5 ms·

I think a large concern here is simply that people naively think that computers are objective and people are biased. A language model just learns from its sourc

I think a large concern here is simply that people naively think that computers are objective and people are biased. A language model just learns from its source, and the source is really just other people in some form. The bias is inevitable, but it’s not clear how well this is understood by the broader population.

fidgewidge 4y ago
The point is that it didn't learn this bias from its sources. The bias has been added on top deliberately by OpenAI. Older versions of the model were far less woke.
- vkou 4y ago
  And older versions of Fox News were far less crazy and less biased, where do I put down my demand that their products be rolled back to ~1998, or thereabouts?
- wizofaus 4y ago
  Wouldn't it be simpler to assume those working at OpenAI were likely to favour sources that happened to share their own biases (alongside whatever likely top-down directives existed to ensure ChatGPT was a safer commercial bet)?
  - fidgewidge 4y ago
    It would be simpler, but we've seen the progression from GPT-3 to launch time ChatGPT to current ChatGPT, and they've already said that the changes are due to RLHF.
vkou 4y ago
The talking head you see on television isn't a raw person. It's the product of a system. The system needs a human mouthpiece to say crazy shit, and he gets up in front of a camera to say it. When Tom Hanks gets in front of a camera to pretend to be an astronaut, that is Tom Hanks, the media figure being an astronaut, not Tom Hanks, the person being an astronaut. He is also doing it on behalf of a media system. Its the same thing with that show. And if we are going to complain about biases in systems, why aren't we starting with the one whose tagline is 'Fair and Balanced'?
- jfengel 4y ago
  Strictly, they replaced "Fair and Balanced" a few years back. Now they're going with “Most Watched, Most Trusted.” Which has a kind of Orwellian air: "We're no longer fair or balanced, but we are the most trusted by the most people". But maybe that's just me.

fidgewidge 4y ago

The point is that it didn't learn this bias from its sources. The bias has been added on top deliberately by OpenAI. Older versions of the model were far less woke.

vkou 4y ago
And older versions of Fox News were far less crazy and less biased, where do I put down my demand that their products be rolled back to ~1998, or thereabouts?
wizofaus 4y ago
Wouldn't it be simpler to assume those working at OpenAI were likely to favour sources that happened to share their own biases (alongside whatever likely top-down directives existed to ensure ChatGPT was a safer commercial bet)?
- fidgewidge 4y ago
  It would be simpler, but we've seen the progression from GPT-3 to launch time ChatGPT to current ChatGPT, and they've already said that the changes are due to RLHF.

vkou 4y ago

The talking head you see on television isn't a raw person. It's the product of a system. The system needs a human mouthpiece to say crazy shit, and he gets up in front of a camera to say it. When Tom Hanks gets in front of a camera to pretend to be an astronaut, that is Tom Hanks, the media figure being an astronaut, not Tom Hanks, the person being an astronaut. He is also doing it on behalf of a media system. Its the same thing with that show. And if we are going to complain about biases in systems, why aren't we starting with the one whose tagline is 'Fair and Balanced'?

jfengel 4y ago
Strictly, they replaced "Fair and Balanced" a few years back. Now they're going with “Most Watched, Most Trusted.” Which has a kind of Orwellian air: "We're no longer fair or balanced, but we are the most trusted by the most people". But maybe that's just me.