models open source
Meta's 30B Agent Weights Fit a 24GB GPU
Muse Glimmer's 4-bit weights fit under 20GB for 24GB hardware; benchmark gaps make it a pilot, not a cloud replacement.
The Weighted Average
Minimum weight-footprint reduction from full precision to 4-bit ≥63.6% — Meta's 30B Agent Weights Fit a 24GB GPU Expected net proceeds from Intel's priced equity offering $19.7B — Intel's $20B Raise Is a Capacity Test JetBrains AI development-expense increase in six months 10× — JetBrains' 10× AI Bill Needs a Traffic Cop Anthropic-reported lower bound for zeta zeros on the critical line 67.2% — Claude's Math Result Needs a Reproducibility Gate Proposed CISA cut as a share of the SLCGP's four-year authorization 70.7% — California Puts an AI Cyber Officer in Every Agency
The Wire
agentic engineering
Claude Code Auto Mode Needs Hard Denies
Auto mode caught 6.5× as many planted dangerous commands as people. Teams should replace prompt theater with explicit policy.
enterprise ai work
OfficeQA Makes the Harness the Enterprise Moat
Databricks lifted matched-model accuracy from 37.5% to 52.8%. Document-agent buyers should test parsing and orchestration together.
Quoted
“Local inference is no longer a toy demo; it is a procurement option with a memory budget.”
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.”