Meta Accelerates Toward Superintelligence: RL Environment and Unprecedented Compute Growth

Meta is investing in superintelligence development through a partnership with an RL environment startup and unprecedented growth in computing power. According to SemiAnalysis analysts' assessment, this is the most aggressive compute ramp in the history of the industry, and the company is clearly playing a long game for artificial general intelligence.

RL environment: the key to AGI

Meta has backed or acquired a startup that developed a reinforcement learning environment—a space where AI agents learn through trial and error, like a human learning to ride a bicycle or play a video game. This differs from the traditional LLMs we know, which simply predict the next text token.

Why does this matter? Because superintelligence likely cannot be achieved through text tokens alone. An agent must interact with the environment, receive feedback, try different strategies. RL environments are exactly what's needed for such learning.

Scaling RL environments is one of the most difficult engineering challenges in AI. Meta clearly invested in this startup because it sees it as a critical component of future superintelligence architecture.

Computational leap: unprecedented scale

Meta is scaling computing power so aggressively that SemiAnalysis, accustomed to rapid industry development, calls it a phenomenon. Infrastructure is being deployed on a geographic scale exceeding 2,000 kilometers—more than the distance from Moscow to Vladivostok.

Such scale implies distributed training: GPU clusters scattered across different regions and even countries must synchronize in real time. This requires solving classical problems:

• Network latency between nodes
• Data consistency during node failures
• Bandwidth optimization
• Cooling and power supply at this scale

Meta's engineers must also rethink the model architecture itself to make it resilient to network delays. This is not just scaling—it's a qualitatively new level of infrastructure complexity.

What experts say

SemiAnalysis notes that this movement signals to other labs: the race for superintelligence is not just about algorithms and data, it's about computing power. Those who cannot scale computation as aggressively as Meta risk falling years behind.

The article also offers advice for Google DeepMind, which currently takes RL environments less seriously: this could be a strategic mistake. Google DeepMind once brought AlphaGo to the world—an RL agent that played go better than world champions. But in recent years, DeepMind's focus has shifted to large language models. If Google misses the wave of RL scaling, it could lose the race for AGI.

What does this mean

Text-based LLMs may turn out to be a transitional stage rather than the final AGI architecture. Meta is betting that the future belongs to RL agents learning to interact with the real (or simulated) world. This requires enormous computation, but the potential reward—AGI—justifies the investment.

The message to the rest of the industry is clear: if you are serious about AGI, computation is not a side concern but the center of development. Meta has already reached this point of realization and is acting accordingly.

*Meta has been recognized as an extremist organization and is banned in the Russian Federation.