Can ChatGPT Fix Bugs? Some developers believe that machine learning technology has become advanced enough to enable code to heal itself from common errors. This week, a “pseudonymous” developer, known as BioBootloader, showcased a project called Wolverine, which uses GPT-4 to automatically fix bugs in Python code.
At first glance, the demonstration showed a cleverly designed prompt designed to help with programming. But it goes much deeper as the technology used provides potential for a program to make repairs beyond runtime errors, and instead deliver required results. It remains to be seen whether Wolverine is a bridge to the future, or simply a dead end.
BioBootloader created Wolverine as an AI-based tool to allow Python scripts to self-heal. In the demonstration, he injects mistakes into a custom calculator script and then runs it, only to have it crash. With the help of Wolverine, GPT-4 is able to identify and fix the errors by applying proposed changes to the code, re-running it, and then making more changes if needed. This could create a future where apps are able to fix their own bugs, and even detect and fix issues that may arise after deployment. Questions about safety, though, need to be addressed first.
The technology behind Wolverine is the result of a complex process of prompt engineering from BioBootloader. While currently a primitive prototype, the program is a good example of AI technology continuing to improve and stabilize. Donald Papp, who broke the story, labeled it as a “self-healing” ability, whereas andreas-motzek reminds users that chatbots don’t always understand the project requirements. Adrian Knarr is even more skeptical, arguing that the use of this technology will only encourage bad programming habits, whereas programming in compiled languages can catch errors that might normally manifest during runtime.
BioBootloader has opened up the project for contributions, and encourages people to explore possible ways to extend the program. He concedes that ChatGPT has limits in terms of the number of tokens it can store in a running memory, and may “forget” details as conversations may switch between different paths – but he also notes that it works well for API integrations, helping developers cut down on their development time and effort.
Whether or not ChatGPT can fully replace a human programmer is yet to be seen, but the evolution of the AI-powered tool serves as an exciting reminder of the evolution of technology. The enterprise and open source world has already started to embrace AI-enhanced tools and services, and ChatGPT combined with Wolverine has the potential to be the next big step towards automated development.
