4 ms·

The thing I most value chatGPT so far is that it is pretty good at answering fairly basic programming questions I have throughout the day. I've asked to to gene

The thing I most value chatGPT so far is that it is pretty good at answering fairly basic programming questions I have throughout the day. I've asked to to generate more advanced code / ideas with some pretty weird results.

naet 4y ago
Stack Overflow recently put a ban on chatGPT generated answers because they often sound correct but end up being wrong. https://meta.stackoverflow.com/questions/421831/temporary-policy-chatgpt-is-banned https://meta.stackoverflow.com/questions/421831/temporary-po...
- ericbarrett 4y ago
  How will they know? I don't think they'll be able to enforce this except the most obvious cases (where answers are posted by bots without any human review). Cat's out of the bag.
  - rashkov 4y ago
    It’d be neat if chatgpt encoded a signature via steganography of some sort, sort of like a watermark
    - DonHopkins 4y ago
      It should always include a duck. https://news.ycombinator.com/item?id=9137736 https://news.ycombinator.com/item?id=9137736 https://rachelbythebay.com/w/2013/06/05/duck/ https://rachelbythebay.com/w/2013/06/05/duck/
  - robocat 4y ago
    Stack Overflow can firstly warn users that using machine generated text will get you banned. SO can train a model to detect GPT3 output from different models, and then run it over all recently submitted questions and answers. https://huggingface.co/openai-detector https://huggingface.co/openai-detector is already fairly good at detecting ChatGPT generated text: so far my experience is that it says it is fake then it is fake, but it can be fooled into thinking generated text is “real”. Although perhaps a little unfair on those people using GPT to tidy up their spelling and grammar.
    - hxugufjfjf 4y ago
      A text written by a human but tidied up with GPT is not detected as GPT generated text. I've tested this extensively with large volumes of human written text.
- Winsaucerer 4y ago
  When asking ChatGPT yourself, I presume most of us would be aware that its answers could be wildly wrong. Given that we know this and it's in the front of our mind when asking ChatGPT, I think it has the potential to be very useful. Putting those answers on Stack Overflow is perhaps concerning for the reasons outlined in that link, in a way that it isn't when we deliberately ask ChatGPT ourselves.
- dragonwriter 4y ago
  Human generated answers on SO have the same problem; they frequently are persuasively, confidently written and wrong. Isn’t that what comments, voting, etc., exists to help sort out?
  - BeefWellington 4y ago
    If you read the post about it, the issue is the ease with which an AI can be abused to write up a few thousand confidently incorrect posts. The problem exists on both ends but there's a rough limit to the amount of people-hours dedicated to writing confidently incorrect answers on SO, whereas someone could task a machine to pump out answers 24/7 and inundate the site quickly.
    - visarga 4y ago
      > there's a rough limit to the amount of people-hours dedicated to writing confidently incorrect answers A personal limit, sure. But in aggregate they add up.
- theteapot 4y ago
  Isn't the main issue attribution? I mean you've got guys who put in hours a week over years to get rep and demonstrate their knowledge base then some script kiddy blows them out of the water. I'd be kind pissed. I think AI generated answers would be more acceptable if they were attributed to AI.
- pbhjpbhj 4y ago
  I wonder how soon well be able to make a fingerprint of a running computer system in a way that an answer can be automatically tested against a [sanitised] virtualized machine? Within a limited domain, like Ubuntu installs, this seems doable -- though maybe in a future of essentially unlimited bandwidth and storage we can just upload a virtualized version of our computers. Like, here's my computer, my Chat AI keeps swearing at me, how do I fix it. Then solutions posted on StackOverflew can be interpreted and tested by an AI against the virtualized copy of the machine.
  - xiphias2 4y ago
    I started using AI to write unit tests for me and it gives better results for writing unit tests than code, because it's simpler. I believe it will be getting better together with TDD.
onionisafruit 4y ago
I’ve been using it the same. So far it’s a net negative because using it distracts me and sucks me into trying to find its boundaries. But that will eventually change.
- sterlind 4y ago
  it's been useful for me as a way to talk through my ideas. like rubber ducking with a disturbingly-intelligent duck. what's really nice is I can describe an algorithm and it will generate and run through examples. and often works even if I only describe the gist.
  - visarga 4y ago
    It's great to generate ideas and helps finding if you skipped some obvious ones.

