Both.

Capabilities get overhyped as leadership of each of the large AI companies - and some of the smaller ones - spend disgusting amounts of fiat that need to be justified to current and future shareholders. They will gladly ride doomer panic waves to exaggerate the capabilities of their products and land higher valuations, attract more investment and skim off the top.

At the same time, the industry is not even teenage or infantile; it's not even truly born yet. The I in AI is currently not really there. It's just an extrapolation framework that emulates conversation (or action), not intelligence. We need a dozen fundamental breakthroughs for that I to really take form. Just realize that even a small LLM like llama3.2 has multiples more encodings of words and their relations than you have, it can emulate knowledge about tons more subjects. But it doesn't have actual knowledge. It's a hack, to deceive you, and it does that by following patterns of conversation that seem as a conversation to you, but at the core, there is nothing there that can be as insightful as even the youngest, pre-school child.

LLMs are databases, with a fuzzy search engine built-in. They aren't intelligence.

I'm at the point where I recognize the usefulness of AI in a multitude of situations. However, I still don't like the fact that it's used for medical treatments or laboratories. I believe that human beings deserve to be cured or treated by other human beings!

My honest take is that AI has a lot of potential, but very different from what is being hyped, and perhaps not useful to the vast majority of people who just seem to want a way to use their brain less and less.

My specific case is related with research in mathematical physics. Even last year, all llms were basically useless to me. This year, I constantly use them for brainstorming and producing boilerplate proofs of simple (but useful) lemmas. Of course, I supervise the output carefully, but I am now sufficiently able to optimise my pipeline in such a way that checking is faster than producing everything from scratch myself.

I strongly believe that there are brilliant people doing research in AI now, and this will result in AI (not only llms) being more and more useful for creative people.

The elephant in the room, however, is the issue of regulating the use of AI and the companies developing and selling AI solutions. This, I reckon, will be a total shitshow.

I believe AI is best used for processing a huge chunk of repeating data, that's definitely where it excels. Also not half-bad for programming, but it'd probably make spaghetti code in the long run.

Good for saving hours on a task that anyone can do but nobody wants to do, but it's not reliable enough (yet) to do tasks that people actually STRUGGLE to do. I could ask AI to arrange my Excel spreadsheet and it'll do it completely fine, I could also ask AI to program me a rudimentary script on Python that does something very specific and small. Ask any more or less, and it'll barf up a random answer confidently.