AI 的最大瓶颈：能源、计算与 AGI 的未来

AI's Biggest Bottleneck: Energy, Compute, and the Future of AGI

格雷格·布罗克曼 Greg Brockman · Matthew Berman · 2025-10-08 · 约 44 分钟 · 原视频 ↗

本期速览 · Overview

OpenAI 的 Greg Brockman 讨论扩展挑战、计算的作用，以及 AGI 从终点到持续过程的演变。

OpenAI's Greg Brockman discusses scaling challenges, the role of compute, and the evolution of AGI from a destination to a continuous process.

算力是 AI 创新的根本驱动力和瓶颈。
Compute is the fundamental driver and bottleneck for AI innovation.
AGI 是一个持续的过程，而非单一终点。
AGI is a continuous journey, not a single destination.
能源和物理基础设施是最大瓶颈，而非算法。
Energy and physical infrastructure are the biggest bottlenecks, not algorithms.

算力是 AI 创新的根本驱动力，并将成为巨大瓶颈。
Compute is the fundamental driver of AI innovation and will become a massive bottleneck.
基础模型蕴含无限可能；后训练将其提炼为一致的行为。
Base models contain a universe of possibilities; post-training refines them into consistent behaviors.
AGI 是一个持续过程，而非终点；旅程与里程碑同样重要。
AGI is a continuous process, not a destination; the journey matters as much as the milestone.
AI 会改变许多工作，但也会创造新岗位；人际连接和品味仍不可替代。
AI will change many jobs but also create new ones; human connection and taste remain irreplaceable.
互联网体验正从静态网站转向由 AI 中介的动态交互。
The internet experience is shifting from static websites to dynamic, AI-mediated interactions.
能够长时间思考的主动式 AI 将变得更常见，用于解决难题。
Proactive AI that can think for long periods will become more common, solving hard problems.

语言模型可能拥有世界模型，尽管缺乏完整感官数据，空间推理能力证明了这一点。
Language models may have a world model despite lacking full sensory data, as shown by spatial reasoning.
非 GPU 架构的实际构建难度远超 2017 年的预期。
Non-GPU architectures have proven much harder to build than expected in 2017.
软件最终将完全实时生成，包括 UI，从头开始，无需遗留代码。
Software will eventually be fully generated in real time, including UI, from scratch without legacy code.
AI 进步的最大瓶颈是能源和物理基础设施，而非算法。
The biggest bottleneck for AI progress is energy and physical infrastructure, not algorithms.
基于广告的变现模式可能衰退，因为 AI 智能体代表用户浏览。
Advertising-based monetization may decline as AI agents browse on behalf of users.
OpenAI 内部的算力分配是一个痛苦且由人驱动的过程，并非完全自动化。
OpenAI's internal compute allocation is a painful, human-driven process, not fully automated.