I think Jiayi Weng already dropped some hints about heuristic learning (HL) during his previous podcast. To me, the logic behind HL feels very practical and down-to-earth. After all, it's the distilled experience of an AI infra expert from OpenAI.
Theoretically speaking, it elevates the meta-learner from the algorithmic level to a higher abstraction layer, where the agent itself becomes the module.
Siegel upper space = complex symmetric matrices with positive-definite imaginary part
Siegel upper space or disk space (generalizes Poincaré disk) via Cayley transform is a non-positive curvature (NPC) complex matrix manifold
👉https://t.co/CdAOsgVaAn
Hugely impressive patchset from "mkver" porting old MMX code to hand written SSE2 assembly, making sure old platforms like 32-bit still work.
Up to 10x faster than C.
https://t.co/B5SZJNf0oN
Here’s the Haskell code that falls out of my magical mystery tour through pseudofunctors, the 2-Grothendieck construction and homotopy categories. Reckon I deserve a slow clap