[lovehashbrowns]

I made myself a silly ai-chungus it works as a series of containers that communicate over mqtt. I have an ollama shim for other services to talk to a model on my other machine, telegram shim that acts as a ui, a study component that will give me a random subject for me to study over the course of a week and I give a proper implementation (let’s say a ring buffer) and it reviews the code I wrote. It has some minor gen things using comfyui like a card pull system with cards, card text, rarity system, and special card effects like holo. And my favorite bit has been tying my todo system into it so I have a thermal receipt printer which will print a task I have to do for the day, prints a barcode, and when I finish the task I scan it with an iOS shortcut. It’s beautiful. What’s been the most fun was designing the mqtt topics in a way that makes everything else flow perfectly. Oh and there’s a tts system that uses kitten-tts which will produce audio for certain messages and an esp32 that gets those sound files and plays them on a speaker, or I can play the messages on telegram. Like if I do an /overview command on telegram i get an overview of my incomplete tasks and the ollama model helps prioritize. It’s my favorite use of ai junk at the moment

[andai]

This is great, you should do a write up. With photos!

Can you explain the study thing in more detail? Or give an example of how you use it?

[lovehashbrowns]

Sure so there’s a command in telegram you do /study and you can either tell it a subject or it’ll randomly generate one. I did /study ring buffers last time. So then opencode / GLM 5.1 or whatever write an entire site for me explaining the concepts, tradeoffs between things like infinite read write pointers, memory mirroring, how to tell if the writer looped the reader when the pointers are not monotonically increasing and instead get modulo’d, and bit masking with powers of 2 etc etc. then it wrote a ring buffer with various features in pseudo code so I can’t cheat. I also got papers to read from ACM and such. So I pickled it all over a week and came back with an Odin implementation, submitted to the site running on docker, and the ai reviewed it for me, telling me mistakes I made or areas where code isn’t really professional or I missed something. It’s genuinely super helpful. And the idea is whatever the AI tells me isn’t my only source of info so for example if the ai says infinitely increasing pointers will overflow but for example if I’m using Python, Python integers are basically bounded by memory and practically on most systems “infinite” so shut up AI I can do that or if it gives wrong info about something I’m getting published journals or websites as reading material that (hopefully lol) isn’t hallucinated. But really it’s the classic college style learn it then do a lab type of loop. Except it’s topics I’m genuinely interested in in the moment so the energy flows perfectly into it. My only wish would be to have it be cooperative working actively with others for the more accurate college experience bouncing ideas off other people seeing where we’re all confused or just me etc