Notes from the boundary of math and AI.
Essays by Lifan Hu on model architecture, production judgment, topology, and the connections between research and engineering.
- We Are Too Impatient to Hack — Why modern hackathons often confuse rapid assembly with hacking—and why the real thing begins with obsession, uncertainty, and a willingness to stay stuck.
- Kimi K3 — Open Source Miracle? — A close look at Kimi K3’s mixture-of-experts design, delta attention, and the practical meaning of openness at trillion-parameter scale.
- On the Epistemology of Production — What three technical rounds at TikTok clarified about judgment, uncertainty, and the difference between knowing a model and operating a system.
- How DeepSeek-V3 Made a 671B Model Surprisingly Affordable — Four engineering ideas behind DeepSeek-V3’s efficiency: latent attention, expert routing, pipeline overlap, and disaggregated inference.
- From Open Covers to Open Source — How algebraic topology becomes a practical way to think about invariants, refactoring, manifolds, and systems that must change without breaking.