Meta's new AI agent helps with baking and gym reviews

Meta CEO Mark Zuckerberg described using the company's new AI agent, Muse, for family chores and fitness feedback, highlighting both its utility and its occasional lack of practical sense.
Mark Zuckerberg has revealed that he is testing Meta’s new personal AI agent, Muse, with a surprisingly domestic assignment: baking with his three-year-old daughter. Speaking on the Sources podcast, the Meta CEO explained that he asked the system to plan weekend baking projects that are manageable for a toddler and an adult with limited culinary experience. The agent is tasked with selecting recipes and even coordinating ingredient delivery through services like Instacart to ensure everything is ready before they start.
However, the early results suggest the system may need to better understand the concept of difficulty. Zuckerberg noted that Muse initially suggested making cake pops, a task he described as surprisingly difficult for their skill level. "Don't start with cake pops," he advised, indicating that the AI struggled to calibrate its recommendations to the user's actual ability. He reports back to the agent after each session so it can learn from what worked and what did not, aiming to refine its future suggestions.
Muse handles long-term family tasks
Beyond the kitchen, Zuckerberg uses the agent for various family-related responsibilities. He mentioned that Muse monitors availability for mountain-climbing permits for his older daughter and alerts him when one opens up. The AI also created a guide to the video game Civilization for another child, subsequently suggesting how the gameplay could be turned into a history lesson. These examples illustrate Meta's broader claim that Muse is designed to pursue longer-term goals and retain user preferences, distinguishing it from standard chatbots that only respond to immediate prompts.
The agent’s role extends to personal fitness as well. Zuckerberg uses it to review footage from cameras in his mixed martial arts gym. The feedback can be blunt, with the agent once telling him, "It looks like you really gave up." He acknowledged the accuracy of the critique, noting he was indeed tired during that round. This demonstrates the system's ability to analyze video data and provide specific, if harsh, performance feedback.
Privacy concerns and access controls
Meta plans to offer Muse with a large free allowance of about 100 million tokens per week, alongside a paid subscription option. The company envisions the agent helping users run small businesses and manage ad campaigns by accessing outside services. However, this ambition requires users to trust the system with sensitive data and account access. To address this, Zuckerberg said Meta built a confidential virtual-machine system that the company itself cannot inspect, along with monitoring agents and approval prompts for actions such as payments and sending emails.
According to reporting by GN technics/ai (en-US), these features highlight the trade-off between convenience and privacy inherent in autonomous AI agents. While the system can organize complex tasks and interact with external services, it requires a high level of user trust. For now, the agent's performance in Zuckerberg's household exposes a potential limitation: it can efficiently organize ingredients and schedule tasks, but it may not always possess the practical wisdom to know that cake pops are an ambitious place to start for a novice baker.






