Projects in my repos done with the help of "AI" (Average Intelligence) tools have "Sloppy" in their name.
"AI" is a great to get stuff done fast, but it's dum as hell and have to be extremely carefully guided.
Some people compare it to a parrot with a gigantic brain. But it's still just a parrot facerolling on the keyboard.
LLM is not even a neural network, it's an autocomplete dictionary for T9 text prediction from old Nokia phones.
Repeatedly tap on your phone's text predictions - this is the current state of "AI". Transformers arch is garbage.
Though autocomplete was always a good tool anyway. Now with proper expectations you're ready to start building.
Oh, BTW. Stop using cloud services, start with your own local LMStudio/ComfyUI machine. Save monies.
You need 16+ GB VRAM on Nvidia card (preferably 24-32GB), 32+ GB RAM (pref 64-128GB), any half-decent CPU.
3 weeks of pure suffering and you're ready for a true AI future (without quotemarks), it'll pay off in less than a year.
Our videocards now can not only run games, but write somewhat useful code. That's pretty cool right?
And if part of your job or pipeline can actually be replaced by a parrot - maybe it should be replaced.
Think of writing and updating tests. If you're blank-staring at the wall right now, you get it.
Wasserman 56k - unsloth/gemma-4-31b-it@iq4_xs (for 24GB GPU, heavy and reliable, temperature 0.3, top k 64)
Pentester 64k - xortron.criminalcomputing.2026.27b.next@q5_k_m (Qwen3.5 for 24GB GPU, temperature 0.3, top k 40)
Crackhead 150k/32k - ex0bit/gemma4-26b-a4b-prism-pro-dq (16GB GPU + 32GB RAM, temperature 0.3, top k 64)
Use "<|channel>" as your thought start string for Gemma 4, not "<|channel>thought".
On 16GB VRAM card you will have 32k context window while computing 8 layers on CPU, and it's still a Crackhead model. It's basically for testing before buying hardware for fat models.
I use low temperature to prevent tool use typos/screwups, it's a very common problem.
Put Thoughts and responses MUST be terse and short. in a rule or use my portable caveman prompt.
Quantize your vision .mmproj files to Q8_0 so you don't have to blind the model completely.
Don't use uncensored/abliterated crap, every bit of KL divergence makes a huge difference.
Never use Q8_0 KV Cache, it kills the tool calls because it introduces typos and lobotomizes the model.
When short on memory, always disable Unified KV Cache and set Max Concurrent Prediction to 1.
Use OpenAI-compatible API to connect to LM Studio. Best open-source agentic IDE atm seems to be https://zed.dev/
Here are jinja templates for LM Studio and Zed. Very tedious to get right.
- ComfyUI-Enhancement-Utils - PC resource monitor and execution follower
- ComfyUI-SloppyAudio - Audio editing tools based on SoX and BS-RoFormer
- smol-caveman - Portable Caveman prompt designed for local LLMs. Read less slop and get much better results.
- ComfyUI-SloppyInstall.bat - Simplified pip install -r "requirements.txt" for custom nodes in portable ComfyUI.
- SloppyServer.bat - Single file local/Wi-Fi server for debugging multithreaded mobile Unity WebGL builds and other apps