I think about AI a lot. Probably too much. The ideas here are mine. The writing involves a language model and a lot of arguing with it.
March's agent launches made the shift obvious: once an LLM can read Jira, browse links, and open pull requests, the real engineering problem is no longer prompt wording. It's blast radius.
A step-by-step trace through the training pipeline that turns a punctuation mark into a writing tic. Tokenisers, reward models, and the labellers who accidentally taught GPT to love polish.
AI outages reveal how much thinking we've outsourced - and why resilience means using the speed without losing the skill.
I'm trying to think through what it means to build with AI. Not the hype, not the panic. The quieter questions: what happens to your skills when you stop practising them, why models do strange things, whether any of this is making software better. I don't have most of the answers. But I'm working on it.
Built with Astro, caffeine, and approximately 14 million tokens of back-and-forth.