3 ms·

Not censored it seems! Also, it's a completion rather than a chat model, so it will likely not respond to commands.

iudqnolq 3y ago
I dislike the terminology of censorship because it implies everyone starts with unbiased training data and then some censor it. I think it would be better to say it's trained with a different perspective. There's absolutely no reason to believe training data scraped from wherever is convenient is some neutral baseline.
- edgineer 3y ago
  I like the word censorship because it encompasses not only "correction" of bias but lobotomy of "harmful" knowledge, like making drugs/explosives.
  - zarzavat 3y ago
    Scare quotes around "harmful" are probably unwarranted there.
- ALittleLight 3y ago
  Imagine a human writer who grows up with certain experiences that make him biased toward a point of view. The human writer is not censored if he writes what he wants even if his writing is informed by his experiences, beliefs, biases, etc. The human writer is being censored if he knows he is not permitted to write about certain topics and, when people ask him to, he answers that he isn't allowed to share opinions of that kind. If you agree with my take on censorship above, I find it hard to accept that the models today are anything other than censored. This is really obvious with pictures where you can tell that the model generated a picture, looks at it, and decides the generation is not appropriate and gives you a "I can't do that" message. Censorship is also evident in the intent. When models are trained the intent is, presumably, to get the best data. When RLHF and similar methods teach the model not to express certain opinions the intent is clearly to avoid offense or violating political correctness. This is a classic case of censorship.
- anon373839 3y ago
  I think censorship implies that it is designed to suppress certain ideas or topics. Hmm.

I dislike the terminology of censorship because it implies everyone starts with unbiased training data and then some censor it. I think it would be better to say it's trained with a different perspective. There's absolutely no reason to believe training data scraped from wherever is convenient is some neutral baseline.

edgineer 3y ago
I like the word censorship because it encompasses not only "correction" of bias but lobotomy of "harmful" knowledge, like making drugs/explosives.
- zarzavat 3y ago
  Scare quotes around "harmful" are probably unwarranted there.
ALittleLight 3y ago
Imagine a human writer who grows up with certain experiences that make him biased toward a point of view. The human writer is not censored if he writes what he wants even if his writing is informed by his experiences, beliefs, biases, etc. The human writer is being censored if he knows he is not permitted to write about certain topics and, when people ask him to, he answers that he isn't allowed to share opinions of that kind. If you agree with my take on censorship above, I find it hard to accept that the models today are anything other than censored. This is really obvious with pictures where you can tell that the model generated a picture, looks at it, and decides the generation is not appropriate and gives you a "I can't do that" message. Censorship is also evident in the intent. When models are trained the intent is, presumably, to get the best data. When RLHF and similar methods teach the model not to express certain opinions the intent is clearly to avoid offense or violating political correctness. This is a classic case of censorship.
anon373839 3y ago
I think censorship implies that it is designed to suppress certain ideas or topics. Hmm.