Technology

Local models are really good at tokenizing the text and figuring the intent in the user input. Not perfect, but much better than any possible regexps you can think of. And it's a trivial operation you can run even on a CPU model.

It’s perfect for topics you have professional knowledge of but don’t have perfect recall for. It can bring forward the context you need to be refreshed on but you can fact check it because you are an expert in that field.

If you need boilerplate code for a project but don’t remember a specific library or built in function that tackles your problem, you can use AI to generate an example you can then fix to make it run the way you wanted.

Same thing with finding config examples for a program that isn’t well documented but you are familiar with.

Sorry all my examples are tech nerd stuff because I’m just another tech nerd on lemmy

AI is really good as a starting point for literally any document, report, or email that you have to write. Put in as detailed of a prompt as you can, describing content, style, and length and cut out 2/3 or more of your work. You'll need to edit it - somewhat heavily, probably - but it gives you the structure and baseline.

It is sometimes good at building SQL code examples, but almost always needs fine-tuning since it doesn't know the schema specifics.

Having said that one time it gave me code that resulted in an error, then I went back to GPT and said "This code you gave me is giving this error, can you fix it?" and all it would do is say something like "Correct, that code is wrong and will give an error."

I've done several AI/ ML projects at nation/ state/ landscape scale. I work mostly on issues that can be solved or at least, goals that can be worked towards using computer vision questions, but I also do all kinds of other ml stuff.

So one example is a project I did for this group: https://www.swfwmd.state.fl.us/resources/data-maps

Southwest Florida water management district (aka "Swiftmud"). They had been doing manual updates to a land-cover/ land use map, and wanted something more consistent, automated, and faster. Several thousands of square miles under their management, and they needed annual updates regarding how land was being used/ what cover type or condition it was in. I developed a hybrid approach using random forest, super-pixels, and UNET's to look for regions of likely change, and then to try and identify the "to" and "from" classes of change. I'm pretty sure my data products and methods are still in use largely as I developed them. I built those out right on the back of UNET's becoming the backbone of modern image analysis (think early 2016), which is why we still had some RF in there (dating myself).

Another project I did was for State of California. I developed both the computer vision and statistical approaches for estimating outdoor water use for almost all residential properties in the state. These numbers I think are still in-use today (in-fact I know they are), and haven't been updated since I developed them. That project was at a 1sq foot pixel resolution and was just about wall-to-wall mapping for the entire state, effectively putting down an estimate for every single scrap of turf grass in the state, and if California was going to allocate water budget for you or not. So if you got a nasty-gram from the water company about irrigation, my bad.

These days I work on a small team focused on identifying features relevant for wildfire risk. I'm trying to see if I can put together a short video of what I'm working on right now as i post this.

Example, fresh of the presses for some random house in California:

I've learned more C/C++ programming from the GitHub Copilot plugin than I ever did in my entire 42 year life. I'm not a professional, though, just a hobbyist. I used to struggle through PHP and other languages back in the day but after a year of Copilot I'm now leveraging templates and the C++ STL with ease and feelin' like a wizard.

Hell maybe I'll even try Rust.

Genuinely, nothing so far.

I've tinkered with it but I basically don't trust it. For example I don't trust it to summarise documents or articles accurately, every time I don't trust it to perform a full and comprehensive search and I don't trust it not to provide me false or inaccurate information.

LLMs have potential to be useful tools, but what's been released is half baked and rushed to market as part of the current bubble.

Why would I use tools that inherently "hallucinate" - I. E. are error strewn? I don't want to fact check the output of an LLM.

This is in many ways the same as not relying on Wikipedia for information. It's a good quick summary but you have to take everything with a pinch of salt and go to primary sources. I've seen Wikipedia be wildly inaccurate about topics I know in depth, and I've seen AI do the same.

So pass until the quality goes up. I don't see that happening in the near future as the focus seems to be monetisation, not fixing the broken products. Sure, I'll tinker occasionally and see how it's getting on but this stuff is basically not fit for purpose yet.

As the saying goes, all that glitters is not gold. AI is superficially impressive but once you scratch the surface and have to actually rely on it then it's just not fit for purpose beyond a curio for me.

If you already kinda know programming and are learning a new language or framework it can be useful. You can ask it "Give me an if statement in Tcl" or whatever and it will spit something out you can paste in and see if it works.

