5 ms·

It wouldn't surprise me at all if ChatGPT was trained on data originating from Stack Overflow. Not familiar with deep learning algorithms but I can't imagine th

It wouldn't surprise me at all if ChatGPT was trained on data originating from Stack Overflow. Not familiar with deep learning algorithms but I can't imagine that it would be a good idea to have an unintended training data loop.

msla 4y ago
An intended double-loop might be interesting: Train one GPT on the output of another.
- fragmede 4y ago
  OpenAI isn't giving out this service from the goodness of their hearts just for the fun of it. I bet they've gotten it to a place where they can feed its output back into itself for self improvement.
  - maxbond 4y ago
    I dunno, is that really possible/sustainable? I suppose we add information to the system by poking it with prompts, and to some extent that gets encoded in the outputs. But I feel like if you're feeding it things that don't contain a new, fresh signal, you're ultimately going to fit to noise; the more of it's own input it ingests, the less meaningful the network would be, and presumably the more nonsensical the output. But I don't actually know how this stuff works, I'm genuinely asking. I agree OpenAI have something going on here, I'd assumed it was that they wanted to know the strategies we came up with for bypassing their filters, and to get into a cat and mouse game with us so that they could improve them to the point of being productized. Also, marketing, both sales, recruitment, and mindshare. ETA: I think I underestimated before the extent to which the output is often just your own words reflected back to you, in which case, it may well be adding more signal then noise. https://news.ycombinator.com/item?id=33863413 https://news.ycombinator.com/item?id=33863413 Still, some prompts return a walk of text from a single sentence, and I have to imagine those are worse then useless for training.
    - phyphy 4y ago
      If there's a discriminator in ChatGPT then I guess it would probably know when an input data is AI generated.

msla 4y ago

An intended double-loop might be interesting: Train one GPT on the output of another.

fragmede 4y ago
OpenAI isn't giving out this service from the goodness of their hearts just for the fun of it. I bet they've gotten it to a place where they can feed its output back into itself for self improvement.
- maxbond 4y ago
  I dunno, is that really possible/sustainable? I suppose we add information to the system by poking it with prompts, and to some extent that gets encoded in the outputs. But I feel like if you're feeding it things that don't contain a new, fresh signal, you're ultimately going to fit to noise; the more of it's own input it ingests, the less meaningful the network would be, and presumably the more nonsensical the output. But I don't actually know how this stuff works, I'm genuinely asking. I agree OpenAI have something going on here, I'd assumed it was that they wanted to know the strategies we came up with for bypassing their filters, and to get into a cat and mouse game with us so that they could improve them to the point of being productized. Also, marketing, both sales, recruitment, and mindshare. ETA: I think I underestimated before the extent to which the output is often just your own words reflected back to you, in which case, it may well be adding more signal then noise. https://news.ycombinator.com/item?id=33863413 https://news.ycombinator.com/item?id=33863413 Still, some prompts return a walk of text from a single sentence, and I have to imagine those are worse then useless for training.
  - phyphy 4y ago
    If there's a discriminator in ChatGPT then I guess it would probably know when an input data is AI generated.