Meta’s AI strategy remains unclear
Facebook parent company considers rivals’ models as internal AI push falters
Mark Zuckerberg is clear: Meta has no intention of slowing down its investments in generative artificial intelligence, even if it means spending far more than might be reasonable. “If we end up misspending a couple of hundred billion dollars, that is going to be very unfortunate obviously. But the think the risk is higher if you build too slowly on what is going to be the most important technology in history”, he said on a recent podcast.
Behind the lofty rhetoric — reminiscent of the years he spent evangelizing the metaverse — the picture is less flattering. Despite tens of billions already invested and a wave of high-profile hires, the parent company of Facebook and Instagram is still steering without a clear course. So much so that it is now considering a radical shift: using models from OpenAI or Google rather than building its own.
Meta has long invested in AI, notably through its FAIR research lab, considered one of the world’s finest. Like Google, much of its focus had been on machine learning, algorithms that can learn autonomously, which transformed its social networks in the late 2010s.
Llama 4 falls flat
But the rise of generative AI, sparked by the sensational launch of ChatGPT, changed the game. Playing catch-up, the Menlo Park company rushed to develop its large language models, Llama, while overhauling its computing infrastructure. It poured money into Nvidia’s must-have GPUs. As a result, its capital expenditures will hit around $70 billion this year, up from just $28 billion in 2023.
In contrast to OpenAI and Google, Meta bet boldly on open source. At the time, Mark Zuckerberg insisted it was the best way to foster an ecosystem leading toward “superintelligence”, an AI surpassing human abilities. He was confident open source would quickly close the gap with proprietary systems. Last year, he promised the fourth version of Llama would be “the most advanced model on the market.”
But its April release was a flop. The early versions were met with lukewarm reviews over technical shortcomings. And the most powerful variant was delayed due to disappointing performance. Its development has reportedly since been abandoned.
Nine-figure paychecks
After the setback, Zuckerberg decided to take tighter control. He personally spearheaded the creation of a new research arm, Meta Superintelligence Labs, led by Alexandr Wang. The founder of Scale AI, a US startup specializing in data annotation, Wang was lured away at extraordinary cost: $14.3 billion, in a operation what looks more like an acquisition.
Mark Zuckerberg also poached dozens of engineers and researchers from OpenAI, Google, and Apple, offering some nine-figure compensation packages in salary and stock — up to $250 million over four years, according to The New York Times. “It’s a small team, so every single seat is extremely precious,” he explained.
The early days have been rocky. In just four months, Meta has gone through multiple restructurings, amid tensions with existing teams. Above all, the roadmap remains unclear. Abandoning open source is on the table, as is relying on external models. The company has already announced partnerships with Midjourney and German startup Black Forest to power its image-generation tools, where it lags rivals.
According to The Information, Meta is also considering using Google’s Gemini to run its MetaAI chatbot, while OpenAI’s models are another option. Internally, such choices are seen as temporary stopgaps until the next version of Llama arrives. At least for now…