5 ms·

> Anyone can now put words in a row pretty much as well as I can. This suffers the same fallacy that people who think ChatGPT is somehow going to upend coding

by kodah 4y ago

> Anyone can now put words in a row pretty much as well as I can. This suffers the same fallacy that people who think ChatGPT is somehow going to upend coding suffer from. Can GPT-4 write code and sentences? Yes, it can, but it makes a lot of subtle mistakes. Good programmers are often good at: - decomposition - thoroughness - identifying important components My hypothesis is that AI will make those things summarily more obvious if a programmer doesn't have them. A programmer that's not thorough will not have the inclination or drive to review and understand all output, so the output will harm their work whether that's writing or code. AI can't decompose anything sufficiently complex as a function. It can kind of get close, but again, it misses the objective of decomposition because decomposition isn't a language function, it's a higher order abstract function. Last, AI isn't going to recognize importance or priority, especially if it's on-going. Try giving it a rules system and see if it follows the rules to a T. IME, it does not. AI will make good programmers better, and some okay programmers better, but I have a feeling it will also make bad programmers worse.

seba_dos1 4y ago
I agree. Trying to get some non-trivial code out of GPT that isn't a mindless boilerplate feels like helping a clueless schoolmate in an IT class write their code for the teacher's assignment when, sure, they do try their best, but it's clear that they don't really grasp it and just want to finish the assignment and forget about it as soon as possible. You can get a good result, but you need to try really hard. It only seems useful for when you know exactly what to do, but you don't feel like typing it all out - which happens, but in my experience the most valuable part of programmer's job is when you figure things out that nobody around knows how to do.
- peterangular 4y ago
  > Trying to get some non-trivial code out of GPT that isn't a mindless boilerplate For some of us, the boilerplate code generated by ChatGPT is incredibly valuable. I have been utilizing it to create snippets of code for languages that I am not familiar with in terms of syntax. I understand that I may be criticized for this, however, I am not an expert in all of the languages I am required to support. Thus, I often find myself struggling with the syntax while understanding the overall concept I am attempting to accomplish. Example prompt: "Write me a proof of concept in language X that serializes JSON then POSTS to a HTTP API endpoint."
  - seba_dos1 4y ago
    The thing is - how often do you end up needing such snippets? How much more useful is receiving it from ChatGPT than looking it up in a search engine? Continuing with the theme of your example, at least in my experience most of the programmer's time is spent on figuring out which endpoint do you need, why and when does that endpoint sometimes return something you don't expect and how to deal with that happening. Writing JSON serializing in a language your head got rusty with may be faster with ChatGPT than it would be otherwise, but the overall benefit seems negligible, especially when you need to watch out at every step for convincingly-looking nonsense.
    - peterangular 4y ago
      All in good faith, I believe that receiving assistance from ChatGPT is far more useful than looking up snippets on a search engine. Instead of delving through semi-appropriate snippets, I can concisely describe what I need in one or two sentences and ChatGPT typically provides an accurate result, unless the query is opinion-based, complex, or obscure. I am not expecting perfect code since I take it upon myself to polish the output. I am now experimenting with ChatGPT parsing our own historical documentation, such as API references, and even basic markdown notes. When it comes to API endpoints I can get 80-90% accurate boilerplate. Admittedly, ChatGPT cannot provide assistance in understanding unexpected behavior, unless perhaps it is due to a misunderstanding of the language. And, though it is true that resolving API endpoint issues can be a time-consuming task, this does not negate the time savings that ChatGPT can provide in other situations. I would urge you to try it out in your own workflow before calling shots on if the usefulness is negligible or not. And even then, I'd say not every tool is for everyone. This thing is really helping me out and I'm seeing long-term viability for it in my workflows.
      - seba_dos1 4y ago
        > I would urge you to try it out in your own workflow before calling shots on if the usefulness is negligible or not. What makes you think I didn't? I already stated that: > It only seems useful for when you know exactly what to do, but you don't feel like typing it all out - which happens It does pretty well in these kind of situations when the code in question is trivial, but in my experience the time savings coming from that are minor. In fact, its main value comes from my ADHD mind not being able to focus on boring stuff I already know how to do - in those cases, GPT gets me the boilerplate I can start operating on right away, which sure is helpful as otherwise I'd just end up in a web browser scrolling some stuff. Other than that? It seems like verifying ChatGPT's output is much more time consuming than doing it on your own as soon as you get out of "problems copy'n'pastable from StackOverflow" territory.
andrewchambers 4y ago
Are you talking about future AI? or just AI as it stands today. It is a moving target.

