2月，Anthropic和OpenAI表示，领先的中国实验室使用蒸馏来复制美国模型；Anthropic具体指称DeepSeek、Moonshot和MiniMax创建了24,000个欺诈账户，并进行了超过1,600万次模型交互，约合每个账户667次交互。这些指控延续了OpenAI去年年初的指责，并将模型模仿界定为对高价值输出的大规模提取，而非孤立的滥用。
成本不对称是核心：美国实验室称单次前沿训练运行就可能花费数十亿美元，而摩根大通估计到2030年数据中心投资将达5万亿美元，但中国竞争者仅以数周延迟、并以训练和推理成本的一小部分，就在部分基准上达到可比表现。随着前沿训练从网络规模文本摄取转向计算密集的试错式强化，蒸馏的统计重要性上升，使合成数据收割成为受芯片约束企业更快的路径。
趋势信号指向加速：Anthropic和Google DeepMind都报告，过去一年蒸馏尝试变得更常见，策略包括流量混淆和将任务拆分到数千个账户上，以及关于中国“小作坊产业”协调此类活动的报告。在商业层面，语境中被引用的MIT研究者称，自DeepSeek发布R1以来，中国在开源模型市场的份额快速增长并已超过美国，而尽管存在云访问禁令和更严格芯片管制等选项，近期执法前景仍不确定。
In February, Anthropic and OpenAI said leading Chinese labs used distillation to copy U.S. models; Anthropic specifically alleged that DeepSeek, Moonshot, and MiniMax created 24,000 fraudulent accounts and made over 16 million model interactions, roughly 667 interactions per account. The claims extend OpenAI’s accusation from early last year and frame model-mimicry as large-scale extraction of high-value outputs rather than isolated misuse.
The cost asymmetry is central: U.S. labs report individual frontier training runs can cost billions of dollars, while JPMorgan estimates $5 trillion in data-center investment by 2030, yet Chinese rivals have reached comparable performance on some benchmarks within only a few weeks and at a fraction of training and inference cost. Distillation’s statistical importance has risen as frontier training shifted from web-scale text ingestion to compute-heavy trial-and-error reinforcement, making synthetic data harvesting a faster path for chip-constrained firms.
Trend signals point to acceleration: Anthropic and Google DeepMind both report distillation attempts became more common over the past year, with tactics including traffic obfuscation and task-splitting across thousands of accounts, plus reports of a Chinese “cottage industry” coordinating such activity. Commercially, MIT researchers cited in the context say China’s share of the open-model market has rapidly grown and overtaken America’s since DeepSeek’s R1 release, while near-term enforcement appears uncertain despite options such as cloud access bans and tighter chip controls.
Source: Anthropic says China’s AI tigers are copycats
Subtitle: DeepSeek’s new model has American officials and firms on edge
Dateline: 2月 26, 2026 04:27 上午