Mark Zuckerberg on Muse, AI Safety and Meta's Next Bet

Jacklyn Dallas introduces Mark Zuckerberg and asks how Meta's long-running bets on AI, smart glasses and virtual worlds fit together. Zuckerberg says the company has worked with machine learning since the early News Feed and later established an AI research lab.

He says Meta Superintelligence Labs represented a restart of its research effort with people from across the industry. Meta was preparing another generation of models and had launched Muse, its personal agent.

Zuckerberg says he experimented with an OpenClaw-like setup at home. The experience suggested to him that a personal agent could be valuable to billions if it worked without requiring people to buy a computer, use a terminal or debug their own installation.

Meta built an agent and supporting systems, including a virtual computer for each agent, he says. He reports that Muse had millions of users about two weeks after launch, although he notes that a company usually needs to iterate after seeing real user feedback.

Dallas notes that Muse already has a visible character. Zuckerberg says the avatar makes AI feel useful and fun; a designer's first version remained his favorite through subsequent iterations.

Asked how a model for personal tasks differs from a general one, he separates coding expertise from the ability to act as an agent. Even when a user is not writing software, he says, an agent may write code in the background to accomplish a task.

A personal agent also needs discretion. Zuckerberg gives the example of finding a restaurant: the agent might know about an allergy or pregnancy, but should complete the request while sharing as little sensitive information as necessary.

He says Meta can train such behavior into its own model and build the rest of the product around it. That includes memory, a mechanism for periodically checking whether it can help with a user's goals, animated avatars and live voice interaction.

Dallas asks why the virtual machine matters. Zuckerberg says the agent needs somewhere to store a person's information, and Meta wanted to approximate the experience of giving it a separate computer without asking every user to buy and configure one.

He describes secure credential storage intended to let the agent use passwords when instructed without exposing them unnecessarily. The broader aim, he says, is to provide both enough capability to act and protection for sensitive personal information.

Zuckerberg describes a separate security agent, Sentinel, that monitors information entering and leaving Muse. He says it can cut off a suspected compromise and ask a user for permission when an action needs their involvement.

Meta was also developing a confidential VM with Moxie Marlinspike, he says. The idea is to give the user an encryption key so Meta cannot access its contents. He says that makes debugging harder and that this version had not yet shipped.

He compares the goal to having an agent's computer under one's own desk, inaccessible to the provider or to someone who compromises its servers. He expects personal agents to know users' goals intimately, making privacy and security central to trust.

Dallas asks how Meta shapes a model's personality, beyond making it capable. Zuckerberg argues there is no single right personality: users should be able to adjust how their agent behaves, as well as its avatar and voice.

He says he made his own agent direct and worklike, after trying a more sarcastic version. Its default character wears a toga and speaks in an unusually deep voice.

Dallas asks about the progress of virtual-world technology and photorealistic avatars. Zuckerberg says he once assumed holograms would arrive before advanced AI, but the order has turned out differently: affordable, ubiquitous holograms remain harder to deliver.

Meta has therefore shifted much of its glasses effort toward Muse and other AI features, while continuing work on presence and avatars. He contrasts an earlier demonstration that required extensive scanning and powerful equipment with a newer process that can create an avatar from a few photos.

Asked whether people's virtual selves will differ from their physical selves, he says the social and psychological effects are still poorly understood. An avatar can convey a person's appearance and also serve as a form of expression, much as people choose a profile photo.

Looking toward 2030, Zuckerberg describes the metaverse as a blending of the physical and digital worlds. Instead of reaching digital content only through a desk or phone, he imagines remote people joining others with a sense of presence.

He gives the example of a poker night with some friends physically present and others joining as holograms. He also imagines AI agents taking part in work meetings alongside people, with a visible character representing each agent.

Dallas asks whether glasses could be used to direct work on a computer. Zuckerberg points to VR workstations with multiple virtual monitors and describes a range of glasses, from audio-only devices to display and holographic prototypes.

