3 ms·

I've always interpreted AI "safety" as being more about ensuring ideological consistency and a particular narrative than anything about actually keeping humanit

I've always interpreted AI "safety" as being more about ensuring ideological consistency and a particular narrative than anything about actually keeping humanity "safe". And I'm not even talking about the culture war, just yesterday I tried having DALLE generate a picture for a DND character and it kept refusing to (and no it wasn't some sort of weird hentai thing) because it felt like it was not an appropriate image.

phowat 3y ago
One thing I noticed recently and was able to reproduce a bunch of times is asking dall-e the exact same prompt where the only difference is the subject being described as a male or female and it refuses to generate the female one. I don't know what to think of it but it's weird.
- jiggawatts 3y ago
  It’s refusing to be sexist by treating women as special.
ekidd 3y ago
> And I'm not even talking about the culture war, just yesterday I tried having DALLE generate a picture for a DND character and it kept refusing to (and no it wasn't some sort of weird hentai thing) because it felt like it was not an appropriate image. I think this may be a problem in DallE's training data. I tried using it to generate character portraits for a local gaming group. And even when it did generate images of female characters, nearly half of them were NSFW. It took quite a lot of prompt engineering to minimize this, and I still wasn't happy about the results. Oh, and DallE kept getting upset at me for its own policy violations. Other than that, DallE 3 has been terrific. But I think you're hitting some kind of significant failure mode. > anything about actually keeping humanity "safe". I've said this before and I'll say it again—lots of the people working on this stuff actually do fear that we're getting close to a point where we might build an actual working AI. Whether or not they're correct is another question, but many of them do fear that. Also, early pre-release versions of GPT4 were apparently pretty disturbing. I've heard them compared to talking to Hannibal Lector.

phowat 3y ago

One thing I noticed recently and was able to reproduce a bunch of times is asking dall-e the exact same prompt where the only difference is the subject being described as a male or female and it refuses to generate the female one. I don't know what to think of it but it's weird.

jiggawatts 3y ago
It’s refusing to be sexist by treating women as special.

ekidd 3y ago

> And I'm not even talking about the culture war, just yesterday I tried having DALLE generate a picture for a DND character and it kept refusing to (and no it wasn't some sort of weird hentai thing) because it felt like it was not an appropriate image. I think this may be a problem in DallE's training data. I tried using it to generate character portraits for a local gaming group. And even when it did generate images of female characters, nearly half of them were NSFW. It took quite a lot of prompt engineering to minimize this, and I still wasn't happy about the results. Oh, and DallE kept getting upset at me for its own policy violations. Other than that, DallE 3 has been terrific. But I think you're hitting some kind of significant failure mode. > anything about actually keeping humanity "safe". I've said this before and I'll say it again—lots of the people working on this stuff actually do fear that we're getting close to a point where we might build an actual working AI. Whether or not they're correct is another question, but many of them do fear that. Also, early pre-release versions of GPT4 were apparently pretty disturbing. I've heard them compared to talking to Hannibal Lector.