naet 4y ago

Stack Overflow recently put a ban on chatGPT generated answers because they often sound correct but end up being wrong. https://meta.stackoverflow.com/questions/421831/temporary-policy-chatgpt-is-banned https://meta.stackoverflow.com/questions/421831/temporary-po...

ericbarrett 4y ago
How will they know? I don't think they'll be able to enforce this except the most obvious cases (where answers are posted by bots without any human review). Cat's out of the bag.
- rashkov 4y ago
  It’d be neat if chatgpt encoded a signature via steganography of some sort, sort of like a watermark
  - DonHopkins 4y ago
    It should always include a duck. https://news.ycombinator.com/item?id=9137736 https://news.ycombinator.com/item?id=9137736 https://rachelbythebay.com/w/2013/06/05/duck/ https://rachelbythebay.com/w/2013/06/05/duck/
- robocat 4y ago
  Stack Overflow can firstly warn users that using machine generated text will get you banned. SO can train a model to detect GPT3 output from different models, and then run it over all recently submitted questions and answers. https://huggingface.co/openai-detector https://huggingface.co/openai-detector is already fairly good at detecting ChatGPT generated text: so far my experience is that it says it is fake then it is fake, but it can be fooled into thinking generated text is “real”. Although perhaps a little unfair on those people using GPT to tidy up their spelling and grammar.
  - hxugufjfjf 4y ago
    A text written by a human but tidied up with GPT is not detected as GPT generated text. I've tested this extensively with large volumes of human written text.
Winsaucerer 4y ago
When asking ChatGPT yourself, I presume most of us would be aware that its answers could be wildly wrong. Given that we know this and it's in the front of our mind when asking ChatGPT, I think it has the potential to be very useful. Putting those answers on Stack Overflow is perhaps concerning for the reasons outlined in that link, in a way that it isn't when we deliberately ask ChatGPT ourselves.
dragonwriter 4y ago
Human generated answers on SO have the same problem; they frequently are persuasively, confidently written and wrong. Isn’t that what comments, voting, etc., exists to help sort out?
- BeefWellington 4y ago
  If you read the post about it, the issue is the ease with which an AI can be abused to write up a few thousand confidently incorrect posts. The problem exists on both ends but there's a rough limit to the amount of people-hours dedicated to writing confidently incorrect answers on SO, whereas someone could task a machine to pump out answers 24/7 and inundate the site quickly.
  - visarga 4y ago
    > there's a rough limit to the amount of people-hours dedicated to writing confidently incorrect answers A personal limit, sure. But in aggregate they add up.
theteapot 4y ago
Isn't the main issue attribution? I mean you've got guys who put in hours a week over years to get rep and demonstrate their knowledge base then some script kiddy blows them out of the water. I'd be kind pissed. I think AI generated answers would be more acceptable if they were attributed to AI.
pbhjpbhj 4y ago
I wonder how soon well be able to make a fingerprint of a running computer system in a way that an answer can be automatically tested against a [sanitised] virtualized machine? Within a limited domain, like Ubuntu installs, this seems doable -- though maybe in a future of essentially unlimited bandwidth and storage we can just upload a virtualized version of our computers. Like, here's my computer, my Chat AI keeps swearing at me, how do I fix it. Then solutions posted on StackOverflew can be interpreted and tested by an AI against the virtualized copy of the machine.
- xiphias2 4y ago
  I started using AI to write unit tests for me and it gives better results for writing unit tests than code, because it's simpler. I believe it will be getting better together with TDD.

onionisafruit 4y ago

I’ve been using it the same. So far it’s a net negative because using it distracts me and sucks me into trying to find its boundaries. But that will eventually change.

sterlind 4y ago
it's been useful for me as a way to talk through my ideas. like rubber ducking with a disturbingly-intelligent duck. what's really nice is I can describe an algorithm and it will generate and run through examples. and often works even if I only describe the gist.
- visarga 4y ago
  It's great to generate ideas and helps finding if you skipped some obvious ones.