LLMs are just tools
LLMs are just tools
Posted Jun 4, 2026 0:31 UTC (Thu) by gerdesj (subscriber, #5446)Parent article: Tridgell: rsync and outrage
The sheer cost of running a fully tooled up LLM that knows "everything" and thinking you can sell it to everyone at cost plus margin is as daft as thinking you can sell an encyclopedia to every household.
The real utility (I think) of these beasts is in smaller, focused efforts. Qwen 3.6 35B (quantized to 4 bit) can run on an elderly NVIDIA RTX A4000 with 16GB of GPU RAM using llama.cpp. I found another use for my Veeam and others backup boxes - RAM etc was rather cheaper a few years back!
I treat it exactly the same as I do carpentry. I can cut a 10' line through, say, 18mm plywood to within better than a 1mm (0.5mm either side) tolerance with a panel saw, if I am careful. With a table saw I can do better but the cut will be wider and with a rotary saw I will generally want a long guide and it will also be wider. The table saw and rotary saw will always best me for vertical, when set correctly. I do my best - I use my dominant eye to sight and I set myself physically as "square" as possible to the work piece. A panel saw is pretty good at staying square - if it starts to flex then you are going off track!
Hopefully the analogy is not lost on anyone.
When you use a tool you need to know how to get the best out of it. Ignore "magic" thinking (prompt optimisation and other wankery) and instead accept that the answers you get back might be a bit wrong.
