"According to the company behind ChatGPT, DeepSeek is systematically attempting to extract knowledge from leading American AI systems in order to improve its own models. In the memo, which OpenAI sent to the U.S. House Select Committee on Strategic Competition between the U.S. and the Chinese Communist Party, OpenAI outlines attempts to circumvent technical and access restrictions. The company claims that accounts linked to DeepSeek employees have developed methods to access AI models via external, obfuscated network routes."
"In addition, scripts have reportedly been written to automatically collect output for use in so-called distillation. Distillation is a method in which a more powerful, better-trained model is used to evaluate and refine the performance of a newer model. By analyzing responses from an advanced system, a less advanced model can learn faster and achieve better results without going through the entire training process itself."
"DeepSeek, based in Hangzhou, previously attracted international attention with models that, according to various experts, can compete with leading American alternatives. This led to concerns in the United States that China is rapidly gaining ground in the global AI race despite trade restrictions and export controls. DeepSeek's parent company, High-Flyer, has not yet publicly responded to the allegations. DeepSeek itself has also not provided any substantive explanation."
An internal memo states DeepSeek systematically attempts to extract knowledge from leading American AI systems to enhance its own models. Accounts linked to DeepSeek employees allegedly used external, obfuscated network routes to access AI models, and scripts reportedly collected model outputs automatically for distillation. Distillation leverages a more powerful, better-trained model to evaluate and refine a newer model by analyzing advanced-system responses, enabling faster learning without full retraining. OpenAI characterizes these practices as freeloading on U.S. AI investments and says it acts against suspected exploiters. DeepSeek, based in Hangzhou, and parent High-Flyer have not publicly responded, while some DeepSeek models draw technical praise.
Read at Techzine Global
Unable to calculate read time
Collection
[
|
...
]