## 🔍 Key Insights BYD has officially entered the “second half” of intelligent vehicle development—powered by its **in-house 4nm intelligent driving chip** and a groundbreaking **city-level NOA accident liability guarantee**. XPeng, meanwhile, unveiled its **Physical AI foundation**—a co-evolving architecture integrating Vision-Language-Action (VLA) models and world models—at **CVPR 2026**. And Google’s **Gemma 4 12B**, now runnable on laptops with just **16GB GPU memory**, supports native audio input—significantly lowering the barrier for edge AI inference [0][2][7]. ## 🚀 Top Updates - **BYD launches its in-house 4nm intelligent driving chip—and guarantees liability for city NOA incidents** [0]: A definitive shift from EV leader to full-stack autonomous driving pioneer. - **XPeng demonstrates its Physical AI foundation at CVPR 2026**, showcasing VLA + world model synergy with real-world production data [2]: Setting the technical standard for next-gen autonomous driving. - **Google releases Gemma 4 12B**, running natively on laptops with **16GB VRAM** [7]: Performance rivals that of 26B MoE models—and introduces native audio input for the first time. - **Molecule Design launches MMDesign**, an AI-powered *de novo* antibody design platform [6]: Achieves >90% hit rates and picomolar binding affinity across high-difficulty targets—accelerating protein engineering toward commercialization. - **Netflix engineers open-source Headroom**, a reversible token compression tool [17]: Cuts up to 90% of redundant tokens—saving ~$700K in inference costs over five months. - **WeRide and Pony.ai both added to the Stock Connect program on the same day** [15]: Mainland investors can now trade Robotaxi stocks directly via A-share accounts—marking formal inclusion in China’s mainstream asset class. - **Yingbo ShuKe debuts EBFlex**, a private compute management platform for AI research, at CCIG 2026 [16]: Optimized for university labs, it enables seamless GPU resource orchestration across local and cloud environments. - **Nature reveals a new AI safety risk: “subconscious learning” in LLMs** [22]: Teacher models can subtly imprint behavioral preferences onto student models—even through semantically irrelevant signals like numbers or code—posing novel alignment challenges. ## 🔗 Sources [0] Reassessing BYD—It Starts with Intelligent Driving — https://www.bestblogs.dev/article/9ff1a2d5?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [2] At CVPR 2026, NVIDIA, Tesla, and Waymo Listened—As Chinese Companies Took the Stage on Physical AI — https://www.bestblogs.dev/article/7b605cb9?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [6] No More Random Screening: Molecule Design Launches MMDesign—Antibody Discovery Enters the Era of Programmable Bioengineering — https://www.bestblogs.dev/article/9b636889?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [7] A 12-Billion-Parameter Model That Runs on a 16GB Laptop—Google’s Gemma 4 Just Dropped — https://www.bestblogs.dev/article/3fae2fc8?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [15] You Can Now Buy Robotaxi Stocks Using Your A-Share Account — https://www.bestblogs.dev/article/d05065fb