In ChatGPT answer for a question “Please provide example of calculation of parameters for polynomial model” used numeric values for demonstration, but those numeric values sometimes are not actually applied to mentioned equations.
See attached image with screenshot of issue answer. Result of matrix multiplication in sub-step 2 “Compute {X transposed} * Y” is incorrect. Correct resulting matrix should be [212, 82, 35].
I do not know how the resulting matrix was guessed, but actual values weren’t applied to the actual equation of matrix multiplication.
To fix this in the future, I propose OpenAI to build a non-AI algorithmic system that will contain an axiomatic space of knowledgeable math, that will actually perform math operations. AI models (and AGI models) will use this algorithmic system as a back-end for math operations to avoid errors. Anyways it’s too big to be done by a single person in reasonable time.
If OpenAi is interested in this - I can add links to articles with thoughts on this subject, and links to PoCs on Kaggle that demonstrates the possibility of a non-AI axiomatic system in Euclidean space to resolve geometry tasks. (do not want to add unnecessary links without invitation).