3Blue1Brown's neural network video is the one visual explanation that actually makes the concept stick

The digit recognition example is the smart choice of entry point. It is concrete enough that you can follow exactly what the network is doing at each layer, abstract enough that the generalisation to other tasks becomes obvious, and visually demonstrable in a way that makes the mechanics intuitive rather than just credible.

The specific thing the video makes clear that most text explanations miss: a neural network does not have a programmed rule for recognising a digit. It has weights across millions of connections that were adjusted through training until the pattern of activations reliably produced correct outputs. That shift from rules to weights is the conceptual leap that makes modern AI feel qualitatively different from earlier software.

Worth watching in sequence with the backpropagation video if you want to understand training rather than just inference.

Does having a visual mental model for how neural networks work change how you think about their limitations, or do you find the abstraction level still too far removed from how you actually use the tools?