Recent Chinese Artificial Intelligence Companies DeepSeek For launching an inexpensive and comparable performance OpenAI The AI models for the flagship product sent shockwaves around the world. However, OpenAI suspects that these models were developed based on its data.
OpenAI and Microsoft are investigating whether DeepSeek integrated OpenAI's AI models into DeepSeek's own models through OpenAI's APIs, Bloomberg reported. Microsoft security researchers discovered in late 2024 that large amounts of data were being exported through OpenAI developer accounts that were thought to be linked to DeepSeek, the sources said.
OpenAI told the Financial Times that they have found evidence that DeepSeek uses "distillation" techniques. According to 1AI, "distillation" is a common development technique where developers train their AI models by extracting data from larger, more powerful models. This technique allows small models to be trained efficiently at a much lower cost than the $100 million or more that OpenAI spent to train GPT-4.While developers can integrate their AI technology into their applications through OpenAI's APIs, utilizing the output data to build competitive models violates OpenAI's Terms of Service. OpenAI, however, did not disclose the specific details of the evidence it found.
TheVerge says this is full of irony; after all, OpenAI itself has pushed its GPT models by massively crawling the web for textual information (without permission).
In a statement, OpenAI said, "We are well aware that companies in countries such as China, as well as a number of others, have been attempting to distill the models of leading U.S. AI companies. As a leader in AI, we have taken countermeasures to protect our intellectual property, which includes carefully choosing which cutting-edge features to include when releasing our models. We believe it is critical to work closely with the U.S. government going forward to prevent adversaries and competitors from stealing U.S. technology and to protect state-of-the-art models."