TensorAtlas — A visual atlas of machine learning

Field notes

The first lesson constructs a bigram model from three repeated words, adding START and END markers before counting every neighboring pair. The page makes the probabilities concrete: after a, the model chooses t, r, or n with 60%, 30%, and 10% likelihoods, while the visible controls allow a reader to choose the next token or generate a word. The atlas then links this foundation to 21 further introductions, including attention, convolution, clustering, and agents.