Asklemmy

For image gen I don’t have a good use but it is very complex which means sometimes I can just lose an hour or 2 fumbling around in a complex network of nodes.
What I found fascinating was how strangely good the results were when I created an image, then fed the result back as input, and repeated that process.
The only useful thing I used image gen for was creating references for an artist to create a PFP for me that looks rad as hell.

As for LLMs, also not really. I think about 90% of the time LLMs either give a useless or just wrong answer. I can’t seem to find the thing that LLMs are supposed to be good for. One thing all LLMs I tried have failed consistently at was finding a movie from a vague description I gave.

I've been using it recently for generating alt text for images (my bots on Mastodon and Aunty Madge on !yoursinclair@retrolemmy.com specifically). It's pretty good at that, although does sometimes give weirdly wrong details - especially the TED Music Bot, if it gets the usual +4 startup screen it says it's +4 on key F1, instead of 3-plus-1, and tells me the wrong colours for the text and background (I think it may be getting it confused with the C64, bit the colours are right there on the image!). It's infinitely preferable to having no alt text, which would be the alternative.

The other thing it's really good at is summarising articles.

I've also used it when I've had an error in my code I can't track down, or a bracket missing that I can't figure out. It quite often gives nonsense but I've had some success. Usually a normal web search is perfectly adequate though!

I used it to extract thousands of rows from tables in PDFs and generate enumerations for them in various programming languages. I had to do some pre-processing with the Python script, and review all the output to make sure it didn't screw up, but it saved me a lot of time.

I took AI courses in college and it was fun to learn about then when it was a bunch of toy examples that showed the potential of these systems, but it was clear enough to anyone in those classes or doing that research how not ready they were for real applications because of all the known flaws with how model training worked. And then some ceos just ignored all that and started blowing up the bubble.

So my answer is the research models that could play video games kinda good. Everything after that was getting ahead of ourselves.

I think there are many thousands of folk in fields beyond IT that use it all the time. It's by no means perfect, but for many of us managing teams or doing boring AF admin, working with procurement, writing user documentation or trying to navigate basic system configs then it's immensely useful.

That stupid song about the Windows registry. Don’t look it up, or at least don’t come crying to me after listening to it 50 times. It has no right being as catchy as it is. I fear for younger people and how God generated music will sound in a year or two.

I used to ask why classical music wasn’t performed by Metal bands. Now I can have AI make it for me and it shouldn’t even be that hard given the songs are already written. Now if I were to say “compose a new symphony in the style of Beethoven and make it sound like Iron Maiden are playing it… that would be different.

A couple so far. I have a local copy of Stable Diffusion. It's handy for upscaling some kinds of images. I'll also use it to flesh out my worldbuilding project with landscapes and scenes.

Less often, I'll consult ChatGPT, without logging in. For the times when a search engine doesn't cut it but a forum post would be too much. I'm usually skeptical of AI summaries, but I find it justified for boiling down poorly-written stuff that I have to read, but isn't worth my time in long form.

I actually just view it as the latest abstraction of search. Yahoo in the 90's did not give you a blurb summary of links or would do math equations for you or give answers to simple questions like what time is it or whats the capital of alaska.

I've tried learning digital drawing before, but my programmer brain finds prompt engineering to be much more intuitive, so i've been doing that a lot lately.

Also its surprisingly good at upscaling in "image-to-image, 0.1 strength" mode. I thought I would need a dedicated upscaling mode for that. The result looks noticeably better than with normal bicubic upscaling.

bash scripting, light python programs, fixing software and hardware issues, learning Japanese, learning other things, writing applications, all the boring stuff and stuff need extreme repetition or data mining. Ideas comes from me, work is done by machine.

As a teacher, it's good at making assignments for students. Though I don't use it out of principal.

Best use for me has been feeding it logs and pcaps, incorporating it into Wireshark sounds far more useful to me than summarizing a two-line email.

I had ai write a python script for me that checked a directory for zip/rar files that don't also have video files, unzip them and if it fails retry at a doubling interval. It would have taken me an hour or so, took the ai 5min+15min of me fixing it. I also had it add logging which is definitely something I wouldn't have been arsed to add otherwise.

It isnt very helpful at my day job writing professional code though, and I dont think someone who didn't already know how to do it themselves could make use of it.

I use notebooklm to act as my research assistant when it comes to labour law and union contract. Claude has helped significantly with compiling the chapters for my upcoming book. Deepseek has created several courses for me to further refine my skills and education.

I find it's good as a "search engine of last resort". I would pay $2 monthly at most for one, and less if it's American.

Despite claims to the contrary ChatGPT is very accurate with its responses when such responses involve web searches. Where it falls apart is complex multi step things like coding questions.

I make heavy use of it to skip past all the clutter of Google search results and end up with clear summaries that answer my questions. That's all I'd really use it for as anything more than that its output is highly variable in quality.

I only really use LLMs for project ideas, naming things, or recipes, but it’s great for those. For recipes especially, trusting the gaslight machine adds an extra layer of suspense and fun to cooking or baking.

It’s also fun, but not useful, to ask Linux questions - chatgpt specifically told me to restart dbus while in an active GUI session when I asked it about a simple case of unintended behaviour. It’s probably the dumbest Linux advice I’ve heard to date, so I got a lot of enjoyment out of trying it to see how bad it would fuck things up