p.pocket transformer
ENTIRELY IN YOUR BROWSER
An experiment in small intelligence / 001

A small model, in plain sight.

A real transformer, predicting one character at a time. Watch it think, follow the probabilities, and give it a new beginning.

NO CLOUD REQUIRED100% local
2layers
2attention heads
24model width
48context characters
—parameters

Live generation

GENERATING  
Alice was beginning to wonder
Starting local inference…48-character rolling context

What comes next?

NEXT-CHARACTER PROBABILITY

The six likeliest characters, plus all remaining possibilities. Bars show the full distribution, totaling 100%.

Give it a beginning

21 PROMPTS / CLICK TO RESEED

The character trail

LATEST FIRST
STEPCHARACTERPROBABILITY

Little model. Real machinery.

Trained on a public-domain excerpt from Lewis Carroll’s Alice’s Adventures in Wonderland (1865). Each character passes through embeddings, two layers of masked self-attention, and a tiny neural network. It learns fragments of language, not facts: expect curious spelling and wandering sentences. Unseen characters remain in your prompt and become spaces inside the model.