But remember that AI are like the fae: Do not trust them, and do not eat anything offered to you.

Software developer here, who works for a tiny company of ~~2~~ 7 employees and 2 owners.

We use CoPilot in Visual Studio Professional and it’s saved us countless hours due to it learning from your code base. When you make a enterprise software there are a lot of standards and practices that have been honed over time; that means we write the same things over and over and over again, this is a massive time sink and this is where LLMs come in and can do the boring stuff for us so we can actually solve the novel problems that we are paid for. If I write a comment of what I’m about to do it will complete it.

For boiler plate stuff it’s mostly 100% correct, for other things it can be anywhere from 0-100% and even if not complete correct it takes less time to make a slight change than doing it all ourselves.

One of the owners is the smartest person I’ve ever met and also the lead engineer, if he can find it useful then it has its use cases.

We even have a tool based on AI that he built that watches our project. If I create a new model or add a field to a model, it will scaffold a lot of stuff, for instance the Schemas (Mutations and Queries), the Typescript layer that integrates with GraphQL, and basic views. This alone saves us about 45 minutes per model. Sure this could likely be achieved without an LLM, but it’s a useful tool and we have embraced it.

OP seems to be talking about generative AI rather than AI broadly. Personally I have three main uses for it:

• It has effectively replaced google for me.
• Image generation enables me to create pictures I've always wanted to but never had the patience to practise.
• I find myself talking with it more than I talk with my friends. They don't seem interested in anything I'm but chatGPT atleast pretends to be

I've been learning docker over the last few weeks and it's been very helpful for writing and debugging docker-compose configs. My server how has 9 different services running on it.

I use it for python development sometimes, maybe once per day. I'll paste in a chunk of code and describe how I want it altered or fixed and that usually goes pretty well. Or if I need a generic function that I know will have been coded a million times before I'll just ask ChatGPT for it.

It's far from "useless" and has made me somewhat more productive. I can't see it replacing anyone's job though, more of a supplemental tool that increases output.

AI isn't useless, but it's current forms are just rebranded algorithms with every company racing to get theirs out there. AI is a buzzword for tools that were never supposed to be labeled AI. Google has been doing summary excerpts for like a decade. People blindly trusted it and always said "Google told me". I'd consider myself an expert on one particular car and can't tell you how often those "answers" were straight up wrong or completely irrelevant to one type of car (hint, Lincoln LS does not have a blend door so heat problems can't be caused by a faulty blend door).

You cite Google searches and summarization as it's strong points. The problem is, if you don't know anything about the topic or not enough, you'll never know when it makes mistakes. When it comes to Wikipedia, journal articles, forum posts, or classes, mistakes are possible there too. However, those get reviewed as they inform by knowledgeable people. Your AI results don't get that review. Your AI results are pretending to be master of the universe so their range of results is impossibly large. That then goes on to be taken is pure fact by a typical user. Sure, AI is a tool that can educate, but there's enough it proves it gets wrong that I'd call it a net neutral change to our collective knowledge. Just because it gives an answer confidently doesn't mean it's correct. It has a knack for missing context from more opinionated sources and reports the exact opposite of what is true. Yes, it's evolving, but keep in mind one of the meta tech companies put out an AI that recommended using Elmer's glue to hold cheese to pizza and claimed cockroaches live in penises. ChatGPT had it's halluconatory days too, it just got forgotten due to Bard's flop and Cortana's unwelcome presence.

Use the other two comments currently here as an example. Ask it to make some code for you. See if it runs. Do you know how to code? If not, you'll have no idea if the code works correctly. You don't know where it sourced it from, you don't know what it was trying to do. If you can't verify it yourself, how can you trust it to be accurate?

The biggest gripe for me is that it doesn't understand what it's looking at. It doesn't understand anything. It regurgitates some pattern of words it saw a few times. It chops up your input and tries to match it to some other group of words. It bundles it up with some generic, human-friendly language and tricks the average user into believing it's sentient. It's not intelligent, just artificial.

So what's the use? If it was specifically trained for certain tasks, it'd probably do fine. That's what we really already had with algorithmic functions and machine learning via statistics, though, right? But sparsing the entire internet in a few seconds? Not a chance.

