Fmllm: 4mb training data, 100mb model, Fibonacci embeddings, near-coherent. WTF? | Not Hacker News!