What insights have neural networks revealed about chess?

I think NN's brought us 3 main types of knowledge about chess: Specific good types of positions, that playing positionally is viable at a high level, and that we are nowhere near creating a perfect chess engine.

1. Fawn pawns (6th rank pawns that are safely nested in opponent pawns) are the clearest example here. Another example would be the strength of a pair of opposite color bishops which most AB engines underestimated.

2. AlphaZero/LC0 both play in a much more positional manner. While hard to quantify, this basically means playing to maximize space and mobility rather than pure material.

3. The fact that LC0 plays so differently than SF and is about at the same level means that neither are close to perfect play. The intuition here is that as you approach optimal strength, all strategies should converge, so the difference between play styles implies that there is a lot of ability for both to grow stronger.