Alternative hypothesis: it's hard to define what "smart" and "dumb" mean, and the metrics they're using are optimizing for a different definition than what othe

Alternative hypothesis: it's hard to define what "smart" and "dumb" mean, and the metrics they're using are optimizing for a different definition than what other people are using. For example if their definition of "smart" means refusing to answer questions on delicate or specialized topics, then in some way they've made the model smarter but in ways they made it dumber.