Hi 👋

I’m a PhD student in Computer Science at Stanford, advised by the amazing Chris Ré as part of Hazy Research and the Stanford AI Lab. I’m also an ML researcher at Cursor, where I am building Composer and Grok.

My research focuses on ML systems. I design hardware-aware algorithms and abstractions to accelerate AI model training and inference. Some of my work includes megakernels, efficient MXFP8/NVFP4 training, GPU networking, and ThunderKittens.

Prior to Stanford and Cursor, I co-founded Blux, an AI B2B startup specializing in empowering Korean e-commerce through AI-driven personalization. Blux raised $3M+ and is personalizing over 10 million Korean users’ online journeys monthly.

In addition, I am an electric guitarist and composer. I have released two albums and performed as lead guitarist and producer for multiple rock bands since 2008.

For more on my background, see my LinkedIn.

Writing

• Mixture-of-Kittens: our open-source MoE megakernel for NVL72s (Aug 2026)
• ThunderKittens 2.0: Even Faster Kernels for Your GPUs (Feb 2026)
• Loads and Loads of Fluffy Kittens (Nov 2025)
• ParallelKittens: Simple and Fast Multi-GPU AI Kernels (Nov 2025)
• We Bought the Whole GPU, So We’re Damn Well Going to Use the Whole GPU (Sep 2025)
• How Many Llamas Can Dance in the Span of a Kernel? (Sep 2025)
• One Kernel for All Your GPUs (Sep 2025)
• 1.5x Faster MoE Training with Custom MXFP8 Kernels Built From Scratch (Aug 2025)
• Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B (May 2025)

Open Source

• Mixture-of-Kittens (New): Mixture-of-experts training megakernel for NVL72s.
• ThunderKittens/ParallelKittens (3,600+ stars): Tile primitives for fast AI kernels. Adopted by Cursor, Together AI, Jump Trading, Modular, TileLang, and Nvidia CuTe 4.0.
• Megakernels (700+ stars): Framework for running an entire LLM as a single fused GPU kernel (“megakernel”). Includes implementations of FlashMLA, Llama 3 1B, and 8-GPU Llama 3 70B.
• GPU Experiments: A collection of GPU experiments and benchmarks that I use for my personal understanding and research.
• MERCI: Fast embedding reduction algorithm for deep learning recommendation models (DLRMs) and other systems with very large embedding tables.
• ELF32 Dynamic Linker for Raspberry Pi: ELF dynamic linker on bare metal, allowing you to port shared libraries. It is taught as a lab at Stanford: CS 240LX: ELF and Dynamic Linker.
• Co-Chuck: Collaborative online environment that allows you to “code” your music.