Mathematics and ChatGPT 3.5 Turbo

[martin.merck]

It seems that GPT 3.5 Turbo is not able to turn to polar coordinates when solving the integral of a Gaussian function. Any clue if Chat GPT 4 is able to do that? I am not sure if this a legitimate question. If not, please delete it!

[EricGT]

Welcome to the forum!

@martin.merck

Any clue if Chat GPT 4 is able to do that?

I find that if you ask ChatGPT to list the rules of Calculus needed to solve the problem in Prolog then ask it to use chain of thought by showing each step it does much better.

For a post with more detail see

Chat GPT 4 and calculus: Experiences?

Don’t be so fast to shoot ChatGPT down or LLMs for that matter. Prompt Create the rules of calculus derivatives for Prolog then show step by step how to calculate the derivative of 2-(2/t) Completion Creating a Prolog program to handle calculus derivatives involves defining rules that capture the basic principles of differentiation. In Prolog, we can define rules for the derivative of constants, power functions, addition, subtraction, multiplication, division, and more complex functions as …

---

There are two basic yet powerful ideas behind this that work in other situations.

1. Converting the part that needs to be translated to something the LLM is trained. For example many find that translating the question to English then asking the question and translating back to the original language is better than prompting in the original language. Also programmers find that translating the code to Python, asking for the Python code to solve the problem then translating the Python back to original language works better at times.

2. Give the LLM time to think as some would say or to reason out the steps by using Chain of Thought. In this case by listing the Calculus rules needed first one can see that the LLM knows the correct rules and then can use those rules to solve the problem.

While I know Prolog and what and how the rules of Calculus in Prolog work, many may not and so my adaptation is not suitable for all. However these two ideas do seem to fit well in certain situations.