Pangram CEO Max Spero on why AI writing is so bloated: "I think most AI models, they're not actually optimized for information density. They're optimized, at least standard RLHF is, we pose a question, the assistant answers, and then we're looking at this answer and saying, does the response contain the answer?" "So by default, this is actually going to reward a longer response, because a longer response is more likely to include the answer." "I think obviously today big labs are doing a lot…