Recently, with the progress of a case concerning artificial intelligence copyright — the Kadrey v. Meta case, internal messages from Meta have been unsealed by the court, revealing that company executives are determined to surpass OpenAI's GPT-4 model in the development of Llama3.
Ahmad Al-Dahle, Meta's Vice President of Generative AI, mentioned in a message from October 2023: “Honestly, our goal must be GPT-4. We have 64,000 GPUs! We need to learn how to build cutting-edge technology and win this competition.”
Although Meta has released open AI models, the company's AI leadership is clearly more focused on competitors that do not publicly share model weights, such as Anthropic and OpenAI, viewing their Claude and GPT-4 as benchmarks. Despite frequent mentions of the French AI startup Mistral, Meta executives appear to dismiss it. Al-Dahle stated in a message: “Mistral is a piece of cake for us; we should be able to do better.”
In the AI field, major companies are competing to launch advanced AI models, and these court documents reveal a highly tense atmosphere at Meta in this competition. In several messages, Meta's AI leaders noted that they are “very proactive” in acquiring the data needed to train Llama. One executive even said: “Llama3 is the only thing I care about.” They discussed how to improve the dataset to enhance Llama3's performance.
However, the prosecutor in the case accused Meta executives of possibly overlooking data usage in their eagerness to launch AI models, involving some copyrighted books. Touvron mentioned that the dataset combination for Llama2 “performed poorly” and explored how to use better data sources to improve Llama3. Al-Dahle asked: “Do we have the right datasets? Is there anything we can't use for stupid reasons?”
Meta CEO Mark Zuckerberg previously stated that efforts are being made to narrow the performance gap between the Llama models and the closed-source models from companies like OpenAI and Google. These internal messages show that Meta is under significant pressure to achieve this goal. In a letter from July 2024, Zuckerberg mentioned: “This year, Llama3 is competitive among the state-of-the-art models and leads in certain areas.”
In April 2024, Meta finally released Llama3, which performed excellently in the competition, surpassing the open options from Mistral, but the data used for training this model — which reportedly received Zuckerberg's approval — is under scrutiny from multiple lawsuits.
Key Points:
🌟 Meta executives are focused on surpassing OpenAI's GPT-4 model in the development of Llama3.
💡 The company is proactive in data acquisition but faces accusations of using copyrighted data.
📈 Zuckerberg looks forward to future Llama models becoming the most advanced choice in the industry.
