Meta’s top executives have reportedly considered “de-investing” in the company’s Llama generative AI, according to a New York Times report on Friday. Instead, the Facebook parent may turn to AI models from rival companies.
Why would Meta shift focus away from its own AI products?
According to The New York Times, Meta CEO Mark Zuckerberg, Head of Product Chris Cox, and Chief Technology Officer Andrew Bosworth have privately discussed adopting AI models from OpenAI or Anthropic instead of their own Llama models. One major differentiator between the products Meta is considering is that Llama is open source, while the others are proprietary and closed to all but their original developers. The Times said Meta has not made a final decision on whether to scale back its investment in Llama.
Company officials said they “remain fully committed to developing Llama and plan to have multiple additional releases this year alone,” a Meta spokesperson said in a statement to The New York Times.
The heightened competition comes in part from the Silicon Valley trend of pursuing “superintelligence,” the theoretical human-level AI that companies like OpenAI were created to build.
More must-read AI coverage
Meta is aggressively headhunting top AI talent
Meta has scrambled to live up to its AI ambitions in recent weeks. Once aiming to compete with OpenAI and make a premier voice-based chatbot, Meta is now headhunting from competing AI companies and offering nine-figure compensation packages. Zuckerberg reportedly recognized that Meta’s own AI wasn’t ready to be the next great product, people close to him told the Times.
The nine-figure payments have attracted some talent. According to the Times, four OpenAI researchers have accepted deals to join Meta. In the past month, Meta has hired OpenAI researchers Trapit Bansal Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai, as well as former Scale AI CEO Alexandr Wang.
The latest Meta AI release was Llama 4, which uses a “mixture of experts” architecture to improve resource efficiency. Two versions were released in April. A third version, known internally as Behemoth, was intended to launch around the same time to train other models; however, it was delayed as doubts emerged about whether Meta could deliver a model with a significant improvement over current offerings.
Generative AI use and excitement overall appeared to plateau in late 2024, according to a survey from Slack.
Read about how an AI bug-hunting tool on HackerOne outperformed human security researchers in a real-world setting for the first time.
