Meta would consider a major strategic turning point in its artificial intelligence policy, according to several sources close to the file cited by The New York Timesthe management of its new “Superintelligence” laboratory would have discussed the abandonment of Behemoth, its most advanced open source model, in favor of the development of a closed model. This reversal, still under study, questions both the limits of openness as about the group’s real ambitions in terms of general artificial intelligence (AG).
A cultural rocking in the IA ecosystem
For several years, Meta has been distinguished from other players in the sector by publishing its Llama language models under open license. This strategy, defended by figures like Yann Lecun, is based on the idea that openness promotes collective innovation and allows wider control over the uses of AI. The researcher still recently stated the idea that the platform that will prevail will be the one open.
But this doctrine seems today weakened, the Behemoth model, presented internally as a “frontier model”, ended its training phase but was not deemed satisfactory. No public version has been published, and the teams in charge of the project would have suspended their tests since the announcement of the creation of the new laboratory in June.
A new generation of leaders
The rise of Alexandr Wang, founder of Scale Ai, who became Chief Ai Officer de Meta, participates in this change. In June, Meta invested $ 14.3 billion in Scale AI, taking 49 % of its capital. Part of Scale’s leaders joined the group to take the lead in a pole now baptized “Meta Superintelligence Labs”, directly attached to Mark Zuckerberg.
Open source or technological sovereignty?
The temptation of the closed model is part of a context where opening begins to be perceived as a strategic risk. In April, the Chinese startup Deepseek published a high -performance model in part based on the open source work of Meta. The risk of seeing competitors take advantage of its advances for free, without contributing to their development, obliges Meta to reassess his position as quickly as possible.
A separate laboratory, internal tensions
The new laboratory directed by Wang operates on the sidelines of the rest of the IA division and is installed in a separate space, closest to the office of Mark Zuckerberg in Menlo Park. It brings together a dozen handpicked researchers, whose organization is close to the structures adopted by Openai or Google Deepmind. The purpose of this team is to accelerate the development of an AG model while securing intellectual property.
Of course, this centralization is not without creating tensions, and departures are anticipated during the next jacket period, in August, especially among AI employees not integrated into the new team. If Meta persists in this path, she will then have to arbitrate between internal cohesion and technological acceleration. A next episode to follow this summer?
