Probability X-ray
A real language model running entirely in your browser — no server, no API key, nothing you type leaves this page. But a chat box hides the interesting part. This one doesn't: every word it generates is colored by how sure it was, and you can hover any word to see the alternatives it nearly said instead, with the exact probabilities. You are looking straight at the distribution a model samples from, on text you wrote, computed live on your own GPU.
How to read it: at every step the model produces a probability for all ~49,000 possible next tokens. It samples one, and that becomes the word. Here, each generated token is tinted by the probability the model assigned to the word it chose — calm green when it was confident, hot red when it was basically guessing. Hover a word to open the distribution it sampled from: the chosen token plus the runners-up it was weighing. Numbers are the model's raw probabilities (temperature 1), so they are honest, not cosmetic. Related demo: Watch a Model Think steps through the same lens on GPT-2. Here it runs on a 4-bit Transformers.js model on your hardware. Frontier models are ~1000× bigger; the architecture is identical, the difference is scale.