a lot of the LLMs and models-of-this-approach blow out when they go beyond window length (and similar-strain cases), yeah, but I wonder if this is them trying to do this because of that or because of other bits

I could also see this being done as “lowering liability” (which is a question that’s going to start happening as all the long-known issues of these things start amplifying as more and more dipshits over-rely on them)