Public transcript

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

By: Cheeky Pint
Length: 2 h 14 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Mark Zuckerberg and Alexandr Wang of Meta Superintelligence Labs sit down with Stripe's Patrick Collison to introduce Muse Wonder, their new frontier-class model, and discuss how Meta rebuilt its AI effort after Llama 4. They walk through how the model was trained, the tens of billions of dollars and multi-gigawatt clusters behind it, and why they believe personal agents are the defining AI product. The conversation also covers agentic commerce, why the frontier has stayed competitive, and how the MSL team was assembled around talent density.

Key points

Muse Wonder is Meta's new frontier-class model, available in two versions: Muse Wonder and the higher-end Muse Wonder Full, with the paid subscription tier getting Wonder first.
The model is claimed to be state-of-the-art on personal and professional agent tasks, with highlights in active memory, active focus (long-horizon persistence), domain expertise in areas like legal and finance, and frontier coding performance on benchmarks like DeepSWE and SWE-Marathon.
Zuckerberg uses Muse to coach him in MMA by watching gym cameras and giving feedback between rounds, which has become an internal eval for the product.
Meta rebuilt its AI team after being unhappy with the Llama 4 trajectory, with Zuckerberg recruiting Alexandr Wang and standing up MSL roughly a year before this conversation.
Training a frontier model now costs tens of billions of dollars, with a gigawatt cluster costing around $15 billion a year to operate and $50-60 billion in CapEx to build.
Meta is building the Prometheus cluster in Ohio (over a gigawatt) and the Hyperion cluster in Louisiana (scaling to 5 gigawatts).
The executives describe the decision to commit $100-200 billion before seeing the models, with CFO Susan Li stressing to Zuckerberg how bad failure would be.
They argue the frontier has stayed competitive because of uneven execution across labs, the anti-hegemonic force of distillation, and industry-wide preference for a diverse ecosystem.
Agentic commerce is going vertical, with Muse users fully delegating purchases; Zuckerberg became a new regular Instacart user through Muse-assisted baking with his daughter.
Meta plans to monetize through subscriptions, commerce revenue sharing, an API, and a new Meta Enterprise Platform led by CJ Desai.

Questions it answers

00:00Muse use cases

What are people actually doing with Muse?

Zuckerberg's Muse sources obscure teas weekly from farms in Taiwan and Japan. Wang's favorite use is MMA coaching: cameras in his gym let Muse watch his sparring and give feedback during 60-second breaks between rounds, which has become an internal eval for the product.

Muse may be the fastest-growing agent or AI app to date per third-party reports.
The MMA coach use case forces Muse to build its own video-processing pipeline under time pressure.
Zuckerberg sees everyone implicitly writing tons of code through agents without knowing it.

04:36Model as product

Why do AI winning products look simple and convergent?

The guests push back on the idea that AI products collapse into one agreed shape. They argue the differentiation now lives in the model itself — personality, proactivity, discretion, privacy and security — rather than in traditional product specs, so 'the model is the product.'

Meta trains models end-to-end for personal superintelligence, weighting multimodality heavily.
Dedicated VMs for agents matter for privacy and security confidence.
Muse has no model selector or effort selector, deliberately, for simplicity.

14:58Muse Wonder launch

What is Muse Wonder and what does it do well?

Muse Wonder is Meta's new frontier-class model, significantly bigger than Muse Spark, trained via a tiered ladder of increasingly large pre-trains. It ships as Muse Wonder and Muse Wonder Full, with highlights in active memory, active focus, domain expertise (legal, finance, health, tax), and state-of-the-art coding on benchmarks like DeepSWE and SWE-Marathon.

Wonder is served first in paid subscriptions; Spark remains the free tier.
Active memory is built into the model itself, not an external harness.
Meta expects the time between bigger model classes to compress going forward.

24:30Benchmarks and evals

Are external benchmarks still useful for grading models?

Benchmarks saturate quickly, so Meta increasingly steers by internal evals tied to their product focus. They quantify 'big model smell' as the gap between in-distribution and out-of-distribution performance, and see a huge step up from Spark to Wonder. Relying only on public benchmarks means outsourcing product thinking to whichever PhD student made a viral benchmark.

Discretion — getting things done while disclosing as little as possible — matters more for personal agents than coding agents.
The most important metric is whether people have a great experience using the model.
Internal evals detail dozens of small capabilities that each matter ~2% but aggregate into magic.

30:30Training pipeline

What actually goes into training a frontier model?