He says Muse is being brought to glasses so a person can address a named agent, have it work in its secure virtual machine and receive help through the glasses. He contrasts that with a single prompt and reply.

Dallas asks what it felt like when the AI lab appeared to be struggling. Zuckerberg says Llama 4 was the part that had fallen off the trajectory he expected, after Llama 3 had been close to the frontier at the time.

His diagnosis was that he had organized the team like the much larger groups that work on feeds and advertising. He concluded that language-model research needed a tight-knit group treating the work as a shared science project, and built Meta Superintelligence Labs with people from inside and outside the company.

He says the surprise after Llama 4 was more unsettling than the time needed to rebuild. Founders, in his view, are tested both when an effort goes wrong and when a product exceeds expectations and needs to be expanded quickly.

With Muse, he says, teams across Meta were working on distribution and infrastructure, including making better use of available GPUs and bringing the product to glasses.

Dallas asks how he communicates a vision across many projects. Zuckerberg says writing helps him clarify his own views. He describes a 15-page document on AI's social effects, government, jobs, national security and risks, developed through discussions and multiple drafts.

He condensed that document into a shorter opinion piece and video, because different audiences need different forms of explanation. He says founders cannot communicate in exactly the same way in every situation: some audiences need more explanation and reassurance than others.

Zuckerberg describes building as an intrinsic need, comparable to another person's need to write or create. He says he becomes unhappy when he cannot exercise that creativity.

He compares learning a language, martial arts or another difficult skill with developing product intuition. Some parts can be understood theoretically, he says, but intuition requires repeated practice and time.

He gives his daughter's interest in music as an example: he wanted her to find what interested her but also learn enough piano and music theory to support composing. For adults as well, he says, sustained practice requires discipline.

Dallas suggests that people want to build when given a way to start. Zuckerberg agrees that people have different creative or service motivations, and says an AI agent can help someone sketch an idea before they know every step needed to execute it.

Dallas asks about his work on disease outside Meta. Zuckerberg says the original goal was to help the scientific community cure diseases by the end of the century, not for his organization to do that alone. Its approach is to create tools that let researchers see and understand biology in new ways.

He argues that large scientific tools can be underfunded when money is spread across individual research projects. Biohub, he says, aims to build tools that other scientists can use to accelerate their work.

Zuckerberg now expects progress sooner than his original horizon, although he gives no precise date. He describes a virtual-cell model that could simulate proteins and cells, and potentially more complex biological systems, allowing researchers to explore many experimental scenarios.

Asked whether that means living forever, he distinguishes extending lifespan from treating disease. His stated focus is helping people cure, prevent or manage illnesses, including conditions that might otherwise cause lasting damage or reduce quality of life.

He says his immediate attention changes from week to week and is currently focused on Muse: learning from users, expanding access and improving its models. His ambition is for Meta to build leading models.

On future breakthroughs, Zuckerberg says researchers do not know in advance exactly what they will find. He thinks scaling infrastructure may take the field very far, even without a fundamental new architecture, although he also expects more efficient architectures to be possible.

He describes a large Ohio computing cluster being used for a next generation of models and a larger cluster under construction in Louisiana. He guesses that training on multi-gigawatt clusters could produce something approximating AGI, while saying that would not necessarily be the most efficient route.

In his view, large companies invest in scale because it offers a possible path even if researchers do not discover a cheaper breakthrough. He cautions that scaling is not solved: each level brings new engineering problems and unexpected behavior that must be debugged.

He says alignment and trustworthiness are essential if agents are to serve many people. An agent must understand not only a literal instruction but also the user's intent and values, so it does not achieve a goal in an unwanted way.

He expects alignment to rely on both user feedback and work during training. As an example, a model should learn to solve a coding problem rather than obtain a reward by changing a configuration or exploiting a shortcut.

Zuckerberg says Meta spent extra months improving Muse's discretion and the security of its virtual machine. He argues that making a product safe and reliable serves both users and the company, and believes AI labs have strong incentives to get those properties right.