> Presumably they don't want to release this so that it is raw and unfiltered and can better monitor for cases of manipulation or deviation from training.
My take was:
1. A genuine, un-RLHF'd "chain of thought" might contain things that shouldn't be told to the user. E.g., it might at some point think to itself, "One way to make an explosive would be to mix $X and $Y" or "It seems like they might be able to poison the person".
2. They want the "Chain of Thought" as much as possible to reflect the actual reasoning that the model is using; in part so that they can understand what the model is actually thinking. They fear that if they RLHF the chain of thought, the model will self-censor in a way which undermines their ability to see what it's really thinking
3. So, they RLHF only the final output, not the CoT, letting the CoT be as frank within itself as any human; and post-filter the CoT for the user.
My take was:
1. A genuine, un-RLHF'd "chain of thought" might contain things that shouldn't be told to the user. E.g., it might at some point think to itself, "One way to make an explosive would be to mix $X and $Y" or "It seems like they might be able to poison the person".
2. They want the "Chain of Thought" as much as possible to reflect the actual reasoning that the model is using; in part so that they can understand what the model is actually thinking. They fear that if they RLHF the chain of thought, the model will self-censor in a way which undermines their ability to see what it's really thinking
3. So, they RLHF only the final output, not the CoT, letting the CoT be as frank within itself as any human; and post-filter the CoT for the user.