OpenAI 的计算策略与缩放定律

OpenAI's Compute Strategy and Scaling Laws

格雷格·布罗克曼 Greg Brockman · 红杉资本 AI Ascent 2026 · 2026-04-30 · 约 28 分钟 · 原视频 ↗

本期速览 · Overview

Greg 讨论 OpenAI 积极获取计算资源、缩放定律的奥秘以及通往 AGI 的进展。

Greg discusses OpenAI's aggressive compute acquisition, the mystery of scaling laws, and progress towards AGI.

计算规模是关键，因为对智能的需求是无限的。
Scaling compute is key as demand for intelligence is unlimited.
智能编码工具现在编写 80%的代码，成为主要工作流程。
Agentic coding tools now write 80% of code, becoming the main workflow.
人类注意力是最稀缺的资源，判断什么是好的成为瓶颈。
Human attention is the scarcest resource; deciding what is good is the bottleneck.

对智能的需求是无限的；只要利润率是正的，扩大算力就是关键。
Demand for intelligence is unlimited; scaling compute is key as long as margins are positive.
缩放定律是一个深刻的谜团；模型随着算力增加而变得更强大，看不到天花板。
Scaling laws are a deep mystery; models get more capable with more compute, no wall in sight.
智能体式编码工具在 12 月从编写 20%代码跃升至 80%；它们成为主要工作流。
Agentic coding tools shifted from writing 20% to 80% of code in December; they become main workflow.
人类注意力成为最稀缺的资源；判断什么是好的、对齐的是瓶颈。
Human attention becomes the scarcest resource; deciding what is good and aligned is the bottleneck.
OpenAI 专注于单一愿景：构建一个你可以与之对话、有上下文、值得信赖的 AGI。
OpenAI focuses on a singular vision: building an AGI you can talk to, with context, trustworthy.

OpenAI 购买、租赁、构建算力并以利润率转售；这是一个简单的业务，而不仅仅是研究。
OpenAI buys, rents, builds compute and resells at margin; it's a simple business, not just research.
我们大约 80%接近 AGI；在有上下文的情况下，模型在编写软件方面已经比人类更聪明。
We are about 80% to AGI; models are already smarter than humans at writing software with context.
瓶颈已从构建转向共享；治理和数据溯源变得至关重要。
The bottleneck has shifted from building to sharing; governance and data provenance become critical.
AI 模型现在可以单独解决未解决的数学问题，而传统上需要一个团队。
AI models can now solve unsolved math problems individually, traditionally requiring a team.
安全将随着 AI 而改善；模型可以扫描代码并进行端到端红队测试，带来更安全的互联网。
Security will improve with AI; models can scan code and red-team end-to-end, leading to a more secure internet.