Comment by wccrawford

I'm still getting there, but I've seen this. I spent hours a couple days ago trying to get pi, llamacpp, and qwen3.6 working well. I had them working before, but not well. I Googled, and I followed it's instructions and following the links it found. But everything was a partial answer, and nothing was clear.

I finally decided to ask Claude to do it. It discovered things I didn't even know to look for, immediately making it run twice as fast, and then also helped me do the changes I wanted, like per-model settings.

I knew it was settings.json, but I had it in the wrong spot (thanks, Google) and what I didn't realize is that the 2 settings.json files are different. 1 is global-only, and 1 is per-model-only. If you put the wrong settings in, it'll just ignore them silently. Which left me wondering why it didn't work.

Then it tested my settings. Then I had to use the official webpages to set the proper settings and it tested those, too. It made suggestions about removing or changing some settings that also made things work better and faster.

In less than a hour, it had vastly outstripped everything I'd tried, and it got it right and verified it.

And with how wrong I've seen Google's AI be (and how wrong others have said it is) I can't trust it to even bring me the basics about things. It's worse than useless, it's distracting.