Ggml: X2 speed for WASM by optimizing SIMD (github.com)

If I'm reading this correctly, Deepseek R1, after prompting, sped up its own web implementation by a factor of 2. Thus demonstrating that self-optimization is possible. Further optimizations seem likely to be findable.

already being discussed at https://news.ycombinator.com/item?id=42852866