机器学习研究员 Nathan Lambert 分享其走访中国顶尖 AI 实验室的一手见闻，探讨文化组织差异、算力限制与开源生态。
Machine learning researcher Nathan Lambert shares firsthand insights from his trip to China's leading AI labs, discussing cultural and organizational differences, compute constraints, and the open-source ecosystem.
要点 · TL;DR
中国 AI 实验室擅长工程执行，而非范式发明。 Chinese AI labs excel at engineering execution, not paradigm invention.
算力限制将随时间扩大中美 AI 差距。 Compute constraints will widen the US-China AI gap over time.
开放权重模型帮助中国实验室在美国获得影响力。 Open-weight models help Chinese labs gain influence in the US.
核心观点 · Key points
中国 AI 实验室专注于精细的工程和执行，而非范式创新。 Chinese AI labs focus on meticulous engineering and execution, not paradigm invention.
开放权重模型帮助中国实验室在美国市场获得影响力。 Open-weight models help Chinese labs gain influence in the US market.
预训练的缩放定律仍然成立，但算力限制将扩大中美差距。 Pre-training scaling laws still hold, but compute constraints will widen the US-China gap.
蒸馏是灰色地带；并非所有 API 使用都是知识产权盗窃。 Distillation is a gray area; not all API use is IP theft.
中国实验室在基准测试上落后 6-9 个月，但在知识工作上的差距可能扩大。 Chinese labs are 6-9 months behind on benchmarks, but gap may grow on knowledge work.
反共识 · Contrarian takes
中国实验室并未受到政府大量补贴；支持是间接且竞争性的。 Chinese labs are not heavily government-subsidized; support is indirect and competitive.
由于路径依赖，DeepSeek 并非其他中国实验室的基础层。 DeepSeek is not a base layer for other Chinese labs due to path dependency.
中国研究人员使用 Claude 而非 Codex，与西方媒体叙事相悖。 Chinese researchers use Claude, not Codex, contradicting Western media narratives.
华为芯片可用于推理但非训练；实验室购买但未使用。 Huawei chips work for inference but not training; labs buy them but don't use.
由于闭源软件的盈利能力，开放模型长期不可持续。 Open models are unsustainable long-term due to closed software profitability.
本期章节 · Chapters（共 16）
引言与嘉宾背景Introduction and Guest Background
建立关系与尊重的重要性Building relationships and the importance of respect
中美研究文化差异Differences in Research Culture Between China and US
东亚教育与文化变迁Education and Cultural Shift in East Asia
中国 AI 生态与专业化China's AI Ecosystem and Specialization
中美国有化与开放模型Nationalization and Open Models in China vs US
中国主要 AI 实验室点评High-Level Commentary on Major Chinese Labs
中国 AI 公司访问与文化Chinese AI company visits and culture
市场整合与开放封闭模型Market consolidation and open vs closed models
蒸馏与性能差距Distillation and Performance Gap
中美算力与数据限制Compute and Data Constraints in China vs US
中国 AI 芯片格局：英伟达 vs 华为China's AI chip landscape and Nvidia vs Huawei
中国 AI 变现：SaaS 与推理支出Monetizing AI in China: SaaS vs inference spend
政府角色与补贴Government roles and subsidies in Chinese AI