Blake Ledden
AI engineer with a focus on systems who owns the stack end-to-end, from the GPU kernel to the eval.
Day to day: GPU kernels and compiler backends (CUDA / ROCm / Metal), inference servers, multi-model orchestration, model training and fine-tuning, and evals. I publish what I measure, including what didn't work.
GPU Kernels & Compilers
- Writing PTX for Consumer Blackwell. 280 hand-written PTX kernels. FP8 flash attention from 30 to 108 TFLOPS via a transposed-V layout, and where hand-written PTX loses to the compiler.
- Four Months Upstreaming Consumer Blackwell. SM120/SM121 enabled across flash-attention, FlashInfer, vLLM, CUTLASS, Triton, and SGLang, with merged forward/backward/varlen attention kernels and the dispatch fixes under them.
- Same Source, Same Bytes.
A Metal backend for OpenAI Triton with a correct-or-refuse contract. The same
@triton.jitsource, verified byte-identical against NVIDIA. - tridec. One Triton kernel source for quantum-LDPC decoders, validated on NVIDIA (CUDA), AMD (ROCm), and Apple silicon (Metal). On PyPI.
- 341× Faster. Optimizing a Rust vector database from 52 to 17,746 QPS, top-7 on ANN-Benchmarks at 0.999 recall.
LLM Training Dynamics
A series of multi-seed experiments on the Tinker platform: Noisy Student for LLMs (consensus self-training, +3.9pp GSM8K), The Distribution Cliff (off-policy distillation collapse), Open Character Training (+39% alignment), SL vs RL efficiency (an information-theoretic gap, validated), Constitutional AI from base models, and GAN-style joke training (when metrics lie), all synthesized in Seven Patterns from 300+ Evaluation Runs.
Tools
- Recall. A local, cross-session memory plugin for Claude Code, measured at about 1% of my token budget across 470,000 requests.
Background
Previously 4+ years at Apple on authentication and developer tooling.