3 ms·

Are you implying OpenAi is running most of it's API queries through humans? Like theranos did with it's tests? Because that's just ludicrous, the gpt architectu

Are you implying OpenAi is running most of it's API queries through humans? Like theranos did with it's tests? Because that's just ludicrous, the gpt architecture is well known and has had a few independent implementations. We know it's real, and even if this story was accurate, a few humans tuning the model is nothing unusual. But what you get from the API now is not generated or tweaked by humans. That only happens on the training data or when they are testing the model. (Edit: In this case they seem to be hard-coding some answers to prevent abusive/newsworthy outputs but again that is completely irrelevant to the performance of GPT itself. It's just a filter) The comparison to theranos makes no sense and it's becoming a lazy meme at this point.

josefx 5y ago
> the gpt architecture is well known and has had a few independent implementations. We know it's real Wasn't the claim that all the secret sauce was in the model? Also there are systems with a lab in a box style similar to Theranos they just didn't sell themselves as miracle machines capable of identifying everything from a drop of blood. > But what you get from the API now is not generated or tweaked by humans. From what I get from the other answers the API basically returns some random crap unless you fine tune some heat parameter. So it isn't surprising that it returns a different answer with every query. I don't know why this is done but I just hope nobody tries to use a system with random output for anything important without first putting it through a human filter.
- moconnor 5y ago
  No, you do not get random crap from the GPT-3 API and you don’t have to tune the heat parameter, defaults are generally fine. You can get a unique formulation each time; this is generally seen as a feature rather than a bug, humans are typically the same (different words, same meaning). The big difference is that you can also get a different meaning / opinion each time - coupled with the human-like language this is disconcerting and unexpected for many people. GPT-3 is more understandable if you think of it as sampling an opinion from the internet rather than a coherent entity with its own opinions.

josefx 5y ago

> the gpt architecture is well known and has had a few independent implementations. We know it's real Wasn't the claim that all the secret sauce was in the model? Also there are systems with a lab in a box style similar to Theranos they just didn't sell themselves as miracle machines capable of identifying everything from a drop of blood. > But what you get from the API now is not generated or tweaked by humans. From what I get from the other answers the API basically returns some random crap unless you fine tune some heat parameter. So it isn't surprising that it returns a different answer with every query. I don't know why this is done but I just hope nobody tries to use a system with random output for anything important without first putting it through a human filter.

moconnor 5y ago
No, you do not get random crap from the GPT-3 API and you don’t have to tune the heat parameter, defaults are generally fine. You can get a unique formulation each time; this is generally seen as a feature rather than a bug, humans are typically the same (different words, same meaning). The big difference is that you can also get a different meaning / opinion each time - coupled with the human-like language this is disconcerting and unexpected for many people. GPT-3 is more understandable if you think of it as sampling an opinion from the internet rather than a coherent entity with its own opinions.