The AI Architect
Articles about The AI Architect
STOP Trusting AI Leaderboards — Bench the Part You Actually Own
I ran Terminal-Bench 2.1 against my own serving stacks. GLM and Qwen swapped places, 19 points apart. The part that flipped the order is the part you own.
Before I write about this publicly
Hey — quick heads-up before anything goes out to the wider list. For the past year I’ve been building a private AI lab at home: dual NVIDIA nodes, near-frontier-class local models, benchmark cards published for every claim. It started as a personal project and got out of hand, as these things do. A few readers have asked if they could buy time with it, so I’m opening a few engagements before year-end: The model-downgrade eval — $5k flat. You send prompts and traces;
What It Costs to Run Your Own AI Now
The live October 2026 price ladder, from a used 3090 to a two-Spark cluster to the M5 Ultra, and the one physics rule that picks your rung.
My Agent Spent 32,524 Tokens Before It Even Read a File
I metered four agent harnesses in five configurations on four of my own jobs, one model held constant. The bill starts before the work does.
I Ran 60 AI Agents for a Full Day on Jev, the Model That Can’t Talk. It Cost 35 Cents
A day of silent agents on TypeSafe's Jev: 13,200 decisions, $0.35, and the prompt change that fixed the night shift.
I Cut My AI Coding Bill From $200 to $20 by Routing Jobs, Not Downgrading
The cheap-token era is ending. The method for deciding which jobs need a premium model, the tools that do the routing, and where to start.