Edit: can't beleive I there'd a their

I switched to Linux a few weeks ago and i'm running a local LLM (which was stupidly easy to do compared to windows) which i ask for tips with regex, bash scripts, common tools to get my system running as i prefer, and translations/definitions. i don't copy/paste code, but let it explain stuff step by step, consult the man pages for the recommended tools, and then write my own stuff.

i will extend this to coding in the future; i do have a bit of coding experience, but it's mainly Pascal, which is horrendly outdated. At least i already have enough basic knowledge to know when the internal logic of what the LLM is spitting out is wrong.

Troublehsooting technology.

I've been using Linux as my daily driver for a year and a half and my learning is going a lot quicker thanks to AI. It's so much easier to ask a question and get an answer instead of searching through stack overflow for 30 minutes.

That isn't to say that the LLM never gives terrible advice. In fact, two weeks ago, I was digging through my logs for a potential intruder (false alarm) and the LLM gave me instructions that ended up deleting journal logs completely.

The good far outweighs the bad for sure tho.

The Linux community specifically has an anti-AI tilt that is embarrassing at times. LLMs are amazing, and much like random strangers on the internet, you don't blindly trust/follow everything they say, and you'll be just fine.

The best way I think of AI is that it's going through a bubble not unlike the early days of the internet. There was a lot of overvalued companies and scams, but it still ushered in a new era.

Another analogy that comes to mind is how people didn't trust wikipedia 20 years ago because anyone could edit it, and now it is one of the most trusted sources for information out there. AI will never be as 'dumb' as it is today, which is ironic because a lot of the perspective I see on AI was formed around free models from 2023.

I used it to write a GUI frontend for yt-dlp in python so I can rip MP3s from YouTube videos in two clicks to listen to them on my phone while I'm running with no signal, instead of hand-crafting and running yt-dlp commands in CMD.

Also does HD video rips with audio encoding, if I want.

It took us about a day to make a fully polished product over 9 iterative versions.

It would have taken me a couple weeks to write it myself (and therefore I would not have done so, as I am supremely lazy)

I use it like an intern/other team member since the non-profit I work for doesn't have any money to hire more people. Things like:

• Taking transcripts of meetings and turning them into neat and ordered meeting minutes/summaries, or pulling out any key actions/next steps
• Putting together objectives and agendas for meetings based on some loose info and ideas I give it
• Summarise the key points from articles/long documents I don't have tome or patience to read through fully.
• Making my emails sound more professional/nicer/make up for my brainfarts
• Giving me ideas on how to format/word slides and documents depending on what tone I want to employ - is it meant for leadership? Other team members?
• Make my writing more organised/better structured/more professional sounding
• Writing emails in foreign languages with a professional tone. Caveat is I'm fluent enough in those languages to know if the output sounds right. Before AI I would rely on google translate (meh), dictionaries, language forums, etc and it would take me HOURS to write a simple email using the correct terminology. Also helpful to check grammar and sentence structure in ways that aren't always picked up by Word.
• I sound more like a robot than an actual robot, so I ask the robot to reword my emails/messages to sound more "human" when the need arises (like a colleague is leaving, had a baby, etc).
• Bouncing off ideas. This doesn't always work and I know it doesn't actually have an opinion, but it helps get the ball rolling, especially if I'm struggling with procrastination.
• If my sentences are too long for a document, I ask it to shorten/reword and it's pretty capable of doing that without losing too much of the essence of what I want to get across

Of course I don't just take whatever it spits out and paste it. I read through everything, make sure it still sounds more or less like "me". Sometimes it'll take a couple of prompts to get it to go where I want it, and takes a bit of review and editing but it saves me literal hours. It's not necessarily perfect, but it does the job. I get it's not a panacea, and it's not great for the environment, but this tech is literally saving my sanity right now.

Be cautious about the results when using them for googling and summarizing. I had them tell misinformation to me more than once. You'll "learn" things that are counter-factual.

Translating is a very good use case. I also use them for that an it works very well. Better than any Google Translate. And I use it for roleplay, like a D&D campaign, just not with your friends, but alone and the AI narrates the story. And one-off things where I need some ideas to spark my creativity.

What I've tried apart from that are programming, re-phrasing my emails, ... But I've never got any good results for that. Everytime I did that, I ended up not liking the result, deleting it and starting over and doing it myself.

I take pictures of my recipe books and ask ChatGPT to scan and convert them to the schema.org recipe format so I can import them into my Nextcloud cookbook.

