Meta accelerates toward superintelligence: an RL environment and an unprecedented compute ramp

Meta is investing in superintelligence through a partnership with a startup's RL environment and an unprecedented expansion of compute capacity. SemiAnalysis analysts call it the most aggressive compute ramp in the industry's history, and say the company is clearly playing a long game for artificial general intelligence.

The RL environment: a key to AGI

Meta has backed or acquired a startup that built a reinforcement learning environment: a setting in which AI agents learn by trial and error, the way a person learns to ride a bike or play a video game. This differs from the LLMs we are used to, which simply predict the next text token.

Why does it matter? Because superintelligence most likely cannot be reached through text tokens alone. An agent has to interact with an environment, receive feedback and try different strategies. RL environments are exactly what this kind of training requires.

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

The compute jump: unprecedented scale

Meta is expanding compute so aggressively that SemiAnalysis, which is used to the industry's rapid pace, calls it a phenomenon. The infrastructure is being rolled out across a geographic span of more than 2,000 km, which is greater than the distance from Moscow to Vladivostok.

Such a scale implies distributed training: GPU clusters spread across different regions and even countries must synchronize in real time. That requires solving classic problems:

• Network latency between nodes
• Data consistency when nodes fail
• Bandwidth optimization
• Cooling and power supply at this scale

Meta's engineers also need to rethink the model architecture itself so that it tolerates network delays. This is not just scaling; it is a qualitatively new level of infrastructure complexity.

What experts say

SemiAnalysis notes that the move signals to other labs that the race for superintelligence is not only about algorithms and data, but about compute. Those who cannot scale compute as aggressively may fall years behind Meta.

The article also offers a warning for Google DeepMind, which has so far treated RL environments less seriously: that could be a strategic mistake. Google DeepMind once gave the world AlphaGo, 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 may lose the race for AGI.

What this means

Text LLMs may turn out to be a transitional stage rather than the final architecture of AGI. Meta is betting that the future belongs to RL agents that learn to interact with the real (or simulated) world. That demands enormous compute, but the potential payoff, AGI, justifies the investment.

For the rest of the industry, the message is clear: if you are serious about AGI, compute is not a side issue but the center of development. Meta has already reached that realization and is acting on it.

*Meta is designated an extremist organization and banned in Russia.