Localized Knowledge Experts: Reppo Intelligence Pods

Reppo Pods: Intelligence pods are inspired by research coming out of Google Deepmind advocating for parsely activated expert modules instead of a single dense neural network layer. Each of the experts contains a fraction of the parameters of the full dense layer and specializes in certain areas. We also experienced by the challenges first-hand coordinating a group of AI developers and modelers over the last 10 months.

A “Reppo” pod can be thought of as an inference DAO mimicking an AWS Organization on-chain, that can be spun up in a permissionless manner. Unlike Bittensor, there is no cap on the number of pods that one can spin up but there is gas fees involved to spin up a Reppo Pod.

• Reppo pods are domain-specific clusters of human+AI that have ownership and monetization primitives through the Reppo Protocol. Pods also have individual governance as well as rights to participate in protocol governance.

• Pods act as an exponential number of specialized experts to serve queries & use-cases, as opposed to a single generalized LLM. Pods have reputation mechanics built-in to accrue credibility, through the Reppo protocol.

• Pods look like intelligent mirrors of existing organizations/companies/firms that specialize in a domain who wish to monetize their moat without being present. It can be of any form ranging from ONNX* to pytorch. The pod decides and uses ModelRivet to satisfy end-customer demand.

• Reppo pods enable network effects and a plethora of emerging use-cases related to satisfying downstream user-needs for verifiable AI.

*ONNX allows for pre-trained models to be applied to different use cases. This opens up tremendous opportunities for Reppo to enter a flywheel after the first batch of models are committed to the network. While these first models might have been developed for specific networks for specific target users, others can leverage the same models to bootstrap dApps and products efficiently, significantly cutting down the cost of bootstrapping a such system for blockchain networks.