I know the reputation that AI has on Lemmy, however I’ve found that some users (like myself) have found that LLMs can be useful tools.
What are fellow AI users using these tools for? Furthermore, what models are you using that find the most useful?
You must log in or register to comment.
I’ve enjoyed messing with Perplexity and Duck AI.
None
LLMs can be useful in hyperfocused , contained environments where the models are trained on a specific data set to provide a service for a specific function only. So it won’t be able to answer random questions you throw at it, but it can be helpful on the only thing it’s trained to do.
Also known as “narrow AI”. You know like a traffic camera that can put a rectangle on every car in the picture, but nothing else. Those kinds of narrow applications have been around for decades already.
The one that the other department tried, and which failed to meet expectations dramatically. Gave management a healthy dose of reality on “AI”.
I use predictive AI for certain classification tasks daily at work, however I call that Deep Learning and not AI. I don’t want to be too specific, but you can imagine we are classifying certain objects - is this a traffic light, is this a tree etc. It is a task that cannot be solved geometrically very good, so Deep Learning is the perfect use case there.
I know people dislike and complain about it, but I absolutely love Suno. LOVE IT. I’ve created what I think are some really cool songs. Will they ever be hits on the radio? Nope. Will anyone else listen to them besides me? Probably not. But boy, after tweaking, I’d rather listen to some of the songs I’ve created than the garbage on the radio!
There’s also Udio.com and Producer.ai out there, and possibly some others - music generation is becoming fairly widespread. I didn’t mention any of this in my list of recommendations though because OP specifically asked for LLMs. :)
I don’t know how Suno has become so much more popular than Udio. Every Suno track I’ve heard has sounded like the same generic pop, and the vocals always have this noticeable “synthy” quality.
Have you heard the stuff from the new v4 model? The vocals are so much clearer and the instrumentation gets pretty varied (ymmv depending on how specific you get with the styles though)
I have – I had high hopes for it, but the vocals are still pretty bad. They all have this metallic, tinny quality, liked they’re subtly vocoded or being sung into those little toy spring microphones for kids. It’s a constant reminder that it’s artificial and just completely takes me out if it.
Holy crow, that freaked me out. That’s really impressive. Pretty uncanny valley, but I can definitely see the appeal.
If we are talking chatbots I see them as another level of abstraction to search and is useful but I have concerns on the energy use. Other uses I have encountered is just sorta a convenience thing. Where it can do a bunch of things that individual software can do but at a one stop shop. I have not directly been involved in other aspects but im aware how they are baked into things like facial recognition and tracking and such.
Dont use AI chat as a replacement for search except on popular subjects with broad consensus. Which unfortunately is when you generally don’t need it.
its fine as long as it gives references to check out. I mean its not fine because of the energy usage but if that is solved I would use it for search. again as long as it tells me sources.
If what you want is sources, a regular Google search will do a better job if it’s a popular subject.
Chatgpt and it’s kin will be inexplicably creative in its choice of sources, and in its summary thereof. And in the sources themselves sometimes.
If it’s something you care to get right, just skip AI.
If it’s meaningless, then it’s harmless in its potential inaccuracySee just like a normal search its up to you to evaluate it. ai search wise is as I said another abstraction. Not using it is like turning off the little snipets search engines do nowadays and going back to just clicking an each and every link. The problem is people just taking the response as gospel with no critical thought.
Like a normal search, except it only provides like 4 links, it’s choice of links is even worse than Google SEO, and it provides inaccurate summaries of them rather than relevant text snippets.
So yes, people just taking the response as gospel is bad. But also it’s just worse than search if you’re using it as a search.
