Effective use-cases for LLMs

17 points by markerz

nugget

A local LLM has effectively ended false-positive alarms from my security cameras (Frigate NVR) by analyzing the video of motion-trigger alerts to discriminate between real and false triggers.

Ive taught it how to identify when a specific gate is open or closed just from the camera images, and I'm working on teaching it to identify specific vehicles and animals when analyzing events with modest success.

Vaelatern

Is this default frigate behavior, or are you doing something special on top?
vittorioromeo

Which model?

nwjsmith

Mutation testing is made significantly easier to implement with LLMs.

joelgrus

I gave my coding agent skills to query cloudwatch logs and dynamodb, and now it can root-cause prod errors in 5 minutes that would seriously take me at least an hour to figure out.

(Hopefully my skills aren't atrophying as a result, but it's really an incredible tool for this.)

zladuric

Mine definitely are; do it manually occasionally just go refresh your knowledge!

bartz

I sometimes use small models* to replace simple google searches where I'd only look at the AI overview anyways. Stuff like "neccessary or necessary", or "NYC metro population"

*small enough to be fast with just 16GB RAM, like Gemma 4 E4B at Q4 quantization

justinpombrio

Be careful with this sort of thing: that query is pretty close to the query "how many buildings are in NYC", for which an LLM gave an answer to me that was off by more than an order of magnitude.
abareplace

On a Mac/iPhone, you have an offline dictionary; as a non-native speaker, I also use a spell checker. And I would not trust a small LLM for factual information like NYC metro population.
markerz

I also like LLMs for Google searches where I’d take the “average” result and highlight outliers. For example, I often load the top 5 recipes and see the range of ingredients and quantities I could use, and notice when a recipe is an outlier.
- bartz
  
  That's a very clever use case!
kraxen72
interesting! for both of these, i'd probably do a google search (or, searx ng in my case), or for spellng, i use the krunner spellcheck plugin, or wordhippo for synonyms/antonyms (i have a custom search engine, e.g. whs emerge -> synonyms for emerge. wha similarly for antonyms).

what i do find small/dumb/fast llms good at is somehing right between a google search and a full chat question: stuff that'd likely fail as a generic keyword google search, or too niche to yield results, but not stuff that i'd bother opening a new llm convesation for (no files as context needed, no "thinking" mode, no follow-up messages)

e.g.
- latex paragraph don't tab it in but hug left
- dnf cli can i see a column when a package was updated/published as this version?
- svelte 5 named slot alternative in this case <paste in my component>
- css selector badge-shape.ytBadgeShapeCommerce > .ytBadgeShapeText that has both the text and .ytBadgeShapeIcon
right now i use google's ai mode which runs gemini 3.5 flash, but i'm i have used a bunch of other providers in the past. i don't really care as long as 1) i can pass the query string in the url, so i can set it up as a custom search engine and 2) it's fast, like enter to response under 5 seconds.
pyfisch

These types of queries are where small LLMs with tool calling should shine, if you integrate them with web search or a knowledge base. Unfortunately I don't know a good system to self-host this.
- markerz
  
  Personally I self-host Open WebUI as my ChatGPT.com-like interface. There’s some performance issues but otherwise it feels good. I tried integrating with Kagi Search but it’s starting to feel a bit expensive with both a monthly subscription AND a per-API call cost. Been using Exa.ai for their free tier.
  
  For automations and webhooks, I self-host n8n. That’s it’s own set of headaches but I find it quite useful.