Research and writing on small models.https://oropis.com/https://oropis.com/posts/emergent-behavior-self-play/https://oropis.com/posts/emergent-behavior-self-play/In multi-agent RL the environment includes the other learners. Self-play turns that instability into a curriculum: hide-and-seek agents invent pursuit, walls, and tool use that no reward function names.Thu, 13 Aug 2026 00:00:00 GMThttps://oropis.com/posts/nnue/https://oropis.com/posts/nnue/How a 41,024 × 256 × 32 × 1 fixed-point network, recomputed only where it changed, let a CPU engine run a 10-million-node search with a learned evaluation, and what it teaches about low-bit inference.Thu, 13 Aug 2026 00:00:00 GMThttps://oropis.com/posts/era-of-1-bit-llms/https://oropis.com/posts/era-of-1-bit-llms/Why a weight that can say 'a bit negative' might be the biggest efficiency lever in LLM inference, from BitNet b1.58 and Bitnet.cpp to the two-bit frontier.Tue, 28 Jul 2026 00:00:00 GMThttps://oropis.com/posts/tetranet/https://oropis.com/posts/tetranet/A symmetric quaternary weight grid {−1, −c, c, 1} with power-of-two intermediates reclaims the fourth state of a two-bit register while keeping the dot product multiply-free.Fri, 17 Jul 2026 00:00:00 GMT