LLMs (what you mean by "AI" with the scare quotes) are great at spamming and scamming people.

I’d generalize that somewhat to say that LLMs are good where you can cheaply verify the results relative to the cost of being wrong. A lot of scamming is already structured around most people ignoring their messages, and they don’t support anything so it’s just a question of whether they lose more potential marks than doing it manually.

(This is also why coding is one of the areas where they perform best: the cost of structural validation is low and there’s a human in the loop verifying the behavior)

I've tinkered with LLMs, because I keep thinking it'd be cool to have an "assistant" that could research and boil down information for me. The problem is coming up with a question that is A) too complicated to be answered in a single web search that I can do myself as fast as asking the question, and B) not too complicated or important for me to accept the answer without redoing the research myself to verify the answers are correct.

For simple questions (How long does the moon take to orbit the earth), a search engine will give the answer right in the results; I don't even have to click through to a page. An LLM can't save me any time there, so I'd only be using it to be using "AI" (which is what I see people around me doing).

For difficult or currently controversial questions (What's the best hosting service for my new subscription site, taking into account price, reliability, location, and hardware support?) there's no way I could trust it not to be making shit up. By the time I checked all its work, I might as well have done it myself.

So I'd like to usefully use it, but I can't figure out how to use it as more than a curious toy.

My experience with LLMs is that they are vastly superior to google at the moment for doing research on topics you know very little about. They make doing research fun and productive again, in a way that google has failed at for over 15 years. I used to be able to google a topic I only vaguely knew existed, and after a few tries and scolling through 10’s of pages of results, I could find a term that would bring me much closer to the term of art that would crack open the topic for me and give me good results. That is nearly impossible today. Google simply doesn’t populate page 4 of a search term, and it doesn’t have booleans to filter down either. But with an LLM, I can poke and prod it until it gives me something that I can then use to reference real sources.

LLMs also seems to be really good at giving sort of the average (modal?) answer on a topic, so I find it useful for trying to get an average understanding of something. One example I’ve found useful is getting it to interpret ambiguously worded datasheets. I don’t necessarily trust its answers, but it may show that I’m totally thinking about something wrong.

Asking an LLM straight up to recommend a hosting service is not going to return anything useful.

A better approach would be to ask it to help define what it is you want from a hosting service. You can give it a fuzzy poorly written brain dump of what you are looking for ask it to “restate what my requirements are and expand upon them (but don’t solve yet)”. Let it help you understand and clarify your own thinking!

It’s a dialog. Asking LLM’s for answers isn’t really a good use of the tool. Especially something so incredibly broad like “list top ten best providers”. They don’t fucking know! They will just regurgitate what they were trained on. Except maybe ChatGPT’s “deep research” which does a bunch of websearches and compiles and interprets the results.

One of the “problems” with LLM’s is people simply not knowing how to use them or understanding what they are good at and what they aren’t so good at. It’s hard to know because it’s both masked in a hype and honestly I don’t think we have a good understanding of their edges yet.