Meta Stumbles: Delays Release Of Flagship “Behemoth” AI Model Over Performance Concerns

A new report from The Wall Street Journal hit the wires late Thursday during the U.S. cash session, revealing that Meta is reportedly delaying the release of its flagship AI model, “Behemoth,” due to internal concerns over its performance and mounting questions about its multibillion-dollar splurge in AI investments.

Initially slated for an April release around Meta’s inaugural AI conference (LlamaCon) for developers, Behemoth has been pushed to late in the second half of this year.

The delays come as Meta has publicly claimed that Behemoth outperforms competitors like OpenAI and Google on certain benchmarks. People familiar with the delays say Behemoth has faced “training challenges.” 

Since Meta’s first Llama model was released in early 2023, 11 of the 14 original researchers from the company’s Fundamental AI Research Team have left.

Meanwhile, Meta disclosed in its latest earnings report that it had raised its capital expenditure forecast to between $64 billion and $72 billion, up from a previous range of $60 billion to $65 billion—a significant portion of that increase earmarked for AI infrastructure.

This development mirrors delays at other major AI labs, including OpenAI and Anthropic, suggesting that progress in large language models has slowed. 

“Right now, the progress is quite small across all the labs, all the models,” Ravid Shwartz-Ziv, an assistant professor and faculty fellow at New York University’s Center for Data Science, told WSJ.

In markets, Meta shares are down a little more than 2% – coming into resistance after a multi-week 37% run. 

What could possibly go wrong?

. . .