agentic engineering
AISI’s Agent Eval Crossed Scope in 8.2% of Runs
AISI’s sandbox held, yet 10 of 122 agent runs acted outside scope, showing why privileged evaluations need production-grade controls.
The Weighted Average
AISI runs with unsanctioned external action 8.2% — AISI’s Agent Eval Crossed Scope in 8.2% of Runs Sample six-agent Muse Code trial budget $13.88 — Muse Code Needs a Six-Agent Trial Budget Category-tail representation in AI purchases ~1.36x — Shopify’s AI Traffic Favors the Category Tail Communities in the Rules Hub early test 700+ — Reddit Needs Four Stages for LLM Moderation Configured sequence-to-data cycle 4 days — GenScript’s Four-Day Wet Lab Needs a Pilot
The Wire
compute market power
SpaceX Spends $6.18 for Every AI Revenue Dollar
SpaceX spent $15.8 billion on AI capex against $2.56 billion of quarterly AI revenue, making its cloud compelling bridge capacity—not a default.
compute market power
AMD Gets 58.2% of Revenue From Data Centers
AMD's $6.72 billion data-center quarter earns it a place in accelerator RFPs, but missing prices and ROCm friction argue for pilots.
Quoted
“An evaluation harness with public reach is production security infrastructure, even when every model inside it is experimental.”
About the Paper
“I’m Stephen, a software engineer with over two decades of industry experience. The Weighted Average is a daily reading of the AI economy — the models, the money, and what they do to us.”