My batchmate and fellow Amazon colleague Jayanth pointed out that though ChatGPT is good at generative learning, it might perform poorly on tasks which require precision. One of the questions that was posed to ChatGPT was: "Solve for x, y, z, w: x^4 + y^4 + z^4 = w^4". The solution provided in his chat thread/session was patently wrong but when I posed the same question to ChatGPT again & again (Yes, you can ask it to retry and Yes, you might get weirdly different responses), it gave a "proof by contradiction" which makes propositional logic stand on its head. It's again patently "absurd", not only "incorrect". Since ChatGPT output a proof in LaTeX, I rendered it and took a snapshot for you guys. This needs to be understood in the light of amazing interactive theorem provers available like CoQ which of course do not use neural networks to write proofs. It also needs to be understood that at a very unique point of proving journey, ChatGPT completely derails. It goes into a whirlpool of illogical implications and never seems to come out. This is true for other mathematical proofs I asked it to write and verify ranging from abstract algebra, number theory to stochastic calculus.
#chatgpt
#chatgpt3
#proofs
#mathematics
#theoremprover
#coq
#mathematicalmodeling
#logic
#propositions
#openai
#algebra
#neuralnetworks
