Stanford AI Project Plagiarism: Meet the 26-Year-Old CTO of the Controversial Large Model Company, Once Celebrated as a ‘Young Genius’

Recently, the Llama3-V model, an open-source project led by Stanford University's AI team, has faced controversy due to allegations of plagiarism involving the MiniCPM-Llama3-V2.5 model developed by Tsinghua University and Mianbi Technologies. This incident has ignited significant online discussions. On June 3, Mianbi's CEO, Li Dahai, and co-founder Liu Zhiyuan publicly expressed their disappointment, emphasizing that while international recognition is valuable, it should not compromise integrity. They stressed the need for a collaborative and trustworthy community, stating, "We hope our team's hard work receives the recognition it deserves, but not in this manner."

In light of these events, Stanford authors Siddharth Sharma and Aksh Garg issued a formal apology on social media for the academic misconduct and withdrew the Llama3-V model from circulation.

Mianbi Technologies is emerging as a notable player in the AI landscape, having recently secured a funding round amounting to hundreds of millions. The company, part of the "Tsinghua Venture Capital" ecosystem, was founded in August 2022 with a focus on AI model innovation and practical applications. Zeng Guoyang, the 26-year-old technical leader and legal representative, is recognized as a young talent in the field.

Founded by Liu Zhiyuan, a tenured professor at Tsinghua University, Mianbi has drawn talent primarily from the Tsinghua University NLP lab. Zeng, showing early promise, began programming at age eight and excelled in national and international competitions before interning at Megvii, a leading AI firm in China.

In interviews, Zeng emphasized the significance of crafting strategies based on their unique strengths rather than simply following trends related to model parameter scale. "Pursuing only model size is not a sustainable path; efficiency is key for large models," he stated.

Mianbi's core technical team comprises graduates from top Chinese NLP research labs, averaging 28 years in age, with 80% hailing from elite universities and including talent from organizations like Alibaba and ByteDance. Co-founder and CEO Li Dahai, a Peking University graduate, has prior experience with Google and served as CTO of Zhihu. Liu Zhiyuan, the Chief Scientist, is a respected researcher with over 200 published papers and numerous prestigious awards.

In April, Mianbi announced significant funding and featured speakers like Professor Andrew Ng at the Sequoia AI Ascent 2024 event, discussing the prospects of multi-agent systems. One such system, ChatDev, co-developed by Mianbi and Tsinghua's NLP lab, showcases automated software development capabilities.

Mianbi also revealed its AI agent applications and the MiniCPM model at the Zhongguancun Forum. On May 28, Li Dahai announced the launch of MiniCPM-Llama3-V2.5, identified as the world’s most powerful edge-side multimodal model. With just 8 billion parameters, it outperformed other models, like Gemini Pro and GPT-4V, in multimodal performance. This next-generation model operates efficiently on consumer-grade hardware, delivering impressive speed and functionality, including OCR capabilities for mobile use.

The MiniCPM-V series quickly gained popularity, accumulating over 130,000 downloads shortly after its release and topping the trending charts on platforms like Hugging Face and GitHub.