You go to the sites just like you would with the snippets and if they don’t pan out you can rephrase or just go back to a normal search. This is what I meant by a level of abstraction. You can skip the scroll and check what it gives you and if its good you save time (much like the snippets saved time) and if not you are no worse and you correct or go slightly older school. As much as I agree the chatbots can be wrong I don’t find them to usually be off base. They generally find pretty decent resources. Now when they are wrong they can be really wrong but its no different from someone searching and just using the first return without going through the results and evaluating each link. It reminds me when a neighbor in the dorms was explaining html to me as he was making a site and I was like. Why? Just use gopher.
But like I said, if you’re using it like a search then it’s just worse.
It’s way slower, it’s way fewer results, and the summaries can never be as accurate as verbatim quotes of the pages themselves.The only reason to use AI is to get the summary, and then you’re not using it as a search. Maybe it provides some references you can use to fact check, but that’s still not a search.
You have to remember, LLM’s are literally just auto-complete. They don’t have the goal of giving you an answer or resources, they have the goal of providing text that would complete the conversation in a way that looks similar to what they’ve seen before. If they can give you a biased answer supported by cherry picked references, that’s just as valid of a completion, because it looks like how such a conversation might be completed.
In situations where the response can be inherently assessed for correctness (eg art, but that’s a whole other ethical issue) or correctness can be automatically verified (eg programming) then there is some value in limited use.
Although I personally think the implications of putting it in the hands of business owners isn’t worth it, that another topic.
Uh, kinda, maybe? Most use cases are things that I don’t really see the use of, or have found to produce more flawed results than previous ways of doing things.
However, a couple years ago, someone on reddit said their perfect use case was using AI to hunt down things you enjoyed but have forgotten the titles of - books, movies, tv series, songs, videogames, etc.
Well, I have a few of those half-forgotten items, where I’ve remembered snippets of things but have no idea what they actually were called. I’ve tried looking them up over the years with regular search engines, with no luck. And a few times in the past couple years, I’ve used random AI engines to try to track these titles down.
And the thing is, AI absolutely has not been able to tell me what the titles to anything was. However, in trying to come up with more details to pass to AI, I’ve accidentally found other webpages that helped me find what I was looking for. Like, one of the things I was looking for was a horribly bad 1970’s tv movie and, in my latest search for the title, on like page 8 of my google/duckduckgo results trying to find something to feed the AI from what little I remembered, I ran across a website that lists the cast and plot of, like, every tv movie. Not just top ten, or people that became big stars later, or horror movies or whatever, but every movie from like the 50s /60s on. And I sat on that website and read through the high-level plotline of every tv movie from, like 1968 on, and eventually found the movie.
There was a book where I remembered the first name of the main character and a very specific scene, even some of the exact words, and the three AI engines I tried couldn’t tell me anything. But in searching for more details (and I had tried serving for it before), I eventually ran across a book site that helped me out. Interesting thing: when I passed direct quotes to AI, they couldn’t tell me, when I asked what books had that plot, they couldn’t tell me, but if I asked if a specific scene happened in the book, they said it was there.
I have one game that I’m still searching for, but AI engines have inadvertently helped me find most of the rest of my wishlist.
GPT to write emails
I’m going to need you to elaborate on this one…
Not the person you asked, but I have a similar use-case.
I write a lot of emails for work. Most of them are written from templates that I’ll use dozens of times a day, and some of those templates are just large blocks of text full of information that are ugly and hard to read.
I’ll sometimes take these templates, plug them into ChatGPT, and ask it to reword the email. Perhaps I want it to have a more empathetic tone for an emotionally-elevated user, or maybe I need it to sound more technical for a more knowledgeable user, or simplify the explanation for a less knowledgeable user, etc. I’ll then use that output as a base to write my own version from there.
None of the GPT output goes into my actual emails, though. I’m mostly using it for inspiration purposes, to help me write my own messages with verbiage or perspectives I may not have originally considered. It’s super useful when you have a user who just isn’t understanding your instructions and you need to word it differently, or if you just need a fresh take on some stale templates.
Exactly ! Same here
Thank you for the response, and I can totally relate.
