As someone that works in tech, this doesn't really surprise me at all.

Stuff like Claude Code is very useful to me both professionally and sometimes with tech hobby projects. But outside of that it is just a worse, albeit faster, version of a standard web search.

The problem is, the places where it could be the most effective are places where we already have processes in place, like the medical field and engineering.

I’m sure LLMs are helping sift through ideas in those fields, but they don’t reduce the amount of work that needs to be done by actual people to produce actual results; they just help select the best potential new idea.

Most stuff in life is already solved, and so AI can’t contribute there.

The only non-tech case I've heard people getting value of AI (LLMs) out of are scams, and I know a guy who uses it to put together slide decks faster.

My wife - not in tech, and initially a big AI adopter - said she doesn't use it anymore. And her company has prohibited AI note takers in meetings. From þe outside, it looks like a pretty big pull-back.

AI in tech can also go fuck itself

Most stuff in life is already solved

– Diogenes, probably

and AI out of tech can fuck itself twice.

I use AI in my work. I have an exclusive use case: when I get admin documents that I have to fill in, and that nobody will ever read, I have AI fill them in for me.

But any time I do anything else, I use my brain. I tried getting LLMs to write research articles for me, and I found that there was no part of the process that they could actually help with. Getting the wall of text or bad list of references meant that I had just doubled my job.

It seems like there should be a good application for these systems, but I don't know what it is.

I work in tech and the biggest use case for AI for me is searching for information about the codebase. Essentially a fuzzy search and finding related information (ex: how does X service do Y operation).

I find if I use it for coding tasks it frequently does a worse job and ultimately slower than me doing it myself. But I also have over a decade of experience just in the language I'm using. Sometimes I use it to write a shell script of repetitive work, or writing some unit tests, but rarely larger more complex things. I really try to avoid using it for things that I can do myself in vim or whatnot because of the environmental impact that a LLM has.

This is what's been annoying me since the beginning. Anytime someone comes up with an example of where AI can be useful it's always coding. It's like they think that literally everyone programs things all the time and does nothing else.

Can you elaborate on this? What does not help for research you reckon?

scientists/ at least some publishers are getting in trouble for having AI written research paper, if not partially written. stem people were already irked, that science papers already were kinda slop or fluff pieces(i suspect its to boosts one's CV for academia hiring)

too bad medical systems, use AI to "Transcribe" i looked at one that seems to come from a really bad model(im not in tech) poor grammer, and spelling, i was trying to decipher what the MD even summarizing.

i have a family member that is a big AI adopter, and still is. probably because he spends so much time in it. the own industry that laid people off for AI in 2023, im surprised people were willing to use it.

i wouldnt even use that as "Faster web search" just a quick glance, of where the source is.

Because research requires precision. To do background reading, I have to throw out everything that isn't relevant to my particular study. My study is new, so the main driver is my concept of what I want to do. The AI is good at finding things with the same keywords, but not at excluding what isn't relevant to my thought process.

I use Google Scholar for searching for literature. I put in some keywords, then scan through the titles. Most of them are not relevant, but I can click through hundreds of titles in a few minutes to focus my attention where it matters, then use the references/citations of those papers to build out what I need to focus on.

AI will give me a massive block of text that includes 5-6 references, with long explanations of what each is about. It is an attempt to do the same thing, but I have to throw out most or all of them, re-prompt, and then get another similar answer. Sometimes it throws things out that aren't peer reviewed.

It also loses the happenstance of stumbling on something irrelevant but important. That's why I still look at the monthly journal publication lists to see what is happening.

Writing is similar. I need to condense several peer reviewed papers to the core ideas that are relevant to what I am saying, which might mean 1-2 sentences. AI will always give paragraphs that are not really getting to what I want, so I still have to do the work myself anyway. It is just adding steps to do the same process.

I recently rejected an AI written paper, and left a nasty note to the authors about how they were wasting my time.

Interesting this is different than my experience with it. I find that it removes the process of understanding and as such it feels empty or vapid. It has helped me finding papers that I would not have found because google offers a bubble in scholar that I cannot get removed from.

Last time I visited a doctor in person was a couple of weeks ago, and þey asked if I'd consent to an AI transciptor. I said no and she brought someone in for it.

I believe it evolved from search engine algorithms, and so honestly this is probably one of the few helpful uses of AI.

Still ignores the ethics, but hey at least it isnt all smoke and mirrors

I could see locally hosted transcriptions being helpful, but its scary how willing some doctors were to throw confidential patient information at the black box spy company.

I think it's because LLMs in single shot scenarios are always liars, fabulists, or both.

The only reason (and only way) it works with coding is because you can hook it up to tools to run static analysis on its bullshit and it can guess enough derivatives in a loop to eventually build something that kinda sorta works.

This doesn't apply for many fields outside of writing software. There just aren't many fields where it applies to have a machine produce permutations and then automatically run validation on the result.

Imagine if you put it in charge of manufacturing or something, you'd have so much wasted raw material before you got to an acceptable result that you'd go out of business.