Decentralized Inference and Apple's Private AI Cloud Compute

Earlier this week, Apple showcased a slew of new AI features and services for its hardware product line. The announcement also included broad-stroke details around Apple’s Private Cloud Compute (PCC), which highlights an ever-growing sector of importance within the AI x Crypto stack — decentralized inference.

Within the AI x Crypto context, decentralized inference refers to the ability to add a layer of accountability to an AI model. Users need some assurance that, for some given AI model and input data, the output truly is what is returned to the user. These assurances are difficult to grant when the models are owned and operated by centralized entities. Generally, the “assurances” boil down to simply trusting the model operators.

The above-described trust model is the standard for the Web2 setting. However, as onchain applications begin to leverage AI systems, trust must be replaced by some form of accountability and verifiability. Numerous teams are now working within the AI x Crypto vertical to achieve these ends, and they are leveraging familiar tools — ZK proofs and cryptoeconomics.

Decentralized and Verifiable Inference

To use an AI-powered product, the underlying model is inferenced. When you ask ChatGPT to summarize an article, you are inferencing OpenAI’s model (e.g., GPT-4o, GPT-3.5, etc). That model is hosted on some server at a large data center, which enables it to scale up and down to meet the current levels of demand. Once the summarized article is returned, that completes the cycle of a centralized inference workflow.

In the above workflow, OpenAI represents the centralized point of trust. Users must trust that the model they are paying for is actually processing the inference. Additionally, there is trust that OpenAI will not tamper with the output. Currently, most users accept the premise that OpenAI is acting honestly. However, as AI becomes less of a toy or fun product and more akin to mission-critical infrastructure for individuals, institutions, and governments, removing points of centralized trust will be of paramount importance.

Decentralized Inference Methodologies – ZK, Cryptoeconomic, and TEEs