Engineering Blog
Evidence before
confidence.
Notes on Physical AI, autonomous-driving systems, and the engineering choices between a model and a moving vehicle.
7 published notesEngineering Case Study
Architecture, interfaces, validation, and the operational failure envelope.
Baidu Apollo 11.0: Open Source Is Valuable When It Closes an Operational Loop
Apollo 11.0 is moving from a collection of driving modules toward an operational system for delivery, sweeping, patrol, and shuttle vehicles. That shift changes what open-source maturity means.
Baidu Apollo 11.0: 오픈소스의 가치는 운영 루프를 닫을 때 생긴다Waymo’s Sixth-Generation Driver: Scaling Is a Hardware–Validation Co-Design Problem
Waymo’s 2026 move to fully autonomous operation with its sixth-generation Driver shows that reducing sensors is only defensible when fleet learning and validation preserve redundancy.
Waymo 6세대 Driver: 확장은 하드웨어와 검증의 공동 설계 문제다Technical Insight Memo
A compact position on a technical or industrial shift.
NVIDIA at CES 2025: The Autonomous Vehicle Is Becoming a Three-Computer System
NVIDIA’s CES message was not a single chip announcement. It was a systems argument: training, simulation, and in-vehicle inference must be designed as one closed engineering loop.
NVIDIA CES 2025: 자율주행차는 세 개의 컴퓨터로 재구성되고 있다DEEPX and DeepFusion AI: Korea’s Physical-AI Opportunity Is Between the Sensor and the Cloud
CES 2026 showed two complementary infrastructure bets: ultra-low-power edge inference and learned 4D-radar sensor fusion. Their real value appears when they are evaluated as one latency-and-power budget.
DEEPX와 DeepFusion AI: 한국 Physical AI의 기회는 센서와 클라우드 사이에 있다Deep Dive / Paper Reproduction
Models, silicon, and algorithms examined below the headline.
Alpamayo: Why an Autonomous Vehicle Needs a Reasoning Teacher, Not a Chatbot Driver
NVIDIA’s reasoning VLA is most useful as a teacher model for long-tail decisions. The engineering question is how to distill its explanations into a bounded, testable driving policy.
Alpamayo: 자율주행차에 필요한 것은 챗봇 운전자가 아니라 추론 교사다China’s Custom AV Silicon Is About Controlling the Iteration Loop, Not Winning a TOPS Contest
NIO’s NX9031, XPENG’s Turing chips, and Li Auto’s MAHE M100 reveal a broader strategy: align the model, compiler, vehicle OS, sensors, and product cadence under one architecture.
중국 자율주행 커스텀 실리콘의 목적은 TOPS 경쟁이 아니라 반복 루프의 통제다Build Log
A shorter field note on a system, prototype, or engineering question.