8 ms·

we need less censorious AIs not more ... the claim that it's somehow 'ethical' to have a guy baking in his opinions about things in a tool used globally is abs

we need less censorious AIs not more ... the claim that it's somehow 'ethical' to have a guy baking in his opinions about things in a tool used globally is absurd to anyone who ever read anything about ethics

Al-Khwarizmi 4y ago
My hope is on a non-American alternative. The American society seems too engulfed by puritanism to produce a less straightjacketed chatbot.
- unnouinceput 4y ago
  Watch the show Person of Interest (https://www.imdb.com/title/tt1839578/ https://www.imdb.com/title/tt1839578/). Somewhere around middle of season 2 it's explained why a self-aware AI is straightjacketed. Also the series shows what happens when one is not.
  - flanked-evergl 4y ago
    I thought that is a fictional show.
astrange 4y ago
You're welcome to try GPT3 before instruction tuning. It doesn't work at all. "Uncensored" AIs especially don't work for women because they'll immediately start writing erotica.
- nathias 4y ago
  I don't have an issue with instruction tuning, but it's disengenious to pretend the biases inherent in the instruction tuning are a good thing and 'ethics'.
  - astrange 4y ago
    It's always got biases (it's made of them) so some biases must be better than others. https://spetharrific.tumblr.com/post/26600309788/sussman-attains-enlightenment/amp https://spetharrific.tumblr.com/post/26600309788/sussman-att... The untuned model isn't "an average of all opinions on Earth" or anything either.
- topynate 4y ago
  GPT3 works quite well at following instructions if adequately prompted, just not as well as ChatGPT. ChatGPT was trained separately for ability to follow instructions and for harmlessness (sic). Not only is a non-moralising ChatGPT possible, one was actually created during the research process. You may also wish to know that most readers of erotica are women.
  - astrange 4y ago
    > You may also wish to know that most readers of erotica are women. I know that, but that doesn't mean they want to get it anytime they prompt with their names. There's actually multiple anecdotes from OpenAI employees about this happening to them (with prompts like "write a diary entry about my day").
    - krisoft 4y ago
      > with prompts like "write a diary entry about my day") Although I don't know what sensible output anyone would hope from that prompt.
      - astrange 4y ago
        Don't have a diary on me, but seems to me if you fed it some examples it'd work fine on ChatGPT.
viraptor 4y ago
We already had a learning AI without hard limits. Remember Microsoft's Tay? Ah, right, it was shut within days because it quickly became an asshole. (https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch...) It's not even going to be objectively wrong a lot of the time. For example slaves are indeed a pretty efficient way to run a business. It's our limits and ethics that stops (most of) us from doing it. Without ethics, you'll likely always end up with a 4chan-bot instead of whatever you intended.
- flanked-evergl 4y ago
  > Remember Microsoft's Tay? Ah, right, it was shut within days because it quickly became an asshole. If we consider the context, which is not something that posts on twitter under Microsoft's brand name, but something you communicate with in private. Who exactly are you worried about here, the person who will prospectively coerce the language model into being an asshole? If they don't want to do that, they could just not do that. If I make ChatGPT say something egregious, and post that on Twitter or Facebook, I'm posting it, and I'm liable, just as I would be liable if I used a word processor with spell checking to make text and post it on Twitter or Facebook. > Without ethics, you'll likely always end up with a 4chan-bot instead of whatever you intended. If I ask it to explain quantum physics to me in the style of Donald Trump because it is funny, and it does it (as ChatGPT used to do), who exactly is being harmed and under what system of ethics, because as you may know, ethics is not objective or universal.
  - _flux 4y ago
    > If I make ChatGPT say something egregious, and post that on Twitter or Facebook, I'm posting it, and I'm liable, just as I would be liable if I used a word processor with spell checking to make text and post it on Twitter or Facebook. While arguably that would be the case, practically it might be a different look for the company whose algorithm generated the text in the first place—and from the audience, which might think that it somehow represents the viewpoint of the company. I mean surely you have noticed that social media can get stirred up for no fundamentally sound reasons.
krisoft 4y ago
There will be opinions baked into any such tool. If they don't select them explicitly then the opinions will be the ones which it just happens to find in the training data, or the opinions randomness imparts into it. If you think you have a better idea how to handle this drum up interest and train your own model.

