Twitter user investigating potential Fable distillation in DeepSeek V4 Pro (twitter.com)

A recent investigation led by a Twitter user has uncovered potential evidence that DeepSeek is conducting a large-scale distillation operation using outputs from proprietary models, particularly Claude Fable 5. By utilizing Deepseek V4 through their official API, the user found that for complex prompts, such as those related to 3D games, the outputs bore striking similarities to those produced by Fable 5. Interestingly, the model’s structure and coherence of thought (CoT) also deviated significantly from typical Deepseek outputs, suggesting a possible integration or routing to Fable's classifiers. This discovery is pivotal for the AI/ML community as it raises questions about the ethical implications of model distillation and proprietary data usage. The investigation highlighted that while Deepseek V4 performs well on complex coding tasks, its performance degrades significantly when trying to address queries related to sensitive tasks like cyber or biological issues, potentially indicating that these requests were routed back to Fable after encountering classifiers. This could reflect deeper insights into how AI systems interact with and leverage proprietary models, prompting further scrutiny and discussion on distillation practices and data traffic within the AI ecosystem.