A software engineer taught AI to hunt bugs by interfacing an LLM with debugging tools and has released the open source code: 'It's like going from hunting with a stone spear to using a guided missile'

AI is an incredibly contentious space in the tech world, but that's mostly because of a misappropriation of all things involved with it. It all starts with the name, Artificial Intelligence or AI, because almost nothing we call AI in today's landscape qualifies as such. These aren't intelligences, artificial or otherwise.

Generally, they're language models trained on huge data sets noticing patterns and predicting next steps. As such they're terrible at lots of things we see them used for and just won't stop halucinating, but they can be useful tools when put to the right work. Debugging Windows and analysing crash data is a perfect example of exactly the right work for an AI.

Tom's Hardware reports that Sven Scharmentke (AKA Svnscha) a software engineer more than familiar with debugging Windows crashes has released a language model that can essentially do it for you. The mcp-windbg tool gives language models the ability to interface with WinDBG, Windows own multipurpose debugging tool. The result is a crash analysis tool that can be interacted with in a natural language that hunts down crash points for you.

Scharmentke put his findings in a blog post, including a video which shows an example of the debugger at work. You can watch as Copilot is asked in very plain natural english to help find the problems, and it does so. It's able to read the crash dump, find relevant codes within it to pinpoint the problem, and then analyse the original code for problems and provide solutions. It can even help you find the crash dump in the first place.

From the example it looks like Scharmentke has successfully taken a rather complicated process that would once require a trained professional to ascertain, and turned it into something even I could do. Better yet it's a frustrating task, that's far more easily handled by a machine than a human, and not many people would want to have to do anyway. That's a perfect use for an LLM, and that's just such a breath of fresh air.

Usually this kind of debugging takes a lot of time, an encyclopaedic knowledge of various codes and pointers, an understanding of how the code runs, and a dogmatic level of determination. Now, it's a few minutes of casual conversation with your friendly computerised helper. As Scharmentke says "It's like going from hunting with a stone spear to using a guided missile,".

But of course this comes with the same standard warning all AI should. It doesn't actually think, and its answers should always be taken with some scepticism involved. Scharmentke also reminds us that this isn't a magical cure-all tool, and instead is just a "simple Python wrapper around CDB that relies on the LLM's WinDBG expertise." Still if you want to use this LLM and put this new style of debugging interface to the test, you can download it from Github and give it a try.