Mark Zuckerberg and Alexandr Wang
Muse Wonder and the $100B AI Bet

Muse writes and improves the code needed for combat sports coaching behind the scenes. To build personal agents like this, Meta is focusing on the ability to understand not only text, but also video and audio.

The difference between AI products comes less from their appearance than from how the model behaves and speaks. Teaching features and personality to the model is now part of building the product itself.

Meta says Muse Wonder can compete with the best models available. It is designed to remember more over time and complete complex tasks from start to finish, with a focus on personal and workplace agents.

Strong models are built by evaluating them against real-world uses, not just external benchmark scores. Developers must expand the data and computing used for pretraining while refining the tasks and methods used in reinforcement learning.

Developing cutting-edge AI is a long-term effort that requires betting huge sums before the results are clear. Meta believes it needs its own superintelligence and the facilities to support it if it wants to build the products of the future itself.

Agents can take the hassle out of buying things and shop on a person’s behalf based on their individual needs. This will not just make existing purchases easier. It could create new demand and expand the commerce market.

A sponsor message read by John Collison. OpenRouter, which Stripe has agreed to acquire, connects many AI models through one API and routes around outages.

AI is moving beyond delivering information toward taking action, receiving instructions by voice and showing results through vision. Smart glasses and home devices will become new ways to use these systems.

Meta plans to fund its AI investments through Muse subscriptions, commerce, and enterprise services. It believes that even modestly increasing growth in its existing core businesses with AI could provide much of the money it needs.

AI models will advance for different kinds of work, while distillation can help later entrants catch up with leaders. Companies and developers also have reasons to favor a world where many models coexist instead of one company controlling the market.

Meta Superintelligence Labs began by recruiting exceptional talent and shifting the company toward an AI-first approach. Its leader needs sound judgment, flexibility, and the ability to work directly with the technology.

When Mark Zuckerberg becomes convinced by new information, he can change course quickly and decisively. Meta is willing to invest heavily in changes that have not yet appeared in its results, showing its belief in AI through action rather than words.