#tech #idea
# [[Epistemic status]]
#shower-thought #to-digest
# Related
# TODO
> [!TODO] TODO
# Tech idea
Tech idea with no revenue in mind.
## Problem
## Solution
Simulate an ecosystem with some microscopic basic [[Evolution|evolution]] like organism same way as https://github.com/The-Tensox/Erutan-unity
But in addition organism are powered by a [[Large language model|LLM]] where probably some batching inference across all organisms is done to allow using a 7b more on consumer hardware
I thought about this while watching random people talking in a cafe
People get inputs from the world and it triggers their completion like they see the sun they talk about the weather, there is food everywhere so they talk about food etc
Typically weather, food, conversation as it is the most common [[Conversation|conversation]] topics in our llm distribution thanks to the [[Biology/Evolution/Monkey brain/Monkey Brain|Monkey Brain]]
Back to the simulation, I’m wondering if you could also use the llm to create the evolution part (average of the genes + mutation/error)
Not sure it’s so smart but why not :D
So the version 0 would look like:
A bunch of organisms moving around according to their local inputs and their outputs would also be driven by their local inputs (ie two organisms meeting each other trigger a conversation)
A hack to make it work v0 would be to hard code a kind of Markov process where the llm could decide in which state to be in according to its inputs (hungry->seek food)
In order to make each individual “unique” we could use [[Embedbase]] instead of painful fine tuning where we would store the history of the individual in his own dataset
An iteration would create fine tuned model on each emerging species where it would create a natural intelligence where [[Embedbase]] would provide a “nurtural” intelligence