The "Moons larger than Mercury" example is interesting. ChatGPT first gives the wrong answer, but then goes on to actually explain the opposite (the correct ans

The "Moons larger than Mercury" example is interesting. ChatGPT first gives the wrong answer, but then goes on to actually explain the opposite (the correct answer). It seems it got tricked by doing chain of thought in the wrong order. If it had done the thinking first, it would probably have answered correctly. Another option would be to use a multi-pass algorithm. It would be easy for a language model reading the first output to generate a new answer correct answer. I wonder if this kind of boosted model (similar to diffusion) could be the future of text generators.