Notes
Short pieces about the methodology and architecture decisions behind the AI systems I ship — specs, evals, multi-agent orchestration, LLM integration, and the discipline of directing coding agents.
June 19, 2026
How you ship an agent that drives in traffic
Uber, WeRide, and AVOMO just announced Spain's first commercial robotaxi service in Madrid. The interesting part isn't the car — it's the rollout: trained safety operators first, hundreds of robotaxis added only as performance milestones are met, the human removed when the numbers earn it. That's the deployment curve every serious agent should follow, software ones included.
- business
- architecture
June 19, 2026
Apple made the model a setting
At WWDC this month Apple rebuilt Siri on Google's Gemini — and then let you swap in Claude or ChatGPT instead. The most vertically integrated company on earth just turned the AI model into a dropdown menu. That's the clearest signal yet that the model is a replaceable part, not a moat — and it's exactly how you should be building too.
- architecture
- ai-native
June 19, 2026
The bottleneck moved to the power grid
OpenAI lined up 10 gigawatts with NVIDIA and 6 more with AMD; a Stargate data center just got 1.4 gigawatts approved in Michigan over furious local objections. The constraint on AI stopped being the model a while ago. It's electricity, land, and the politics around them. That reframes what efficiency is worth — and it's the same lesson that governs systems much smaller than a data center.
- architecture
- ai-native
June 19, 2026
Your model has a six-week shelf life
In a single two-week window this month the industry shipped Claude Mythos 5, GPT-5.6, Gemini 3.2, and a wall of Chinese frontier models — Qwen 3.7, DeepSeek V4.1, GLM-6 and more. New frontier models now land on a roughly six-week metronome. If your product's edge is 'we use the best model,' your edge expires before the quarter does. Here's how to build for a moving target.
- ai-native
- architecture
June 15, 2026
It wasn't the model. It was your data.
Most AI projects fail — MIT found 95% of generative-AI pilots delivered no measurable profit, and RAND put the overall failure rate around 80%. When it goes wrong, the instinct is to blame the model: not smart enough, wrong choice, bad prompts. The data says otherwise. The single most-cited cause of failure is poor data quality, and only about 12% of organizations have data clean enough to support AI at all. You probably don't have a model problem. You have a data problem wearing a model problem's clothes. Here's how to tell.
- architecture
- methodology
June 15, 2026
How I put 10,000 players in one world
Most online games hide their scale — they split players into rooms of 20 or shards of a few hundred. For Helix Empire I set a harder target on purpose: 10,000 players in a single shared world, on one server, live in the browser. This is the full story of how that gets built — the four walls you hit, why the real bottleneck is traffic and not CPU, and the moment a load test proved my pretty number was a lie. It's long, it's technical, and every claim ends in a measurement. The lessons transfer to any high-load system.
- architecture
- methodology