I like that GPT can help me brainstorm and then I take the reigns from there. Not everything in my career is crystal clear and more often than not, my boss is asking me for solutions because he doesn’t already have answers.
I find that things like GPT really help as jumping off points that help to get the conversation started in real life situations.
removed by mod
I like this. It’s a fun use situation that can enhance creativity as well as perform as an entertainment platform.
I used GPT to help me plan a 2 week long road trip with my family. It was pretty fucking awesome at finding cool places to stop and activities for my kids to do.
It definitely made some stupid ass suggestions that would have routed us far off our course, or suggested stopping at places 15 minutes into our trip, but sifting through the slop was still a lot quicker than doing all of the research myself.
I also use GPT to make birthday cards. Have it generate an image of some kind of inside joke etc. I used to do these by hand, and this makes it way quicker.
I also use it at work for sending out communications and stuff. It can take the information I have and format it and professionalize it really quick.
I also use it for Powershell scripting here and there, but it does some really wacky stuff sometimes that I have to go in and fix. Or it halucinates entire modules that don’t exist and when I point it out it’s like “good catch! That doesn’t exist!” and it always gives me a little chuckle. My rule with AI and Powershell is that I don’t ask it to do things that I don’t already know how to do. I like to learn things and be good at my job, but I don’t mind using GPT to help with some of the busy work.
I got an email once from HR that said I got a bike commuter benefit I didn’t know about, and couldn’t find more information about in the attachment, so I emailed HR and it turns out they used AI to write the email, and wouldn’t be giving out any corrections or bike commuter benefits. Bullshit.
https://notebooklm.google.com/ is really handy for various things, you can throw a bunch of documents into it and then ask questions and chat interactively about their contents. I’ve got a notebook for a roleplaying campaign I’m running where I’ve thrown the various sourcebook PDFs, as well as the “setting bible” for my homebrew campaign, and even transcripts of the actual sessions. I can ask it what happened in previous episodes that I might have forgotten, or to come up with stats for whatever monster I might need off the cuff, or questions about how the rules work.
Copilot has been a fantastic programming buddy. For those going a little more in depth who don’t want to spring for a full blown GitHub Copilot subscription and Visual Studio integration, there’s https://voideditor.com/ - I’ve hooked it up to the free Gemini APIs and it works great, though it runs out of tokens pretty quickly if you use it heavily.
To be honest, this is the only thing Google did right about AI IMO.
https://notebooklm.google.com/ is really handy for various things, you can throw a bunch of documents into it and then ask questions and chat interactively about their contents.
Nice, thanks! I’ve been looking for something I can stuff a bunch of technical manuals into and ask it to recite specifications or procedures. It even gave me the document and pages it got the information from so I could verify. That’s really all I ever wanted from “AI”.
Supermaven code autocomplete in vscode is really nice
Copilot in VScode is something you’d have to tear out of my cold, dead hands. Pressing Tab to auto complete is so useful. I use the GPT 4.1 model or whatever it is called. I tried Gemini but for some reason it’s complete ass when doing code. Android Studio Gemini is worse than the free tier on the website.
However, I’ve found the Gemini Pro model on the website is incredibly good for information assistance. To give an idea of my current uses, I have two chats pinned on it: fact checking and programming advice. I use the former for general research that would take more than a few minutes of Googling but need an answer now, and the latter for brainstorming code design or technical tutorials (recently had it help me set up a VM in WSL).
One tool I wish I could use is ElevenLabs. Had a friend on the free tier of it make some really cool and convincing voice lines (I forgot what character it was) a long time ago. Looks easy to use too. I can’t justify spending money just to play with it but if I had a purpose for it, I would.
Just today I was tinkering with Continue.dev extension for VSCode. Locally running the models and not having sensitive proprietary source code sent over the wire to a 3rd party service was a big requirement for me to even consider bringing AI into my IDE.
Tabby is a locally ran one that I’ve been really enjoying too
