Blog
Blog
Long-form essays on AI products, eval / benchmark methodology, and design.
-
Borrowed Eyes
Wiring a design-review eye onto a text-only model took a fraction of the time. The rest went into one question — is the borrowed eye fit to be a judge? Plus the three times I thought I had verified something and had verified nothing at all.
-
The Feeling Is Real, the Explanation Usually Isn't
Three days after DeepSeek Harness went open source, I sat down to write my first plugin. What drew me in was everything-is-a-plugin, because self-evolution needs one hard precondition: an agent has to have a self that can be modified. And when I said cold start takes about forty seconds, that was a real feeling. I timed it three times. Three seconds. The missing thirty-seven were my own waiting.
-
FDE: The Engineer Who Puts AI to Work
AI demos are easy now. The hard part is connecting one to a real business and owning what happens after launch. That is the gap FDEs fill.
-
Shipping an Agent Is 10× Harder Than the Demo
A convincing agent demo is easy. Production requires the system to verify outputs, constrain failures, and remain accountable after launch.
-
Beyond Five Archetypes, What Must PMs Decide?
In the AI era, PMs create value by making choices where answers are unclear and turning verifiable judgment into rules people and agents can inherit.
-
Loop Engineering Is Inference-Time RL
Loop Engineering is reinforcement learning moved to inference time with frozen weights. This mapping explains its design rules and failure modes.