​ I've been working on a project called Leo / PSCLS. The goal isn’t to build yet another Transformer with a different name. I’ve been trying to explore a different question: «What if we built an AI architecture around persistent neural state, sparse connectivity, recurrent processing, and memory — instead of making attention and large dense parameter matrices the core building blocks?» Leo is still very early and nowhere near a fluent language model. But we’ve reached a point wh...