Claude 2 appears to be explicit about learning from input. Since my first point of comparison is GPT-4, and Claude 2 is asking for correction, it occurs to me t

Claude 2 appears to be explicit about learning from input. Since my first point of comparison is GPT-4, and Claude 2 is asking for correction, it occurs to me that this could be an interesting strategy to close the gap — allow your users to equalize the models through reinforcement learning from your competitor’s model. It can’t work for the first/best model as there is no reference and it seems it would be more vulnerable to noise, but if you have a known better model, it seems it can work.