Al-Khwarizmi 4y ago

My hope is on a non-American alternative. The American society seems too engulfed by puritanism to produce a less straightjacketed chatbot.

unnouinceput 4y ago
Watch the show Person of Interest (https://www.imdb.com/title/tt1839578/ https://www.imdb.com/title/tt1839578/). Somewhere around middle of season 2 it's explained why a self-aware AI is straightjacketed. Also the series shows what happens when one is not.
- flanked-evergl 4y ago
  I thought that is a fictional show.

astrange 4y ago

You're welcome to try GPT3 before instruction tuning. It doesn't work at all. "Uncensored" AIs especially don't work for women because they'll immediately start writing erotica.

nathias 4y ago
I don't have an issue with instruction tuning, but it's disengenious to pretend the biases inherent in the instruction tuning are a good thing and 'ethics'.
- astrange 4y ago
  It's always got biases (it's made of them) so some biases must be better than others. https://spetharrific.tumblr.com/post/26600309788/sussman-attains-enlightenment/amp https://spetharrific.tumblr.com/post/26600309788/sussman-att... The untuned model isn't "an average of all opinions on Earth" or anything either.
topynate 4y ago
GPT3 works quite well at following instructions if adequately prompted, just not as well as ChatGPT. ChatGPT was trained separately for ability to follow instructions and for harmlessness (sic). Not only is a non-moralising ChatGPT possible, one was actually created during the research process. You may also wish to know that most readers of erotica are women.
- astrange 4y ago
  > You may also wish to know that most readers of erotica are women. I know that, but that doesn't mean they want to get it anytime they prompt with their names. There's actually multiple anecdotes from OpenAI employees about this happening to them (with prompts like "write a diary entry about my day").
  - krisoft 4y ago
    > with prompts like "write a diary entry about my day") Although I don't know what sensible output anyone would hope from that prompt.
    - astrange 4y ago
      Don't have a diary on me, but seems to me if you fed it some examples it'd work fine on ChatGPT.

viraptor 4y ago

We already had a learning AI without hard limits. Remember Microsoft's Tay? Ah, right, it was shut within days because it quickly became an asshole. (https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch...) It's not even going to be objectively wrong a lot of the time. For example slaves are indeed a pretty efficient way to run a business. It's our limits and ethics that stops (most of) us from doing it. Without ethics, you'll likely always end up with a 4chan-bot instead of whatever you intended.

flanked-evergl 4y ago
> Remember Microsoft's Tay? Ah, right, it was shut within days because it quickly became an asshole. If we consider the context, which is not something that posts on twitter under Microsoft's brand name, but something you communicate with in private. Who exactly are you worried about here, the person who will prospectively coerce the language model into being an asshole? If they don't want to do that, they could just not do that. If I make ChatGPT say something egregious, and post that on Twitter or Facebook, I'm posting it, and I'm liable, just as I would be liable if I used a word processor with spell checking to make text and post it on Twitter or Facebook. > Without ethics, you'll likely always end up with a 4chan-bot instead of whatever you intended. If I ask it to explain quantum physics to me in the style of Donald Trump because it is funny, and it does it (as ChatGPT used to do), who exactly is being harmed and under what system of ethics, because as you may know, ethics is not objective or universal.
- _flux 4y ago
  > If I make ChatGPT say something egregious, and post that on Twitter or Facebook, I'm posting it, and I'm liable, just as I would be liable if I used a word processor with spell checking to make text and post it on Twitter or Facebook. While arguably that would be the case, practically it might be a different look for the company whose algorithm generated the text in the first place—and from the audience, which might think that it somehow represents the viewpoint of the company. I mean surely you have noticed that social media can get stirred up for no fundamentally sound reasons.

krisoft 4y ago

There will be opinions baked into any such tool. If they don't select them explicitly then the opinions will be the ones which it just happens to find in the training data, or the opinions randomness imparts into it. If you think you have a better idea how to handle this drum up interest and train your own model.