How can it be specifically trained on benchmarks when it is leading on blind chatbot tests?
The post you quoted is not a Grok problem if other LLMs are also failing, it seems, to me, to be a fundamental failure in the current approach to AI model development.