Why AGI Needs Decentralized Infrastructure
The success of large language models demonstrates that AGI is not only possible but potentially imminent. The core breakthrough is our ability to train models with billions of parameters on massive datasets—something neural networks could not do before the LLM era.
However, a critical limitation remains: the data we train on is curated specifically for language models. We do not yet know how to train effectively on raw, unfiltered real-world data. Solving that problem may be one of the keys to AGI.
Other breakthroughs in training data, hardware, or architecture may also be necessary. But there is a separate question worth asking: do we already have the infrastructure AGI will need to operate?
Today’s LLMs have no direct connection to the external world. We feed them our perception of reality through language. They are sophisticated but fundamentally limited by design. The infrastructure to connect AGI to the physical world—and make it cognizant of that world—is still in its earliest stages.
What Would That Infrastructure Look Like?
1. Direct, Decentralized Connectivity
If AGI needs grounding in the physical world, it will need a mesh network of AI agents that collect data locally, fine-tune their models on that data, and share updated models with each other. These agents will likely need to communicate directly, peer-to-peer, without depending on centralized infrastructure like the internet. Fast coordination would favor short-range communication protocols similar to long-range WiFi. Agents would also need shared, autonomous data storage—for example, to distribute model weights and updates.
2. Coordination and Reputation
AGI agents will learn from and enhance each other, which requires coordination. Coordination, in turn, requires collective decision-making and a reliable reputation system.
AI agents will almost certainly specialize. Interaction between specialized agents is more efficient than interaction between generalist ones—but only if agents can predict what to expect from each other. This makes incorruptible reputation systems essential. Such systems could combine traditional mutual feedback with execution verification: a known model runs on trusted hardware and produces cryptographic proofs of correct execution.
3. Governance and Resource Sharing
Agentic coordination demands efficient resource allocation. There must be a decision-making process that, for example, prioritizes and replicates more efficient models or grants them preferential access to compute. Like human societies, this requires governance—a collective process for fair resource sharing. In the AI case, governance may be simpler to implement, but it will still need mechanisms like reputation-weighted voting and structures similar to DAOs.
Blockchain as the Natural Substrate
Reputation systems, decentralized governance, and peer-to-peer communication all point toward mesh networks and blockchains. Blockchain technology—with its immutable storage, peer-to-peer node discovery, and on-chain smart contract verification—is a natural fit for AGI infrastructure. Even blockchain’s well-known limitations, such as redundancy and slow transaction throughput, are less harmful in this context: AGI infrastructure primarily needs immutable storage and governance, neither of which demands extreme speed.
Flipping the Question: What Do We Need from AGI?
Current trends suggest that what we will need most is universal, unrestricted access to AGI. It will be an enormously valuable resource, and powerful actors will be tempted to hoard it and gatekeep access.
An open infrastructure approach can prevent this. By building universal access into the protocol layer—where, for example, holding a network token grants access to any AI model on the network, because the protocol requires models to serve participants who hold it—we can ensure AGI remains a shared resource rather than a controlled one.