9 Comments
User's avatar
Alex Lastovetskiy's avatar

The shift from "what should I say to the AI?" to "what should the AI know?" feels like the maturation of the field. Thanks for the practical framework.

Nir Diamant's avatar

well described :)

you are welcome

Skairipa's avatar

Last year I had to create or update hundreds of IT risk and security descriptions. Instead of telling the AI to act as an IT risk & security expert, I instead gave it specific current or new risk and control descriptions as its context as well as the structure of the new or updated descriptions it was to draft. I started first with telling the AI what I was doing and what my goal was, then iterated through each item. Not only did I end cutting about 300+ of potentially 900 hours of work due to the quality of the drafts produced by the AI, but the majority of my final, edited drafts have replaced the old corporate descriptions going forward. To me, context has always been of paramount importance when promoting.

Nir Diamant's avatar

300 hours, wow. That's a solid win.

Yeah you nailed it - giving it actual examples and structure vs just "be an expert" makes all the difference. Most people don't bother with that setup work and then complain when they get generic output.

Cool that your stuff is actually replacing the old descriptions too. That's when you know it worked.

Skairipa's avatar

Prompting not promoting. Dang that Google keyboard AI autospell

Riswan Ahamed's avatar

Well, How does we develop a LLm models with contexual understanding,? example: if suppose i have to develop a model for my startup with the documents, after 1 months or 3 months , the llm needs to by updated of the new info, doe i Need to train the model with new info and ask it to forget the older ones ??