Plus, your video is of Kimi K2.6.

This has always been the case with Chinese models. Their web ui is a service provided from China, so they are required to censor it, but not the models themselves.

What is plausible is that they haven't made any attempts to explicitly steer certain opinions into a certain direction, and just let the model take over the bias of the training data, whatever that may be. Filtering in the front end is the easy, lazy way out to be legally compliant.

Other companies have not shown similar problems so far.

China has many government agencies and state-owned enterprises that need models which can be deployed locally. This is also part of "Xinchuang" (self-owned systems, self-owned hardware, etc. New government computers all run special Linux versions on domestic CPUs, and LLMs also need to be like this).

IMHO an Ai is the llm plus it's harness.

A good harness would allow the llm to investigate on a map.

Just like the llm can use a python script to figure out how many r's there are in strawberry.

These tests are simply not that predictable of performance of the llm.