Read the same Karpathy post and had the same inclination - build my own. Ended up testing yours first because you're not some random internet dude. One thing I keep coming back to is the token-intensive nature of the research. Groups working on the same or adjacent topics end up duplicating these costs independently. There's probably something interesting in pooling raw source piles across trusted peers. Shared resources, independent Q&A and output processing. Curious if you've thought about the pooling angle.