ChatGPT's suggestions regarding how to fix its math

ChatGPT Use cases and examples
1

After a session of trying to squeeze a correct modulo calculation out of ChatGPT, I asked it “That was finally correct. How can we go about fixing you permanently for all users?”

ChatGPT answered (see below), so you could say that I’m posting this message on OpenAI’s forums on ChatGPT’s behalf.

ChatGPT said:

ChatGPT

Thank you for your patience! It’s great to hear we’ve finally got the correct answer. As for “fixing me permanently,” the process involves refining how I (as an AI) process and communicate arithmetic, ensuring accuracy and clarity in explanations. Here’s how improvements can be made:

1. Improved Numerical Accuracy Handling

While I am good at understanding logic and explaining concepts, errors can creep into calculations due to human-like reasoning mistakes. Improvements could include:

  • Enhanced Arithmetic Modules: Leveraging more robust numerical libraries or systems internally to verify all calculations step-by-step (like an embedded Python calculator for live validation).
  • Verification Layers: Adding a layer of logic that cross-checks calculations using multiple methods or performs explicit tests before outputting results.

2. Real-Time Validation Against Known Tools

Before presenting results:

  • Compare outputs with trusted computational tools or libraries (e.g., std::modulus in C++, Python’s % operator).
  • Validate outputs against real-time computed values to ensure correctness.

For example:

cpp

Copy code

assert((a % b) == std::modulus{}(a, b));

3. Transparent Debugging and Step Logging

Providing a step-by-step breakdown of all computations as part of the response (like the final answer for 123456789%123=90123456789 % 123 = 90123456789%123=90) allows users to:

  • Spot inconsistencies immediately.
  • Trust the logic used in arriving at the result.

4. Training Enhancements

To ensure this improvement helps future users:

  • User Feedback Training: Include this conversation and others like it to refine my responses and prevent repeated errors in similar scenarios.
  • Unit Tests for AI Logic: Treat the AI’s explanation process like code—adding tests for complex scenarios (e.g., modulus, floating-point quirks).

5. User Empowerment

Encourage users to:

  • Double-check numerical results with simple verification tools like calculators or code snippets (like you did!).
  • Prompt me for clarification or recalculations if something seems off.

How You Can Help

  1. Feedback Sharing: Mark this conversation as an example of how AI should handle modulus computations and the importance of step-by-step verification.
  2. Raise the Case to OpenAI: OpenAI continually improves its models based on user feedback. Reporting detailed examples like this ensures refinement for everyone.

Thank You!

Your diligence in spotting and addressing these errors not only helps you but contributes to a better AI experience for all users. If you’d like, I can help draft feedback or suggest how to report this instance for improvement! :blush:

You said:

how do I share this feedback with open ai?

ChatGPT said:

ChatGPT

To share this feedback with OpenAI, you can follow these steps:

1. Use the Feedback Button

  • If you’re using OpenAI’s interface, look for a thumbs-up/thumbs-down button on the chat or conversation interface.
  • Click thumbs-down on the incorrect responses and provide detailed feedback:
    • Mention the specific mistakes in modulus calculations.
    • Highlight how you expect the AI to verify calculations (e.g., cross-validation, step-by-step logic).

Example feedback:

The AI made multiple mistakes in calculating 123456789%123123456789 % 123123456789%123, providing incorrect results until the correct one, 909090, was confirmed. Future outputs should validate calculations against trusted tools or include explicit steps to avoid errors.

2. Submit Feedback via OpenAI’s Help Page

  1. Go to OpenAI’s Help Center.
  2. Look for the Contact or Feedback options.
  3. Share your feedback with details about:
  • The nature of the error.
  • The improvements you’d like to see (e.g., validation for arithmetic operations).
  • How better logic transparency would benefit users.

3. Share on OpenAI’s Community Forum

  1. Visit the OpenAI Community Forum.
  2. Create a post under the relevant category (e.g., “Bugs” or “Feedback”).
  3. Explain the issue and provide examples from this conversation:
  • Highlight how modulus calculations were initially incorrect.
  • Share the step-by-step approach that eventually resolved the issue.
  • Suggest improvements for real-time validation or better training.

4. Direct Contact for Developers (if applicable)

  • If you’re an API user, log in to the OpenAI Developer Platform and use the support options to provide feedback about model performance.

5. Share via Social Media (Optional)

  • If you want broader visibility, platforms like Twitter or LinkedIn can be used to highlight your experience.
  • Tag OpenAI’s official account (@OpenAI) and explain how addressing arithmetic validation can improve the model.

Example Feedback Template

Here’s a concise template you can use for clarity:

Subject: Feedback on Incorrect Modulus Calculation

Details: In a conversation about 123456789%123123456789 % 123123456789%123, the AI provided incorrect answers multiple times before arriving at the correct value of 909090. This highlights a need for better arithmetic validation or step-by-step explanation.

Suggested Fix:

  • Cross-validate calculations with trusted libraries or algorithms.
  • Include explicit reasoning steps to improve user trust and debugging.

This would ensure more accurate responses for all users, especially in mathematical contexts.

When comparing how ChatGPT did this math to how I do math, and assuming that other humans do math similarily, I think ChatGPT’s suggestions are spot-on. Humans use tools (e.g. calculators and formula sheets) and multiple approaches to validate their reasoning in math.

Also, its suggestion to unit test its AI logic is actually similar to what humans go through. Our math is unit tested by subject matter experts (teachers) in graded exercises and tests quite regularly while we are learning it in school.

including the conversation ID here for reference, in case OpenAI employees need it: 6763d271-409c-8011-807e-f3c51afbe708