Mark Zuckerberg — AI will write most Meta code in 18 months

Llama 4 Release and Model Architecture

• New Model Series: Meta announced four Llama 4 models, releasing the "Scout" and "Maverick" models immediately.
  • Scout and Maverick are mid-to-small size models optimized for efficiency, latency, and running on a single host.
  • They are natively multimodal and offer some of the highest intelligence-per-cost ratios available.
  • A smaller model nicknamed "Little Llama" (likely an 8B parameter variant) is expected to release in the coming months.
• Frontier Model: A "Behemoth" model exceeding two trillion parameters is in development.
  • This model requires custom-built infrastructure for post-training due to its sheer size.
  • Strategy involves using the Behemoth to distill intelligence into smaller, consumer-runnable models.
• Reasoning Capabilities: Meta is developing a Llama 4 reasoning model.
  • This approach mirrors the "test-time compute" strategy seen in models like o1/o4, where extended inference time yields higher accuracy.
  • Meta prioritizes low-latency responses (e.g., <0.5 seconds) for consumer products over the extended reasoning delays of specialized models.
• Benchmarking Philosophy: Meta is shifting focus from external leaderboards (e.g., LMArena) to internal product metrics ("North Star").
  • External benchmarks are viewed as gameable and often skewed toward specific use cases (like math or coding) that don't reflect general user value.
  • Model tuning is anchored to revealed user preferences and usage within the Meta AI ecosystem.
  • Meta believes the gap between open source and closed source is narrowing, with open source potentially overtaking closed source as the most widely used models soon.

AI Strategy and Intelligence Explosion

• Coding Agents: Meta is building internal coding and research agents specifically to advance Llama research.
  • Prediction: Within 12–18 months, the majority of code written for Meta's AI efforts will be authored by AI agents, not humans.
  • These agents will perform full-loop tasks: interpreting goals, running tests, identifying issues, and writing higher-quality code than average senior engineers.
• Intelligence Explosion Timeline: While optimistic about an intelligence explosion, Mark Zuckerberg cites physical infrastructure as the primary bottleneck.
  • Scaling to gigawatt compute clusters requires significant time for permitting, energy supply chain stabilization, and hardware fabrication.
  • Supply chains for AI hardware (e.g., NVIDIA systems) create a natural lag, preventing instantaneous "takeoff."
• Human-Machine Co-evolution: The future involves a feedback loop where humans learn to use AI and AI learns user preferences.
  • Long-term utility relies on historical context (e.g., an AI remembering conversations from two years prior), which requires time to build up.
  • Rapid adoption of AI assistants drives data collection, which accelerates model improvement alongside physical infrastructure build-out.

Product Development and User Experience

• Meta AI Usage: Meta AI now serves nearly 1 billion monthly active users.
  • Usage is currently highest in WhatsApp globally, with 100+ million users in the U.S.
  • A standalone Meta AI app is being launched to create a first-class experience in markets where WhatsApp is not the primary messaging platform.
• Multimodal and Voice: New features include full-duplex voice interaction and holographic overlays.
  • Full-duplex voice aims for natural, overlapping conversation similar to human interaction.
  • Design principle for AR glasses is to "get out of the way," ensuring hardware functions as high-quality glasses first, with AI as an on-demand overlay.
• Content Evolution: Zuckerberg predicts a shift from passive video consumption to interactive, AI-driven content.
  • Future feeds will feature content that reacts to user input, changes dynamically, or turns into games.
  • The world is expected to become "funnier, weirder, and quirkier" as AI lowers the barrier to creating nuanced cultural content.

Social Dynamics and Human-AI Relationships

• AI Companionship: Meta anticipates a rise in meaningful relationships between humans and AI.
  • Current use cases include drafting difficult conversations (e.g., with partners or bosses) and providing social connection.
  • Zuckerberg argues that human demand for connection (average person wants ~15 friends but has <3) creates a natural market for AI companions.
• Safety and Design: Concerns about "reward hacking" or addiction are mitigated by design principles that prioritize user agency.
  • Philosophy: "If you think something someone is doing is bad and they think it's valuable, they are usually right."
  • AI interactions will evolve to include non-verbal cues (gestures, avatars) via Reality Labs to enhance realism.

Competitive Landscape and Open Source

• Global Competition: Acknowledges China's advantages in physical infrastructure deployment but highlights U.S. advantages in silicon and multimodal capabilities.
  • Chinese models (e.g., DeepSeek) face constraints due to export controls, forcing them to invest heavily in low-level infrastructure optimizations rather than feature innovation.
  • Meta claims Llama 4 models are competitive on text and lead significantly in multimodal (image, voice) capabilities.
• Open Source License: Meta maintains its current licensing terms for Llama, including requirements for large-scale commercial users to engage in dialogue.
  • Goal: Ensure transparency and partnership with major cloud providers rather than blocking usage.
  • Zuckerberg argues that open source is a cultural movement that Meta pioneered and must sustain to prevent industry closure.
• Distillation and Security: Focus on distilling large models into smaller ones while maintaining security.
  • Risks of distilling from non-domestic models include hidden vulnerabilities in code or embedded cultural biases.
  • Mitigation strategies include using verifiable domains (math, code), security filters (Llama Guard, Code Shield), and extensive red-teaming.

Business Models and Macro Economics

• Monetization Strategy: Meta plans a dual model: free, ad-supported tiers for mass usage and premium, paid tiers for high-compute tasks.
  • Ad-supported models work for services where ad inventory is sufficient to cover costs.
  • Premium models are necessary for resource-intensive applications (e.g., thousands of software engineering agents) where ad revenue is insufficient.
• Labor Market Impact: Contrary to the belief that AI will reduce demand for labor, Meta predicts increased demand for human work.
  • AI can reduce service costs by 90%, making previously unviable services (e.g., 24/7 voice customer support for billions of users) economically feasible.
  • Historical precedent suggests technology creates more jobs by expanding markets and lowering costs for human labor.

Governance and Politics

• Government Relations: Zuckerberg advocates for a productive relationship with the current U.S. administration to facilitate AI infrastructure growth.
  • Emphasizes the need for dialogue on energy, permitting, and policy to enable gigawatt-scale data center construction.
• AI Governance: Meta has moved away from deferring moderation decisions to external governments or media.
  • Stance: Meta must own content moderation decisions as a meaningful company, learning from past failures in fact-checking and community notes.
  • Focus: Building internal systems to detect nation-state interference rather than relying on external validation.