Inside Meta's grind to turn Muse from a clunky chatbot into an AI sensation

Meta CEO Mark Zuckerberg hypes up its Muse AI assistant at the company's Connect event. AP Photo/Jeff Chiu Meta's top AI staff fretted about Muse days before launch as it struggled to speak good English. Ex-GitHub CEO Nat Friedman's team worked around the clock to turn the AI assistant into a hit. After a "long ass" grind, Muse is now a "banger," says Meta's 29-year-old AI chief, Alexandr Wang. A year ago, Meta was struggling to convince Silicon Valley that it could keep up in the AI race. Now it has one of the industry's hottest products. Muse, Meta's new personal AI assistant, has rocketed to the top of the App Store and become a victory lap for CEO Mark Zuckerberg's massive AI spending. Meta's Connect conference this week was all about the app, which AI chief Alexandr Wang — who is 29 and has been posting risqué memes of Muse's mascot all week — called a "banger." Getting there required months of painstaking work, a team led by former GitHub CEO Nat Friedman, and thousands of employees pulled from across Meta to teach the agent how to actually get things done, employees told Business Insider. Meta also snapped up top AI agent talent through acqui-hires, including the leaders of Dreamer, an AI agent startup. Whether Meta can sustain Muse's momentum is another question. Competitors like Instinct are thriving, and OpenAI is widely expected to launch a rival. Privacy concerns have also arisen, as Muse requires broad access to users' accounts to function. Meta employees are still reeling from a year of mass layoffs, employee surveillance, and low morale. For now, at least, Muse's success — and a more than 20% increase in Meta's stock price since its launch — has become a morale boost, employees said. It also proves that Meta remains firmly in the AI race after being seen as a laggard a year ago. "Muse is busy working for millions of people because of the hard work of a bunch of people at Meta — and we're still cooking," a Meta spokesperson said. Nat Friedman grinds to turn Muse into a hit Muse's success was far from inevitable. Since September 2025, Friedman, who heads the superintelligence unit's Products and Applied Research team, has been working hard on Muse, then codenamed "Hatch." As recently as two weeks before Muse's launch this month, the chatbot struggled with speaking good English, a person directly familiar with the matter said — and it came as a huge relief when it finally started scoring well on coherence tests. Earlier this year, some of Meta's top AI employees were also worried about Muse, an employee told Business Insider. They had tried it out and found it working poorly — it felt like just another chatbot, but with a worse underlying AI model and extra steps, that person added. Friedman's team eventually pulled it off. Tens of thousands of Meta employees tested Muse for months, coming up with a range of use cases for the agent, which became key to the effort. That data helped Friedman's team improve Muse's ability to do things like book meetings and cancel plane tickets: the painful work of getting better at increasingly complex tasks, or "hill-climbing," as Silicon Valley calls it. The team trained Muse to be considerate so that it didn't overshare, like telling people someone couldn't make a meeting because they were hungover, the person familiar said. Meta also focused on ensuring it could interact with the internet seamlessly by using tools it hadn't used before. "Our researchers did a ton of work teaching agents how to choose the right tool for a job," Wang said this week at Connect. Friedman's role was critical. Zuckerberg brought in the executive last year as part of a deal to buy a stake in his venture capital fund, NFDG. Last year, Friedman and Wang cowrote a memo outlining a "superintelligent personal sidekick" that would act as a trusted friend and advisor. Friedman is unabashedly technical and hands-on. Internally, the executive is opinionated and proud of it, chafing with other teams across Meta's sprawling operations, one of the employees said. Meta's communications chief, Andy Stone, recently called him "Nat the Truthteller" after he derided a report about Muse's security as written by "dummies." He also moves fast. His bare-bones personal website spells out his philosophy in characteristically blunt terms: "slow is fake" and "it's important to do things fast," it says. Friedman has pushed Meta's AI staff to ditch slow internal tools in favor of faster external ones, such as Vercel. When the independent AI agent OpenClaw took the tech world by storm this winter, Friedman acted quickly. The free desktop-based tool was popular with developers, but mystifying to everyday users. Friedman aimed to build a far more accessible version from scratch. He issued a two-page internal memo titled "The Claw Is The Law," praising the product, and also bought hundreds of Mac Minis for Meta's superintelligence team to run the agent. While Muse is user-friendly now, making it all happen took "months of blood, sweat, and tears," the person familiar said. Although it took a "long ass time," now it's something that anyone's "grandma can use," Wang later posted. Meta goes on an AI hiring spree Meta also went on a hiring binge for researchers and engineers in agentic AI. It hired David Singleton and Hugo Barra, the cofounders of Dreamer, a startup helping people build their own agents, along with the rest of their team. Singleton has worked on Muse since he began at Meta in March, helping design the core technology that enables the app to store sensitive information securely, he posted on X. The efforts worked. Although the arrival of the viral AI agent Instinct initially upstaged Meta, Muse took off by tapping into Meta's enormous built-in audience across its apps — including among people who had never used AI agents before. Meta wasn't the only one on a hiring spree. OpenAI hired OpenClaw's creator, Pete Steinberger, in a coup earlier this year. Meta was especially worried that it would be upstaged by a large competitor, the person familiar with the matter said, and the team was shocked as the months went by without such an entrant. More to come after Muse Wall Street is cheering on Muse, seeing it as a way for Meta to become more than an ad company. Zuckerberg told Alex Heath's Sources earlier this month that he wants people to shop and save money on Muse, which Meta plans to make money from by taking a small cut of each transaction. "We see the Meta bull case increasingly playing out," Mizuho Americas analyst Lloyd Walmsley wrote in a note. Users have praised how easy it is to set up Muse. Hadi Javeed, the cofounder of healthcare startup RevelAi, said his dad saw an ad for Muse in his Facebook feed and downloaded it, promptly using it to buy a new pair of shoes. "They completely deleted all the setup that was required," Javeed said. At the same time, Javeed, like many others, has raised data privacy concerns about Muse. By design, an app like Muse creates a new order of security and privacy risks, such as the app buying things without a user's permission. Privacy and trust remain key roadblocks to broader adoption of Muse, Bank of America analysts wrote. Zuckerberg has said that the company held back Muse's launch due to safety and product liability concerns. He and Friedman also personally reached out to Signal creator Moxie Marlinspike to help co-design Muse's security architecture, Zuckerberg told Heath. For now, Meta is on a high. The company has hinted that much more will come as it works on a new AI model, Watermelon. It says the model has already caught up with OpenAI. Wang says Meta has other ideas that are "just as big" as Muse. "The kitchen is just getting warmed up," he posted on X. Have a tip? Contact Meta reporter Charles Rollet securely on Signal at charlesrollet.12 or +1-628-282-2811. Use a personal email address and a nonwork device; here's our guide to sharing information securely. Additional reporting by Pranav Dixit. Read the original article on Business Insider

Read it at Business Insider ↗

← Front page