Meta execs obsessed over beating OpenAI’s GPT-4 internally, court filings reveal

Meta execs aimed to beat OpenAI's GPT-4 with Llama 3.

: Meta's AI leaders were focused on outperforming OpenAI's GPT-4, particularly while working on Llama 3, as revealed by court filings. Internal messages show an intense focus on securing resources and improving datasets for AI training, despite legal concerns. With 64,000 GPUs, executives like Ahmad Al-Dahle emphasized staying competitive in the AI race. Llama 3, released in April 2024, achieved competitiveness with top closed models.

Meta’s executives and researchers were determined to outperform OpenAI’s GPT-4 model during the development of Llama 3, as shown by internal messages revealed in court filings from the Kadrey v. Meta case. These messages detail efforts by figures like Ahmad Al-Dahle, Meta’s VP of Generative AI, to rally resources, including 64,000 GPUs, to stay ahead of competitors like Anthropic and OpenAI, despite dismissing the French startup Mistral.

Within the company, there was a significant push to obtain the best data for training Llama, even mentioning potential legal issues with datasets containing copyrighted materials from publishers like Cengage Learning and Pearson Education. An aggressive stance on data acquisition was evident, with executives focusing fervently on Llama 3's success, showcasing their strategy to close the performance gap with closed model leaders like OpenAI and Google.

By April 2024, when Llama 3 was released, the open AI model had achieved competitiveness with leading closed models from Google, OpenAI, and Anthropic while surpassing some open model competitors. However, the datasets used, reportedly approved by Mark Zuckerberg, are under legal scrutiny due to copyright concerns, highlighting the legal challenges Meta faces in its quest to lead AI innovation.