For example, classical statistical learning is very well-posed and mathematically careful, and has been used all the time in communications, web applications, bioinformatics, weather prediction, etc. Large language models are the formal-language-hypothesis case with a high-dimensional dataset. LLMs are not as well understood yet. Furthermore, the LLMs are usually part of a larger system. So if you've mucked around with LLMs in Python on a powerful home computer...you haven't seen the whole story.

Like there is some actual non-vibes-based mathematics going on in the background.

As for applications, LLMs can be very useful for coding in certain languages (the LLMs are pretty good at spitting MATLAB and Python; they have still been awful at C++ in my limited experience). And they can be excellent for machine translation. (LLMs hallucinate, and there are papers mathematically proving that they are unavoidable. Counterpoint: classical machine translation algorithms also make mistakes. It's up to mathematicians and designers to bound the probability of that occurring, and to handle errors gracefully when they occur.)

Emphasis here is on can be. If you need a perfectly correct answer absolutely all the time...machine learning and its derivatives are just not the right tools for the job. The mathematics says this right from the start. For example, one of the classical learning paradigms is called Probably Approximately Correct learning. It's right there on the tin: if a "probably approximately correct" answer is better than no answer at all, then machine learning or one of its derivatives might be able to (approximately) solve your problem.

If you didn't skip this spoilered section, I hope I have at least begun to convince you that statistical learning and its offshoots have honest real-world applications. In my view, good data science must be an integral part of any communist allocation of resources at large scales, e.g. production forecasting, weather prediction, logistics, etc.

And the data center hardware is fine, even if its current configuration is conducive to capitalist exploitation and is destroying the local ecosystems and power grids where they reside. Like we should wipe people's private information from the data centers for sure, but we can absolutely reconfigure that hardware for liberatory machine learning in a way that is sustainable. Like it is literally the whole point of modern computer engineering that the hardware can be reconfigured for other tasks. Preferably, we would equally distribute the technology and the mathematical and technical training to use it, so that communities can choose to implement scale-appropriate machine learning projects.

Raschka's book is good if you don't care about the theory, although it doesn't focus on LLMs and predates ChatGPT. HuggingFace does have a lot of gratis (free as in beer) open-weight LLM models and free (as in speech...and beer) open-source courses for how to use them (well, the models are open-weight, not open-source, but I think the rest of their courses are actually open-source). I was able to run most of the courses locally on my modest hardware.