Hi there 👋

Prompt Engineering + RAG has been all the rage. Even now, when I google “Prompt Engineering” I get over 250 million results with a heavy mixture of both sales pitches and earnest explorations on no shot or few shot learning. Admittedly this blog title is extremely controversial, and I’m not here to deny that no shot or few shot learning exists. What I want is to temper the discourse and promises of what prompt engineering CAN get you for various problems....

I recently got into a lively discussion on an exploratory client call about LLM applications and where they are useful. As you may have read from me before, the first thing I like to establish on projects is how precise the use case for this LLM application needs to be. Let’s just say we got sidetracked on the question of whether or not LLM’s are even non-deterministic! As someone dealing with mostly gpt-4-turbo on a day to day basis, I have seen in plenty of my experimental iterations widely variable outputs even at temperature = 0....