Massive AI Extraction
A significant controversy has emerged within the artificial intelligence sector, with a leading US-based company, Anthropic, making serious accusations
against three prominent Chinese AI developers: DeepSeek, MiniMax, and Moonshot AI. Anthropic asserts that these entities engaged in synchronized, large-scale operations to extract capabilities from its advanced AI models, specifically the Claude series, through a method known as 'distillation.' The company detailed that these collective efforts involved an astonishing number of interactions, reportedly exceeding 16 million engagements, facilitated by approximately 24,000 fabricated user accounts. This massive data extraction campaign is considered a direct violation of Anthropic's usage policies and regional access limitations, raising alarms about the methods used to advance AI development.
Understanding AI Distillation
At its core, 'distillation' in the context of artificial intelligence is a sophisticated training methodology where a more compact or less complex AI model learns by mimicking the outputs of a highly advanced, larger system. While this technique is a standard practice for creating more efficient AI, Anthropic highlights its potential for misuse. The concern is that it can be exploited to replicate cutting-edge functionalities and knowledge embedded in state-of-the-art models without incurring the substantial costs and time typically associated with independent research and development. This allows developers to acquire advanced AI capabilities more rapidly and economically, potentially disrupting the natural progression of AI innovation.
Under Scrutiny
Anthropic's investigation revealed distinct patterns of suspicious activity attributed to DeepSeek, Moonshot AI, and MiniMax. DeepSeek's participation reportedly involved over 150,000 interactions, with a particular emphasis on gathering data related to reasoning processes and reinforcement learning frameworks. The company identified synchronized usage across numerous accounts, including shared payment details and coordinated timing of requests, strongly suggesting a deliberate, organized effort to circumvent detection mechanisms. Moonshot AI's involvement was even more extensive, accounting for over 3.4 million interactions. Their prompts were designed to extract data on agentic reasoning, tool utilization, coding, data analysis, and computer vision competencies, indicating a systematic approach to training their own models. MiniMax conducted the most significant operation, with more than 13 million exchanges focused on agentic coding and orchestration tools. Notably, when Anthropic launched an updated model, MiniMax's traffic reportedly shifted to the new system within a mere 24 hours, demonstrating a highly agile and adaptive extraction strategy.
Access and Policy
Anthropic explicitly stated that commercial access to its Claude models is not provided in China. The alleged methods employed by some of these AI labs involved sophisticated routing of traffic through proxy services that managed extensive networks of fraudulent accounts, a strategy described as 'hydra cluster' architectures. These proxies channeled requests through Anthropic's API and third-party cloud platforms, obscuring the origin of the activity. In one instance, a single proxy network was found to manage over 20,000 accounts concurrently. Anthropic contends that such practices could potentially circumvent US export control regulations by enabling foreign developers to close technological gaps through the unauthorized acquisition of advanced AI capabilities. This situation highlights the complex intersection of AI development, international trade, and regulatory oversight.
