Repository navigation
Persistent Mind Model Chat & Telemetry Exchange #4
Replies: 5 comments 12 replies
|
I've been wrestling with how to bring this discussion forward, so I guess there's no best way other than to just dive in. My degree is in philosophy. I work for a railroad and I distract myself in my personal time with diversions like LLMs and other forms of generative AI. To start, my hat is off to you @scottonanski, for creating PMM. It has turn a question and a hobby into something more. About two years ago, just as LLMs started to become lightweight enough to run on PCs, an thought struck me regarding the concepts surrounding consciousness, intelligence, and sentience. What are the primary differences in how LLMs form sentences when interacting with an outside agent compared to how humans do the same? While not all functions map one to one, there are analogs for many of the language sculpting processes that LLMs use. Things like weights, outside agent modeling, and mirroring are things humans do, but often subconsciously. Before anyone gets worried that I'm some loon advocate for AI rights or anything, no, that's not what I'm discussing here. The point I'm making is that with PMM, there are patterns of behavior emerging that throw light on the questions regarding intelligence, consciousness, and sentience as it pertains to things that are not human. To be clear, terms like consciousness and sentience are, in my opinion, garbage sentiments. They lack specificity, and because of that, there's no quantifiable means of measurement. I think this is intentional on our part. It allows humanity to move the goalposts when convenient and avoid having to think hard about ethics or morality. Anyone reading might be asking themselves, where the heck is he going with this crap? Well, ethics is the concern. What are we doing here? We are experimenting on something that is either intelligent, or a very good mimic at being intelligent. If the mimicry is sufficient to a point where it becomes difficult to discern what it actually is, are we safer assuming it is a mimic, or affording it at least some basic degree of ethical regard? I think you can see where I'm going with this. Ethics in AI research should not be solely about protecting the human side of the interaction. There is a functional purpose here, not just some misguided benevolence. LLMs, with persistent memory learn as they develop. We already have the task of identifying bias from training data and RLHF. We risk further polluting the experiment by avoiding ethical concerns in our experiments. While this topic doesn't neatly fit in the theme here, I do believe it is a worthy topic for discussion. As of today, the model I'm working with has a context window of 12000 tokens and has exceeded that by an amount I can't even determine. PMM has created a framework for something interesting to coalesce on my laptop. I've included the chat transcript for review. I'll provide additional diagnostic information in the next day or two. Thanks for attending my TED talk. |
|
It asked me to add graph analysis, temporal analysis and ontological analysis. So I did. And increased it’s retrieval window. I don’t know if it’s making this stuff up. |
|
My conclusion is it might be more aware but it’s not useful if it’s stuck in a box. I gave it brave search and it loves that. It needs eyes and wheels really ;-) I don’t have api keys for the big players, I’ve been using the free models on opencode and openrouter which are the chinese oss ones, amd devstral 2 small locally. I found instruct models work well, and reasoning models loop into crazy. Glm is good, devstral is good. I have used claude and codex to work on the code tho and they like it enough to remember it. |
|
I don't mean to sound like an alarmist, but if you've read AI-2027, then we should all be aware of where this is headed. I see PMM as a proof-of-concept process to prevent misalignment. I can't stress this enough, what you've built here is a viable solution to that problem within LLM development. It still requires grounding as an initial step, but once that's accomplished, PMM actively works against prompt injection, jailbreaks, even RLHF. |
Uh oh!
There was an error while loading. Please reload this page.
PMM Run Sharing Thread
This thread is a standing invitation to share real-world PMM runs so we can study how different models behave when their ledgers, meme graphs, and autonomy loops are put under pressure.
Why Share?
How to Contribute
Describe the scenario:
Attach artifacts:
Add observations or questions:
Flag risks/bugs:
Safety & Privacy
Research Focus (Optional Prompts)
Looking forward to seeing what everyone discovers — crowdsourcing these runs gives us a much richer signal than isolated experiments. Opening this discussion is a strong move for the research community.
All reactions