The Full Picture of Meta's AGI Strategy: Llama and Its Future Roadmap
Meta (formerly Facebook) has set a long-term vision for achieving AGI (Artificial General Intelligence) and is driving technology development and an open-source strategy centered on its proprietary large language model, the "Llama" series. With the release of "Llama 4" in 2025, the company has assembled a broad lineup ranging from small and medium-sized models like Scout and Maverick to the large-scale model "Behemoth," and is proceeding with integration into Meta AI applications that serve 1 billion monthly users. This article provides a detailed explanation of Meta's AGI plan from the following six perspectives.
1. Current Status of Meta AI and the Evolution of Llama 4
1-1. Release Status and User Scale
Llama 4 Scout/Maverick were released in the first quarter of 2025, significantly improving "intelligence per cost" compared to existing models.
Meta AI is used within WhatsApp and Messenger, with monthly active users exceeding 1 billion. This is a result of seamless integration with existing messaging platforms, not just the Meta AI app alone.
1-2. Model Characteristics and Design Philosophy
Scout/Maverick: Mid-to-small-sized models designed for low latency, capable of running on a single host.
Little Llama: Currently under development as the successor to the 8B parameter model.
Behemoth: A frontier model in the 2-trillion parameter class. The company is also focusing on technology to "distill" large-scale models into sizes that are easier to use in actual applications.
2. Open Source Strategy and Community
2-1. Licensing Policy
Meta has adopted the "Llama License," maintaining a requirement for prior consultation for commercial use by entities with over 700 million users. This is intended to prevent other cloud companies from reselling the models without permission while fostering a healthy open-source community.
2-2. Collaboration with the Community
Leveraging the advantages of open source, external researchers and developers are "distilling" and "fine-tuning" the models to jointly create application-specific models.
The company has released security tools such as "Llama Guard" and "Code Shield" to establish an environment where external users can utilize the models safely.
3. AI Product Metrics and Benchmarks
3-1. Making Product Value the North Star
Meta adopts a policy of prioritizing "value actually felt by the user" as the most important metric, rather than optimizing solely for scores on external benchmarks like the Chatbot Arena. It emphasizes "repeat rates," "feedback," and "retention" within the product, positioning these as the North Star for model design and evaluation.
3-2. The Pitfalls of Benchmark Optimization
Benchmark optimization can easily be achieved through tuning tailored to the metrics, and it has been pointed out that this can lead to cases that diverge from the actual user experience.
Meta intentionally releases pure models that do not aim for top benchmark rankings, prioritizing transparency and reproducibility.
4. Future Vision and the Path to AGI
4-1. Intelligence Explosion and Infrastructure Challenges
To achieve an "intelligence explosion," the automation of software engineering and AI research is key.
However, physical infrastructure (data center construction, power procurement, chip manufacturing) requires years to build, and this serves as a real-world bottleneck.
4-2. Personal Assistants and Daily Life
Enhancing full-duplex voice interaction and personalization features, while also expanding to new devices such as smart glasses for the metaverse.
Envisioning a world where people "talk to AI daily to receive information and task support," aiming to break away from dependence on smartphone and PC screens.
5. Ethics, Governance, and Social Impact
5-1. Designing Relationships with AI
As long as users recognize it as "valuable," interactions with AI and virtual human relationships are respected. Zuckerberg himself states, "Even if I think an act is 'bad,' if the other party judges it to be valuable, there must be value there that I cannot yet see"¹.
At the same time, care is taken in UI/UX and ecosystem design to avoid excessive "reward hacking."
5-2. Collaboration with Government and Regulators
In AI governance, Meta emphasizes "constructive dialogue" with governments and regulatory authorities. Drawing on past experience with content moderation, the company demonstrates a stance of contributing to policy formation for AI safety and privacy protection.
6. Economic Models and Business Strategy
6-1. Diversity in Monetization
Free services follow an advertising revenue model, while subscriptions and enterprise-level billing are being considered for premium use cases and large-scale compute usage.
By creating new customer touchpoints, such as "automating customer support using AI," it becomes possible to provide services to areas that were previously not profitable.
6-2. The Role of the CEO and Organizational Management
Zuckerberg positions the four main responsibilities of the CEO as "securing top talent," "inter-team collaboration," "product quality management," and "infrastructure investment decisions."
In the AI field, where research and product development are driven by model design priorities, the CEO himself is deeply involved in setting development parameters and roadmaps.
Meta's AGI plan is supported by open-source-driven technological innovation, an evaluation framework that prioritizes user value, and the dual pillars of physical infrastructure and software development. Looking ahead, numerous long-term roadmap items await, including the realization of an intelligence explosion, the integration of AI into daily life, and addressing social and ethical challenges. The future that Meta envisions, where 'AI makes people's lives richer,' will only be achievable through the fusion of technology, business, and governance.
Related Articles
