I was motivated to understand the latent space of models. Afterreading a few good fire papers, they represent that concepts are organized as manifolds in a latent space hypothesis. In their experiments they have shown a few circular structures, tree based structres. Some of their papers also talk about how circular operations are used for computation inside model.

This begs the question, if we assume that concepts are organized as manifold blobs in hyperspace, then how do they grow from the origins ? Are there any building blocks for intelligence ? if we have a recipie of the building blocks of intelligence, then its enough to construct some highly trained small blobs and then rewire them to construct a highly customized / capable models ? Can we identify the building blocks of intelligence ?

lets say we have these building blocks Language Math Vision Planning Memory Physics Social reasoning