OpenAI says China's DeepSeek trained its AI by distilling US models

Key Concepts

• Distillation (知識蒸餾): A technique in AI model training where a smaller, newer model learns from the outputs of a larger, more established model.
• Frontier AI Labs (前沿人工智能實驗室): Leading AI research and development companies pushing the boundaries of AI capabilities.
• Free Riding (搭便車): Utilizing the work and capabilities of others without contributing proportionally.
• Strategic Competition (戰略競爭): The geopolitical rivalry between the US and China, particularly in technology.

---

OpenAI’s Concerns Regarding Deepseek’s Model Replication

OpenAI, under the leadership of Sam Altman, has formally alerted US lawmakers to alleged efforts by Chinese AI startup Deepseek to replicate the capabilities of OpenAI’s models, including ChatGPT, and those of other leading US AI companies. This warning was delivered via a memo to the US House Select Committee on Strategic Competition between the US and the Chinese Communist Party on Thursday.

The core of OpenAI’s concern centers around a technique called “distillation.” This process doesn’t involve directly copying the code or architecture of OpenAI’s models. Instead, Deepseek is reportedly utilizing distillation to leverage the outputs of OpenAI’s models for its own training purposes. Specifically, a more powerful, pre-existing AI model (like those developed by OpenAI) assesses the quality of responses generated by Deepseek’s newer model. This evaluation process effectively transfers the knowledge and learnings from the established model to the newer one, allowing Deepseek to improve its model’s performance without undertaking the same extensive and costly initial training. OpenAI characterizes this as “ongoing efforts to free ride on the capabilities developed by OpenAI and other US frontier labs.”

Deepseek’s Emergence and Impact

Deepseek, along with its parent company Highflyier, gained significant attention in early 2023 when it released a suite of AI models that demonstrated performance comparable to some of the best models originating from the United States. This rapid advancement fueled anxieties within Washington regarding China’s potential to quickly close the gap in the AI race, despite existing export controls and restrictions on technology transfer.

Geopolitical Implications & Lack of Response

The memo highlights the potential for strategic competition in the AI sector. OpenAI’s warning suggests a concern that Deepseek’s approach, if successful, could undermine the competitive advantage currently held by US AI companies. The fact that Deepseek and Highflyier did not respond to requests for comment from Reuters further underscores the sensitivity of the issue.

Data & Statistics (Implied)

While the transcript doesn’t provide specific numerical data, it implies a significant investment by Deepseek and Highflyier, evidenced by their ability to develop models rivaling US offerings in a relatively short timeframe. The concern also suggests a substantial financial and resource commitment by OpenAI and other US “frontier AI labs” to develop the foundational models Deepseek is allegedly leveraging.

Notable Quote

“Ongoing efforts to free ride on the capabilities developed by OpenAI and other US frontier labs” – Sam Altman (via OpenAI memo). This quote encapsulates the central argument: Deepseek is unfairly benefiting from the innovation and investment of US companies.

Synthesis

The core takeaway is that OpenAI perceives Deepseek’s use of distillation as a potentially damaging tactic that allows a Chinese competitor to rapidly advance its AI capabilities by leveraging the work of US companies. This raises concerns about strategic competition and the potential erosion of US leadership in the AI field. The situation warrants further investigation and potentially policy responses to address the risks associated with this type of model replication.