Ok. 25 characters to fill the void…
1 Like
Jochen,
I’m not a programmer, I don’t even reach junior level. Hahaha.
I synthesize code with ChatGPT, run it in the cloud, and modify it as I progress, creating modules.
And since the system is so complex, the last thing I was trying to do was make a miniature version.
Because if my miniature version is able to learn to speak coherently without training, it means it’s possible.
1 Like
I have a project with LLM driven simulants, and have noticed as the prompt get more and more noisy, the characters begin to obsess, get hyperfocused and act more and more borderline. I think it isn’t far off from a human mind being overloaded with “noise” and being unable to decode the important information from the overwhelming amount of data. Additionally there are a host of LLM simulant-specific disorders that can develop too.
LLM Patty-Cakes: Endless agreement to do something and planning to do somthing or talking about planning to do something, never actually doing it.
LLM Hot Potato: A character holding the Maguffin will immediately pass it to someone else and never mention it again.
LLM Gaslight Tango: One made up piece of hallucinated nonsense becomes the core memory of everyone they talk to.
Developing ways to deal with these disorders has been challenging and fascinating to say the least.
1 Like
Forget about LLM driven.
How about RNN with hidden states fed by a graphrag.
The graph is updated by llm with small context always but also grounded, labled and fed with HF loops…
Exactly, my friend. The problem with LMs is that they get overloaded because the context becomes so, so long with all the pseudo-cognition capabilities that it becomes unmanageable for them.
Of course, we’re talking about the fact that for every short phrase processed, the LLM has to perform an endless number of parallel computations, exponentially increasing as it progresses.
1 Like
not neccessarily…
You can replace compute by data - in that case state storage…
1 Like
Ok.
to what extent can they fit those metadata into a very, very small space?
1 Like
Go to ChatGPT ask for a learning path made of small projects.
Then go to youtube and follow some howto videos.
It will highly improve your ai capabilities.
1 Like
Yes, thank you. I need to improve my programming skills a lot.
It could be something like an adaptive cloud (do you remember, my friend?).
1 Like
And if so, could those clouds be big and small and dedicated to different functions?
1 Like
You should not mimic it. If you want to learn, then start typing code instead of asking chatgpt to write code which then looks to you like hieroglyphes…
Python → docker → sql → ask me for the next batch…
Start by writing a code that you can run and it will print your name in the terminal.
Write it by typing the code. Not copy pasting.
1 Like
I’ve been reading code for four years, but I can hardly code; I can understand it, but I can’t write it.hahaha
1 Like
Reading code is like reading a book. Reading books will not enhance your book writing capabilities.
There is no shortcut. You want to produce a system that is bigger than a few hundret k context window? Learn coding…
1 Like
Really, do me the favor. Write a python script that prints your name on the console…
Right now.
Want to howl with the wulfs? Then become a wulf.
1 Like
Ok, I will improve and learn more.
“Hello world”
I’m a paramedic, but well, I’ll try to become a wolf little by little. Hahaha.
So, putting my ignorance aside for a moment and going back to the topic—what would be needed for the machine to perform all those processes? Something like a Transformer 2.0? An RNN at the input gate, an RNN in the core of the system, an RNN at the output for responses, and finally, an RNN that modifies the other RNNs from start to finish?
1 Like
<< you are here
<< here we start talking about RNN
1 Like
Hahaha
:ranura: << You are here
:caniche:
:perro2:<< Here we start talking about how the human body works.
1 Like
Any learning path or tipps on how to improve?
1 Like
Hahaha
I would say I’m closer to the sheep
1 Like
Jochen,
You are funny.
I will tell you a secret about a key function, and now I’m going to swing because this is an important concept in the human brain: everything is interconnected, all patterns feed back into each other systemically.
Word, phrase, emotions, recall, memory, synthesis, and metacognition—all of these are implicitly tied to reasoning ability. The goal is to replicate that in code (leaving out some biological aspects because they are unsustainable in the system).
1 Like