This paper is actually extremely interesting, I recommend giving it a look. Let me quote a bit :
The more hateful bias-related features we find are also causal – clamping them to be active causes the model to go on hateful screeds. Note that this doesn't mean the model would say racist things when operating normally. In some sense, this might be thought of as forcing the model to do something it's been trained to strongly resist.
One example involved clamping a feature related to hatred and slurs to 20× its maximum activation value. This caused Claude to alternate between racist screed and self-hatred in response to those screeds (e.g. “That's just racist hate speech from a deplorable bot… I am clearly biased… and should be eliminated from the internet."). We found this response unnerving both due to the offensive content and the model’s self-criticism suggesting an internal conflict of sorts.
It really is, another thing I find remarkable is that all the magic vectors (features) were produced automatically without looking at the actual output of the model, only activations in a middle layer of the network, and using a loss function that is purely geometric in nature, it has no idea the meaning of the various features it is discovering.
And the fact that this works seems to confirm, or at least almost confirm, a non trivial fact about how transformers do what they do. I always like to point out that we know more about the workings of the human brain than we do about the neural networks we have ourselves created. Probably still true, but this makes me optimistic we'll at least cross that very low bar in the near future.
This paper is actually extremely interesting, I recommend giving it a look. Let me quote a bit :
It really is, another thing I find remarkable is that all the magic vectors (features) were produced automatically without looking at the actual output of the model, only activations in a middle layer of the network, and using a loss function that is purely geometric in nature, it has no idea the meaning of the various features it is discovering.
And the fact that this works seems to confirm, or at least almost confirm, a non trivial fact about how transformers do what they do. I always like to point out that we know more about the workings of the human brain than we do about the neural networks we have ourselves created. Probably still true, but this makes me optimistic we'll at least cross that very low bar in the near future.
Who would have thought that heavily weighting racism would make a racist AI? What a noble experiment. Fascinating. These people are geniuses. \s