4 ms·

How do you know what part of the answer is wrong?

dizhn 4y ago
I asked chatgpt to give me the equivalent "ss" command to "netstat -nap". It did and it was correct because I already knew the answer and tried it anyway. Then I asked the equivalent of "netstat -rn" (if any), it gave me "ss -r" something, which is wrong. So you either know enough to spot the error or you verify all of it. Another example. I had a long session with it where I was trying to get it to write a gitlab ci config file where the artifacts to be deployed to Netlify were already present at the root of the repo. As an addition it had a requirement to use 'sed' to rewrite a domain name recursively. This is the first and only time chatgpt was useful to me although it made a lot of mistakes that needed correcting.It was more useful than a google search because I wasn't able to find the correct format with it as most examples assumed a build step. One thing that was frustrating is how I would give it requirements and it would incorporate them only to lose a prior requirement when fixing a latter one. I ended up hand fixing one of the later configs rather than using the final one it gave me. For example I needed to use a node image because netlify-cli was required (the only custom thing necessary from the image) but it would revert to using an Alpine image and point to a now archived repo with a nonexistent release url for gitlab's cli. When it fixed something else it would break the find .. sed combination and add nonsensical things like *. html to it rather than working on all files recursively. In the end, through a combination of AI helping and the user (me) chiming in and trying to reason about it I was able to get somewhere. Not ideal but not useless.
klibertp 4y ago
Just like with a junior coder who, when asked about something they don't know, says whatever in hopes it's close enough to the real answer. If the conversation is over Slack and you don't have out-of-band information, like what facial expression they have, it's basically 100% equivalent to ChatGPT. Only, ChatGPT is actually better at lying, because there's no noticeable difference between the style of real vs. wrong responses, so you need to be a bit more careful. In other words, follow-up questions and cross-referencing are a must. You can't take any of ChatGPT answers as true without verification. You also need to be really careful when asking questions, because if you give it too many assumptions, it's absolutely great at confirming your biases for you. I find it to be a great "rubber duck" that can help you see and understand the relationships between facts that you might have missed otherwise. OTOH, it'll be many, many years before I'm comfortable asking it about things I'm not already mostly familiar with or can verify easily.

I asked chatgpt to give me the equivalent "ss" command to "netstat -nap". It did and it was correct because I already knew the answer and tried it anyway. Then I asked the equivalent of "netstat -rn" (if any), it gave me "ss -r" something, which is wrong. So you either know enough to spot the error or you verify all of it. Another example. I had a long session with it where I was trying to get it to write a gitlab ci config file where the artifacts to be deployed to Netlify were already present at the root of the repo. As an addition it had a requirement to use 'sed' to rewrite a domain name recursively. This is the first and only time chatgpt was useful to me although it made a lot of mistakes that needed correcting.It was more useful than a google search because I wasn't able to find the correct format with it as most examples assumed a build step. One thing that was frustrating is how I would give it requirements and it would incorporate them only to lose a prior requirement when fixing a latter one. I ended up hand fixing one of the later configs rather than using the final one it gave me. For example I needed to use a node image because netlify-cli was required (the only custom thing necessary from the image) but it would revert to using an Alpine image and point to a now archived repo with a nonexistent release url for gitlab's cli. When it fixed something else it would break the find .. sed combination and add nonsensical things like *. html to it rather than working on all files recursively. In the end, through a combination of AI helping and the user (me) chiming in and trying to reason about it I was able to get somewhere. Not ideal but not useless.

klibertp 4y ago

Just like with a junior coder who, when asked about something they don't know, says whatever in hopes it's close enough to the real answer. If the conversation is over Slack and you don't have out-of-band information, like what facial expression they have, it's basically 100% equivalent to ChatGPT. Only, ChatGPT is actually better at lying, because there's no noticeable difference between the style of real vs. wrong responses, so you need to be a bit more careful. In other words, follow-up questions and cross-referencing are a must. You can't take any of ChatGPT answers as true without verification. You also need to be really careful when asking questions, because if you give it too many assumptions, it's absolutely great at confirming your biases for you. I find it to be a great "rubber duck" that can help you see and understand the relationships between facts that you might have missed otherwise. OTOH, it'll be many, many years before I'm comfortable asking it about things I'm not already mostly familiar with or can verify easily.