nguyen's herbarium

herbariumn.latinherba
a collection of dried plants, pressed and named.

notes

  1. Optimizing GEMM on Apple's M5 from scratch, Part 1: Basics47 min · systemnearly slop
  2. Optimizing GEMM on Apple's M5 from scratch, Part 2: Outperforming MPS's GEMM (in big 2026)26 min · systemnearly slop
  3. A prologue to protein structure prediction20 min · biologyslop
  4. Note on Sutton & Barto — Chapter 3: Finite Markov Decision Process (MDP)21 min · reinforcement learningnearly slop
  5. A Little Bit About "Numbers" and "Mathematics"13 min · mathematicsnearly slop
  6. Residual Stream is Key to Transformer Interpretability17 min · llmnearly slop

posts

  1. Why I want to learn to code again?8 minnonslop
  2. Một chút về Transformers (Phần 2): Transformers11 minnonslop
  3. Một chút về Transformers (Phần 1): Attention, Attention, Attention12 minnonslop
  4. Về một vài câu hỏi mà mình gặp khi interview C++10 minnonslop

research

  1. A Rank-One Representation Edit Recovers Much of RMU’s WMDP Accuracy Loss28 min · mechanistic interpretabilityslop

papers

  1. Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention6 notes · Katharopoulos et al., 2020

projects

  1. comtamc++20, metalA tiny eager-mode deep learning framework in C++20 and Metal for Apple GPUs
  2. cuda-labcuda, pytorchA minimal CUDA learning lab for correctness, roofline reasoning, and LLM inference kernels
  3. banhxeotriton, pytorchA tiny deep learning compiler that compiles to Triton kernels
  4. pho-codeagent harnessA small custom agent harness based on Pi SDK and built with 100% vibing
  5. bami-hubAn "always free" learning and sharing hub for AI community in Vietnam