SingularityNET drops self-learning proto-AGI into Minecraft

Most video game non-player characters are puppets on invisible strings, following strict scripts hardcoded by developers. But drop an AI into a sandbox with zero pre-programmed instructions and let it figure out how the world works on its own, and you get something far more interesting.

SingularityNET and the Artificial Superintelligence Alliance have launched AIRIS inside Minecraft, describing it as a self-learning proto-AGI.

Short for Autonomous Intelligent Reinforcement Inferred Symbolism, AIRIS does not rely on static rule sets. Instead, it observes in-game events in real time, generates its own internal logic, and constantly refines how it navigates, avoids hazards, and tackles unexpected terrain.

Learning without a rulebook

According to Berick Cook, the AI developer at SingularityNET who built AIRIS, the agent takes the wheel in Minecraft through a tool called the Vereya Mod. The mod acts as a conduit that lets external AI algorithms steer player characters, while keeping things clean on the privacy front by only transmitting game state data and storing zero user data.

An ASI Alliance spokesperson noted that Minecraft is effectively the gold standard for testing reinforcement learning algorithms. Because it is an open-ended 3D sandbox with established academic benchmarks, the team can directly measure how AIRIS stacks up against traditional navigation and decision-making models.

SingularityNET CEO Ben Goertzel weighed in on a recent podcast, arguing that today's leading LLMs like ChatGPT fall short of true AGI because they cannot venture outside their training datasets. In his view, real general intelligence requires the distinctly human ability to leap beyond what you have been directly taught.

Why it matters

Testing an autonomous learning framework in a sandbox world is a calculated, cost-effective move. Real-world robotics is punishingly expensive and physically risky, but Minecraft offers an infinitely regenerable testbed where an agent can fail thousands of times without consequences. For developers working on autonomous systems and robotics, AIRIS offers a glimpse into how agents might build functional world models on the fly instead of relying on brittle, manually tuned rules.

We think the symbolic inference approach is a refreshing departure from simply throwing more brute-force compute at transformer models. If an agent can dynamically synthesize new operational rules from raw environmental feedback, it could solve stubborn navigation and spatial reasoning bottlenecks in smart automation.

However, a blocky sandbox remains a heavily simplified abstraction of reality. Surviving a Minecraft biome is not the same as managing a busy industrial warehouse or piloting a delivery drone through inclement weather. Goertzel’s commitment to decentralized infrastructure gives the initiative an open alternative to Big Tech monopolies, but closing the engineering gap between toy environments and real-world deployment remains the real hurdle.