- ai
- building
- devtools
Google Made Flash 4x Faster and Changed Our Latency Budget
A feature we shelved in September because 900ms TTFT felt dead came back off the shelf eight months later when Gemini 3.5 Flash hit 210ms in production.
A feature we shelved in September because 900ms TTFT felt dead came back off the shelf eight months later when Gemini 3.5 Flash hit 210ms in production.
Cursor 3 shipped the Agents Window and we immediately learned the hard way what parallel agents actually need: exclusion lists, worktrees, and honest accounting of what gets traded away.
Vendors silently roll model versions and your production output drifts. Lock the version, log the cost, ship a fallback. The playbook in TypeScript.
We migrated 47 functions off Lambda and our bill dropped 79%. Here is the math, the August 2025 INIT change nobody talked about, and why we are not coming back.
Three overlapping bugs caused four weeks of subtle Claude Code regressions that no standard test suite would have caught.
A Vercel employee clicked Allow All on an AI app's OAuth consent screen. Three weeks later, customer environment variables were on a hacker's drive with a two-million-dollar asking price.
For every product we ship, we kill two. The pattern in the ones that never made it out is almost always the same: we thought too hard before building.