Muse Wonder: Zuckerberg and Wang's $100B AI

Summary · 5 min read instead of 2 h 15 m

Mark Zuckerberg is the founder and CEO of Meta, which builds its own data centers, chips and infrastructure and is scaling AI compute past five gigawatts. Alexandr Wang is Chief AI Officer of Meta Superintelligence Labs (MSL), which he joined from a data business in which, per Collison, Meta took a 49% stake. MSL launched the Muse agent and its new frontier model, Muse Wonder.

Zuckerberg and Wang say Muse Wonder is the first MSL model at the frontier, tuned for personal and professional agents with active memory, persistence, domain expertise and stronger coding. They argue that in AI the model is the product, so product decisions become training decisions, and that internal evals matter more than public benchmarks. Zuckerberg describes committing $100 billion or more to compute before the model was proven, and says the risk was cushioned because compute is useful to others if AI works. Both expect a few labs, not one winner, and describe MSL's reboot around talent density, a flat technical culture and no deadlines. They also see agentic commerce, new devices and enterprise sales funding the build-out.

Minute by minute

1. 0:36

   Personal agents already doing real tasks. Wang says his Muse, named Euler, buys obscure teas from farms in Taiwan and Japan each week. Zuckerberg has Muse watch cameras in his gym and coach him in MMA during one-minute breaks between rounds, building its own video processing pipeline and speeding itself up when told. He says this shows how everyone will write lots of code without knowing it.

2. 4:43

   Multimodal ability is central to Meta's models. Zuckerberg says Meta trains its models end to end to be good personal agents, with heavy visual perception data. Wang says the industry consensus had been that only coding mattered, but consumer products depend on strong multimodal skills. Zuckerberg expects glasses to become a purely multimodal product, possibly without an app store.

3. 7:24

   The model is the product. Collison suggests winning AI products converge on simple, shared designs. Zuckerberg disagrees, pointing to choices like giving each agent a dedicated VM for privacy and security. Wang argues that AI product decisions are expressed as model behavior, personality and proactivity, set through training data and evals. Both favor no model or effort selector, which Wang says confuses ordinary users.

4. 14:31

   Muse Wonder reaches the frontier. Zuckerberg says Muse Wonder comes from a significantly bigger pre-train than Muse Spark, with a higher-end Muse Wonder Full version. Wang highlights active memory, active focus for long tasks, domain expertise and frontier coding, citing state-of-the-art results on DeepSwee and Swee Marathon. Wonder starts mainly in paid subscriptions, while Spark stays free, with API, OpenRouter and Muse Code access.

5. 24:47

   Internal evals matter more than public benchmarks. Zuckerberg says public benchmarks saturate quickly, and Meta increasingly steers by its own evals. Wang describes a metric for big model smell, measured as the gap between in-distribution and out-of-distribution performance, which improved sharply from Spark to Wonder. Zuckerberg cites discretion, such as booking a restaurant without disclosing a medical condition, as a capability personal agents need more than coding agents.

6. 30:10

   Small team, heavy compute per researcher. Zuckerberg says Wang's first plan argued that talent density matters more than headcount. The research team is under 200 people, close to 100, with infrastructure staff beyond that. Compute is concentrated rather than spread across competing efforts. Pre-training pulls in as much public, proprietary and multimodal data as possible, and Zuckerberg says scaling data alongside compute is one of the harder problems.

7. 40:58

   Post-training and alignment to the user. Wang explains that supervised fine-tuning instills behaviors and reinforcement learning raises how often they appear, and that in RL less data can be more. He says RL environments must be curated so models are never rewarded for misaligned behavior. Zuckerberg says Meta treats alignment mainly as following the intentions and values of the person being served, without moralizing, while gracefully refusing catastrophic requests.

8. 48:23

   Betting $100B before the model works. Muse Wonder was planned since the previous August. Zuckerberg cites the Prometheus cluster in Ohio, scaling past a gigawatt, and Hyperion in Louisiana, scaling to 5 gigawatts. He puts operating a gigawatt for a year at about $15 billion and building one at $50 or $60 billion. Wang recalls CFO Susan Li warning him that failure would be very bad. Zuckerberg says the compute would still be valuable to others if AI works.

9. 1:03:31

   Owning the full AI stack. Zuckerberg says Meta is a deep technology company, not an app company, and expects a small handful of firms able to build frontier superintelligence to drive the agenda. He argues others are focused on automating knowledge work, leaving the personal consumer lane open, and that the highest use of superintelligence is invention and creation, not automation.

10. 1:07:16

    Agentic commerce expands the market. Zuckerberg says Muse uses browser use as a backstop while partners integrate through direct APIs, and few businesses have tried to block it. He describes planning weekend baking with his daughter through Muse, which orders Instacart ingredients and once added a watermelon to meet a delivery minimum. He became a regular Instacart user, and suggests agents holding accounts could greatly enlarge commerce.

11. 1:20:48

    Voice, glasses and new gadgets. Zuckerberg says voice assistants struggle with conversational turn-taking, and that hearing information is less efficient than seeing it. Wang notes voice input is exploding, with college students dictating 10 to 15 minutes of stream of consciousness for AI to organize. Zuckerberg describes display glasses, Muse gadgets that phone makers want to integrate, and a home link device that controls lights and speakers.

12. 1:26:56

    Revenue streams to fund the build-out. Zuckerberg lists subscriptions, small cuts of money Muse saves or makes users, a business-in-a-box for small firms, the API and a new Meta enterprise platform. He adds that models accelerate a core business worth hundreds of billions. Collison and Zuckerberg note owning data centers lets Meta depreciate capex, and Zuckerberg says even an H100 is worth more now than last year.

13. 1:33:57

    Why AI is not winner-take-all. Wang says frontier models are diverging in focus, so OpenRouter could help developers understand subtle differences. He attributes the lack of a runaway leader to uneven execution across labs, the power of distillation, and many people choosing a pluralistic outcome. Zuckerberg says a diversity of widely available models is a safer configuration than a few models held by a couple of companies.

14. 1:40:36

    Rebooting the lab from scratch. Zuckerberg says he began calling people for advice weeks after Llama 4 launched in April 2025, and Wang replied with a plan two days after their first call. The full reboot emerged during recruiting, to make the lab attractive. Wang recalls a former lead saying Meta did not need a reasoning model. Both stress a flat structure, no non-technical managers and no deadlines.

15. 2:08:07

    Zuckerberg as a Bayesian boss. Wang says Zuckerberg fully updates his worldview on new information and then reorganizes the company around it, and is unusually bold in sizing bets. He argues Meta treats major shifts seriously before they show in metrics. Zuckerberg says he prefers authentic communication, even when it misfires, over being overly scripted, and finds Wang's posting mostly funny.

Published Oct 12, 2026.