Post AS5zGT0VB1zxAwPVB2 by bcrypt@infosec.exchange
 (DIR) More posts by bcrypt@infosec.exchange
 (DIR) Post #AS5zGQ4I6UFM3p9zAO by simon@fedi.simonwillison.net
       2023-01-28T06:23:38Z
       
       0 likes, 1 repeats
       
       ChatGPT is really useful to help design CLI tools. I wanted to add a new option that meant "fail with a non-zero exit code if the tool sees an HTTP 4xx or 5xx error", so I had a quick conversation with ChatGPT to come up with "--fail" as a good name for that option, consistent with other tools.
       
 (DIR) Post #AS5zGSG3xgqqqugPnE by simon@fedi.simonwillison.net
       2023-01-28T06:26:02Z
       
       0 likes, 0 repeats
       
       This is one of those use-cases where the stakes are so low that the risk of hallucination really isn't a problem at all
       
 (DIR) Post #AS5zGT0VB1zxAwPVB2 by bcrypt@infosec.exchange
       2023-01-28T07:58:35Z
       
       0 likes, 0 repeats
       
       @simon gotta be honest, i read this out of context and thought it was about taking psychedelics at work
       
 (DIR) Post #ASEHuU4ZzDicVIEDUu by pixelistik@chaos.social
       2023-01-28T06:35:51Z
       
       0 likes, 0 repeats
       
       @simon yes, but this also illustrates a weakness of ChatGPT: the wget example is just plain wrong. --no-clobber means something else.
       
 (DIR) Post #ASEHuUoJFCIYn7cjmC by tomnardi@hackaday.social
       2023-01-28T07:19:55Z
       
       0 likes, 0 repeats
       
       @pixelistik @simon Good catch. httpie/http doesn't appear to have a --fail either, it just made that up as well.So in reality, the output of ChatGPT in this case was useless other than to return the information which easily could have been pulled faster/easier from looking at the curl man page.Perhaps more to the point, I don't think an AI was necessary to figure out a logical command switch for "fail" was...believe it or not "fail".
       
 (DIR) Post #ASEHuVQawqvIhrXJS4 by simon@fedi.simonwillison.net
       2023-01-28T07:40:23Z
       
       0 likes, 0 repeats
       
       @tomnardi I disagree: the hallucinations didn't matter for this, and it was much faster for me to have this conversation with ChatGPT than it was to rack my brain for commands they might have a similar option - I can't look at the curl man page if I've forgotten that curl has this kind of option!
       
 (DIR) Post #ASEHuW1oiShIZIx2TA by tomnardi@hackaday.social
       2023-01-28T07:54:10Z
       
       0 likes, 0 repeats
       
       @simonAlright, but that's literally the opposite of what you said in the first post.You praised ChatGPT as a way to find option names that were "consistent with other tools", but in fact 3/4 of the examples it provided were false. It made you THINK you were following best practices, but it was an illusion.That's exactly the danger of relying on these AI tools, they convince the user they have some special insight, but sometimes it's just random.
       
 (DIR) Post #ASEHuWYmjt4KDYNMrA by simon@fedi.simonwillison.net
       2023-01-28T08:07:29Z
       
       0 likes, 0 repeats
       
       @tomnardi that's fair - it lies to me in a way that made me think that the option name was more common than it actually wasI'm still happy with the design decision I made with its assistance though
       
 (DIR) Post #ASEHuX2YxAt7huJ9Gq by bitbear@icosahedron.website
       2023-02-01T08:06:03Z
       
       0 likes, 0 repeats
       
       @simon @tomnardi I feel like the lying is a problem. It distorts the perception of reality, requiring you to do the work ChatGPT was used to alleviate in the first place, only this time with a heavy bias towards whatever bullshit ChatGPT conjured.“Mansplaining as a service” at work (hat tip to @cwebber).