Meta is building a giant AI model to power its ‘entire video ecosystem,’ exec says

Meta’s hefty investment in artificial intelligence includes development of an AI system designed to power Facebook’s entire video recommendation engine across all its platforms, a company executive said Wednesday.

Tom Alison, the head of Facebook, said part of Meta’s “technology roadmap that goes to 2026” involves developing an AI recommendation model that can power both the company’s TikTok-like Reels short video service and more traditional, longer videos.

To date, Meta has typically used a separate model for each of its products, such as Reels, Groups and the core Facebook Feed, Alison said onstage at Morgan Stanley’s tech conference in San Francisco.

As part of Meta’s ambitious foray into AI, the company has been spending billions of dollars on Nvidia graphics processing units, or GPUs. They’ve become the primary chips used by AI researchers for training the types of large language models used to power OpenAI’s popular ChatGPT chatbot and other generative AI models.

Alison said “phase 1” of Meta’s tech roadmap involved switching the company’s current recommendation systems to GPUs from more traditional computer chips, helping to improve the overall performance of products.

As interest in LLMs exploded last year, Meta executives were struck by how these big AI models could “handle lots of data and all kinds of very general-purpose types of activities like chatting,” Alison said. Meta came to see the possibility of a giant recommendation model that could be used across products, and by last year, built “this kind of new model architecture,” Alison said, adding that the company tested it on Reels.

This new “model architecture” helped Facebook obtain “an 8% to 10% gain in Reels watch time” on the core Facebook app, which Alison said helped prove that the model was “learning from the data much more efficiently than the previous generation.”

“We’ve really focused on kind of investing more in making sure that we can…