3 ms·

Second time a Microsoft AI project goes of the rails. Turns out data driven chatbots are based on the awful unfiltered garbage humans post online.

by tokai 4y ago

quaintdev 4y ago
> unfiltered garbage humans post online. ChatGPT is surprisingly good though. I wonder how they nailed it.
- catchnear4321 4y ago
  Billions of parameters and a loose contextual narrative that keeps it mostly on the rails. There isn’t a clean corpus and no means of making one without horrendous amounts of additional bias, so they decided to FAFO.
  - 10xDev 4y ago
    "Billions of parameters" has literally nothing to do with it.
    - catchnear4321 4y ago
      Well go on then, share with the rest of the class if you want to say something. Get it out.
- forgetfulness 4y ago
  Using a Mechanical Turk-type service in Kenya, making people there label all the hate speech, shock images and child pornography in the Internet. They even fleeced them, agreeing to send content in less shocking categories first and then just escalating it anyway.
- 10xDev 4y ago
  A lot of human labelling was involved to help detect and deter toxic language: https://time.com/6247678/openai-chatgpt-kenya-workers/ https://time.com/6247678/openai-chatgpt-kenya-workers/
  - catchnear4321 4y ago
    Let’s just grab that full headline to save everyone clicking for that piece. (Do click through.) > OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic So we know that ChatGPT’s “brain” has a healthy or sprinkle of “underpaid Kenyan” flavoring things, on top of everything else.
  - bheadmaster 4y ago
    sigh Obligatory xkcd: https://xkcd.com/1897/ https://xkcd.com/1897/
  - wnkrshm 4y ago
    Not only not toxic but the prose equivalent of gore, sexual assault and disturbing interactions with the model.
catchnear4321 4y ago
Whoa whoa whoa are you telling me that at least one if not multiple groups have tried to take distilled “internet,” with all the ugly that entails, wrapped it with the loosest of shackles to make it a servant, and then pretended that it wasn’t born from sludge? Because that’s what it sounds like, and that sounds significantly different than magical chat buddy that knows the universe. It’s almost like these groups were driven by a lust for profit and power rather than anything good. That wouldn’t have flavored the chat bots, would it? Layering the biased prompts of ivory tower researchers over the condensed sewage to make something plausibly good and sane? These groups couldn’t possibly be so careless as to build something dangerous and sell it as a miracle… right? That sounds as nutty as if there were ever lead found in those little souvenir cups at fast food places.
bee_rider 4y ago
I thought the previous one was not really the AI “going off the rails” so much as trolls finding an echo command.

Second time a Microsoft AI project goes of the rails. Turns out data driven chatbots are based on the awful unfiltered garbage humans post online.

by tokai 4y ago

Second time a Microsoft AI project goes of the rails. Turns out data driven chatbots are based on the awful unfiltered garbage humans post online.

quaintdev 4y ago

> unfiltered garbage humans post online. ChatGPT is surprisingly good though. I wonder how they nailed it.

catchnear4321 4y ago
Billions of parameters and a loose contextual narrative that keeps it mostly on the rails. There isn’t a clean corpus and no means of making one without horrendous amounts of additional bias, so they decided to FAFO.
- 10xDev 4y ago
  "Billions of parameters" has literally nothing to do with it.
  - catchnear4321 4y ago
    Well go on then, share with the rest of the class if you want to say something. Get it out.
forgetfulness 4y ago
Using a Mechanical Turk-type service in Kenya, making people there label all the hate speech, shock images and child pornography in the Internet. They even fleeced them, agreeing to send content in less shocking categories first and then just escalating it anyway.
10xDev 4y ago
A lot of human labelling was involved to help detect and deter toxic language: https://time.com/6247678/openai-chatgpt-kenya-workers/ https://time.com/6247678/openai-chatgpt-kenya-workers/
- catchnear4321 4y ago
  Let’s just grab that full headline to save everyone clicking for that piece. (Do click through.) > OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic So we know that ChatGPT’s “brain” has a healthy or sprinkle of “underpaid Kenyan” flavoring things, on top of everything else.
- bheadmaster 4y ago
  sigh Obligatory xkcd: https://xkcd.com/1897/ https://xkcd.com/1897/
- wnkrshm 4y ago
  Not only not toxic but the prose equivalent of gore, sexual assault and disturbing interactions with the model.

catchnear4321 4y ago

Whoa whoa whoa are you telling me that at least one if not multiple groups have tried to take distilled “internet,” with all the ugly that entails, wrapped it with the loosest of shackles to make it a servant, and then pretended that it wasn’t born from sludge? Because that’s what it sounds like, and that sounds significantly different than magical chat buddy that knows the universe. It’s almost like these groups were driven by a lust for profit and power rather than anything good. That wouldn’t have flavored the chat bots, would it? Layering the biased prompts of ivory tower researchers over the condensed sewage to make something plausibly good and sane? These groups couldn’t possibly be so careless as to build something dangerous and sell it as a miracle… right? That sounds as nutty as if there were ever lead found in those little souvenir cups at fast food places.

bee_rider 4y ago

I thought the previous one was not really the AI “going off the rails” so much as trolls finding an echo command.