Raw LLM Responses

To be honest I don’t really see why this is surprising. With machine learning (and life more broadly), everything comes with a cost; you want your model to give you safer answers? This will come at a cost in some way to accuracy. A very similar tradeoff exists when trying to design attack resistance for machine learning models; you can make your model resistant to a broad spectrum of attacks, but if you do, the accuracy suffers because of it. The real question is whether the tradeoff is worth it. I think the general discussion about this has become ‘why would they do this to us’ when in reality the better question is ‘was it worth it’, and I think there’s a good discussion to be had there with good points for both sides.