Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Mark Zuckerberg is building up the voice capabilities of Meta’s artificial intelligence this year, as the social media giant pushes forward with plans to generate revenues from the fast-developing technology.
Meta is planning to introduce improved voice features into its latest open-source large language model, Llama 4, expected in the coming weeks, said people familiar with the matter, as it bets that future so-called AI-powered agents will be conversational rather than text-led.
The company has been particularly focused on making the conversation between a user and its voice model closer to a two-way natural dialogue, allowing for interruptions from the user rather than a more rigid question and answer format, one person said.
The voice push comes as Zuckerberg, chief executive, has outlined bold plans to make the $1.7tn Silicon Valley company the “AI leader”, calling 2025 a make-or-break year for many of its AI products, as the group races against rivals such as OpenAI, Microsoft and Google to commercialise the technology.
This has led the company to look at trialling premium subscriptions for its AI assistant Meta AI, for agentic tasks such as booking reservations and video creation, said two people familiar with the matter. It is also considering introducing paid advertising or sponsored posts into the search results of its AI assistant, one of the people said.
Zuckerberg this year revealed plans to build an AI engineering agent that has the coding and problem-solving abilities of a mid-level engineer, which he said has a potentially “very large market”.
Meta declined to comment.
The group’s chief product officer Chris Cox on Wednesday highlighted some of its plans for Llama 4, saying it would be an “omni model” whereby speech would “be native . . . rather than translating voice into text, sending text to the LLM, getting text out, and turning that back into speech”.
Speaking at the Morgan Stanley technology, media & telecom conference, he added: “I believe it’s a huge deal for the interface product, the idea that you can talk to the internet and just ask it anything. I think we are still wrapping our heads around how powerful that is.”
Meta has also been discussing the guardrails that the newest Llama model should have around what it can output and whether to lower them, two people familiar with the matter said.
The discussions come amid a flurry of launches from rivals and warnings from newly appointed ‘AI tsar’ David Sacks, a Silicon Valley venture capitalist, who has said he wants to ensure US AI models are not politically biased or “woke”.
OpenAI released its voice mode last year and has focused on giving it distinct personalities, while Grok 3, created by Elon Musk’s xAI and available on the X platform, rolled out its voice features to select users late last month.
The Grok model was specifically designed to have fewer guardrails, including an “unhinged mode” that deliberately responds in ways intended to be “objectionable, inappropriate, and offensive”, according to the company.
Meta last year unveiled a less “sanctimonious” version of its AI model for its third Llama iteration, following criticism that Llama 2 was refusing to answer innocent questions.
Allowing users to interact with an AI assistant using voice commands is a major feature of Meta’s Ray Bans smart glasses, which have recently become a big hit among consumers. The group has accelerated its plans to build lightweight headsets that can usurp the smartphone as consumers’ main computing device.
Additional reporting by Melissa Heikkilä in London
