5 ms·

You basically picked the worse possible task for ChatGPT and are applying it as a broad conclusion. ChatGPT is _very_ good at taking human-language description

You basically picked the worse possible task for ChatGPT and are applying it as a broad conclusion. ChatGPT is _very_ good at taking human-language descriptions and transforming that into code. It works even better when instructions are concise and clear. I theoretically believe this stems from the fact that there’s a lot of human description of code for the LLM to train on. However, there’s relatively very little machine to machine code available. The tasks you’re looking to accomplish will likely have much better success if you break it into two “human-to-code” steps. 1. Ask ChatGPT to explain what the Python code is doing, line-by-line 2. Then ask ChatGPT to write the equivalent Lisp code for the description.

ssl-3 2y ago
Yep. It often "reasons", as much as it does, best with regular language. And sometimes, just like with real people, it works better if one or more intermediate steps are interposed. But it doesn't necessarily "know" this -- just as real people mightn't. It's not unfair to say "Yeah, well if the bot was so smart then why didn't it just take care of that intermediate step for me?" But it's also not unfair to retort "Well, if you were so smart then why did you expect perfect results from an imperfect tool without expending any real effort to understand and/or explain the thing on your part?"
- Jensson 2y ago
  > But it's also not unfair to retort "Well, if you were so smart... But it was the tool being evaluated, not the person. If the person has to code in natural language then why not just code it directly?
ptx 2y ago
Like XMLVM[0] with English instead of XML. [0] http://www.xmlvm.org/overview/ http://www.xmlvm.org/overview/
chrisjj 2y ago
> You basically picked the worse possible task for ChatGPT Wow, what were the odds on that?? :) I'd be interested in verification, so will start a new thread for that. > The tasks you’re looking to accomplish will likely have much better success if you break it into two “human-to-code” steps. > 1. Ask ChatGPT to explain what the Python code is doing, line-by-line The answer is "However, there's a logical flaw and a few issues in the implementation:" followed by points every one of which is in error, showing ChaptGPT has deeply mis"understood" the code. > 2. Then ask ChatGPT to write the equivalent Lisp code for the description The result fails with "odd number of &KEY arguments".

ssl-3 2y ago

Yep. It often "reasons", as much as it does, best with regular language. And sometimes, just like with real people, it works better if one or more intermediate steps are interposed. But it doesn't necessarily "know" this -- just as real people mightn't. It's not unfair to say "Yeah, well if the bot was so smart then why didn't it just take care of that intermediate step for me?" But it's also not unfair to retort "Well, if you were so smart then why did you expect perfect results from an imperfect tool without expending any real effort to understand and/or explain the thing on your part?"

Jensson 2y ago
> But it's also not unfair to retort "Well, if you were so smart... But it was the tool being evaluated, not the person. If the person has to code in natural language then why not just code it directly?

ptx 2y ago

Like XMLVM[0] with English instead of XML. [0] http://www.xmlvm.org/overview/ http://www.xmlvm.org/overview/

chrisjj 2y ago

> You basically picked the worse possible task for ChatGPT Wow, what were the odds on that?? :) I'd be interested in verification, so will start a new thread for that. > The tasks you’re looking to accomplish will likely have much better success if you break it into two “human-to-code” steps. > 1. Ask ChatGPT to explain what the Python code is doing, line-by-line The answer is "However, there's a logical flaw and a few issues in the implementation:" followed by points every one of which is in error, showing ChaptGPT has deeply mis"understood" the code. > 2. Then ask ChatGPT to write the equivalent Lisp code for the description The result fails with "odd number of &KEY arguments".