
You'd coax it, trick it, wrap your instructions in elaborate rituals. Please return JSON. Only return JSON. Do not return anything other than JSON.

The combination of good system prompts and well-structured RAG has covered every use case we've thrown at it. That's not to say fine-tuning is useless. It isn't. But it solves a different problem than most people think

The thing ran beautifully. But when I stepped back and looked at what it was actually doing, I realised I'd built a cathedral where a shed would have done.

For the first year or so, I treated it like a very fast intern. I'd give it tasks, it would do them, I'd review the output and either accept it or send it back with corrections

Hallucinations aren't a bug. They're not a flaw that will get fixed in the next model release. They're the nature of how this technology works, and understanding that changes how you work with it.

Language models are probability engines. They know a lot about everything and very little about something specific.

Tool calling has been around for a while now. In AI terms, that means about two years, which feels like a decade.