Executives and researchers leading Meta's artificial intelligence efforts were obsessed with beating OpenAI's GPT while developing Llama 3, according to internal information released by a court on Tuesday in one of the company's ongoing AI copyright cases, Kadrey v. Meta. 4 models.

"Honestly... our target needs to be GPT-4," Ahmad Al-Dahle, Meta's vice president of generative artificial intelligence, said in an October 2023 message to Meta researcher Hugo Touvron. "We're coming to 64k GPUs! We need to learn how to push the frontier and win this race."

Although Meta releases open AI models, the company's AI leaders are more focused on beating competitors like Anthropic and OpenAI that typically don't release model weights, instead keeping them behind APIs. Meta executives and researchers point to Anthropic's Claude and OpenAI's GPT-4 as the gold standards for their efforts.

French artificial intelligence startup Mistral is one of Meta's biggest public competitors. The company has been mentioned several times in internal messages, but in a dismissive tone.

"Mistral is insignificant to us," Al-Dal said in a message. "We should be able to do better," he said later.

Today, tech companies are competing to steal each other's thunder with cutting-edge AI models, but these court documents reveal how competitive Meta's AI leader really was — and appears to still be. In several parts of the message exchange, Meta's AI leadership talked about how they were "very aggressive" in getting the right data to train Llama; at one point, one executive even stated in a letter to a colleague, " What I really care about is Llama 3."

In the case, prosecutors allege that Meta executives occasionally cut corners in the frantic race to ship artificial intelligence models and were trained on copyrighted books in the process.

Touvron noted in a message that the combination of data sets used by Llama 2 was "terrible" and talked about how Meta could improve Llama 3 with a better combination of data sources. Touvron and Al-Dahle then talked about cleaning up the usage paths LibGen dataset, which contains copyrighted work from Cengage Learning, Macmillan Learning, McGraw Hill and Pearson Education.

“Do we have the right data set there(?),” Al-Dahle said. "Is there anything you want to use but can't for some stupid reason?"

Meta CEO Mark Zuckerberg has previously said he is working to close the performance gap between Llama's AI models and closed models from OpenAI, Google and others. Insider sources reveal the intense pressure within the company to do so.

Zuckerberg said in a letter in July 2024: "This year, Llama 3 is competitive with the most advanced models and leads in some areas. Starting next year, we expect that future Llama models will Become the most advanced vehicle in the industry.”

When Meta finally releases Llama 3 in April 2024, the open AI model will be competitive with leading closed models from Google, OpenAI, and Anthropic, and better than Mistral's open options. However, the data Meta used to train the model — which Zuckerberg reportedly approved for use despite owning the copyright — is facing scrutiny in several ongoing lawsuits.