This page is Shannon's idea model. For question answering over your own text, see Shannon QA.
A small model that has to think in ideas
Shannon reads up to 128 word-pieces and squeezes them into 32 idea codes, each one of 4,096 entries in a learned codebook. Its decoder then writes the next 128 word-pieces from those codes alone: it never sees your text. Everything below runs on your device with onnxruntime-web; nothing is sent to a server.
Model: loading model list…
Try it
Sampling settings
Temperature 0 = greedy. Same seed + same ideas = same output.
Ideas
Each chip is one idea: the code number and the 4 word-pieces it was made from. Tap a chip to see what the code means and swap in a neighbour.
Load the model, then press "Text → ideas".
Shannon's continuation
What the real text said next
Idea map
All 4,096 codebook vectors, laid out in 2-D with UMAP (cosine). Nearby codes have similar vectors. Colours are 28 k-means regions, labelled by the words their codes over-represent. After "Text → ideas", the codes of your text are ringed and joined in slot order: that path is the thought Shannon passes to its decoder.
Idea dictionary
What each code stands for, read off held-out validation windows: the 4-word-piece spans it encoded most, words it over-represents, and sample contexts. Many codes are fuzzy; that is an honest picture of a model this small.
Pick a code.