Thoughts on software engineering, AI, side projects, and the things I learn along the way.
A small narrative write-up of the first mempool experiment: training a tiny Qwen3 router to choose between larger worker models, publishing the first dataset/model artifacts, and writing down what still feels brittle.
Small language models running on consumer hardware as Zero Cost Intelligence, and how a system of them could handle non-realtime tasks.
I built a computer-use agent by splitting it in two: a small reasoning model that plans, and a vision-grounding model that turns 'click the blue Sign In button' into pixel coordinates. Cheaper, faster, and weirdly more reliable than a single big multimodal call.
A rough guide to what NATs actually mean in LLM training loss, because I kept seeing this term everywhere and finally decided to understand it.
I spent a day fine-tuning Qwen3-VL-2B for PDF-to-markdown conversion using SFT + GRPO. Total cost: $4. Here's what worked, what didn't, and why GRPO alone fails on vision models.
Three weeks running Claude 24/7 taught me how to make coding agents actually work: verification loops, team standards, and the right tooling setup.
Scattered thoughts on the nature of trust in LLMs, how context gives meaning, and language as a facilitator of senses.
Striving for the best in the age of Gen AI, and preparing for the unknown future.
Notes on the DeepSeek-R1 paper — how pure reinforcement learning with GRPO enables emergent reasoning in LLMs without supervised data.