Meta built its AI reputation on openness — that may be changing

Members of Meta's newly established Superintelligence Lab are currently debating a significant shift in their AI approach, moving away from their renowned open-source AI project, Behemoth, towards a more closed model, according to reports from The New York Times. Sources indicate that while training on Behemoth has been completed, its launch has been postponed due to disappointing internal results. Following the lab's inception, evaluations of the model have been reportedly put on hold. These discussions within Meta remain preliminary, as any potential changes would require the approval of CEO Mark Zuckerberg. A spokesperson for the company reassured TechCrunch that their commitment to open-source AI remains steadfast. "We plan to continue releasing leading open source models," the spokesperson stated, emphasizing that Meta has historically held back on releasing all its developments and intends to maintain a balance between open and closed model training. If Meta does decide to focus on closed-source models, it would signal a profound ideological shift for the organization. Although Meta currently utilizes advanced closed models for internal applications, such as those driving its Meta AI assistant, Zuckerberg has previously championed open-source initiatives, presenting the Llama family of models as a key competitive advantage against companies like OpenAI, which he criticized for becoming more insular following its partnership with Microsoft. Amidst mounting pressure to diversify revenue streams beyond advertising, Meta has invested heavily in AI, including competitive salaries and signing bonuses to attract top talent, expanding data center capabilities, and covering the substantial costs tied to developing artificial general intelligence (AGI), or 'superintelligence.' Despite possessing one of the leading AI research facilities globally, Meta finds itself trailing behind competitors like OpenAI, Anthropic, Google DeepMind, and xAI in terms of monetizing its AI advancements. Should Meta prioritize closed models, it could imply that their previous commitment to openness was more of a strategic maneuver rather than a genuine ideological stance. Past remarks from Zuckerberg suggest a hesitance to fully embrace the open-source model, indicating that while he supports open initiatives, he has not pledged to release every development. He noted, "If at some point, however, there’s some qualitative change in what the thing is capable of, and we feel like it’s not responsible to open source it, then we won’t." Adopting closed models would empower Meta with greater control and monetization opportunities, particularly if they believe their newly acquired talent can achieve superior performance. This transformation could potentially alter the AI landscape, slowing the momentum of open-source initiatives that have been largely propelled by Meta and its Llama models. The balance of power may shift back in favor of major players with closed ecosystems, while grassroots open-source development may persist among smaller entities focused on fine-tuning and model alignment. On a broader scale, Meta's retreat from open-source strategies might create an opening for countries like China, which have actively pursued open-source AI solutions, such as DeepSeek and Moonshot AI, to enhance their domestic capabilities and extend their global influence.