[chipdart]

> Another big area of hype is "prompt engineering." That one seems to have calmed down slightly, but for a while, there were large swaths of the Internet who were amazed that the set intersection of "talk like a decent human being" and "be precise in your communication" could generally lead to good results.

I think your comment conveys your obliviousness of the problem domain.

The main driving need for prompt engineering is not an inability to "talk like a decent human being". That's just your personal need to insult and demean people who are interested in a problem domain you know nothing about.

The main driving need for prompt engineering is aspects like not being able to control how context is formed and persisted in a particular model, and how to form the necessary and sufficient context to get a model to output anything interesting. Some applications require computationally expensive and time-consuming runs, and knowing what inputs to provide to a system which by it's very mature is open-ended is a critical skill to adequately use the system in professional settings.

Let's put it like this: GitHub copilot is a LLM service which is extremely narrow in what are their applications and use cases. Yet, you can't even get it to add unit tests to a function following a specific style without putting the effort to build up the context it needs to output what you expect.