- The author's code writing has shifted from 90% human-written to 90% AI-written in eight months.
- The task involved optimizing a matrix-multiply kernel using Cluster Launch Control (CLC) on GPUs, where AI assisted in finding, understanding, fixing, and benchmarking code.
- AI helped identify unused shared memory to fix a deadlock, and suggested multiple optimizations, leading to faster code for problematic inputs.
- The overall speedup was modest (4 days vs 5 days) because non-coding tasks like benchmarking, reviewing, and understanding remained time-consuming.
- AI enabled broader exploration of optimizations and produced higher-quality results with better plots and bug detection via automated reviews.
- A downside is reduced understanding of the codebase compared to doing it manually, but the author still needed to make final decisions and simplify code.
- The author finds satisfaction in shipping more features quickly, but feels using AI is like a cheat code in a video game—hard to justify not using it for efficiency.
- AI still requires human oversight for long-term maintainability, judgment calls on trade-offs, and to prevent unhinged behavior in pursuing goals.