A language model with 10.8 million numbers, trained on one laptop on the eight Sherlock Holmes books. It does exactly one thing: guess the next letter. It runs right here in your browser. Nothing you type is sent anywhere.
These are the model's real odds for the next character. It picks one, adds it, and repeats. That loop is the whole trick, in any language model.
Watch it learn
Real writing samples, saved during training. At step 0 the numbers are random; nobody taught it spelling, grammar, or who Holmes is.
surprise while training (bits per character, lower = better)
chance it writes “s” after “Sherlock Holme”
What's inside
Letter by letter. Big chatbots read chunks of words called tokens; this one reads single characters (76 different ones), like Markov counting letters in 1913.
Same recipe as the big ones. A small transformer: attention and processing layers, stacked six deep, the architecture of GPT-style models, just tiny.
Trained by guessing. Shown real Holmes text, it guessed the next letter, measured its surprise, and nudged every number a little, 6,000 times.
It will make things up. It has only read eight books and knows nothing else. Everything it writes is its best guess at what sounds like Conan Doyle.