4 ms·

I'm curious about what topics you have knowledge of where it's failed. Does it seem like there's a pattern to the failures? I've been using GPT for coding hel

by lcuff 3y ago

I'm curious about what topics you have knowledge of where it's failed. Does it seem like there's a pattern to the failures? I've been using GPT for coding help, and it it is very helpful in ruby and bash, though it often delivers buggy software: Badly handled non-happy-path conditions, mostly, which when I tell it to handle the case, it may. It's a huge help for me finding gems and showing me standard ruby library syntax. On the other hand, it's been useless when I try to get it to write Applescript for me. I believe that says more about AppleScript than about GPT. Sigh.

notamy 3y ago
> Does it seem like there's a pattern to the failures? Personally, trying to use it to write code (primarily Elixir backend and Rust systems/CLI), I tend to run into: - Hallucinating APIs that don't exist - Hallucinating entire libraries, despite being told repeatedly they don't exist - Saying it will make requested changes and not doing so - Not being anywhere close to idiomatic code - Not being able to explain code it writes - Running out of "memory" (I can't remember the right term. Context?) in the middle of generating code, then telling me I never prompted it when I ask it to continue On the other hand, I've found that it's good at cleaning up ugly data. I can copy/paste in a table with bad formatting, ask it to turn it into code, and it does it near-perfectly. That's been the best use-case for it I've found so far. I use boring normal free ChatGPT so maybe it's on me for not using GPT-4 or some other model, but either way, imo it's not been very impressive in the problem spaces I find myself in.
- mrtranscendence 3y ago
  > Not being able to explain code it writes I've actually found it (ChatGPT-4) reasonably good at explaining other people's code, for what it's worth. It does require providing some context for the code it's explaining. ("This function is part of library X that does Y, and it seems to be about Z. I can't figure out the purpose of the loop in the middle, though. Can you explain it?") I'm with you on it writing code, though. It makes things up too frequently to be really useful.
- simonw 3y ago
  Yeah, GPT-4 is so, so much better at not doing the things you're describing there. ChatGPT is also much better for Python and JavaScript than Rust and Elixir, presumably because it's seen an order of magnitude more example code for those languages.
- gwd 3y ago
  > I use boring normal free ChatGPT so maybe it's on me for not using GPT-4 or some other model, but either way, imo it's not been very impressive in the problem spaces I find myself in. You can use GPT-4 fairly cheap if you sign up for a developer account on platform.openai.com and then use their playground. There you pay per usage; even though I've been using it fairly heavily, my typical monthly usage is still way under $20. Here's an example that both impressed me and saved me a load of time just yesterday: https://gitlab.com/-/snippets/2535443 https://gitlab.com/-/snippets/2535443
devjab 3y ago
I can give you a few examples. Some of the “issues” are more general, as in it giving you dated answers. I asked it so build some ODATA things in Typescript and C#, and it did to varying degrees of success, but some of the code was deprecated. Some ranging to “you should never, ever, do things this way” to “well IActionResult was replaced by ActionResult but it hardly matters”. Others was where it made things up once pressed upon being wrong. I think the two most hilarious situations was when it did its “Sorry, you’re right…” thing and then proceed to give the exact same answer it had just given prior. The other was when it made up a function that had never existed, it has a very convincing name and at first I thought it was again a matter of something deprecated, but it turned out to be completely made up code. Some of this is down to me not prompting it right, but part of it is also worrying. Because what really made me worry about it was when I joked with it. I read (listen to) a lot of Warhammer audiobooks from Black Library, and, when I jokingly asked it something silly about Khorne at one point, it led me down a rabbit hole of discussing books with it. Books it obviously had never read, but would still confidently tell you about wrongly. Maybe it got its knowledge from the internet, maybe it made it up, but what was interesting to me about it was that if it would be so confidently incorrect about those books, then what else would it be confidently incorrect about. This isn’t just a GPT problem of course. If you want to learn something, you need to consider the sources you use. When I went to folkeskolen (school for children aged 6-14) we were taught the pyramids were build by slaves. Later that has been disputed because of how well fed the workers were, but if you wanted to learn about the pyramids and you read my old school books then you wouldn’t get the updated knowledge. Similarly a lot of the sources you can find for learning programming are outright terrible, but outside of GPT you tend to be presented with a myriad of choice to remind you that some sources are better than others. With GPT the source or what it teaches you isn’t obvious. If you wanted to learn about the pyramids, you probably wouldn’t pick a 30 year old school book for children after all.

