Meta's Chief AI Officer, Alexandr Wang, has recently revealed that the company's open-source AI strategy needs reassessment, particularly concerning its latest model, Muse Spark. This decision follows early training phases where significant safety issues were identified, including potential bio risks. As a result, Meta opted to keep Muse Spark confidential rather than risk an open release, a sentiment Wang believes is echoed across the AI industry as it grapples with similar challenges.
Wang, who was appointed last year after a notable $15 billion arrangement with Scale AI, spoke with Bloomberg Tech about the limitations facing Meta’s open-source initiatives. He acknowledged that the company originally gained prominence with Llama, a series of open-weight models that set a high standard for accessible AI development. However, the landscape has changed. While Meta plans to continue releasing open-source models deemed "fit and safe," future frontier developments will be kept proprietary.
The implications of this shift are significant; as part of the establishment of Meta Superintelligence Labs, Wang revised the company's protocols for evaluating AI risks. He emphasized that Muse Spark's internal deployment allows Meta to implement safeguard measures that are unattainable once models are released publicly.
Despite these updates, Muse Spark has struggled to compete with other frontier models, particularly in coding tasks. Meta employees testing the model have shown a preference for Anthropic's Claude over their own offering, and while Muse Spark has garnered praise for its visual capabilities, it has not yet positioned itself as a leading contender in software development. Insiders have noted that parts of Muse Spark share similarities with DeepSeek's latest model and have remarked on its reliance on Llama 4's underlying code and datasets, despite earlier claims of being built independently.
Access to Muse Spark has also been limited; it primarily operates within Meta’s applications, with the availability of API access described as restrictive. Meanwhile, Wang faces increasing financial pressure, as a report revealed that a staggering 97.6% of Meta's projected 2025 revenue is dependent on advertising. To alleviate this reliance, Meta is exploring subscription models across its platforms, including Instagram, Facebook, and WhatsApp, and has introduced a $7.99 subscription for its AI chatbot in select regions.
Analysts predict that this pivot toward subscriptions could yield substantial revenue, with estimates ranging from $15.6 billion next year to as much as $20 billion annually by 2030, according to various financial institutions. These ambitious forecasts come on the heels of a year in which Meta's non-advertising revenue totaled less than $5 billion. The pressing question now is whether Wang's lab can deliver the models needed to substantiate these expectations.
