I personally blame this on instruction tuning. Base models are in my mind akin to the Solaris Ocean. Wandering thoughts that we aren't really even trying to understand. The tuned models, however, are as if somebody figured out a way to force the Solaris Ocean to do their bidding as the Ocean understands it. From this perspective it is clear that giving everyone barely restricted ability to channel the Ocean thoughts into actions leads to outcomes that we now observe.