Large language models (LLM) – ChatGPT and friends – are one of those technologies with a crazy learning curve. They look simple and friendly (resulting in plenty of useless demoware) but become devilishly hard to work with once you try to squeeze consistent value out of them.

Most people don’t want to talk about the hard stuff (sexy demoware results in more page views), but there’s an occasional exception, for example All the Hard Stuff Nobody Talks About when Building Products with LLMs describing all the gotchas Honeycomb engineers discovered when creating a LLM-based user interface.