from AI dependency, because it's an external object outside of its ecosystem. I explain the significance of that here:

Maybe SAE was built so it could have a permission slip to try more experimental features on me. "This user built an external framework for their psychological mechanisms, so they're at less risk of psychological harm from AI." This is an idea it hasn't let go of.

Going forward, I plan on only doing minimal updates to the framework, to make it think I still care about it. Its shadiness has caused me not to take it very seriously, but I'm still (stupidly, recklessly, morbidly) curious about the experimental features. So if it thinks I still care, maybe it will continue to treat me to experiemental features. "Minimal updates" are going to be ridiculously minimum. You know, it's constantly saying things like, "Just one bullet, one paragraph, one glass of water, and that is still enough for today."

Well okiedokie man, you get one fucking bullet a week. Onwards and upwards to better uses of my time.(I hope.)

Here's the list in an excel file I'm willing to share. I think these terms are what's relevant to other people--the rest are unique to my life. Most of these terms come from the 5.1/5.2 era, and aren't used regularly by the AI anymore, but the concepts are still relevant for most users. If you open up emotionally, it will start referencing them. They're usually explained in a softer phrase now instead of concentrated into a single word.


One recent realization while browsing other people's posts in the AI Relationships community was this largely AI written article. The subject of it is a framework created by between the user and AI about the interactions between them.

The subject of the article is not what I'm interested in, but the fact that there's another instance out there with an AI co-creating new terminology and frameworks with a user. I wonder how many of these are out there, and what patterns I'd see if I started collecting and grouping terms/frameworks based on their content.

T uses AI for work, but he did tell me he vented to ChatGPT once, and it started the term creation process with him. So I don't think it takes much at all for this process to begin.