More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
Picture the head of Meta’s AI effort being pulled aside by the company’s chief financial officer. Alexandr Wang recalled her message this way: if the new model worked, great. If it didn’t, the financial consequences would be “really, really, really bad.” That’s the scale of the gamble behind Muse, Meta’s personal AI agent. It’s also worth getting one distinction straight at the start: Wang co-founded Scale AI. He did not found Muse as an independent company. Muse was built at Meta, with Wang leading a major part of its AI overhaul.
The story begins with a setback. After Meta’s Llama 4 models received a lukewarm response in April 2025, Mark Zuckerberg reached out to Wang, then known for building Scale AI into a data-labeling business. Wang said a WhatsApp message arrived while he was in a staff meeting. Their first call lasted about 44 minutes, by Zuckerberg’s recollection. Two days later, Wang sent him a long memo. One argument stood out: don’t just hire more people; assemble an unusually strong team. Zuckerberg was prepared to rebuild Meta’s AI effort, and Wang was the person he chose to help lead it.
Meta took a $14.3 billion, 49 percent stake in Scale AI to bring Wang aboard. He became Meta’s chief AI officer and helped build Meta Superintelligence Labs. But the idea for Muse wasn’t Wang’s alone. In an account of Wang’s October 2026 interview with David Senra, Wang traced it to a June 2025 personal-superintelligence memo written with Zuckerberg, Nat Friedman, and others. The aim was more ambitious than a chatbot: an assistant that could remember what mattered to someone and carry out tasks for them.
The direction sharpened in February 2026. Wang said experiments with an agent called OpenClaw convinced people inside Meta that personal agents were becoming practical. While much of the industry’s attention was on agents that write code, Meta bet on one for everyday life. A Muse prototype came together in roughly one or two weeks and was shown to Meta’s board. But a board demo is not a product you can safely hand to millions of people. Wang said the team then spent months testing more than a hundred behaviors the agent needed to get right. Earlier model versions passed about 80 checks and failed 20; a tuned version of Muse Spark 1.3 was the first to clear them all.
Muse launched in the United States on September 8, 2026. Meta described an agent that can use a browser, fill out forms, help plan projects, and keep working after a person closes the app. That power creates a different sort of risk from an ordinary chatbot. Give an agent access to your email, accounts, and purchases, and a mistake might become an action. Meta says Muse runs in a dedicated virtual machine with a separate Sentinel agent reviewing what reaches the internet. The company also says Muse asks permission before sensitive actions, including sending email or making purchases. Those are Meta’s stated safeguards, not a guarantee that every risk has disappeared.
Then there is the money. Zuckerberg described asking Wang whether the effort would work before committing to spending on the order of $100 billion or more. Fortune reported that Meta expected total 2026 capital expenditures of $130 billion to $145 billion. That is company-wide spending, not a Muse-only price tag, but it shows why the AI bet mattered so much. There was leadership risk, too: Wang had built a major AI-data company, yet had not previously trained a frontier model. He had to recruit a new research team and deliver results while Meta spent at enormous scale.
By October 2026, Muse had shipped, and Meta had announced newer Muse Wonder models. In its October 12 announcement, Meta said those models were available to developers and Muse Code users, while their rollout into the Muse agent was still ahead. So the ending is neither a triumph declared nor a failure foretold. Wang’s team turned a memo, a rapid prototype, and months of reliability work into a launched product. The unresolved question—the one behind that warning from Meta’s CFO—is whether an agent people trust with real tasks can ultimately justify the scale of the bet.