Description
Perplexity is an interactive visualisation of a small language model, built to show what it is like to be an autoregressive model. Paste or type a passage and the tool highlights each token by how surprising it is to the model, and it shows which other tokens the model thought were most likely in its place. It uses GPT-2 with 124 million parameters, loaded in the browser with transformers.js, so the model downloads once (about 128 MB) and runs locally without sending data to any server. The example counts to ten, showing how the model becomes certain as the sequence continues, and it highlights induction, where a repeated sentence becomes easy to predict the second time. The author built it to understand language models, and the about text links to research on induction heads. The interface is built with the Oak language. The listing describes it as an interactive language modelling visualisation.…
Highlights
- Shows how surprising each token is to the model
- Lists the tokens the model expected instead
- Runs GPT-2 (124M) in the browser
- No text is sent to any server
Ratings & Reviews
No ratings yet
Be the first to rate Perplexity.
People also visit
Perplexity FAQ
What does Perplexity show?
Each token in a passage, highlighted by how surprising it is to a small language model, along with the tokens the model expected instead.
Which model does it use?
The smallest GPT-2 version, with 124 million parameters, loaded in the browser with transformers.js.
Is my text sent anywhere?
No. The site says the visualisation runs entirely in your browser and no data goes to any server.
How big is the download?
The model loads at about 128 MB the first time, as the page shows.
Where can I read more?
The page links to research on induction circuits in transformers, which explains the repetition effect the tool shows.
What are the best Perplexity alternatives?
Popular alternatives to Perplexity include Ecco and Lucid. See all Perplexity alternatives