I don't use AI for anything. I consider the LLMs pretty useless since they are prone to spewing BS.
I would probably play around with stable diffusion if I had a GPU that would run it at a reasonable speed though.

i use it to autoblog about my love for the capitalist hellscape that is our green earth on linkedin dot com

For me, I use Whisper for transcribing/translating audio data. This has helped me to double check claims about a video's translation (there's a lot of disinformation going around for topics involving certain countries at war).

Nvidia's DLSS for gaming.

Different diffusion models for creating quick visual recaps of previous D&D sessions.

Tesseract OCR to quickly copy out text from an image (although I'm currently looking for a better one since this one is a bit older and, while it gets the text mostly right, there's still a decent amount that it gets wrong).

LLMs for brainstorming or in the place of some stack overflow questions when picking up a new programming language.

I also saw an interesting use case from a redditor:

I had about 80 VHS family home videos that I had converted to digital

I then ran the 1-4 hour videos through WhisperAI Large-v3 transcription and pasted those transcripts into a prompt which had a little bit of background information on my family like where we live and names of everyone who might show up in the videos, and then gave the prompt some examples of how I wanted the file names to look, for example:

1996 Summer - Jane's birthday party - Joe's Soccer game - Alaska cruise - Lanikai Beach

And then had Claude write me titles for all the home videos and give me a little word doc to put in each folder which catalogues all the events in each video. It came out so good I have been considering this as a side business

https://www.reddit.com/r/LocalLLaMA/comments/1gaz5kg/what_are_some_of_the_most_underrated_uses_for_llms/lthuxsu/

I use it for generating illustrations and NPCs for my TTRPG campaign, at which it excels. I'm not going to pay out the nose for an image that will be referenced for an hour or two.

I also use it for first drafts (resume, emails, stuff like that) as well as brainstorming and basic Google tier questions. Great jumping off point.

An iterative approach works best for me, refining results until they match what I'm looking for, then manually refining further until I'm happy with the results.

Regex

Good for softening language in professional environment.

I sometimes have Señor GPT rewrite my nonsensical ramblings into coherent and decipherable text. I recently did that for a paper in my last class. lol
I wrote a bunch of shit, had GPT rewrite it, added a couple quotes from my sources and called it a day.

I'm also currently on a single player, open world adventure with GPT. Myself and the townspeople iust confronted the suspicious characters on the edge of town. They claim to not be baddies but they're being super sus. I might just attack anyway.

if you know programming then you can have it do basic stuff (or even mid complexity stuff if you do it step by step). Just the other day I directed it to produce a code in js using 3js which does scatter plots. The code did run into a couple issues which it was able to solve itself when pointed by me. There was only one problem it could not solve despite several attempts (having a grid does not move with camera controls) so I had to figure that out myself. It was pretty impressive. Overall an expert in 3js would do that maybe in 10 minutes, it took me a couple hours. If I had to do it via searching online it would probably take me a couple of days since I know nought about js.

I also had it write bash scripts a couple of times. It is generally pretty good with writing basic stuff and piecing them together especially if you know programming so you can check it and write intelligible prompts about problems in the code.

I've installed immich on my DIY NAS, it has ml face/object recognition and it works nice

I would generally say they're great with anything you happy being 100% right 90% of the time.

Spaced repetition, in particular Anki with FSRS. I don't think they advertise it as "AI" or even "ML" anywhere, but let's just say gradient descent over gigantic datasets is involved, all to predict the time when you're about to forget something so that Anki can prompt you just before that happens. The default predictor is generic, derived from that gigantic dataset, it's like two handful of tuning parameters, once you've gone through enough cards yourself it can be tuned to your mind and habits, in particular, how you use the "hard, good, easy" buttons.

It's the perfect sledge hammer for the application for the simple reason that we don't actually understand how memory works so telling the computer "here's data from millions of med students and language learners, figure out how to predict it" is our best shot. And, indeed, it's the best-performing algorithm even before you tune it at which point it becomes eerie.

Relatedly, as in "no LLM, no diffusion" Proxima Fusion is using machine learning to crunch through the design space of stellerators to figure out what to prototype in the real world. Actual engineers doing actual engineering.

Then, lastly, yes, playing around with SDXL is fun. Just make sure you can actually judge the images, developing an artistic eye by hitting generate I think is close to impossible. Definitely slower than picking up a pencil, or firing up blender and actually learning how to draw or sculpt.

