Hacker News Reader

Comment by VectorLock | Hacker News Reader

This is the biggest problem I encounter when trying to use ChatGPT on a daily basis for computer programming tasks. It "hallucinates" plausible looking code that never existed or would never work, especially confusing whats in one module or API for something in another. This is where ChatGPT breaks when pushed a bit further than "make customized StackOverflow snippets."

For example I asked ChatGPT to show me how to use an AWS SDK "waiter" to wait on a notification on an SNS topic. It showed me code that looked right, but was confusing functions in the SQS library for those that would do the thing with SNS (but SNS doesn't support what I wanted)

View on HN

It wrote me a python snippet while my question was about a go library. When prompted it's a go library it wrote similar looking code in go with the same function names that don't actually exist in the library. It's like google search past 2010. It's trying to please everybody too much rather than saying I can't do that. Though when asked to write a new original Koran verse, it does refuse to do that. :)

I guess the issue at core is that it doesn't, and can't know if it can or can't do it. That's not what it's designed to do, even if it does quite well at seeming so.

Have you tried using the code-davinci-002 model instead of ChatGPT?

For example - https://platform.openai.com/playground/p/default-translate-c...

The codex models are intended for doing work with code rather than language and may give better results in that context. https://help.openai.com/en/articles/6195637-getting-started-...

It does indeed sound problematic to use ChatGPT daily for computer programming tasks. ChatGPT is not a snippets manager but text completion.

It may be more helpful to look for better answers on Amazon's help pages for SNS and AWS SDK.

I know the answer. SNS can't do that. But ChatGPT hallucinated it could. Just like the original post about a capability their API doesn't provide.

The problem is compounded by the fact that sometimes it produces really good results. One task, good results. Next task, totally hallucinated result.

That’s what my boss said about me on my last performance review.

Yeah, it quickly breaks down with fine minutiae like the precise API signatures for a random library. It doesn't help that API changes are inevitable while the model retains a memory of all the now outdated documentation from its training.

Reply on news.ycombinator.com