Visual explainers
One idea, told in pictures. The interactive lab stays pinned on screen and moves as you scroll, so you watch the mechanism change one step at a time. Prefer to be shown? Every lab also has a Watch button that plays its tour with captions and optional narration.
How self-attention mixes wordsFollow one word through self-attention, from its vector to queries and keys, dot-product scores, scaling, softmax weights and a blend of values, and see how that lets "it" find the noun it refers to.7 min · from Self-attentionHow backpropagation assigns blameFollow one neuron through a forward pass, a loss, and a backward pass with real numbers, and see how the chain rule hands every weight its share of the blame for one bad prediction.7 min · from BackpropagationUnderfitting, overfitting, and the U-shaped curveWatch a polynomial go from too stiff to too wiggly, see training error fall while validation error turns back up, and find out what more data and more noise do to the sweet spot.6 min · from Overfitting and the bias-variance trade-offHow gradient descent finds the bottomWatch an optimizer walk downhill on a loss surface, see why a learning rate that's too big makes it zigzag and then explode, and how momentum and Adam fix the zigzag.6 min · from Gradient descentWhy long contexts eat GPU memoryWatch the KV cache grow with every token and every user until it outgrows the model itself, then see how grouped-query attention and 8-bit caches win the memory back, and why decode speed is set by memory bandwidth.7 min · from LLM inferenceHow retrieval-augmented generation finds its evidenceFollow one question through a RAG pipeline (chunking, lexical scoring, a vocabulary-mismatch miss, semantic matching, top-k and reranking) and see why no prompt can rescue evidence that was never retrieved.7 min · from The RAG pipeline