Basics of Decentralized AI (DeAI)

Author: Shaun Geer · 2024. Written to support ecosystem education for developers and researchers exploring DeAI infrastructure.

A conceptual overview from 2024. The benefits below depend on the architecture and controls of a particular implementation.

What Is Decentralized AI?

Decentralized AI (DeAI) refers to AI systems whose core infrastructure: model training, inference, data storage, and governance: is distributed across a network of independent nodes rather than controlled by a single entity.

The motivations for DeAI overlap with but are distinct from general Web3 motivations. The primary drivers are:

• ·Privacy: exploring designs that limit which parties can observe sensitive inputs.
• ·Verifiability: evaluating mechanisms for checking execution and output integrity.
• ·Censorship resistance: reducing dependence on a single infrastructure operator.
• ·Access: exploring alternatives to centrally operated, permissioned APIs.

Distributed Inference

Inference is the process of running a trained model against a new input. Deployment choices determine where that computation happens and who operates the infrastructure.

Distributed inference splits this workload across a network of nodes. Models may be sharded (each node holds a portion of the model weights and processes part of the computation) or replicated (multiple nodes run the full model, and results are aggregated or verified against each other).

Privacy-Preserving Computation

Three primary technologies enable private computation over AI inputs:

• ·Secure Multi-Party Computation (MPC): multiple parties jointly compute a function without revealing their private inputs to each other.
• ·Homomorphic Encryption (HE): allows computation to be performed on encrypted data, so that the inference server never sees the plaintext input.
• ·Trusted Execution Environments (TEEs): hardware-enforced secure enclaves that prevent the server operator from observing computation running inside them.

Governance in DeAI Networks

DeAI governance addresses three distinct layers:

• ·Model governance: who can submit models to the network, and what review or verification processes apply.
• ·Infrastructure governance: how node requirements, staking parameters, and slashing conditions are set and updated.
• ·Output governance: how the network handles requests to censor, modify, or roll back model outputs.

DeAI vs. Centralized AI: Key Tradeoffs

• ·Latency: coordination between nodes can add overhead, which should be measured for the intended workload.
• ·Cost: compare hardware utilization, communication overhead, and operating costs for a specific deployment.
• ·Privacy: assess encryption, execution boundaries, and who can observe inputs at each step.
• ·Resilience: assess dependencies on operators, access layers, and network governance.