Writing
Notes on evaluation harnesses, reliability, and the measurements that show whether a change to an AI system actually helped.
-
My Thoughts on Graph Engineering
Four frameworks converged on graph execution years before "graph engineering" existed as a term — and their own engineers argued against the diagram.
-
The American Economy Is Leveraged to a Leaderboard
Kimi K3 didn't erase America's AI lead. It priced how much of current US growth rests on seven stocks and one story holding.
-
Google Built a Faster AI by Making It Worse. When Is That a Good Trade?
DiffusionGemma, dFlash and TurboQuant get sold as one 6x speedup. They're three different tradeoffs, and collapsing them hides the one that matters.
-
My Thoughts on Loop Engineering
Loop engineering is a new label for an old bottleneck. The generator was never the hard part to build — the verifier always was.
-
DeepSeek Killed the Commodity Tier, Not the Frontier
Cheap open-weight models killed single-shot chat pricing. Multi-step agentic reliability is a different market, priced on different logic.
-
Claude Code Dynamic Workflows: What's Actually Changed
What Dynamic Workflows actually changes under the hood, what it costs in tokens, and the tasks it's the wrong tool for.