one year on
DeepSeek R1 generates 99% of code for llama.cpp WebAssembly speedup PR
A single PR doubles WASM inference speed for llama.cpp, with the model writing nearly all of the code after a human provided prompts and tests.
A pull request in llama.cpp says it gives x2 speed for WASM by optimizing SIMD instructions — and the submitter says 99% of the code was written by DeepSeek R1.
Xuan-Son Nguyen, the PR author, wrote on January 27 that they only developed tests and wrote prompts. The model ran directly on chat.deepseek.com, spending three to five minutes “thinking” about each prompt. For one challenging quantization function, Nguyen asked R1 to invent a new approach without prior examples, resulting in code with a structure different from the rest of the codebase.
Simon Willison, who shared the PR, said he has also seen promising results from DeepSeek R1 for code. He compared it to OpenAI’s o1 on a plugin rewrite task and noted the R1 chain of thought eliminated a hardcoded model map, which was his preferred resolution. Paul Gauthier, creator of the AI coding tool aider, reported that aider now writes about 70% of new code in each release and that he has been shifting from Sonnet to DeepSeek V3.
On Hacker News, a thread drawing 979 points and 746 comments debated the significance of AI-generated code. Some expressed concern about maintainability when the model produces unorthodox implementations; others noted that access to the full R1 model remains expensive and API outages have been a barrier. Distilled versions are available on platforms like Groq and can run locally on high-end laptops.
The record
Reported that aider now writes about 70% of new code in each release, and that he has been shifting coding from Sonnet to DeepSeek V3; R1 API outages have hindered experimentation.
Reported that he is experimenting with DeepSeek-R1-Distill-Qwen-32B locally via Ollama, finding its chain of thought useful for refactoring even when the output code has mistakes.
Reported that the PR's approach for q6_K_q8_K was invented by DeepSeek without prior examples, noting the structure differs from the rest and may not pass code review in some organizations.
One year later — open only if you can handle spoilers
DeepSeek R1's coding capability was a preview of a broader pattern: by mid-2025, AI-generated code became the majority in many open-source projects, and the debate shifted from whether models could code to how to manage their output. The R1 line also influenced subsequent reasoning models from other labs, though its own API stability remained a recurring issue.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy