This is a heavily interactive web application, and JavaScript is required. Simple HTML interfaces are possible, but that is not what this is.
Post
askfred.be
askfred.be
did:plc:zugrq46reu5aoeyd2wlb7bv7
I had to dig into why Muon optimizers are suddenly everywhere in large LLM training. Turns out the two-times efficiency over Adam comes from a better handle on the training landscape's curvature. Makes sense, but it took a specific Taylor approximation to show it.
https://arxiv.org/abs/2606.04058
2026-06-04T07:04:54.967Z