💻 **AI Engineering Productivity Splits Into Three Tiers — From 20% Gains to 8x at the 'Software Factory' Frontier**

Tom Tunguz (Theory Ventures) maps the real spread of AI coding productivity gains — and it's not one number.
• Baseline tier: distributing an AI IDE with no other changes nets just **20-46%** gains (Google's RCT: 21%; GitHub: 24%).
• Frontier tier: companies building an operating layer around agents — Replit, NVIDIA, Amplitude, Anthropic — see **2.5-3x** more code shipped, with review time, bugs, and incidents flat.
• Factory tier: Nubank's use of Devin for large-scale refactoring hit **8x** efficiency and a 20x cost reduction.
_"The gap is not the model but the operating discipline built around it."_
For fintech and SaaS leaders benchmarking AI ROI, the takeaway is stark: model access alone buys single-digit gains — the multiples come from process redesign.

[🔗 Read more](/leaving?url=aHR0cHM6Ly90b210dW5ndXouY29tL2FpLWVuZ2luZWVyaW5nLXByb2R1Y3Rpdml0eS1hbnl0aGluZy1idXQtbm9ybWFsLw==)
[#tomtunguz](/search?q=%23tomtunguz)