Reports: Meta training its new AI using rival models; switching to closed models in quest for profits

A pair of new reports about internal struggles at Meta add new information to how Mark Zuckerberg’s hard pivot to AI is going.

The New York Times details some of the friction between the company’s old guard and Alexandr Wang, the 28-year-old upstart who now leads Meta’s AI division.

One detail: Meta asked the company’s longtime CTO, Andrew Bosworth — considered to be one of the Meta’s top executives — to cut $2 billion from the budget of the division he leads, Reality Labs. The segment is responsible for the company’s AR glasses and the metaverse, the feature that the company changed its name in homage to in 2021. The budget cut from Bosworth’s division will go to the AI division, whose leader joined the company in June, though Meta said next year’s budget isn’t final.

A report last week saying the company is planning 30% budget cuts for the money-losing Reality Labs caused Meta’s stock to surge higher.

Another detail from the Times’ reporting is that according to sources, Bosworth and Chris Cox, the company’s chief product officer, wanted Wang’s team to concentrate on using Instagram and Facebook data to help train Meta’s new foundational AI model — known as a “frontier” model — to improve the company’s social media feeds and advertising business.

But Wang, who is developing the model, pushed back. He argued that the goal should be to catch up to rival AI models from OpenAI and Google before focusing on products, the sources said.

Closed is the new open

Separately, a Bloomberg report out today explains Meta’s effort to build not just a “superintelligent” AI model, but one that is also super profitable. Per the report, Zuckerberg “spends much of his time and energy” working day to day with his new team of AI all-stars, known as “TBD Lab.”

The report also has details of how Meta is building its next model, code-named Avocado. The TBD team is reportedly using third-party models to help train Avocado, including those of its rivals Google and OpenAI. The team is “distilling” from Google’s Gemma, OpenAI’s open-weight model gpt-oss, and the Qwen model from Alibaba, per the report. Use of a Chinese model like Qwen for training could complicate Meta’s efforts to sell its AI for use in national security applications.

A major shift away from open-source models toward proprietary closed ones also seems to be part of Meta’s new strategy. This is a notable departure from Zuckerberg’s passionate, repeated praise of open-source AI, as the Meta chief has recently signaled that the company will be using more closed models. A proprietary model would make it easier to charge for Meta’s AI services compared to its previous strategy of giving away its Llama models for free.