https://dustinmorgan.org/blog/Recent content in Blog on Dustin MorganHugoen-usThu, 22 Oct 2026 00:00:00 +0000https://dustinmorgan.org/blog/rebooting-this-site/Thu, 22 Oct 2026 00:00:00 +0000https://dustinmorgan.org/blog/rebooting-this-site/The old dustinmorgan.org was a three.js demo with a bio I wrote in one night. This is the replacement.https://dustinmorgan.org/blog/training-the-harness-instead-of-the-model/Tue, 20 Oct 2026 00:00:00 +0000https://dustinmorgan.org/blog/training-the-harness-instead-of-the-model/I fine-tuned dozens of models to make Agent Neo better at incidents, and none of them beat a better harness. So I started training the harness. It went from 5 to 10 out of 12 hard investigations.https://dustinmorgan.org/blog/coming-soon-cip-and-ctp/Mon, 28 Sep 2026 00:00:00 +0000https://dustinmorgan.org/blog/coming-soon-cip-and-ctp/Two things I’ve been building in private: running big open-weight models across the machines you already have, and training on them too.https://dustinmorgan.org/blog/what-i-got-wrong-about-sq4/Wed, 26 Aug 2026 00:00:00 +0000https://dustinmorgan.org/blog/what-i-got-wrong-about-sq4/I published a 4-bit quantization format and called its quality indistinguishable from FP16. Once I measured it properly, standard 4-bit formats beat it. I’ve retired it.https://dustinmorgan.org/blog/sq4-is-live-llama-cpp-pr-and-whitepaper/Wed, 13 May 2026 00:00:00 +0000https://dustinmorgan.org/blog/sq4-is-live-llama-cpp-pr-and-whitepaper/We published the SQ4 whitepaper and opened a PR to bring percentile-calibrated 4-bit quantization to llama.cpp.https://dustinmorgan.org/blog/agent-neo/Tue, 05 May 2026 00:00:00 +0000https://dustinmorgan.org/blog/agent-neo/Neo is an open-source Go agent that investigates infrastructure incidents with the CLIs you already have. It once filed 28 bug reports about its own ability to file duplicate bug reports.https://dustinmorgan.org/blog/v152-fused-inference-fp16-tokenizer/Fri, 01 May 2026 00:00:00 +0000https://dustinmorgan.org/blog/v152-fused-inference-fp16-tokenizer/Inference +28%. Training +54% at dim≥1024. The tokenizer was O(n²) and nobody noticed until we fed it 48MB.https://dustinmorgan.org/blog/benchmarks-i-owe-pytorch-a-fair-fight/Mon, 27 Apr 2026 00:00:00 +0000https://dustinmorgan.org/blog/benchmarks-i-owe-pytorch-a-fair-fight/The 1.54x number has an asterisk. I ran PyTorch with default settings. Here’s what I still owe.https://dustinmorgan.org/blog/why-i-wrote-a-gpu-training-engine-in-go/Mon, 20 Apr 2026 00:00:00 +0000https://dustinmorgan.org/blog/why-i-wrote-a-gpu-training-engine-in-go/I got tired of Python’s toolchain for fine-tuning. So I wrote my own — and accidentally built something faster.