The sample above gives a cue to how rudimentary reasoning can be done. After GPT answers a question, it asks itself "what are the possible flaws in this answer?

The sample above gives a cue to how rudimentary reasoning can be done. After GPT answers a question, it asks itself "what are the possible flaws in this answer?" then it evalutes each potential flow, up to 3-5, with another question "does this flaw really apply?" It's going to be like the mcts addition that made alphago so strong.