Blog

Months of building in Zig changed what I reach for first when I hit an AI engineering problem. The constraints turn out to rhyme.

I tried to make a code model smarter at inference time. Adaptive decoding, entropy thresholds, the works. It made things worse. Then self-distillation fixed it in four minutes.

Claude Code, Codex, and Cursor all handle context limits differently. I tested an alternative, pointer-based compaction, and it beats summarization by 18 points on grounding after two compactions.

I've tested every multi-agent orchestration setup. Conductor, Cursor, Claude Code Max. I was faster single-threaded. The cognitive science on this runs decades deep.

AI-generated code is flooding pipelines. Main branch success rates hit a five-year low. The Rust rewrite wave accidentally built the verification infrastructure agents need.

Most coverage focused on individual features. Nobody zoomed out to see what Anthropic did across the full week: absorb three categories of SaaS tooling into their own surface area.

There is now a permanent gap between who you are with AI and who you are without it. And nobody knows which one is real anymore.

Two weeks ago, security researchers found over 1,800 exposed OpenClaw instances. Every vulnerability maps to the same failure mode - a code path that didn't hit the check. Petri nets fix this.

There's a formalism from 1962 that solves a problem you hit every month. Nobody told you about it.

Testing engineers without the tools they use every day is a business liability. It actively selects against the future. Here is the fix.

A student revisited an old paper 'just for fun' and accidentally disproved a 40-year-old conjecture. I decoded the maths and built it - 4x faster than Zig's std.HashMap at 99% load.

It's not imposter syndrome. It's being over-leveraged in the code you shipped.

My AI yoga instructor had a 2000-token monolithic prompt and kept forgetting what it was doing. I split it into five agents. Each one got smaller, faster, and more reliable.

I built an AI yoga instructor. The smarter she got at finding existing sequences, the less she created new ones. That's a problem worth thinking about.

I built a platform where LLMs debate each other with custom personalities and adjustable spiciness. A Sassy Drag Queen vs a Noir Detective at Extra Hot is worth watching.

How I learned Go by building a real-time London Underground tracker, featuring concurrent polling, WebSockets, and lots of debugging.