What the site covers
The site explains GPT-style language models: tokens, training versus use, the transformer calculation, and a GPT-2 forward pass.
How a page is organized
The first paragraph defines the idea. A figure shows it. Calculation pages include a short Python block and, where listed, a browser tool.
Which model is inspected
GPT-2 small has 12 layers, a hidden size of 768, 12 heads, a context length of 1024, and a vocabulary size of 50257. The weights are public. The explorer reads a committed trace of one sentence, not a live GPU.
Which models are only described
Closed chat products, and later open architectures, are described on their own pages. Their weights are not loaded by this site in the explorer.
What the site does not do
The site does not host accounts. It does not run a chat with a model. It does not accept private keys, passwords, or documents. The tools do not upload their input.