Tomcat
Professional
- Messages
- 2,689
- Reaction score
- 949
- Points
- 113
A group of researchers from Stanford University was in the center of a loud scandal after the release of a new powerful artificial intelligence model Llama 3-V. The model has attracted worldwide attention for its high performance. When launching Llama 3-V, its creators — computer science students Aksh Garg and Siddharth Sharma, as well as researcher Mustafa Aljaderi-claimed that it could compete with such advanced products as GPT4-V, Gemini Ultra and Claude Opus. At the same time, the cost of her training is much lower — less than $ 500.
Soon after the release, representatives of the AI community began to suspect that the new development was simply copied from another open — source project-MiniCPM-Llama3-V 2.5, created jointly by the Tsinghua University natural language processing laboratory and the Beijing startup ModelBest. Investigation data published by informants on GitHub pointed to the almost identical structure and source code of Llama 3-V and MiniCPM-Llama3-V 2.5.
Garg and Sharma acknowledged the similarity of the architectures and apologized to the creators of the original model. "Our architecture is very similar, and we want to sincerely apologize to the original authors," they said, adding that the original version of Llama 3-V has already been removed.
Liu Zhiyuan, co-founder of ModelBest, expressed almost absolute confidence in the plagiarism of their work when creating Llama 3-V. He noted that MiniCPM-Llama3-V2. 5 has the unique ability to recognize ancient bamboo scripts from the Warring States period around 475-221 BC. e. In 2008, Tsinghua University purchased 2,500 such relic bamboo slats and manually digitized the characters contained on them to create a training dataset. This data is not publicly available, but the Llama 3-V model coped well with text recognition, even repeating the errors of the original markup.
At the same time, Liu stressed that the rapid development of AI is impossible without a global open exchange of algorithms, data and experience. Their own MiniCPM-Llama3-V 2.5 used the latest open source Llama 3 model from Meta as a base. However, the key principles of open source are compliance with protocols, mutual trust between participants and respect for the work of pioneers with recognition of their merits — this, according to Liu, was not taken into account by the Stanford team.
Garg and Sharma fully accepted responsibility for negligence and failure to properly verify the originality of the work, since all the code for Llama 3-V was written by a third party — Aljaderi.
The incident caused a wide response and heated discussions. In China, it has become one of the most discussed topics on social networks. A number of experts noted that such an advanced technology as MiniCPM-Llama3-V 2.5 went almost unnoticed only because it was created in a Chinese laboratory, and not in a prestigious American university.
Nevertheless, experts recognize that China has made an impressive leap over the past decade and has turned from a "technological layman" into one of the leaders in AI innovation. However, the gap with the best Western developments like Sora and GPT-4 is still significant.
• Source: https://www.scmp.com/news/china/sci...r-claims-they-copied-chinese-project-ai-model
Soon after the release, representatives of the AI community began to suspect that the new development was simply copied from another open — source project-MiniCPM-Llama3-V 2.5, created jointly by the Tsinghua University natural language processing laboratory and the Beijing startup ModelBest. Investigation data published by informants on GitHub pointed to the almost identical structure and source code of Llama 3-V and MiniCPM-Llama3-V 2.5.
Garg and Sharma acknowledged the similarity of the architectures and apologized to the creators of the original model. "Our architecture is very similar, and we want to sincerely apologize to the original authors," they said, adding that the original version of Llama 3-V has already been removed.
Liu Zhiyuan, co-founder of ModelBest, expressed almost absolute confidence in the plagiarism of their work when creating Llama 3-V. He noted that MiniCPM-Llama3-V2. 5 has the unique ability to recognize ancient bamboo scripts from the Warring States period around 475-221 BC. e. In 2008, Tsinghua University purchased 2,500 such relic bamboo slats and manually digitized the characters contained on them to create a training dataset. This data is not publicly available, but the Llama 3-V model coped well with text recognition, even repeating the errors of the original markup.
At the same time, Liu stressed that the rapid development of AI is impossible without a global open exchange of algorithms, data and experience. Their own MiniCPM-Llama3-V 2.5 used the latest open source Llama 3 model from Meta as a base. However, the key principles of open source are compliance with protocols, mutual trust between participants and respect for the work of pioneers with recognition of their merits — this, according to Liu, was not taken into account by the Stanford team.
Garg and Sharma fully accepted responsibility for negligence and failure to properly verify the originality of the work, since all the code for Llama 3-V was written by a third party — Aljaderi.
The incident caused a wide response and heated discussions. In China, it has become one of the most discussed topics on social networks. A number of experts noted that such an advanced technology as MiniCPM-Llama3-V 2.5 went almost unnoticed only because it was created in a Chinese laboratory, and not in a prestigious American university.
Nevertheless, experts recognize that China has made an impressive leap over the past decade and has turned from a "technological layman" into one of the leaders in AI innovation. However, the gap with the best Western developments like Sora and GPT-4 is still significant.
• Source: https://www.scmp.com/news/china/sci...r-claims-they-copied-chinese-project-ai-model
