Alexandr Wang on Meta: Superintelligence, AI’s Unfinished Endgame

Why Wang Joined Meta

Alexandr Wang, the founder of Scale AI, says the direct trigger for his move to Meta in 2025 was the misaligned trajectory of Llama 4, which he felt showed that Meta’s AI work had fallen off the cutting‑edge path.

Meta Superintelligence Labs (MSL)

Under Wang’s plan, MSL is organized into a tight structure: the core research lab TBD tackles fundamental technical challenges; the Product & Application Research (PAR) team, led by former GitHub CEO Nat Friedman, focuses on deployment; FAIR continues exploratory research; Daniel Gross heads compute infrastructure; and chief scientist Zhao Shengjia sets the scientific agenda.

Wang emphasizes three “military‑grade” principles: (1) take superintelligence seriously—every decision must be calibrated to the belief that it is imminent; (2) let technical voices dominate over commercial or political pressure; (3) pursue bold, high‑risk bets rather than incremental tweaks.

Compute Stratification

Wang argues that the industry is splitting into two worlds: companies with massive compute can train frontier models and explore new paradigms, while those without are limited to building on others’ models. Meta’s aggressive GPU procurement, led by Daniel Gross, positions it to maintain a long‑term compute advantage.

Muse Spark: An Appetizer

Muse Spark, built from scratch in nine months, is described as an “appetizer” rather than a main course. It matches peer‑lab performance with far fewer tokens, thanks to a clean, rebuilt tech stack, but still lags in agentic programming. Wang hints that larger models are already in development, showing stable scaling in pre‑training, reinforcement learning, and multi‑agent extensions.

The AI Endgame Is Not Yet Here

When asked about the consumer market, Wang says the AI “endgame” has not started. He notes rapid shifts—OpenAI’s early lead, Claude Code’s surge, Gemini’s distribution—demonstrating that competitive barriers are fluid. Meta’s strategy is to create a global agent economy that leverages its billions of users and small‑business ecosystem.

Model Welfare and Safety

Wang explains that Muse Spark was kept closed‑source because it triggered safety checks in areas like biochemistry, cyber‑capabilities, and loss‑of‑control. He stresses that safety is a non‑negotiable baseline; open‑source decisions will depend on passing rigorous security evaluations.

Future Vision

Wang envisions superintelligence as a catalyst for worldwide abundance, linking AI breakthroughs to physical robotics and brain‑computer interfaces (BCI). He cites Meta FAIR’s TRIBE project, which can predict brain responses without personal data, as a step toward integrating digital and neural intelligence.

In summary, Wang frames Llama 4’s deviation as the starting point, MSL’s three principles as the guiding framework, compute stratification as a strategic lens, Muse Spark as a proof‑of‑concept, and the broader AI endgame as an ongoing, open‑ended journey that must also address model welfare and safety.