LenticulumLM assembles a diffusion language model from implicit learners: relations over a few token positions each, combined by product into one model whose every query (continue, predict backwards, fill in) is root finding with the known tokens clamped. It builds on Lenticulum.jl, whose theory vault explains the implicit learners themselves. This vault records what is specific to language: the design, the implementation, and what does not work yet.
Where to start
- Assembling a Language Model from Implicit Learners: the plan, and the decisions ahead
- tokens and sequence: text as a joint space, and an -gram chain assembled from relations, exact in every query direction
- Open Problems in Assembling Language Models: the roadmap, step by step
Everything else
- Map of Content: every note, in reading order.
- Start Here: how the vault is organised, its conventions, and how to read it in Obsidian.
- README: the project pitch.
- API documentation: the Julia package itself, built with Documenter.
- Lenticulum’s vault and the CT-ML wiki: the theory this builds on.