OpenAI-Dynamic-Code-Executor
Originally published as an X Article.
On March 16th, I attended the AGI House hackathon. I’ve been experimenting with LLMs, GPTs, and related technologies since ChatGPT’s release in late 2022. One of the reasons I started programming in 2016 and 2017 was the promise of AI and ML. However, when I realized I didn’t enjoy statistics and math as much, I switched to become a software engineer.
I never thought I would witness something like ChatGPT, and I’m still surprised that the world hasn’t gone mad since writing and talking have been proven as not solely human activities. I believe the Turing test can be fully passed now with even half-assed prompting.
Regarding my March 16th AGI House hackathon experience, I was quite impressed. It was as if suddenly, all the people around me saw the light in this technology. I was no longer just a crazy preacher. All the attendees were up-to-date with the latest releases, buzz, and potential of the technology.
During the hackathon, I met many interesting people, discussed various problems, and had numerous ideas on things to work on. I found getting to know these people more enjoyable than actually hacking something together, which can be done later.
On my flight back, I had some ideas about an agent that could take instructions on what code to write, write it, run it, test it, and store it. This would be a developer agent - a dev/tester agent that could create reusable and improvable code. The interesting part is that this code addition is dynamic, and the code is guaranteed to run because the agent runs it itself.
A problem many agents have is to generate code that is not correct or does not work. This project aims to solve this problem by allowing the agent to execute the code it generates and learn from the results. This code could be used by other agents as a third-party tester or even as a user.
This is only an experiment but is something that was grinding in my mind for a while.
