Notes

The thing AI tools cannot give you is judgment. They can generate, but they cannot yet decide what is worth generating. That gap is where the work still lives, and it is not closing as fast as people think.

Most AI regulation is written by people who don't build. The EU AI Act treats foundation models as products, not infrastructure. Here is what the fine print actually means for developers shipping today, and what you can safely ignore.

The gap between frontier proprietary models and the best open-weight alternatives closed faster than anyone predicted. For most production use cases, you can now make an honest choice between the two. That changes the economics of the whole stack.

Four client engagements shipped in Q1 2026. No employees. One set of tools, used consistently. The tools have crossed a threshold: what used to require a team of five now requires one person who knows how to ask the right questions.

Every client wants flexibility. Every project benefits from constraints. Fixed-scope isn't a sales tactic: it is the only engagement model that reliably produces shipped software. Here is the argument from first principles and eighteen months of data.

The instruction-following jump between Claude 3.5 and Claude 4 is the most significant model improvement I have worked through. It is not smarter in the abstract. It is more reliably correct about what you actually asked. That distinction matters more than benchmark numbers.

Traditional product studios run four-to-six week discovery phases before writing a line of code. AI-native studios compress this to days. Not because they skip the thinking, but because the gap between thinking and prototype has collapsed. Discovery now happens in the artifact.

One-million-token context windows changed how I architect software. No more chunking strategies, no more retrieval pipelines for medium-sized corpora. You give the model everything, and it finds what it needs. This is underrated as an architectural shift.

Most AI agent products in production do not actually run autonomously. They run pipelines with a chatbot on top. The interesting work is in the narrow cases where the loop is genuinely closed. Here is how to tell the difference, and where the real frontier is.

After two years of constant tool churn, the AI development stack is settling. A small set of tools have won the daily workflow and the rest are losing ground fast. This is what the settled stack looks like from inside an active build practice, and why the churn is finally slowing.