@lizecheng

Pydantic Monty: Why AI Agents Need a Rust-Written Python VM Instead of Docker

Mar 19 · 8 min read · Pydantic Monty is a minimal, secure Python bytecode VM written in Rust, designed specifically for executing AI-generated code in agent workflows — with a 6-microsecond cold start versus Docker's 195ms, a zero-access security model, and no CPython dep...

How Google Sashiko Detects Linux Kernel Bugs That Humans Miss: A 5-Stage Multi-Agent Code Review Architecture

Mar 18 · 8 min read · Originally published at lizecheng.net Google's Sashiko is a multi-agent AI system that detected 53% of real Linux kernel bugs — every single one of which had passed complete expert human review before merging. That's not a marginal improvement over h...

Claude Code vs. Cursor vs. GitHub Copilot: The Agentic Architecture Difference That Actually Explains the Benchmark Gap

Mar 16 · 8 min read · Claude Code is a terminal-first AI coding agent that outperforms IDE-embedded tools not because it uses a better model, but because its architecture solves a fundamentally different problem. Originally published at lizecheng.net Eight months after l...

NVIDIA Vera Rubin NVL72: What 10x Lower Inference Cost Per Token Actually Means

Mar 15 · 8 min read · Originally published at lizecheng.net NVIDIA's Vera Rubin platform achieves 10x lower inference cost per token versus Blackwell through four simultaneous architectural shifts — NVFP4 native compute, HBM4's 2.75x memory bandwidth leap, NVLink 6 doubli...

Enterprise AI System Security: The Attack Vectors That Standard Scanners Miss

Mar 15 · 8 min read · Originally published at lizecheng.net Enterprise AI systems introduce a class of vulnerabilities that traditional AppSec tooling was not designed to find — and the McKinsey Lilli breach on February 28, 2026 is the clearest case study we have. An auto...