OpenAI suspects DeepSeek used its AI models to build DeepSeek-R1
text_fields
OpenAI has reportedly raised concerns that its proprietary AI models may have been used to develop DeepSeek-R1, a newly released AI model by Chinese firm DeepSeek.
According to a report by the Financial Times, OpenAI claims to have found evidence of distillation — a process in which AI models learn from the outputs of larger models - potentially involving its technology.
The San Francisco-based AI firm has stated that certain users may have been leveraging its application programming interface (API) to extract and distill information from its models. Suspecting that these activities were linked to DeepSeek, OpenAI, along with its cloud partner Microsoft, conducted an investigation and subsequently blocked the accounts involved.
In a statement to the Financial Times, OpenAI remarked, "We know [China]-based companies — and others — are constantly trying to distill the models of leading US AI companies." The company also noted that it is collaborating with the US government to safeguard its advanced AI models from competitors and potential adversaries.
DeepSeek recently released DeepSeek-R1, an open-source, reasoning-focused AI model, on GitHub and Hugging Face. The model has reportedly outperformed OpenAI's o1 AI models in multiple benchmarks, raising speculation about how it achieved such advancements in a short time.
AI model distillation is a well-known technique where knowledge from a large AI model is transferred to a smaller, more efficient model, improving performance while reducing computational demands. OpenAI’s GPT-4, for example, is believed to have around 1.8 trillion parameters, whereas DeepSeek-R1 operates on 1.5 billion parameters, making it significantly smaller.
Typically, companies apply distillation to their own proprietary models to refine them, as seen with Meta’s Llama 3, which was used to develop coding-specialised models. However, if OpenAI’s claims are accurate, DeepSeek may have extracted OpenAI’s model outputs by using prompt injections on its API, generating large datasets of text that were then converted into training data for DeepSeek-R1.
Despite these allegations, OpenAI has not yet released a public statement on the matter. Interestingly, Sam Altman, OpenAI’s CEO, recently praised DeepSeek for its advancements, acknowledging that its success fosters competition in the AI industry.