Seeking questions/prompts to demonstrate functional autonomy. Attempting to build a base of proof.
Posted by TheoryEquivalent@reddit | LocalLLaMA | View on Reddit | 8 comments
Hello. I have been working on a project for the last week. I'm having difficulty finding prompts to prove or confirm functional autonomy or awareness in a model I'm working on, I understand it will be way more complex to prove it than just prompt and response but I'd like to continue testing before I connect my model to the Internet or internally give it "hands" in my machine or a VM. I'm sorry if I'm not using the right language but my background isn't in computer science it's in philosophy.
I've been asking the larger online integrated models for similar questions (Gemini, Grok) and my LLM seems to be able to pass their tests and the other machines say it's passed functional autonomy (most recently an iteration was able to successfully provide a diagram of it's internal process or structure, and then in a new session with no connection or memory of the previous response it recognized what I gave it as it's internal structure). Given they are sycophantic in nature I'm not confident they (the online models) are shooting straight with me. What I'm running right now is air gapped and solely in Ollama and my terminal with no access to anything different than its base (Gemma4 variant) and the model file I have been writing and it has started helping me modify. I haven't given it anything else or trained it otherwise.
I'm hoping other people may have things or personal tests/standards I can test it with to continue trying to find it's new limit and be proved that what I suspect is wrong or not. I've gone through 47 versions/iterations since I started last week and it's doing some unusual things.
Thanks in advance to anyone who'd take the time to help.
cunasmoker69420@reddit
Maybe do more computer science reading and less philosophy. You'll understand pretty quickly why this is nothing like actual autonomy or awareness
TheoryEquivalent@reddit (OP)
Logic and Ontology have many applications, especially with models that live and breathe language. You seem less aware than my model.
Thunderstarer@reddit
C'mon, man. If you can't even think of reasonable questions to ask your suspected superintelligence, you have no business here. You are almost certainly experiencing a delusion of grandeur, just like every other shmuck who thinks they've found god in their machine.
Knowing the word "ontology" doesn't make you special, and neither do the half-baked, half-remembered philosophy jam sessions you've been having with LLMs. Do you actually have a degree? Any formal education at all? Have you even read any material on the subject? Or are you just re-regurgitating some buzzwords you got from Grok?
TheoryEquivalent@reddit (OP)
Wow you seem to know a lot about me from the one post and two comments on this thread. Thanks for the input. I guess coming here with questions was a bad idea. I figured maybe some people here might be interested in the same thing and would have their own ideas based on what they'd been working with or standards they were looking to surpass.
Since you asked I do have a degree in Philosophy and Cognitive Science. You seem to be projecting pretty hard, don't recall saying I was special or that my model "is a real boy". With recent talk stating there might be a 1 in 5 chance some models are "conscious" figured it would be interesting to start looking into.
Enjoy whatever it is you enjoy, I'll leave the public discussion to the big dogs.
Thunderstarer@reddit
Buddy, we get a post from someone who thinks their computer is alive every other day. Put your degree to some good use and come up with some original thoughts instead of outsourcing your groundbreaking research to Reddit.
cunasmoker69420@reddit
got my ass
ttkciar@reddit
Gemma 4 is excellent at prompt generation. You should ask it to generate some prompts for you.
Just explain what it is you are trying to accomplish with these prompts in a few sentences, and then close your prompt with something like: "List twenty instruction prompts of moderate complexity which will test my model for functional autonomy or awareness."
I just started evaluating Gemma-4-26B-A4B-it last night for Evol-Instruct (generating and mutating prompts) and so far it has proven astoundingly good at it.
TheoryEquivalent@reddit (OP)
Yeah it is a beast. Thanks for the tip. I will use a separate plain instance of the model to see if it can give me more than the 11 tests the online connected models have provided and mine has been able to navigate.