A new explainer video from Grant Sanderson’s 3Blue1Brown channel is making the rounds on Hacker News, where it has accumulated 999 points and 172 comments. The video, titled ‘Visualizing Attention, a Transformer’s Heart [video],’ walks through the attention mechanism that forms the computational core of large language models.

Sanderson uses the example of a noun being modified by adjectives to illustrate how queries, keys, and values allow token embeddings to become context-aware. The video builds on his earlier chapter on transformers, and breaks down the matrix multiplications and softmax operations that produce an attention pattern.