It’s baked into the training. It’s not a simple thing to take it out. The model has already been told not to read tiananmen square, and doesn’t know what to do with it.
Even though it’s magnitudes lower than comparable models, Deepseek still cost millions to train. Unless someone’s willing to invest this just to retrain it from scratch, you’re left with the alignment of its trainers.
Yeah, I guess you could realign it without retraining the whole thing! Dunno what would be the cost though, sometimes this is done with a cohort of human trainers 😅
Question: as i understood it so far, this thing is open source and so is the dataset.
With that, why would it still obey Chinese censorship?
It’s baked into the training. It’s not a simple thing to take it out. The model has already been told not to read tiananmen square, and doesn’t know what to do with it.
Even though it’s magnitudes lower than comparable models, Deepseek still cost millions to train. Unless someone’s willing to invest this just to retrain it from scratch, you’re left with the alignment of its trainers.
Good point.
Is the training set malleable, though? Could you give it some additional rules to basically sidestep this?
Yeah, I guess you could realign it without retraining the whole thing! Dunno what would be the cost though, sometimes this is done with a cohort of human trainers 😅
I feel like we’re talking about a guard dog now…