The mechanistic theory: the found weights implement reusable algorithms — circuits — and features are packed in superposition; emergence is real computation you can sometimes reverse-engineer. The only theory that says what is actually inside the weights, not just that it works. Chris Olah's lineage, 2020→now (the artifacts you folded into TTU1).
the substrate, the structure, the thesis, the evidence
how this book is read
the iconic intro experiment, trained in your browser — superposition you can watch form
▸ TOY MODELS OF SUPERPOSITION — live 5 features → 2 dimensions, swept across sparsity. Dense, the model keeps only the 2 most important and drops the rest; sparse, it packs all five into the plane, overlapping them — and the weight vectors snap from the axes into a pentagon. That is superposition, watched as it forms. A from-scratch JS reimplementation of Anthropic’s notebook (Elhage…Olah 2022, MIT), trained live with Adam — math re-derived, result theirs. open the lab →▸ THE CAPACITY LAB — live the bit-depth question, measured: sweep n_features as powers of two (4,8,16,32,64) into a fixed m dimensions and watch how many actually pack in vs sparsity. Dense → only m fit (the ceiling); sparse → far past m. Superposition as features-per-dimension. run the sweep →▸ THE HOLOGRAM — live, 3D superposition as distributed holographic storage: many features packed into 3 hidden dimensions, rendered as a rotating constellation on a dark plate. Watch them pack past 3 into a polytope — every axis holding fragments of many features, like every patch of holographic film holds the whole scene. rotate the plate →▸ THE INTERFERENCE FIELD — live superposition rendered as a literal hologram: each feature becomes a plane-wave grating and the whole set is superposed into a moiré fringe field — bright and dark where features reinforce and cancel. The smear, made of light. see the fringes →the theses, mechanisms, and dated findings of this theory — each a full .dlw badge (12)
the cited milestones, dated & attributed (render-not-invent)