Writing

Notes from building with AI

What works, what breaks, and the evidence behind it.

Written by Samuel Nnaeto, who builds AI automations and agents for operations teams. More about me →

• 
  Jev scored 96.9% on my agent's commands and still missed the 2 it should have blocked

  28 October 2026 · 9 min read

  I tested Jev, TypeSafe AI's fast decision model, on 256 real agent commands. It scored 96.9% and still let through both that should have been blocked.

• 
  I Just Wanted a Man to Drink Water: What AI Video Still Gets Wrong

  12 October 2026 · 12 min read

  I asked AI video models for 15 seconds of a man drinking water. Faces drifted, the kitchen changed between shots and water ran through the glass. What still goes wrong, with the research.