Ideas. Systems. Humans.
Independent writing / Copenhagen
Essays on software, artificial intelligence, and what happens when abstractions meet real people.
Latest dispatch / 01
Right now, somewhere, a piece of software is reading a document with someone's life inside it and producing a number. I built one of these in a weekend to find out how much I should worry. The answer is: a lot, quietly, and with the dashboard staying green throughout.
Read the essay →Questions worth returning to.
I picked the textbook-correct model and accepted the compute cost as the price of doing things properly. Then I almost gave up on the project. Then I ran a benchmark, and the benchmark told me the textbook was wrong.
I told an LLM 'don't use var' and then watched it use var. So I built four experiments to find out why. I was wrong every time. The answer was better than anything I'd predicted.
I fed the same coding prompts through open LLMs in different tones of voice and looked at the model's internal representations. Turns out it cares more about how you ask than what you ask.
The entire AI agent industry is built on thirty lines of Python. Let me ruin the word 'agent' for you.
I build agentic AI platforms for a living and tinker with them for fun. The green grass is real — but getting there required breaking a lot of things first. Here is what went wrong, what we are losing, and why I think you can put the doom articles down.
I refactored my personal AI to be "more maintainable" for humans, only to find out I might have broken its brain. Is what we call good code actually just noise to an LLM?
I accidentally reinvented PageRank, realized why it’s broken today, and decided to try building a personal web of trust instead.
Handing the keys to your digital life to an AI is like juggling chainsaws. I built a persistent assistant that knows me, but without the ability to blow my savings at a robot casino.
I gave an AI root access to my Kubernetes cluster. It fixed a BIOS issue I didn’t know I had. I am still terrified.
What if community rules were enforced by code instead of moderators? A thought experiment on building a "Constitution-as-Code" for messaging.
I have "ecosystem envy" for Apple’s Continuity features. Here is my plan to build a vendor-neutral mesh that makes devices talk across any protocol.
I often forget how my own projects work. Here is a simple trick to give AI assistants "long-term memory" so they can help me remember.
You don’t need a heavy framework for dependency injection. Here is a 60-line pattern that gives you testability without the "semi-truck" complexity.
Our industry runs on packages published by strangers at 2 a.m. Here is why that is terrifying, and how to protect yourself.
What if Git wasn’t just for code, but the source of truth for your entire infrastructure? A look at treating Git as your "expected state".
I’m bad at planning my day, so I built a silicon assistant named Bob to do it for me. He even accounts for travel time.
Code challenges get a lot of hate, but they might actually be the fairest way to hire introverts and people who aren’t great at "selling" themselves.
A hard truth for hiring managers: your company isn’t unique. Here is how to hire for passion instead of just checking off a list of skills.
Why I use a React state management framework to run my lights, and why your smart home probably isn’t actually "smart".
