Hacker News Reader

Comment by dj_gitmo | Hacker News Reader

We should educate people about what questions you should, and should not, be asking LLMs in the first place. You really should not be asking an LLM

> "Did fighters of the Azov battalion burn an effigy of Trump?”

LLMs are horrible at answering questions about small isolated incidents. Part of modern media literacy should include understanding what LLMs can reliably answer.

View on HN

Media experts struggle to keep up with the latest in generative AI capabilities and shortcomings, educators are unfortunately far behind. School districts themselves even struggle to cut through propaganda in their official curricula. I think things are going to get a lot worse before they get better.

So what questions are LLMs good for?

The ones you can easily answer yourself without the use of an LLM.

> Write me a Github CI skeleton to build and push a docker container to GAR

The kind where the answer is not known, but is easy to verify. "What is the term for...?", "How do I do X with Y library?"

I use google search results including the AI summary for programming. Easy to verify the results too, but I often can't put a good search query together to get what I want, so googles implementation of AI helps here. I don't do overly complicated stuff tho so there's that.

I am also trying to understand the shady limits of LLMs. But your example doesn't give incorrect answers in ChatGPT 4o, Sonnet 3.5 nor Deepseek V2.

Reply on news.ycombinator.com