Ask HN: Designing for AI

2 pointstmaly2 comments

I have noticed when you try to ask an LLM to work with some library, the success will usually depend on the training data and on how often the library has iterated with breaking changes.

Has anyone thought about what a good design for a language or a library would look like for optimal use with an LLM?

View on HN

This is a huge problem right now for new or iterating libraries. I tried using AI to help me with nextauth/auth.js and it took me on some crazy tour of v3, v4, v5 before I just read the docs and got it working. My company’s products are less well-known than say Postgres and the current LLMs aren’t great at using them unless you prompt them strongly to read the docs. I think this is going to hold back the rate of change and favor huge incumbents until it is somehow resolved.

Or the new competition uses libraries that the AI knows better and they leapfrog incumbents that are still struggling to use AI.

Reply on news.ycombinator.com