Thoughts on AI, machine learning, software engineering, and building things that matter.
I trained a small language model from scratch on consumer hardware and achieved 94% exact-match accuracy on CLI command generation. Here's what worked, what failed, and why data quality mattered more than architecture.