LLMs are pretty good at reverse dictionary lookup. If I’m struggling to remember a particular word, I can describe the term very loosely and usually get exactly what I’m looking for. Which makes sense, given how they work under the hood.
I’ve also occasionally used them for study assistance, like creating mnemonics. I always hated the old mnemonic I learned in school for the OSI model because it had absolutely nothing to do with computers or communication; it was some arbitrary mnemonic about pizza. Was able to make an entirely new mnemonic actually related to the subject matter which makes it way easier to remember: “Precise Data Navigation Takes Some Planning Ahead”. Pretty handy.
Great for giving incantatons for ffmpeg, imagemagick, and other power tools.
“Use ffmpeg to get a thumbnail of the fifth second of a video.”
Anything where syntax is complicated, lots of half-baked tutorials exist for the AI to read, and you can immediately confirm if it worked or not. It does hallucinate flags, but fixes if you say “There is no --compress flag” etc.
This is the way.
Tailored boilerplate code
I can write code, but it’s only a skill I’ve picked up out of necessity and I hate doing it. I am not familiar with deep programming concepts or specific language quirks and many projects live or die by how much time I have to invest in learning a language I’ll never use again.
Even self-hosted LLMs are good enough at spitting out boilerplate code in popular languages that I can skip the deep-dive and hit the ground running- you know, be productive.
I do this as well—I’m currently automating a repetitive workflow for work using python. What’s the latest project you’ve generated boilerplate code for?
I’ve done lots of cool things with AI. Image manipulation, sound manipulation, some simple videogames.
I’ve never found anything cool to do with an LLM.
Care to expand on sound manipulation? Are you talking about for removing background noise from recordings or something else?
Some speech recognition work, some selective gain adjustments –not just amplifying certain bands of frequencies, but trying to write a robot that can identify a specific instrument and amplify or mute just that. Also fun with throwing cellular automata at sound files. And with throwing cellular automata at image files to turn them into sound files.
That all sounds pretty neat. Do you do these things locally or is there a cloud service for that?
I did them locally, a long time ago, before cloud was ubiquitous. Some of the project files might still be on my university’s servers, but I doubt I could find them again, at least for the sound editing robots. I know I’ve got some of the image-eating cellular automata around –I was looking at them recently– but the library they depended on is broken.
ChatGPT kind of sucks but is really fast. DeepSeek takes a second but gives really good or hilarious answers. It’s actually good at humor in English and Chinese. Love that it’s actually FOSS too
Legitimately, no. I tried to use it to write code and the code it wrote was dog shit. I tried to use it to write an article and the article it wrote was dog shit. I tried to use it to generate a logo and the logo it generated was both dog shit and raster graphic, so I wouldn’t even have been able to use it.
It’s good at answering some simple things, but sometimes even gets that wrong. It’s like an extremely confident but undeniably stupid friend.
Oh, actually it did do something right. I asked it to help flesh out an idea and turn it into an outline, and it was pretty good at that. So I guess for going from idea to outline and maybe outline to first draft, it’s ok.
My experience is that while it’s useful for creating code from scratch it’s pretty alright if you give it a script and ask it to modify it to do something else.
For instance I have a cron job that runs every 15min and attempts to extract .rar files in a folder and email me if it fails to extract. Problem is if something does go wrong it emails me every 15minutes until I fix it. This is especially annoying if its stuck copying a rar at 99%.
I asked deepseek to store failed file names in a file and have the script ignore those files for an increasing amount of time for each failure. It did a pretty good job, although it changed the name of a variable halfway through (easy fix) and added a comment saying it fixed a typo despite changing nothing about that line. I probably probably would have written almost identical code but it definitely saved me time and effort
The output is only as good as the model being used. If you want to write code then use a model designed for code. Over the weekend I wrote an Android app to be able to connect my phone to my Ollama instance from off my network. I’ve never done any coding beyond scripts, and the AI walked me through setting up the IDE and a git repository before we even got started on the code. 3 hours after I had the idea I had the app installed and working on my phone.
I didn’t say the code didn’t work. I said it was dog shit. Dog shit code can still work, but it will have problems. What it produced looks like an intern wrote it. Nothing against interns, they’re just not gonna be able to write production quality code.
It’s also really unsettling to ask it about my own libraries and have it answer questions about them. It was trained on my code, and I just feel disgusted about that. Like, whatever, they’re not breaking the rules of the license, but it’s still disconcerting to know that they could plagiarize a bunch of my code if someone asked the right prompt.
(And for anyone thinking it, yes, I see the joke about how it was my bad code that it trained on. Funny enough, some of the code I know was in its training data is code I wrote when I was 19, and yeah, it is bad code.)
Crappy but working code has its uses. Code that might or might not work also has its uses. You should primarily use LLMs in situations where you can accept a high error rate. For instance, in situations where output is quick to validate but would take a long time to produce by hand.
It’s good at paraphrasing paragraphs to contain no ‘fifth glyphs’
That’s a big bound forward from last I was looking at it! Avoiding that nasty glyph was notably not in its portfolio of tricks. It would say it was avoiding the fifth, but still slip many through.
Assuming that this discussion is about LLMs, anyway.
I had to instruct it to consult a script to know how many words did contain fifth glyphs, but it did work with that.
Sounds as though your script did all important work, not your AI.
It was vital to call on my LLM, it couldn’t do it on its own, and my script can just count glyphs, not anything to do with words.
Getting my ollama instance to act as Socrates.
It is great for introspection, also not being human, I’m less guarded in my responses, and being local means I’m able to trust it.
Finding specific words in an MP3 log file for our radio station. Free app called Vibe transcribes locally.
I’ve used llms to generate dialogue trees for a game and generate data with coordinates to describe the layout of the game world. in some ways it can replace procedural generation code.
Table top games?
video game
With mixed results I’ve used it for summarising the plots of books if I’m about to go back into a book series I’ve not read for a while.
This is a very rare use case, but one where i definetly found them very useful. Similar to another answer mentioning reverse-dictionary lookup, i used llms for reverse-song/movie lookup. That is, i describe what i know about the song/movie (whatever else, could be many things) and it gives me a list of names that i can then manually check or just directly recognize.
This is useful for me because i tend to not remember names / artists / actor names, etc.
I’ve used them both a good bit for D&D/TTRPG campaigns. The image generation has been great for making NPC portraits and custom magic item images. LLM’s have been pretty handy for practicing my DM-ing and improv, by asking it to act like a player and reacting to what it decides to do. And sometimes in the reverse by asking it to pitch interesting ideas for characters/dungeons/quest lines. I rarely took those in their entirety, but would often have bits and pieces I’d use.
Good for gaining an outside perspective/insight on an argument, discussion, or other form of communication between people. I fed it my friend’s and their ex’s text conversation to it (with permission), and it was able to point out emotional manipulation in the text when asked neutrally about it:
Please analyze this conversation between A and B and tell me what you think of their motivations and character in this conversation. Is there gaslighting? Emotional manipulation? Signs of an abusive communication style? Etc. Or is this an example of a healthy communication?
It is essential not to ask a leading question that frames A or B in particular as the bad or the good guy. For best results, ask neutral questions.
It would have been quite useful for my friend to have this when they were in that relationship. It may be able to spot abusive behaviors from your partner before you and your rose-colored glasses can.
Obvious disclaimers about believing anything it says are obvious. But having an outside perspective analyze your own behavior is useful.
I’ve used LLMs to reverse engineer some recipes.
Do you just try and describe what it tastes like?
I either feed it the list of ingredients or it finds them itself if it’s a popular item. It’s good at guessing the proportions of the ingredients if you’ve got the label.
