Engineering notes from the field.
Essays and field notes on AI, software engineering, design, and the craft of building product teams that ship. Written by the engineers doing the work.
The First Frontier Model Built and Withheld: Reading Anthropic's ASL-4 Call
Anthropic announced a frontier model on April 7 and in the same breath said the public couldn't have it. The first ASL-4 trigger in industry history, and a working preview of how frontier AI deployment actually changes from here.
Opus 4.7, /ultrareview, and the New Floor for AI-Integrated Products
Anthropic shipped a coding model that resolves 87.6% of real GitHub issues and a review command that changes what a production PR pipeline should look like. The bar moved again.
Verifiable Rewards Didn't Kill the Expert-in-the-Loop. They Sharpened It.
RLAIF costs less than a cent per data point. Human feedback costs a dollar. The frontier labs are still spending a billion a year on humans. Here is why the math works.
The Three-Model Routing Playbook: How Product Teams Actually Ship With AI in Q2 2026
The winning pattern isn't picking the best model. It's routing the right workload to each of three models — and having an eval harness honest enough to catch when that changes.
HappyHorse and the Shift Nobody Saw Coming in AI Video
Alibaba quietly took the frontier video-generation crown the same month OpenAI killed Sora. That isn't a leaderboard story — it's a supply-chain one.
Canva's AI 2.0 and the Intent-to-Artifact Era
The headline was conversational design. The actual news is the cost structure — and what it says about every AI feature on every 2026 roadmap.