I Like AI More When It Is Not Driving

I like AI more when it handles the fuzzy language and leaves state, consent, and permissions to normal code. Pocket Nenek got less magical in the implementation and more reliable for the person using it.

It was tempting to send every message through one clever prompt and let the model decide what the user meant, which flow they were in, and what should happen next. That demo felt flexible until one uncertain message made the food trail forget where somebody was.

The useful boundary is ambiguity versus control. A model can classify a messy food memory, suggest a friendly reply, or summarise text, while ordinary code decides whether consent exists and which state transition is allowed.

That boundary also makes fallbacks less dramatic. If the preferred provider is unavailable, the bot can use another provider or a safe response, while known commands keep working because their logic never depended on the model.

I want the model to be helpful without becoming the owner of the conversation. The bot should remember the trail step, make buttons behave predictably, and refuse to save something when the required permission is missing.

This pattern keeps showing up in my projects: small probabilistic islands inside a system with boring, explicit bridges. The user gets a more natural interaction, and I still know which part of the system is allowed to make a decision.