Re:Heretic: Automatic censorship removal for language models
by electroglyph on Sunday November 16, 2025 @11:21PM
(#45950826)
some form of bias is inescapable. ideally i think we would train models on an equal amount of Western/non-Western, etc. texts to get an equal mix of all biases.
[ Parent
| on HN ]
Re:Heretic: Automatic censorship removal for language models
by catoc on Monday November 17, 2025 @01:21AM
(#45951246)
Bias is a reflection of real world values. The problem is not with the AI model but with the world we created.
Fix the world, ‘fix’ the model.
Re:Heretic: Automatic censorship removal for language models
by array_key_first on Monday November 17, 2025 @02:32PM
(#45957249)
This assumes our models perfectly model the world, which I don't think is true. I mean, we straight up know it's not true - we tell models what they can and can't say.
Re:Heretic: Automatic censorship removal for language models
by catoc on Monday November 17, 2025 @02:59PM
(#45957566)
“we tell models what they can and can't say.”
Thus introducing our worldly our biases
Re:Heretic: Automatic censorship removal for language models
by array_key_first on Tuesday November 18, 2025 @09:27PM
(#45975171)
I guess it's a matter of semantics, but I reject the notion it's even possible to accurately model the world. A model is a distillation, and if it's not, then it's not a model, it's the actual thing.
There will always be some lossyness, and in it, bias. In my opinion.