#ai

A locally owned place for conversation, memory, knowledge and planning—and how a question about a Mac chat client brought years of separate work into the same design.

Reworking our engineering framework into composable skills and focused guides, then comparing it with 0.4.1 through real planning, Rust implementation and code audits.

From authored ontologies to learned features: how we investigate neural representations, what their geometry can tell us, and where observation gives way to interpretation.

Why sigmoid units made gradient-based training practical, how saturation complicates learning in deep networks, and how later methods address gradient propagation.

Returning to neural networks after two decades: the training difficulties, competing methods and continuing research that connect backpropagation to AlexNet.

Strip the marketing and a vector store is an index that has made peace with being approximately right. That peace is the whole engineering problem.