黄仁勋与李飞飞，让AI不止于“动嘴”

Core Viewpoint - NVIDIA's CEO Jensen Huang announced a historic shift from "Generative AI" to "Physical AI" and "Reasoning AI" during his keynote at CES 2026, marking a significant evolution in AI technology [1]. Group 1: Product Announcements - The Vera Rubin platform has entered full production, with delivery expected in the second half of 2026. Its inference performance is five times that of the previous Blackwell generation, and training performance is 3.5 times better, while the cost of generating AI tokens has been reduced to one-tenth of the previous cost [1]. - The Cosmos Physical AI platform was introduced, featuring foundational models like Reason2 and Predict2.5, which can learn physical laws through video and telemetry data, enabling AI to understand gravity, collisions, and physical properties of objects [3]. - The Alpamayo reasoning-based autonomous driving model was unveiled as the first model capable of "thinking," allowing it to logically reason through complex scenarios rather than relying solely on preset rules. The first vehicle equipped with this system, the Mercedes-Benz CLA, is set to hit the roads in the U.S. in Q1 2026 [3]. Group 2: Strategic Vision - NVIDIA aims to become the "Android" of general robotics by providing the Isaac platform and open-source models, thereby lowering the barriers to robot development. This signifies a pivotal moment for Physical AI, where embodied intelligence could become as ubiquitous as smartphones [5]. - The introduction of the Alpamayo model indicates a new pathway for autonomous driving, moving away from the reliance on vast amounts of data to a reasoning model that can handle unprecedented extreme situations, suggesting the potential for commercial viability in fully autonomous driving [5]. - Huang highlighted the narrowing gap between open-source and closed-source models, emphasizing NVIDIA's strategy to build a global open ecosystem powered by its chips through the open-source Alpamayo and Cosmos platforms [5]. Group 3: Comparison with Other Initiatives - There is a connection between NVIDIA's Physical AI and Fei-Fei Li's "World Model," both aiming for AI to evolve from a symbolic understanding of text to a three-dimensional understanding of the physical world. However, their paths and commercial ecosystems differ [6][7]. - Li's World Lab focuses on creating a model architecture for understanding the real world with both academic and practical purposes, while NVIDIA aims to establish an operating system for Physical AI [9]. - The competition for defining "Physical AI" exists, but there is also a complementary aspect in hardware, as NVIDIA possesses chips and open-source systems. The distinction lies in NVIDIA's engineering focus compared to Li's more abstract approach [9].