LLM Explorables
Small sketches that show how a language model works, one part at a time. Each lets you change something and watch what happens.
The Explorables
| # | Explorable | What you'll see |
|---|---|---|
| 1 | Text Completions Animation | Next-token prediction. The prompt goes to a completion model, the twenty likeliest next tokens come back with their probabilities, and you pick one — or let the timer pick the top one every second. Presentation mode |
| 2 | Temperature and Sampling | The same twenty candidates rescaled across temperatures from 0 to 2, and a token sampled from them. |
| 3 | Tokenizer Map Reference | A network of every model on OpenRouter and the tokenizer vocabulary it actually uses; click any node for its evidence and source. |
| 4 | Token + TokenID Pairs | The vocabulary: every cl100k_base token as a coloured box; hover one for its id. |
| 5 | Token Probability Field | The completions demo with the model's answer painted onto its whole vocabulary instead of listed: every candidate lights up in its own place among all of its tokens, coloured and sized by its probability, and clicking one appends it. Same prompt, presets, Start and model select; the list display stays the keyboard-accessible one. Presentation mode |
| 6 | Semantic Embeddings Comparison | Two texts as vectors of 1,536 numbers, drawn as coloured lines, with their per-dimension difference and their cosine similarity. |
| 7 | Retrieval-Augmented Generation | How an AI app finds information for a prompt and hands it to a language model. Search a knowledge base by keyword or by meaning with two embedding models, choose how many passages to send and how close they must be, edit the passages, and read the request the model would get. Nothing is generated: the model stays the same, only what it is given changes. Presentation mode |
| 8 | How a Model Gets Made | How a model's behavior comes from how it was made: the dataset, the vocabulary, the training script and the time it trains, each shown on real small models, plus an AI agent's experiments improving the script over a 10-hour session. Load your own session beside them. Presentation mode |
| 9 | Try a Model as It Improves | The same small models at different points in their training: within one ten-minute run, across a 10-hour session as an AI agent improves the training script, and on two datasets. Send a prompt and read both models' replies side by side, beside what each wrote during training. Presentation mode |
Running a Workshop
The Workshop Admin sets up each room: its models and tokenizer, its roster, spend cap and join sheet, and how to run a model yourself.
A Model on Your Own Computer
The completions and temperature demos can also use a model running on your own computer. Choose Connect a model on this computer… at the end of a demo's model list. The Local Model Playground links to the setup steps and checks the connection step by step.