so i've been experimenting with my architecture that aimed to replicate model one's architecture, judging by logprobs.

so i've been experimenting with my architecture that aimed to replicate model one's architecture, judging by logprobs. I've succeeded in replicating it, and I was able to run multimodal llm's such as qwen with the new inference harness, and then to test it out, i switched to a 4b model and ran doom emulator. Since im very limited in compute and memory its heavily labotomized (4b param model lol) but it was able to kill some enemies based off of just the image of the game. looking promising so far, planning on opensourcing this soon