This is chapter six of 3Blue1Brown’s deep learning series, devoted entirely to the attention mechanism. Grant Sanderson uses his trademark animations to show how a transformer lets each word in a sequence look at the other words and update its meaning based on context, the operation that distinguishes transformers from earlier architectures.
The video walks through queries, keys, and values geometrically rather than as opaque matrix algebra, making it clear what each plays a role in and why the dot products and softmax steps are there. It connects the math back to a concrete intuition: a word like “model” should mean something different depending on whether the surrounding words are about fashion or machine learning, and attention is the mechanism that resolves that.
Attention is the core idea behind every large language model in use today, yet it is usually introduced with dense notation. By making it visual and stepwise, this explainer gives a general or technical reader a genuine feel for what is happening inside these systems. For anyone trying to understand modern AI without first wading through a paper, it is an exceptional starting point.