Live at: https://gmunhoz0810.github.io/GPT-Learns-Counting/
In comparisson to the official GPT-4o which still gets this wrong.
GPT Learns Counting is an improved version of the traditional GPT models, which still struggle with letter/word counting tasks.
This new version can correctly:
- Count the occurrences of a specific letter in a word or phrase
- Count the total number of characters in a given text
While counting letters or characters might seem trivial for humans, it's an interesting challenge for AI language models. This project showcases how we can extend the capabilities of large language models like GPT to perform precise counting tasks, which they are not inherently designed to do.
By implementing this functionality, we bridge the gap between natural language processing and exact numerical operations, allowing users to interact with the AI in a more practical and demonstrable way.
The core of this application relies on OpenAI's function calling feature. Here's a brief explanation of how it works:
-
User Input: The user submits a query, like "How many 'a's are in 'banana'?"
-
AI Interpretation: The AI model (GPT-4) interprets the user's intent and determines which function to call.
-
Function Calling: Based on the interpretation, the AI calls one of two custom functions:
count_occurrences(needle, haystack): Counts occurrences of a specific letter or word in a phrase.count_characters(text): Counts the total number of characters in a given text.
-
Result Processing: The function performs the counting operation and returns the result.
-
Response Generation: The application takes the function's result and generates a human-readable response.
-
Display: The response is then displayed to the user.
This approach allows the AI to perform exact counting operations while maintaining a natural language interface for the user.
We welcome feedback and contributions to this project. If you have any suggestions or encounter any issues, please open an issue on our GitHub repository.
Happy counting!