– Pubblicità –
OpenAI recently accused the Chinese AI startup DeepSeek of using their proprietary models to train its app, raising concerns about potential intellectual property violations. The San Francisco-based ChatGpt producer claimed to have found evidence of “distillation,” a technique used to achieve better performance on smaller models using results from larger and more capable models, all at a lower cost. While distillation is common in the industry, the fear was that DeepSeek might be using it to build a competing model, thus violating OpenAI’s terms of service.
DeepSeek’s R1 model surprised markets, investors, and tech companies in Silicon Valley. Last year, OpenAI and Microsoft investigated DeepSeek accounts suspected of distillation using OpenAI’s API and blocked access for violating terms of service. Some experts believe DeepSeek’s model was trained on OpenAI’s Gpt-4 results, potentially breaking the terms of service.
It’s common practice for AI labs in China and the US to use the results of companies like OpenAI, who have invested in teaching their models to produce more human-like responses. DeepSeek claimed to have only used 2,048 Nvidia H800 graphics cards and spent $5.6 million to train its V3 model with 671 billion parameters, a fraction of what OpenAI and Google spent on similar-sized models.
While DeepSeek maintains its innocence, the AI community is closely watching this dispute unfold, raising questions about the boundaries of intellectual property in the rapidly evolving field of artificial intelligence.
