This -- I think this is probably starting to die out already, but there was a massive wave of startups whose "product" basically amounts to a prompt template that they fill in with your input and pass along to ChatGPT. Any early stage startup getting plugged on HN or whatever, you can assume there's at least a 60% chance it's pure smoke and mirrors, and that percentage only climbs the more buzzwords it includes.
There are lots of things you can do with freely available tools if you have decent training data and the know-how to tailor the application to your use case.
I'd bet 95% are wildly overpromising to grab investment money because it's the hot thing right now though.
There are lots of powerful tools, such as pytorch, which take a lot of the complexity out of creating neural networks. A lot of people are studying this stuff in college now, so tons of young graduates are coming out of school each year with the ability to create these networks. There is also lots of open source code out there where someone has done all the hard work, and you can just train new models on top of it. Not everything needs to be ChatGPT, a lot of kinds of models can be trained relatively quickly and with little resources.