by kodah 4y ago

seba_dos1 4y ago

I agree. Trying to get some non-trivial code out of GPT that isn't a mindless boilerplate feels like helping a clueless schoolmate in an IT class write their code for the teacher's assignment when, sure, they do try their best, but it's clear that they don't really grasp it and just want to finish the assignment and forget about it as soon as possible. You can get a good result, but you need to try really hard. It only seems useful for when you know exactly what to do, but you don't feel like typing it all out - which happens, but in my experience the most valuable part of programmer's job is when you figure things out that nobody around knows how to do.

peterangular 4y ago
> Trying to get some non-trivial code out of GPT that isn't a mindless boilerplate For some of us, the boilerplate code generated by ChatGPT is incredibly valuable. I have been utilizing it to create snippets of code for languages that I am not familiar with in terms of syntax. I understand that I may be criticized for this, however, I am not an expert in all of the languages I am required to support. Thus, I often find myself struggling with the syntax while understanding the overall concept I am attempting to accomplish. Example prompt: "Write me a proof of concept in language X that serializes JSON then POSTS to a HTTP API endpoint."
- seba_dos1 4y ago
  The thing is - how often do you end up needing such snippets? How much more useful is receiving it from ChatGPT than looking it up in a search engine? Continuing with the theme of your example, at least in my experience most of the programmer's time is spent on figuring out which endpoint do you need, why and when does that endpoint sometimes return something you don't expect and how to deal with that happening. Writing JSON serializing in a language your head got rusty with may be faster with ChatGPT than it would be otherwise, but the overall benefit seems negligible, especially when you need to watch out at every step for convincingly-looking nonsense.
  - peterangular 4y ago
    All in good faith, I believe that receiving assistance from ChatGPT is far more useful than looking up snippets on a search engine. Instead of delving through semi-appropriate snippets, I can concisely describe what I need in one or two sentences and ChatGPT typically provides an accurate result, unless the query is opinion-based, complex, or obscure. I am not expecting perfect code since I take it upon myself to polish the output. I am now experimenting with ChatGPT parsing our own historical documentation, such as API references, and even basic markdown notes. When it comes to API endpoints I can get 80-90% accurate boilerplate. Admittedly, ChatGPT cannot provide assistance in understanding unexpected behavior, unless perhaps it is due to a misunderstanding of the language. And, though it is true that resolving API endpoint issues can be a time-consuming task, this does not negate the time savings that ChatGPT can provide in other situations. I would urge you to try it out in your own workflow before calling shots on if the usefulness is negligible or not. And even then, I'd say not every tool is for everyone. This thing is really helping me out and I'm seeing long-term viability for it in my workflows.
    - seba_dos1 4y ago
      > I would urge you to try it out in your own workflow before calling shots on if the usefulness is negligible or not. What makes you think I didn't? I already stated that: > It only seems useful for when you know exactly what to do, but you don't feel like typing it all out - which happens It does pretty well in these kind of situations when the code in question is trivial, but in my experience the time savings coming from that are minor. In fact, its main value comes from my ADHD mind not being able to focus on boring stuff I already know how to do - in those cases, GPT gets me the boilerplate I can start operating on right away, which sure is helpful as otherwise I'd just end up in a web browser scrolling some stuff. Other than that? It seems like verifying ChatGPT's output is much more time consuming than doing it on your own as soon as you get out of "problems copy'n'pastable from StackOverflow" territory.

andrewchambers 4y ago

Are you talking about future AI? or just AI as it stands today. It is a moving target.