Omar Marquez | Sopa Images | Light Rocket | Getty Images
meta The significant investment in artificial intelligence includes the development of an AI system designed to power Facebook’s entire video recommendation engine across all of its platforms, company executives said Wednesday.
tom allisonFacebook chiefs said that part of Meta’s “technology roadmap to 2026” includes AI recommendations that could power both the company’s TikTok-like Reels short-form video service and more traditional long-form videos. This includes the development of a dation model.
Allison said on stage at the Morgan Stanley technology conference in San Francisco that in the past, Meta has typically used separate models for each product, such as Reels, Groups and the core Facebook feed. Told.
As part of Meta’s ambitious foray into AI, the company has spent billions of dollars on Nvidia graphics processing units (GPUs). These have become the primary chips used by AI researchers to train the types of large-scale language models used to power OpenAI’s popular ChatGPT chatbot and other generative AI models. Masu.
Allison said “Phase 1” of Meta’s technology roadmap includes switching the company’s current recommendation system from more traditional computer chips to GPUs, improving the product’s overall performance. He said that
As interest in LLM exploded last year, meta executives were shocked to discover that these large AI models “can handle large amounts of data and all kinds of very general-purpose activities, such as chats.” Allison said he received it. Meta realized the potential of huge recommendation models that could be used across products, and by last year had built “this kind of new model architecture,” Allison said, adding that the company had tested it with Reels. Ta.
This new “model architecture” allows Facebook to “increase Reels watch time by 8% to 10%” in the core Facebook app, and allows the model to “learn from data much more efficiently than previous generations.” Allison says it has been proven that ”
“We’re focused on making some kind of additional investment to be able to scale up these models with the right kind of hardware,” he said.
Meta is currently in “Phase 3” of its system rebuild, validating the technology and attempting to apply it to multiple products.
“Instead of just powering Reels, we’re working on a project to power the entire video ecosystem with this single model. We can then add a feed recommendation product and make it available in this model as well. “Do you want to do that?” Allison said. “If we get this right, we think it will not only make recommendations more engaging and relevant, but also make them more responsive.”
Explaining how it would work if successful, Allison said, “When you find something you’re interested in on Reels and go back to your feed, you’ll be able to see more similar content.” .
Allison said Meta has amassed a significant amount of GPUs that will be used to support a wide range of generative AI efforts, including the development of digital assistants.
Some of the generative AI projects Meta is considering include incorporating more sophisticated chat tools into the core feed, which would allow people who saw “suggested posts about Taylor Swift” to You could just click a button and say, ‘Hey, Meta AI, tell me more.'” About what I see now with Taylor Swift. ”
Meta is also experimenting with integrating AI chat tools into groups, so when members of a baking group on Facebook ask a question about dessert, they could potentially get an answer from a digital assistant.
“I think we have an opportunity to bring generative AI into multiplayer consumer environments,” Allison said.
clock: Full CNBC interview with Meta’s Nick Clegg
Don’t miss the next story from CNBC PRO.
