If AI talks so well, why does it still get things wrong?
7:44 · Horizontal
Give your AI a checkup first
Why it can be confidently wrong, and how to head that off ahead of time.
Xiaoming has the AI's "15-day holiday" answer in hand and his suitcase packed when Xiaoting stops him: you'd act on an answer nobody checked?
The two of them act out how AI gets trained: first it reads huge amounts of text to learn how sentences continue, then it learns to follow instructions, and finally human feedback tunes it to be more useful and more honest. Along the way Xiaoming hauls in a giant stack of documents, the cafeteria menu included, and Manman gets buried under the paper. More material isn't always better. Find the few pages that matter first, then have it answer: that's the basic idea behind RAG. Even Xiaoting nearly grabs an outdated version.
You walk away with four quick checkups: ask about a chapter that doesn't exist and see if it makes one up; when you haven't given it any material, ask what its answer is based on; ask the same question five times and see how the answers shift; for exact math, have it write code and actually run it, then double-check with a calculator.
It confidently made up numbers, and they sounded pretty convincing.
Say thisSwap the [ ] part for yours
Look up [what you want to know] for me. After each fact, add the source link and its date. If you can't find something, just say so. Don't guess.
The key points got cut, and what's left is a pile of correct but useless sentences.
Say this
Summarize this in three sentences: first the conclusion, then the most important evidence, then the thing I should watch out for. Don't add anything the original doesn't say.
