Curiously, actual scams also go through "a speculative boom that looked like a scam in the moment", and then they turn out to actually be an overhyped scam that doesn't in fact change the World.

Crypto currencies are a good example.

Your "don't throw the baby out with the bath water" statement makes a lot of sense in the early stages, when we don't really know yet if what's being overhyped might or not be just the beginning of something big, hence one shouldn't just discount a tech because there's a massive hype train on it. The thing is, this was maybe 1 or 2 years ago for things like LLMs, but by now it's becoming obvious that it's a dead end since the speed of improvement and cost relative to improvement ratio have become very bad.

Whilst broader Machine Learning tech is useful, as it was useful already since when it started (back in the 90s Neural Networks were already used to recognized postal codes on mail envelopes for automated sorting), this bubble was never about the broader domain of Machine Learning, it was about a handful of very specific NN architectures with massive numbers of neurons and huge training datasets (generally scrapped from the Internet), and it's those architectures and associated approaches to try and create a machine intelligence that are turning out to not at all deliver what was promised and as they've already reached a point very low incremental returns, seem to be a dead-end in the quest to reach that objective. What they do deliver - an unimaginative text fluff generator - turns out to be mainly useless.

So yeah, if you're betting on the kind of huge neural networks with huge datasets used in the subsection of ML which has been overhyped in this bubble and the kind of things they require such as lots of GPU power, you're going to get burned because that specific Tech pathway isn't going to deliver what was promised, ever.

Does this mean that MLs will stop being useful for things like mail sorting or other forms of image recognition? Of course not, those are completelly different applications of that broad technique which have very little to do with what people now think of as being AI and the bubble around it.

Machine Learning has a bright future, it's just that what was pushed in this bubble wasn't Machine Learning in general but rather very specific architectures within it - just like when the "Revolution in Transportation" which turned out to be the Segway and kind crap thus quickly fizzled didn't destroy the entire concept of transportation, so the blowing up of the LLMs bubble isn't going to destroy the concept of Machine Learning, but in both cases if you went all in into that specific expression a technology (or the artifacts around it, such as massive amounts GPU power for LLMs), that the broader domain will keep going one isn't going to be much comfort to you.