That's the whole problem. We are developing software against LLMs in the same way we use SQL: we write our own instructions as a programmer, then we combine them with input from our users With SQL we can do that safely. With LLMs we can't

→ View original post on X — @simonw,