Hey everyone, I’m the owner of the repo. This started as one of my daily crazy ideas for optimizing LLMs using my software engineering background and I’m just trying to think outside the box when it comes to LLM optimization, even though I’m not an AI engineer or researcher. There’s no academic paper behind this, it’s really just me coming up with the idea and Claude helping prove it out. And tveita is right, there is a real bug, although it doesn’t affect the published benchmark results, I will get that fixed in the next release.