ai safety security
OpenAI’s Astra Pause Makes Release Risk a Control
Astra’s possible one-tier cyber escalation makes dual-model validation and security evidence gates a procurement requirement.
The Weighted Average
Possible Astra capability escalation from High to Critical 1 tier — OpenAI’s Astra Pause Makes Release Risk a Control GW Ranch CO2e permit ceiling per nameplate GW-year ≈3.94 Mt/GW — Amazon’s 7.65 GW Shortcut Carries Permit Risk Spread across typical PowerPoint task credits 5× — ChatGPT PowerPoint’s Free Ride Is Over Work voice credit premium over Voice in Chat 20% — Grounded ChatGPT Voice Needs a Minute Budget Worst-case default wait for hourly agent task pickup <60 min — Veeva’s Workflow Agents Come With an Hourly Queue
The Wire
compute market power
Firmus Puts $2B of Equity Under a $10B Debt Bet
Firmus nearly doubled its valuation above $10.5B, but AI-cloud buyers should contract for energized capacity, not headlines.
compute market power
Lumilens Needs a Dual-Source Optics Qualification
Lumilens raised $700M and is shipping to a hyperscaler, but buyers need yield, watts/Tbps, and field-failure data before switching.
Quoted
“The safe frontier-model launch plan now has two models, one evidence gate, and no promise that capability outranks control.”
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.”