LLMs are kinda dumb and basic, even with all their training.  What really makes the LLM something useful is an agent which enables tool usage and makes it so much more useful.

I had Claude write me an agent, which it’s been doing a very good job at as I’ve been iterating new features.  Before long my own agent and LLM should be able to handle updating itself.  Actually, it probably can do that now.

The LLM I’m using is qwen3-coder-next:q8_0.  It’s producing around 4 tokens/sec on CPU inference only so it’s not blasting any doors off things, but it does work.  If I’m not in a hurry, it’s fine.  It’s still faster than I would be doing all this myself.

This stuff is crazy and cool.

#ai #llm #qwen #qwen3