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

Summary

In an episode of Stripe’s Cheeky Pint interview series, John Collison sits down at a pub with Meta CEO Mark Zuckerberg and Chief AI Officer Alexandr Wang to discuss the release of Meta’s new frontier models, Muse Wonder and Muse Wonder Full. They discuss Meta Superintelligence Labs (MSL), model scaling, multimodal pretraining, evaluation methods, and agentic commerce.

What is shown

• [00:00 - 00:24] Introduction of Mark Zuckerberg, Alexandr Wang, and host John Collison sharing pints of Guinness in a pub setting.
• [01:23 - 03:45] Zuckerberg pulls out his smartphone and demonstrates an eval use case where Muse analyzes video camera footage of his MMA gym training sessions to give posture and technical feedback between rounds.
• [14:41 - 18:41] Discussion of the release of Muse Wonder and Muse Wonder Full (codenamed “Watermelon”), including active memory architecture and frontier coding capabilities.
• [80:02 - 80:50] Zuckerberg places a prototype hardware gadget (a small hockey-puck-shaped device with a button) on the table, presses it, and asks it to order a watermelon via voice command to Stripe’s office via Instacart.
• [133:55 - 134:15] At the end of the interview, Zuckerberg surprises Collison by bringing in a full keg of Guinness shipped directly from Dublin.

---

Claims & numbers

• Adoption & Usage: Wang claims Muse is possibly the fastest-growing agent or AI app in history [00:33].
• Architecture & Features:
  • Wang states Muse Wonder introduces “Active Memory” (in-model compaction rather than external context management) and “Active Focus” (long-horizon multi-step persistence) [18:18 - 21:20].
  • Wang claims Muse Wonder is state-of-the-art across agent benchmarks, specifically citing SWE-bench, DeepSWE, and SWE-marathon, as well as Terminal-Bench 4 [21:28 - 21:40].
  • Wang claims Muse Wonder achieves a large jump over Muse Spark in out-of-distribution vs. in-distribution performance (“big model smell”) [26:17 - 26:55].
• Release & Pricing:
  • Muse Wonder is available via paid consumer subscriptions on Muse and via API (Meta Model API and OpenRouter), while free users receive access to Muse Spark [22:33 - 23:49].
  • Muse Wonder Full is available to subscribers of Muse Code [23:45].
• Compute & Capital Expenditure:
  • Zuckerberg states that Meta built the Prometheus cluster in Ohio (scaling beyond 1 GW) and is constructing the Hyperion cluster in Louisiana (targeting 5 GW) [49:46 - 49:58].
  • Zuckerberg estimates that operating a 1 GW AI cluster costs on the order of $15 billion per year, with capex for building such a facility reaching $50 billion to $60 billion [53:30 - 54:49].
• Organizational Structure:
  • Zuckerberg notes that the core MSL research team consists of fewer than 200 people, closer to 100 researchers, supported by infrastructure teams [34:53 - 35:03].
  • Zuckerberg states that the standing start of MSL began around April 20–21, 2025, after Llama 4 launched on April 4, 2025 [101:06 - 101:23].
• Hardware & Audio:
  • Meta launched “Muse Gadgets” (open-source ESP32 firmware and Linux SDK) and the “Home Link” dongle [86:29 - 86:40].
  • Collison notes that OpenRouter routes over 10 trillion tokens per day, in the context of Stripe’s agreement to acquire OpenRouter [79:47].

---

Notable quotes

• Mark Zuckerberg [03:04]: “Everyone is actually going to be writing tons of code; they’re just not going to know that they’re writing code.”
• Mark Zuckerberg [12:04]: “The model is the product.”
• Alexandr Wang [42:53]: “If I were to say one broad statement about AI, it’s just: scale everything.”

---

Assessment

This is an authentic long-form interview and product launch discussion featuring the heads of Meta and Stripe, including live demonstrations of the Muse smartphone app and a prototype hardware gadget. While the voice-ordering interaction is shown functioning in real time, the high-level benchmark numbers and datacenter spending metrics are verbal executive claims rather than audited technical reports.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.