英伟达 CEO 黄仁勋探讨了 GPU 诞生的关键洞察、当前 AI 爆发的原因，以及他对未来一切移动之物都将实现机器人化的愿景。
Nvidia CEO Jensen Huang discusses the key insights that led to the GPU, the current AI explosion, and his vision for a future where everything that moves will be robotic.
要点 · TL;DR
英伟达在 AI 热潮前豪赌数十亿开发 CUDA，坚信并行计算将成为核心。 Nvidia bet billions on CUDA before AI boom, believing parallel computing would be essential.
AI 规模化有效：更多数据和算力让深度神经网络更智能。 AI scaling works: more data and compute make deep neural networks smarter.
未来十年是 AI 在生物、机器人及个人助手领域的应用科学时代。 Next decade is about applied AI in biology, robotics, and personal assistants.
核心观点 · Key points
并行处理是现代计算的核心，需结合顺序与并行能力。 Parallel processing is essential for modern computing, combining sequential and parallel capabilities.
深度神经网络可通过更多数据和算力扩展，实现广泛的 AI 应用。 Deep neural networks can scale with more data and compute, enabling broad AI applications.
AI 将通过提供个人导师和助手，让每个人都成为超人。 AI will make everyone superhuman by providing personal tutors and assistants.
具身 AI 需要基于物理模拟的世界模型，以确保安全学习。 Physical AI requires world models grounded in physics simulation for safe learning.
计算能效在 8 年内提升了一万倍，这对 AI 进步至关重要。 Energy efficiency in computing has improved 10,000x in 8 years, crucial for AI progress.
未来十年将聚焦 AI 在生物、机器人等领域的应用科学。 The next decade will focus on applied AI across industries like biology and robotics.
反共识 · Contrarian takes
英伟达在 AI 热潮前投入数百亿，基于信念而非即时回报。 Nvidia invested tens of billions before AI boom, based on belief not immediate returns.
GPU 架构应保持通用，而非针对 Transformer 等特定模型专用化。 GPU architecture should remain general, not specialized for specific AI models like Transformers.
AI 安全是工程问题，而不仅是研究挑战。 AI safety is an engineering problem, not just a research challenge.
机器人将在虚拟世界训练后再进入物理世界，颠覆传统顺序。 Robots will be trained in virtual worlds before entering physical ones, reversing traditional order.
AI 最大突破不是新算法，而是认识到 Scaling（规模扩张）有效。 The biggest AI breakthrough was not a new algorithm but the insight that scaling works.
未来 AI 将通过预测而非计算生成大部分像素，以提高效率。 Future AI will generate most pixels via prediction, not computation, for efficiency.
本期章节 · Chapters（共 18）
引言与访谈形式Introduction and Interview Format
CUDA 的起源与愿景CUDA's Origin and Vision
重塑计算堆栈Reinventing the Computing Stack
未来十年：AI 应用科学The Next 10 Years: Application Science of AI
机器人未来与个人 AIRobotic Future and Personal AI
AI 安全问题AI Safety Concerns
技术局限与能效Technological Limitations and Energy Efficiency
CUDA 的可及性与抽象Accessibility and Abstraction with CUDA
AI 架构演进核心信念Core Beliefs on AI Architecture Evolution
物理极限下设计芯片Designing Chips with Physical Limits
未来技术大赌注Big Bets on Future Technologies
预测未来与准备 AIPredicting the Future and Preparing for AI
GeForce RTX 50 系列与 AI 超分辨率GeForce RTX 50 Series and AI Super Resolution