[LIVE] Anthropic Distillation & How Models Cheat (SWE-Bench Dead) | Nathan Lambert & Sebastian Raschka

Anthropic’s recent blog post sparked a heated discussion by exposing a coordinated effort by several Chinese labs to harvest synthetic data from its APIs and use it for model distillation. The post frames the activity as a geopolitical threat, highlighting how GPU shortages push labs toward cheap API access rather than building their own infrastructure. By labeling the practice an "attack," Anthropic underscores the growing tension between open AI services and emerging competitors seeking to shortcut model development.

Model distillation, a long‑standing machine‑learning technique, involves training a compact model on the outputs—logits, text, or other synthetic data—generated by a larger teacher model. In the era of large language models, this process often uses API calls to commercial services like OpenAI, Claude, or DeepSeq, then fine‑tunes a downstream model on the collected responses. Most providers embed terms of service that explicitly prohibit using generated content to create competing models, yet enforcement remains sparse outside the U.S. Detecting illicit distillation is challenging because the same API calls serve legitimate benchmarking, customer‑facing chatbots, and research. Providers typically look for unusually high query volumes, repetitive benchmark patterns, or anomalous token distributions to flag potential abuse.

The broader industry impact is twofold. First, multi‑provider routing platforms such as OpenRouter make it easier to aggregate data from diverse APIs, blurring the line between permissible evaluation and prohibited data harvesting. Second, the race to build efficient, locally runnable models—exemplified by DeepSeq, Minimax, and DeepSeek—creates incentives for aggressive data collection, especially when timing aligns with new model releases. As AI geopolitics intensify, firms must balance open access with robust monitoring, while regulators may need clearer guidelines to protect intellectual property without stifling innovation.