Pre-training hoovers up massive multi-modal data — public internet, books, images, audio, video, plus domain corpora like health and legal. Post-training has two stages: SFT initializes behaviors like personality, then RL dramatically raises the elicitation rate. RL environments must be curated so the model isn't pressured toward misaligned behavior.

The research team is roughly 100-200 people, intentionally small with high compute per researcher.
In RL, less data can be more; quality of environments matters more than volume.
Point releases like Spark 1.1 to 1.3 were pure RL improvements on the same pre-train.

49:10Compute and risk

How much does this cost and how risky was the bet?

A gigawatt cluster costs about $15 billion a year to operate and $50-60 billion in CapEx; training a frontier model overall runs tens of billions. Zuckerberg committed $100-200 billion before seeing the models, with CFO Susan Li stressing how bad failure would be. He argues the downside is hedged because if AI works at all, others would pay well for the compute.

Meta is building Prometheus in Ohio (>1GW) and Hyperion in Louisiana (5GW).
Data centers were planned in 2023-2024, predating MSL.
Zuckerberg frames it as entrepreneurship on hard mode — the opposite of Lean Startup.

1:13:00Agentic commerce

How will agents change how businesses and commerce work?

Muse has good browser use so no integration is required, but Meta is giving important partners direct API access. Every business will eventually have agents, and agentic commerce demand will grow as a share of the pie. Zuckerberg became a new weekly Instacart user through Muse-assisted baking with his daughter, showing agents expand the total market.

Muse is one of the first products where people fully delegate purchases to agents.
Businesses that block agents are betting against consumer benefit and will likely be disrupted.
Agents effectively having their own accounts may dramatically increase the total addressable market for commerce.

1:27:10Why no runaway winner

Why hasn't one lab pulled away from everyone else?

Three forces: execution quality varies across labs and averages out; distillation acts as an anti-hegemonic force packing strong capabilities into small models; and many people and companies actively prefer a diverse ecosystem and put dollars toward it.

If every leader had executed perfectly while in the lead, they might have stayed in the lead.
Small models today beat the biggest models of one or two years ago.
Meta's differentiation is its focus on personal and professional agents over automating knowledge work.

Notes

Muse Wonder Announcement

Meta introduces Muse Wonder and Muse Wonder Full, a new frontier-class model line. Muse Spark (code-named Avocado) was the stepping stone; Wonder (code-named Watermelon) is a significantly bigger pre-train.
Claimed state-of-the-art on personal and professional agent tasks, plus strong coding results on DeepSWE, SWE-Marathon and Terminal-Bench 4.
Key capabilities: active memory (in-model, not a harness), active focus (persistence through long, ambiguous tasks), and domain expertise (state-of-the-art on finance and legal evals).
Wonder goes to paid subscribers first; Spark remains the free tier. Also available via the Meta Model API, OpenRouter, and Muse Code.

Training a Frontier Model

Pipeline: pre-training (massive multi-modal data ingestion) → SFT (supervised fine-tuning, where personality is molded) → RL (reinforcement to elicit behaviors at high rates).
Pre-training is high-stakes and scientific with few samples; the team went through "all 7 stages of grief." RL benefits from quality over quantity, and environments must be curated to avoid driving misaligned behavior.
External benchmarks saturate quickly; Meta increasingly steers by internal evals. They quantify "big model smell" as the gap between in-distribution and out-of-distribution performance.
The research team is roughly 100-200 people — intentionally small, high talent density, with heavy compute per researcher.

Compute and Risk

A gigawatt cluster costs ~$15B/year to operate and ~$50-60B in CapEx. Total frontier model training is "tens of billions of dollars."
Meta is building Prometheus (Ohio, >1GW) and Hyperion (Louisiana, 5GW).
Zuckerberg committed $100-200B before seeing the models; CFO Susan Li repeatedly stressed the stakes. He argues the bet is backstopped because if AI works, others would pay for the compute.

Product and Business

"The model is the product" — personality, proactivity and discretion are trained in, not built as UI.
Agentic commerce is going vertical; Muse users fully delegate purchases. Zuckerberg became a weekly Instacart user through baking with his daughter Aurelia.
Monetization: subscriptions, commerce cuts, API, and the Meta Enterprise Platform under CJ Desai.

MSL Origins

After Llama 4 (April 2025), Zuckerberg concluded a full reboot was needed and recruited Alexandr Wang, betting on his strategic clarity despite Wang never having trained a frontier model.
Culture: flat structure, no non-technical managers, no deadlines, everyone stays technical and uses agents.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.