GenZedong

but I just don’t see the use of it.

How useful it is really depends on what you'd use it for. Which is one of the fair criticisms of it, that capitalism is trying to push it anywhere and everywhere in ways it isn't all that useful or well fit for.

In my assessment of it, text generation AI (to use one example of gen AI) is actually best for fiction, i.e. being a chatbot who can engage people in their fantasies and play them out with them (roleplay, etc.). This is because in fiction, it doesn't matter that much if the AI chooses a wrong token and what should have been purple is pink (e.g. it doesn't matter much if a "fact" is wrong if it's a fictional world to begin with). I could write a whole post just on the subject of chatbots and the pros and cons of them being a thing.

Beyond that, some of the better Instruct models, models tuned with an assistant/user message exchange dynamic in mind, are capable of coding pretty well and explaining concepts about it. I use Deepseek as an assistant for that on occasion, combining it with regular web searches for solutions. Though I am also firm on the idea of its solutions not being something I yoink without understanding because I want to become better at programming, so I generally won't outright copy/paste what it gave me unless I already understand the syntax and just needed the reminder. At times, it helps me in ways similar to rubberducking: https://en.wikipedia.org/wiki/Rubber_duck_debugging

It can make errors in code or say things that are wrong, but when combined with cross-referencing from other sources, it can help with working through things that are complex to search for on their own (especially with search engines getting worse over time instead of better). I've also used it before to get input on areas I can improve at coding and what metrics might point toward quality in the first place; I take it all with a grain of salt, but it can help crystallize things that I'd otherwise be thinking vaguely about.

There are also models that are adept enough to work through mathematical proofs. I'm not sure of to what extent or how useful it is for helping an actual mathematician, but the best of models have come a long way from speaking nonsense a lot. I still wouldn't recommend trusting anything an LLM tells you without cross-referencing, but they're at least less incorrect than they were in the earlier days.

In many ways, people are still figuring out what all use they do have and in what context they are more benefit than they are waste of time. If you do try a text model for free, I'd recommend Deepseek since it's based in China (and also China being based :P) and so if there are privacy concerns about what you say to it, at least you know the data isn't going straight to western corporations to sell. Most AI services are sneaky in the margins about what private really means. They'll often say that things are encrypted, but then have a line about how the data can be used to further tune/train their models, which means at some point it can be looked at by the company unencrypted, even if anonymized, to learn from it.

I only know of like one service that very explicitly puts encryption and privacy first to the point that I'm pretty sure they'd have to change some infrastructural things to even be capable of mass storing unencrypted info like that to train models on it. And they are kind of a niche paid service (no free tier) that lags behind the major corps in model capability because they're running purely on what they can make off of users rather than investor money.