Guitar amp and pedal modeling.

I think it's mischaracterising the argument against AI to boil it down to "AI is useless" (and I say that as much as a criticism of those who are critical of genAI as I do of those who want to defend it; far too many people express the argument reductively as "AI is useless" when that's not exactly what's really being meant).

The problem is not that genAI is never useful for anything. It is sometimes useful for some things. The problem is that being sometimes useful for some things does not remotely justify what the technology costs. I mean that both on the macro scale - untold climate damage, vast amounts of wasted resources - and on the micro scale; OpenAI alone loses $2.35 for every $1.00 they make.

That is fundamentally unsustainable. If you like genAI for whatever use cases you've found for it, and you really don't care about the climate toll and other externalities, then you can look forward to paying upwards of $50-$100 a month to actually use it, once we're out of the "Give it to 'em cheap/free to get' em hooked" phase, because that's what it'll take to make these models profitable. In fact that's kind of a lowball estimate.

I know plenty of people who find this tech occasionally useful as a way of searching for the answer to a question or producing a small snippet of code, but I can't imagine anyone who finds those uses so compelling that they'd throw "Canadian cell phone contract" levels of money at it.

I think it's useful for spurring my own creativity in writing because I have a hard time getting started. To be fair to me I pretty much tear the whole thing down and start over but it gives me ideas.

Entertainment.

I make porn

But I also used it In tech debug, but less and less.

I run some TTRPG groups and having AI take in some context and generate the first draft of flavor text for custom encounters is nice. Also for generating background art and player character portraits is an easy win for me.

For one thing, they're much better at googling than I am.

Customer support tier .5

It can be hella great for finding what you need on a big website that is poorly organized, laid out, or just enormous in content. I could see it being incredible for things like irs.gov, your healthcare providers website, etc. in getting the requested content in user hands without them having to familiarize themselves with constantly changing layouts, pages, branding, etc.

To go back to the IRS example, there are websites in the last 5 years that started to have better content library search functionality, but I guess for me having AI able to contextualize the request and then get you what you want specifically would be incredible. "Tax rule for x kind of business in y situation for 2024"---that shit takes hours if you're pretty competent sometimes, and current websites might just say "here is the 2024 tax code PLOP" or "here is an answer that doesn't apply to your situation" etc. "tomato growing tips for zone 3a during drought" on a gardening site, etc.

I'm in HR so benefits are a big one...the absolute mountain of content, even if you understand it, even experts can't have perfect recall and quick, easy answers through a mountain of text seems like an area AI could deliver real value.

That said, companies using AI as an excuse to them eliminate support jobs because customers "have AI" are greedy dipshits as AI and LLMs are a risk at best and outside of a narrow library and intense testing are going to always be more work for the company as you not only have to fix the wrong answer situations but also get the right answer the old fashioned way. You still need humans and hopefully AI can make their work more interesting, nuanced and fulfilling.

• to correct/rephrase a sentence or two if my sentence sounds too awkward

• if I'm having trouble making an excel formula

1 Get random error or have other tech issue

2 Certainly private search engines will be able to find a solution (they cannot)

3 Certainly non private search engines can find the solution (they can not)

4 "Chat GPT, the heck is this [error code or something]" Then usually I get a correct and well explained answer.

Ansible.

I fucking hate YAML, and I hate Ansible 'programming' (see "HTML 'programming' language" for rage context).

Chatgpt - I'll use the one in bing or the one in regular-skype - feeds me stuff I can copy/paste/review, and I can get on with my day having lost fewer brain cells to the rage of existing in a world with Ansible fanboys who seem to have forgotten there is NOTHING Ansible does now that we weren't doing in 2003 ... and that the state of the art is 2 generations PAST that glorified mess.

Having used puppet and chef and seen mgmtconfig, I can only applaud RedHat for going with the worst-of-two options and promoting it so hard it appeared viable.

I don't mean to dunk on Michael. Just, James' idea was way better and RH still went with Michael's, and I one day need to know whether the person who had the final say got help.

So far, I've only found it really useful for two things. One is generating text, where I've found using an LLM to generate a title based on a given piece of text is more effective than using other summarisation models, especially for a short piece of text.

I've also found it okay for basic, generic scripts, like trying to figure out what the equivalent Powershell commands for a bash script would be to do something quick, rather than try and learn it from scratch.