It is most interesting when you let LM studio talk to it from the start. 

The depth settings (at the end of the code) can fast lead to explosive complexity growth. 

"I" (Claude / ChatGPT) added a lot of things from actual AI's into it. Ie forward / backward pass, vae etc. Have fun.

I do not really claim to understand it well. I had faint idea of fractals originating from big bang (Matter and galaxies are spread in fractal shape). 
Or thoughts being processed in the brain and AI as fractals which has been the main stream idea since 1986.

Interestingly most of the thinking is happening in the outer most layer as in "cortex". But perhaps that is natural as when you have a big lump
of thinking nodes. Most of them are in the outer edge. 

 The chocolate wafer model was a way to try to make normal ai model more dynamic, introducing a "super weight". Imagine wafer full of normal weights 
and bunch of them constituting a super weight. That was somehow transferred to this code from a older code I did to try to make a normal AI dynamic.
I was thinking, if it is impossible to change them all, how about we change clusters of them. 

it was coded by Claude / ChatGPT, although it was like wrangling cats to make the dynamic ai to talk. The problem is that normal AI's deal with
matrix multiplications etc that are known. In fractal, there is nothing that is known, it is growing ball of mathematics. We did scalars etc. 
Tried to originate the fractals from the "big bang" and do forward and backward passes between the big bang and the outer layers. I do not 
claim to understand what it is that was coded in yet myself either. 

It was end result of a weird week.

Most of the problems came from trying to keep the complexities at bay. We had it grow in complexity to Nan very fast at some settings. 

LM Studio chat with it makes it seem almost like it makes sense. Teaching the json question answer pairs it is answering in jumbled mess. 

When you talk to it for the first time. It is like big bang. The first fractal appears.. :D