Who does your assistant serve?

I've commented in the past on the scary reliance on building products around an LLM. Xe also shares their thoughts regarding LLM upgrades.

Most of the time, changing out models is pretty drastic on an AI workflow. In my experience when I've done it I've had to restart from scratch with a new prompt and twiddle things until it worked reliably. The only time switching models has ever been relatively easy for me is when I switch between models in the same family (such as if you go from Qwen 3 30B to Qwen 3 235B). Every other time it's involved a lot of reworking and optimizing so that the model behaves like you'd expect it to.

The full post dives into the ChatGPT-5 boyfriend fiasco and some hard truths on relying on LLMs for counselling.