Why Meta's AI team is falling apart
What's the story
Meta's artificial intelligence (AI) team has witnessed a significant talent exodus, with many key researchers behind the company's open-source Llama models leaving.
The trend has raised concerns over Meta's ability to retain top-tier AI talent.
The company's AI strategy is also being challenged by fast-moving open-source competitors like Mistral, a French start-up that has already lured away several former Meta researchers.
Talent drain
Departure of Llama model creators raises concerns
Of the 14 authors credited on the landmark 2023 paper that introduced Llama, only three remain at Meta: research scientist Hugo Touvron, research engineer Xavier Martinet, and technical program leader Faisal Azhar.
The rest have left the company, many joining or founding emerging rivals.
This exodus is most pronounced at Mistral where former Meta researchers Guillaume Lample and Timothee Lacroix are co-founders.
Competitive landscape
Mistral's rise and Meta's challenges
At Mistral, Lample and Lacroix are building powerful open-source models that directly compete with Meta's flagship AI efforts.
Their departure highlights the growing competition in the AI space as former Meta researchers join forces to create new companies.
Meta has also seen a shake-up with Joelle Pineau, who led the company's Fundamental AI Research group (FAIR) for eight years, announcing her resignation last month.
Model development
Meta's AI model development and market position
Despite investing billions in AI, Meta still lacks a dedicated "reasoning" model for multi-step thinking or problem-solving tasks.
This gap has become more pronounced as other companies like Google and OpenAI focus on these features in their latest models.
The average tenure of the 11 departed authors at Meta was over five years, indicating they weren't short-term hires but researchers deeply embedded in Meta's AI efforts.
Open-source impact
Meta's open-source models and future prospects
The 2023 Llama paper was a major technical milestone that legitimized open-weight large language models with freely available code and parameters.
Meta trained its models on publicly available data and optimized them for efficiency, allowing researchers to run state-of-the-art systems on a single GPU chip.
However, two years later, the company has lost its lead in the open-source AI space to competitors like DeepSeek.
Meta has also delayed its largest-ever AI model, Behemoth, after internal concerns about its performance.