The thing is that the specific technologies behind all of those very practical improvements are not what's being hyped up this bubble. Speech recognition, for example, usually involves a lot of audio preprocessing, followed by some form of RNN/LSTM/Transformer to generate candidates, followed by beam search to score and choose from candidates.

If you are a machine-learning practitioner, you should be familiar with all of those techniques and how they are used so that you can solve practical problems with them. But if you just read about AI in the news and figure you're going to found the next great startup and make a billion off it, you'll probably start by feeding a whole bunch of data into Tensorflow and then getting useless garbage out of it.

This hype bubble is specifically about LLMs, extremely large-parameter transformers that are trained on all the data OpenAI or Google can get their hands on. And then supposedly if you ask them the right questions, you will get useful answers back. For people that put in the time and experimentation to actually find the right questions and the right applications, that will probably be true - but the hype is that this will change everything, and it most certainly will not, just in the same way that beam search is frequently useful but it definitely does not change everything.

But slick promoters will nevertheless manage to use people's lack of knowledge to redirect billions of dollars in capital into their and their employees' pockets, the same way that slick promoters used crypto to redirect billions of dollars in capital into their and their employees' pockets.