Watch language models work, live in your browser.
Fourteen interactive visual explainers with real matrix math, live training loops, and step-by-step walkthroughs. No backend, no canned animations.
Tiny GPT
A 26,592-parameter transformer written from scratch in JavaScript. Watch it train character by character from random noise to coherent sentences, plot attention heads forming, and test overfitting and temperature live.
Transformers and attention math
How Transformers Work
Visual overview of tokenization, causal self-attention heads, and next-token prediction.
/transformer
Transformer Math, Visually
Every matrix multiplication with live shape captions and hoverable dot-product breakdowns.
/transformer-math
One Prediction, Step by Step
Follow one word through all 15 steps from 2D embedding arrows to the final prediction.
/transformer-walkthrough
LLM Layers Diagram
Three zoom levels on a modern decoder stack: RMSNorm, RoPE, grouped-query attention, and SwiGLU.
/llm-layers
MLP as Memory
How feed-forward layers act as key-value memories to store factual associations.
/mlp-memory
Sequence models and tokens
RNN, LSTM, GRU, CNN
Explore hidden state carry-over, vanishing gradients, and why LSTM forget gates stabilized memory.
/seq-models
Seq2seq and BPE
Step through encoder-decoder attention and train a live byte-pair encoding tokenizer on your own text.
/seq2seq-bpe
How LLMs Read New Words
Subword fallback, early-layer detokenization, and in-context definitions.
/new-words
Training, reinforcement learning, and history
Watch Embeddings Learn
Live next-token cross-entropy training showing how similar contexts pull token vectors together.
/training
RL Rollouts, Visually
Simulate GRPO rollouts, group advantages, and why RL sharpens abilities pretraining put there.
/rl
Evolution of LLMs
Architectural timeline from n-grams and word2vec to modern decoder-only frontier models.
/llm-history