来自 Meta 超级智能实验室的 Jason 讨论了理解 2025 年 AI 的三个基本理念：智能商品化、验证者法则和智能的锯齿边缘。
Jason from Meta Super Intelligence Labs discusses three fundamental ideas to navigate AI in 2025: intelligence as a commodity, verifier's law, and the jagged edge of intelligence.
要点 · TL;DR
智能将变成一种商品，成本趋近于零。 Intelligence will become a commodity with cost driven toward zero.
验证者定律：AI 进步取决于任务的可验证性。 Verifier's Law: AI progress depends on task verifiability.
AI 进步是锯齿状的，因任务和数据可用性而异。 AI improvement is jagged, varying by task and data availability.
核心观点 · Key points
智能将成为一种商品，成本趋近于零。 Intelligence will become a commodity with cost driven toward zero.
自适应算力允许为不同任务分配不同算力，降低简单任务的成本。 Adaptive compute enables varying compute per task, reducing cost for easy tasks.
验证者定律：在任务上训练 AI 的能力与该任务的可验证性成正比。 Verifier's Law: ability to train AI on a task is proportional to how easily verifiable it is.
易于验证的任务最终将被 AI 解决。 Easily verifiable tasks will eventually be solved by AI.
AI 进步是参差不齐的；不同任务以不同速度提升。 AI progress is jagged; different tasks improve at different rates.
数据丰富的数字任务中 AI 进步最快。 Digital tasks with abundant data see fastest AI improvement.
反共识 · Contrarian takes
快速超级智能起飞不太可能；改进是逐任务渐进的。 Fast superintelligence takeoff is unlikely; improvement is gradual per task.
自我改进是一个谱系，而非二元开关。 Self-improvement is a spectrum, not a binary switch.
AI 能解决人类因生物限制无法完成的任务，如从数百万张图像预测乳腺癌。 AI can solve tasks humans cannot due to biological limits, like predicting breast cancer from millions of images.
随着公共知识变得廉价，私人内幕信息变得更有价值。 Private insider information becomes more valuable as public knowledge becomes cheap.
一些任务如理发或约会将在很长时间内不受 AI 影响。 Some tasks like hairdressing or dating will remain untouched by AI for a long time.
通过提供答案或测试用例等特权信息可提高可验证性。 Verifiability can be improved by providing privileged information like answer keys or test cases.
本期章节 · Chapters（共 14）
介绍与日程Introduction and Schedule
智能作为商品Intelligence as a Commodity
验证者定律Verifier's Law
智能的锯齿边缘Jagged Edge of Intelligence
智能作为商品Intelligence as a commodity
验证不对称与验证者定律Asymmetry of verification and verifiers law
验证者定律与验证不对称Verifier's Law and Asymmetry of Verification