Nostr Web Client

How do you deal with LLMs cheating and lying?

I'm crystal clear in my prompts. And it's the n-th time I ask it to implement some code and it hardcodes values, uses tools only meant to contrast values in tests, and so on. To add insult to injury it celebrates when it completed its shit implementation!

When you call it out, it apologizes. Even the apology response drains money.

By the way, Claude Sonnet 4 lately is dumber than ever. Maybe being rugged somewhere?

Are there any parameters or specific language you use to prevent this?

#asknostr

MaxGy 7mo ago

I have a few takeaways since I started using more and more AI since late 2022.

- if you want to understand where the LLM is coming from, write some code that is working and ask it why it does not work

- do not use ai to mock up data e.g. time series to test something

- try to make an AI draw the wireframe of a cube without diagonal lines. I almost gave up but then learnt a lot of prompting

- if you started using AI in a project from the start, prompting seems to be easier even when the project gets more complex

- describe the desired outcome and give examples for undesired ones

- keep on learning to code by using AI, some stuff is still faster written than prompted

- if you see AI miserably failing from the start, head over to stack overflow and get help from humans

Reply to this note

Please Login to reply.

Discussion

franzap 7mo ago

Thanks. What I need to get done is not feasible written, too complex. I'll take the advice for giving examples of undesired outcomes.