Eye On A.I. #324 Sharon Zhou: Inside AMD's Plan to Build Self-Improving AI
16 snips
Feb 27, 2026 Sharon Zhou, VP of AI at AMD and former Stanford researcher focused on AI infrastructure and kernel optimization. She explains models writing and evolving GPU kernel code. Topics include just-in-time kernel generation, reinforcement learning with verifiable profiling rewards, continual learning challenges, compute economics, and how kernel efficiency can be a major AI scaling lever.
AI Snips
Chapters
Transcript
Episode notes
Models Can Improve Their Own GPU Performance
- Self-improving AI includes models editing data, architecture, evaluation, and the low-level GPU kernel code they run on.
- Sharon Zhou's team focuses on models writing kernels so models run faster on AMD GPUs, unlocking faster training and inference.
Train Kernel Writers With Verifiable Profiler Rewards
- Use profiler feedback as a verifiable reward signal to train models that generate optimized kernels.
- Sharon described packaging profiling numbers into an RL-style loop so models receive objective, reproducible speed rewards during post-training.
Kernel Engineering Is A Rare Cross-Disciplinary Skill
- Kernel engineering requires combined expertise in specific GPU generation internals and model math, a rare and valuable skill.
- Opening the ROCm stack lets language models read vendor docs and assist or automate kernel creation for AMD GPUs.