Deciding how many nodes and layers to use in a neutral network: many functions can be expressed in 1 or 2 hidden layers, more is often better and faster, and you usually have to experiment.
https://machinelearningmastery.com/how-to-configure-the-number-of-layers-and-nodes-in-a-neural-network/
miniblog.
Related Posts
"Minus 100 points", an article on deciding how to add features C#, remains one of the best introduction to PL design principles I've seen: https://docs.microsoft.com/en-us/archive/blogs/ericgu/minus-100-points
(Design is hard, combinatorial complexity grows easily, saying "no" needs to be a default.)
I've come to realise that the goal of a diff tool isn't to say what's changed. You want it to say what *hasn't* changed.
Unchanged content is crucial for deciding how to group changes into hunks and how to align lines. You need matched content pairs for this.
I've added some new keybindings to deadgrep.el, so I split the shortcut tables into two categories: https://github.com/Wilfred/deadgrep#keybindings
I've done this purely by intuition. Are there good rules or heuristics for deciding when more hierarchy is needed in information, I wonder?