Alexander Wang, whose Scale AI co-founder Mark Zuckerberg was hired as Meta’s chief AI officer in a $15 billion deal last year, acknowledged on Bloomberg Tech that Meta’s longstanding open source approach hit a wall with Muse Spark. The model, released in April, remained proprietary after internal testing pointed to risks that could not be safely contained in an open release, but Wang said rival labs are encountering the same problems as the model scales.“In fact, the early training process triggered some high-risk areas, especially around biorisks, but it also elevated a lot of risks,” Wang told Bloomberg. He added: “I think this is something the entire industry recognizes because the models have improved dramatically over the past year.”
Bet on open source is quietly rewritten
Meta built its AI reputation with Llama, a family of open weight models that briefly established the company as the industry standard for accessible AI. Mr. Wang’s framing is now more cautious. He said Meta will keep undeveloped work locked down and continue to open source models it deems “suitable and safe” for release. Asked if Llama would continue to be the brand for that effort, the company declined to say, “We’re having some exciting discussions internally about branding, but we don’t have anything to share at this time.”Pivoting is key. As part of establishing Meta Superintelligence Labs, Wang updated internal documentation outlining how the company’s advanced AI scaling framework, Meta, assesses and mitigates model risk. He claimed that Muse Spark’s in-product deployment allows Meta to have guardrails that don’t exist once weights are exposed.
Why Muse Spark still lags behind Claude and Gemini when it comes to coding
Despite all that recalibration, Muse Spark has failed to establish itself as a frontier challenger. The Financial Times reported that Meta employees who were asked to test models for software development tasks continued to prefer Anthropic’s Claude. Wang acknowledged that although the model has won praise for its visual understanding, it lags behind rivals when it comes to coding. According to FT, some insiders have compared parts of the system to DeepSeek’s latest model, with others pointing out that Muse Spark relies on Llama 4 code and datasets, even though Wang once said it was “built from the ground up.””Access is also limited. FT explains that this model primarily exists within Meta’s own apps, with limited rollout of private APIs.
The advertising machine still pays for AI.
The broader pressure on Mr. Wang is financial. The Wall Street Journal reported this week that 97.6% of Meta’s 2025 revenue will come from advertising, and that the company’s planned AI capital spending this year is higher than that of Google, Microsoft, or Amazon relative to its size. Zuckerberg is currently testing a $4-a-month subscription tier on Instagram, Facebook and WhatsApp, as well as a $7.99 meta-AI chatbot subscription in some markets, seeking non-ad revenue.Analysts at Trust Securities cited by the Journal predict up to $20 billion in annual subscription opportunities by 2030, while Deutsche Bank has $15.6 billion in reserve next year. Brainless predictions for a company that failed to clear $5 billion in non-advertising revenue last year. Whether Wang’s lab provides a model that justifies these numbers is a question Meta has not yet answered.
