Same underlying question — what happens when you give a model more test-time compute — answered by three small, in-browser demos. Two are opposite transparency choices on a from-scratch toy; the third retrofits the same opaque-loop mechanism onto a real pretrained language model, to show it isn't just a toy-scale trick.
A 153K-parameter recurrent-depth transformer that loops a single block in latent space. No chain-of-thought channel exists in the architecture — you see a loop counter and then the answer, nothing in between.
visible chain-of-thoughtDeepSeek-R1-Distill-Qwen-7B, run client-side via WebLLM. Every reasoning token streams to the screen in plain English before the final answer — the opposite design choice, same class of problem.
real pretrained modelSame prelude/core/coda pattern, retrofitted onto SmolLM2-135M: 29 of 30 layers frozen exactly as shipped, one shared layer fine-tuned and looped. Proof the mechanism isn't toy-scale-only — with an honestly reported training-budget limitation.