by lcuff 3y ago

notamy 3y ago

> Does it seem like there's a pattern to the failures? Personally, trying to use it to write code (primarily Elixir backend and Rust systems/CLI), I tend to run into: - Hallucinating APIs that don't exist - Hallucinating entire libraries, despite being told repeatedly they don't exist - Saying it will make requested changes and not doing so - Not being anywhere close to idiomatic code - Not being able to explain code it writes - Running out of "memory" (I can't remember the right term. Context?) in the middle of generating code, then telling me I never prompted it when I ask it to continue On the other hand, I've found that it's good at cleaning up ugly data. I can copy/paste in a table with bad formatting, ask it to turn it into code, and it does it near-perfectly. That's been the best use-case for it I've found so far. I use boring normal free ChatGPT so maybe it's on me for not using GPT-4 or some other model, but either way, imo it's not been very impressive in the problem spaces I find myself in.

mrtranscendence 3y ago
> Not being able to explain code it writes I've actually found it (ChatGPT-4) reasonably good at explaining other people's code, for what it's worth. It does require providing some context for the code it's explaining. ("This function is part of library X that does Y, and it seems to be about Z. I can't figure out the purpose of the loop in the middle, though. Can you explain it?") I'm with you on it writing code, though. It makes things up too frequently to be really useful.
simonw 3y ago
Yeah, GPT-4 is so, so much better at not doing the things you're describing there. ChatGPT is also much better for Python and JavaScript than Rust and Elixir, presumably because it's seen an order of magnitude more example code for those languages.
gwd 3y ago
> I use boring normal free ChatGPT so maybe it's on me for not using GPT-4 or some other model, but either way, imo it's not been very impressive in the problem spaces I find myself in. You can use GPT-4 fairly cheap if you sign up for a developer account on platform.openai.com and then use their playground. There you pay per usage; even though I've been using it fairly heavily, my typical monthly usage is still way under $20. Here's an example that both impressed me and saved me a load of time just yesterday: https://gitlab.com/-/snippets/2535443 https://gitlab.com/-/snippets/2535443

devjab 3y ago

I can give you a few examples. Some of the “issues” are more general, as in it giving you dated answers. I asked it so build some ODATA things in Typescript and C#, and it did to varying degrees of success, but some of the code was deprecated. Some ranging to “you should never, ever, do things this way” to “well IActionResult was replaced by ActionResult but it hardly matters”. Others was where it made things up once pressed upon being wrong. I think the two most hilarious situations was when it did its “Sorry, you’re right…” thing and then proceed to give the exact same answer it had just given prior. The other was when it made up a function that had never existed, it has a very convincing name and at first I thought it was again a matter of something deprecated, but it turned out to be completely made up code. Some of this is down to me not prompting it right, but part of it is also worrying. Because what really made me worry about it was when I joked with it. I read (listen to) a lot of Warhammer audiobooks from Black Library, and, when I jokingly asked it something silly about Khorne at one point, it led me down a rabbit hole of discussing books with it. Books it obviously had never read, but would still confidently tell you about wrongly. Maybe it got its knowledge from the internet, maybe it made it up, but what was interesting to me about it was that if it would be so confidently incorrect about those books, then what else would it be confidently incorrect about. This isn’t just a GPT problem of course. If you want to learn something, you need to consider the sources you use. When I went to folkeskolen (school for children aged 6-14) we were taught the pyramids were build by slaves. Later that has been disputed because of how well fed the workers were, but if you wanted to learn about the pyramids and you read my old school books then you wouldn’t get the updated knowledge. Similarly a lot of the sources you can find for learning programming are outright terrible, but outside of GPT you tend to be presented with a myriad of choice to remind you that some sources are better than others. With GPT the source or what it teaches you isn’t obvious. If you wanted to learn about the pyramids, you probably wouldn’t pick a 30 year old school book for children after all.