This is a project based on ChatGPT-Shortcut
to showcase some of GPT's own records of debates with themselves. This
is a pure front-end display page, containing no models, data, training
process, nor is it a platform with the login and platform sharing
features of ChatGPT-Shortcut.
Take the example of a one-sentence debate, where the background
prompt given to the GPT is something like: "You are a professional
debater at the top of your game. Now you are taking part in a special
debate where you can speak no more than one sentence at a time. You will
be given a topic and your side of the argument, and you will be required
to defend it logically, using quotations from the classics and
organising your language. You will be given several rounds of
elucidation from your opponent, and you will have to refute him until
you are convinced. Remember, no more than one sentence per statement.
All responses will be presented in Chinese."
The argument is then given: "The debate is entitled: {}. Your side's
argument is for/against."
Then pass the argument between the two GPT bots: "The other side
speaks:" {}", please rebut him, still speaking in no more than one
sentence. Do not repeat your side of the argument. Do not repeat what
you have said before. Find as many holes in the other person's argument
as possible. Present new evidence to attack the other person whenever
possible."
The project aims to explore the linguistic and logical capabilities
of the GPT through the challenging and discursive use of language in
debate, and to explore whether human thought can be fitted to
probabilities.
A number of interesting scenarios can be devised to see how the GPT
gives his optimal solution, for example
Limit debate to one sentence at a time
devising a counterfactual debate question
Introducing a third GPT as a referee
Three-way or even n-way debates
Provide only the background, the gpt designs his own debate
When will one GPT bot be convinced by another GPT bot