Muse Wonder, an AI app, is experiencing rapid growth, prompting a discussion about its underlying strategy. The conversation explores how Meta's AI development prioritizes the model as the product, multimodal capabilities for personal agents, and a focus on 'talent density' over team size. This approach reflects Mark Zuckerberg's leadership, characterized by boldness and a Bayesian method of updating his worldview, which has been crucial for Meta's long-term AI investments and its ambition to become a full-stack technology company.
1. The model is the product
In AI product development, the underlying model itself has become the primary product. This means that much of the complex product work, traditionally focused on user interface and experience, now revolves around the model's capabilities, its 'personality,' and its security and privacy features.
For Muse, significant effort went into designing the privacy and security framework for handling user data and connected services.
2. Outsourcing product thinking to external benchmarks
External benchmarks for AI models quickly become saturated, making them less useful for guiding advanced development. A significant problem with over-reliance on these benchmarks is that it effectively outsources critical product thinking to their creators, who are often PhD students with different objectives than a product company.
This leads to a strategic shift towards internal, use-case-specific benchmarks for more effective model development, ensuring alignment with specific product goals.
3. Multimodal AI for personal agents
Meta's strategy for AI development centers on training models end-to-end to function as highly effective personal agents. This approach emphasizes multimodal capabilities, recognizing that human interaction with the world is inherently multimodal.
This contrasts with an earlier industry consensus that prioritized coding as the paramount skill for AI, demonstrating a shift towards consumer-centric, natural interaction.
4. Talent density over team size
Meta's AI development strategy, particularly for projects like Muse Wonderer, prioritizes 'talent density' over sheer team size. This means focusing on recruiting the best possible individuals and providing them with substantial compute resources.
For the Muse Wonderer project, this involved a deliberate 'reboot' of the effort, pulling top talent from within Meta and aggressively recruiting external experts to create a high-talent-density environment.
But going back to the beginning, I think one of the things that was very interesting and challenging in standing the thing up was like, you really just wanted this like very high talent density effort.
5. Long-term AI investment: entrepreneurship on hard mode
Developing frontier AI models is akin to 'entrepreneurship on hard mode,' demanding extensive long-term planning and massive upfront capital investments in compute infrastructure. For Muse Wonder, planning began in August of the previous year, with significant compute investments made years in advance, such as the Prometheus cluster in Ohio and the Hyperion cluster in Louisiana, scaling to gigawatts of power.
This pre-training phase is high-stakes and scientifically challenging, often leading to periods of uncertainty and 'troughs of sadness' for researchers due to the immense investment and unknown outcomes.
6. Meta's full-stack technology company strategy
Meta aims to be a full-stack technology company, building its own data centers, chips, and low-level infrastructure to control the entire technology stack. This strategy is driven by the belief that controlling the core infrastructure and models is essential for delivering the best user experiences and avoiding reliance on external platforms.
Meta's existing profitable core business provides the financial stability and credit to underwrite these massive, long-term AI investments. Mark Zuckerberg's understanding of 'power laws' further justifies these high-risk investments, as even if a specific AI initiative fails, the underlying compute infrastructure retains significant value and can be utilized or sold.
7. The rise of agentic commerce
AI agents are poised to become a standard business component, much like websites and social media accounts are today. This will lead to a significant shift in commerce, termed 'agentic commerce.'
Muse exemplifies this by enabling users to fully delegate purchases, reducing friction and automatically applying user criteria. This marks a 'step change' in shopping, offering significantly greater personalization through a holistic understanding of the user's life, unlike siloed historical recommendation systems.
Furthermore, AI agents can expand the total addressable market for commerce by creating new demand and converting non-users into active customers, as demonstrated by new Instacart subscribers driven by Muse.
So anyone who wants to do that, we eventually want to get there. and then I think the fullness of time is that, you know, just like today every business has, a website and social media accounts and email addresses and all these things, every business is going to have agents too.
8. Mark Zuckerberg's bayesian leadership and boldness
Mark Zuckerberg's leadership is marked by a unique blend of boldness and a Bayesian approach to updating his worldview. Once he internalizes new information, he rapidly and decisively pivots resources and actions to align with his updated beliefs, even if traditional metrics don't yet validate the change.
This was crucial for the rapid establishment and success of MSL, as he fully committed to the AI initiative, cleared his calendar, and empowered key talent like Alex, demonstrating a 'very high learning rate' and a willingness to make unconventional bets.
9. Staying technical as an AI leader
In rapidly evolving fields like AI, leaders must actively maintain their technical proficiency and direct involvement in coding. Without this, managers can quickly become non-technical and lose relevance, posing a significant risk to effective leadership.
Alongside technical engagement, authenticity in communication is also crucial. While it may sometimes lead to imperfect delivery, genuine communication is preferable to being overly scripted, fostering trust and engagement within dynamic teams.