Technology

Neural nets are a technology which is part of the umbrella term "machine learning". Deep learning is also a term which is part of machine learning, just more specialized towards large NN models.

You can absolutely train NNs on your own machine, after all, that's what I did for my masters before Chatgpt and all that, defining the layers myself, and also what I do right now with CNNs. That said, LLMs do tend to become so large that anyone without a super computer can at most fine tune them.

"Decision tree stuff" would be regular AI, which can be turned into ML by adding a "learning method" like a KNN or neural net, genetic algorithm, etc., which isn't much more than a more complex decision tree where decision thresholds (weights) were automatically estimated by analysis of a dataset. More complex learning methods are even capable of fine tuning themselves during operation (LLMs, KNN, etc.), as you stated.

One big difference from other learning methods and to NN based methods, is that NN likes to add non-weighted layers which, instead of making decisions, transform the data to allow for a more diverse decision process.

EDIT: Some corrections, now that I'm fully awake.

While very similar in structure and function, the NN is indeed no decision tree. It functions much the same as one, as is a basic requirement for most types of AI, but whereas every node in a decision tree has unique branches with their own unique nodes, all of a NN's nodes are interconnected to all nodes of the following layer. This is also one of the strong points of a NN, as something that seemed outrageous to it a moment ago might have become much more plausible when looking at it from a different point of view, such as after a transformative layer.

Also, other learning methods usually don't have layers, or, if one were to define "layer" as "one-shot decision process", they pretty much only have a single or two layers. In contrast, the NN can theoretically have an infinite amount of layers, allowing for pretty much infinite complexity as long as the inputted data is not abstracted beyond reason.

At last, NN don't back-propage by default, though they make it easy to enable such features given enough processing power and optionally enough bandwidth (in the case of chatGPT). LLMs are a little different, as I'm decently sure they implement back-propagation as part of the technologies definition, just like KNN.

This became a little longer than I had hoped, it's just a fascinating topic. I hope you don't mind that I went into more detail than necessary, it was mostly for the random passersby.