The thing is we tried that for decades, using more formal logic to build reasoning engines. And we never got it to be even a fraction as good and generic as learning-based LLMs are today.