Current writing
Notes on frontend work, AI coding tools, and the part where taste has to survive implementation.
Three tests, five models, zero retries. Meta crushed vision and analysis but face-planted on code. ChatGPT wrote the worst code and produced the best output. Meta wins overall — barely.
Google’s latest open-weights model versus Claude Code, head to head on a simple coding task. One took 2 hours. The other took 2 minutes.
I queried 47.5 million items from the complete Hacker News archive to find out what the community really cares about, when to post, and who the power users are.
An interactive cellular automaton — draw a pattern, press play, and watch complexity emerge from four simple rules.
A physics playground disguised as a page of prose. Every letter is a rigid body — click one and watch the words come apart.
A working AI music platform — type a prompt, get a full song with vocals, lyrics, and cover art. One person, open-source models.
In 2023 I fine-tuned GPT-2 355M to write image prompts. Now I retrained the concept with Qwen3-0.6B on Claude-generated data. Same size class, dramatically different results.
I generated 230 AI songs across 10 genres, built a model to rate them, and used the data to 4x the hit rate of producing good tracks. All on a MacBook.
I gave Claude Code and Codex the exact same prompt and deployed whatever came out. No edits, no re-rolls.
20,000 lines of code in 12 hours. It worked. It was also terrible. Here’s what I learned about the gap between “functional” and “shippable.”
Fine-tuned a 355M parameter model on thousands of high-quality Midjourney and Stable Diffusion prompts. Give it a few words, it gives you a masterpiece description.