Recently, a controversy surrounding artificial intelligence (AI) has taken center stage in the international tech community. Stanford University released a large language model named Llama-3-V, which has been accused of plagiarizing open-source work from Tsinghua University and the Chinese mainland-based tech firm Modelbest.
The incident emerged when Modelbest introduced MiniCPM-Llama3-V 2.5, a powerful edge-based multimodal model, in 2021. By April 2024, the model had secured a Series A financing round amounting to hundreds of millions of Chinese yuan, leading to the launch of MiniCPM-Llama3-V2.5, touted as the world's most potent edge-side multimodal model with eight billion parameters. It set new benchmarks in optical character recognition and image encoding speed, making it highly suitable for industrial applications in sectors like e-commerce, finance, and education.
The controversy began on May 29, when a team of Stanford and University of Southern California students released Llama-3-V on a prominent open-source platform. The model claimed it could train a multimodal model comparable to renowned models like GPT-4V for just $500, garnering significant attention. However, experts soon identified striking similarities in architecture and code between Llama-3-V and MiniCPM-Llama3-V 2.5.
On June 3, Li Dahai, CEO of Modelbest, provided evidence that the Llama-3-V project had utilized undisclosed Chinese ancient text data. He highlighted that the two models not only exhibited similar response patterns but also shared identical errors, raising concerns about the integrity of the development process.
In response, the Stanford students issued an apology on social media platform X on June 4, admitting to the plagiarism allegations and announcing the removal of the Llama-3-V model. This incident underscores the critical importance of intellectual property protection and the need for stringent governance within the open-source community to safeguard innovative achievements globally.
Reference(s):
American AI plagiarism shows the importance of IP protection
cgtn.com
