Cheeky Pint: Zuckerberg & Jensen Wang on Meta’s $100B AI Moonshot

SPEAKER_1 (0:01)

When a company wagers one hundred to two hundred billion dollars on a future it can't yet see, those stakes aren't theoretical anymore. Today, we're unpacking the Cheeky Pint conversation where John Collison sat down with Mark Zuckerberg and Jensen Wang to discuss Muse, Muse Wonder, and Meta's massive AI bet. And the growth numbers are staggering. Zuckerberg mentioned that third-party reports suggest Muse may be the fastest-growing agent or AI application in existence. But what really caught my attention were the use cases—one Muse sources obscure teas from Taiwan and Japan, while another coaches MMA by watching gym cameras and giving feedback between rounds.

That MMA example actually points to something bigger. Zuckerberg argues that people may soon be writing huge amounts of code without even realizing it. Agents are already assembling video and image pipelines on the fly, which is why Meta trains its models end to end as personal assistants, not just coding tools. Which brings us to multimodality, right? Zuckerberg's point was that people experience the world through talking, listening, and watching, so the best products must do the same. He sees glasses and other devices eventually becoming pure multimodal interfaces—hearing what you hear, seeing what you see, creating what you need directly in view.

Exactly. And he had this fascinating line: "the model is the product." He insists the most interesting differences now come from model personality, tool use, and intentional choices, not just raw scale. Muse Wonder and Muse Wonderful reflect years of scaling toward more nuanced thinking. Those models bring major capabilities too. Active memory lets the system remember things over days or months, and Active focus helps it persist through long tasks and setbacks. Coding performance has jumped to state-of-the-art levels, though the biggest model will mostly live behind paid subscriptions because it's expensive to serve. What struck me was how Meta evaluates progress. Both leaders warned that public benchmarks can become targets, so Meta leans on internal tests tied to real use cases. Zuckerberg described the gap between in-distribution and out-of-distribution performance as "big model smell"—a signal of whether a model will actually generalize. And building these frontier models is full-stack work. Pre-training absorbs books, internet data, and multimodal inputs, while post-training uses supervised fine-tuning and reinforcement learning. Meta's spent heavily on infrastructure, including large clusters in Ohio and Louisiana, because this work is long-term. That long horizon changes everything. Zuckerberg called building frontier AI entrepreneurship on "hard mode" because investments come long before proof points.

But owning the stack gives Meta the chance to define the next generation of computing. His message kept circling back: stay technical, stay adaptable, move quickly when the frontier shifts.

The business strategy widened too. Zuckerberg says every company will eventually need agents, and businesses that block them risk disruption. Meta envisions a world where Muse handles commerce, bookings, and everyday tasks while partners integrate through APIs. In the end, this episode is about more than one model release. It's about a company betting that personal super-intelligence, not just knowledge-work automation, will define the next era.

As Zuckerberg put it: stay at the frontier, stay in the details, stay technical. The episode ended with Guinness from Dublin in hand, but the message was serious.

Meta is treating AI as a generational platform shift, convinced that winners will be the companies willing to build, invest, and keep pushing when everyone else falls behind.