Xuan-Son Nguyen’s PR for llama.cpp has significantly sped up the performance of WebAssembly (WASM) by optimizing SIMD instructions. What’s surprising is that 99% of the code in the PR was actually written by DeepSeek-R1. Nguyen’s role was mainly to develop tests and write prompts, which were then run by R1 on chat.deepseek.com. The use of DeepSeek R1 led to promising results, with a recent example involving rewriting a plugin to imitate a cached model JSON pattern. During this process, R1 suggested eliminating the hardcoded model_map, dynamically building it from API responses instead. Ultimately, after much consideration, the model_map was eliminated, showcasing the clever insights of R1.
https://simonwillison.net/2025/Jan/27/llamacpp-pr/