Ritual.md

Ritual

Resources

• Introductory Post: Introducing Ritual
• Documentation
• Installation

Tutorials

• Simple logistic classifier with ZK verification
• ML Resources 101

Notes

Intro Post

Introducing Ritual

Goal Statement: Merge the best principles and techniques of cryptography and AI to create a system that enables open and permissionaless creation, distribution and improvement of AI models.

Context: We have recently seen a cambrian explosion in AI models, but we see a few core problems.

Lack of Strong Service Level Agreements (SLAs)

• existing platforms dont offer guarantees on computational integrity, privacy, and censorship resistance.
• permissioned and centralized APIs
• high compute cost and limited hardware resources
• oligopolistic and strcuturally misaligned: current landscape, either close source ur models, or open-source as a donation (no infra to reward their contributions)

Solution

Ritual is an open, modular, sovereign execution layer for AI. Consists of distributed network of nodes with access to compute and model creators, enabling creators to host models on these nodes. Users can then access any model on the network (single uniform API) and the network provides cryptographic guarantees around privacy, and integrity.

Infernet

Infernet is the first evolution of Ritual. It exposes interfaces for smart contracts to access AI model for inference. It's the first building block in a suite of protocols and utilities to enable builders on Ritual and permissionless access to network of model and compute resources.

The grand vision is for Ritual to become schelling point of AI in web3 by tranforming Infernet into a modular suite of execution layers that interop with base layer infra, allowing every protocol and application on an chain to use Ritual as an AI coprocessor.

Team

About Page

• Niraj Pant: GP Polychain, Research Decentralized Systems Lab, CS @ UIUC
• Akilesh Potti: Partner @ Polychan, ML @ Palantir, Quant @ Goldman, ML Research @ MIT, Cornell
• Ricky Moezinia: Founder @ Alta, SE @ Libra, Quant @ Point 72, EECS @ MIT
• Anish Agnihotri: MEV, Research @ Paradigm, Trading @ Polychain, CS @ UWaterloo

Advisors

• Illia Polosukhin: Co-Founder NEAR and Co-Creator of Transformers ("Attention is all you need")
• Sreeram Kannan: Founder Eigenlayer and Associate Prof of CS @ University of Washington
• Tarun Chitra: Founder/CEO Gauntlet and GP @ Robot Ventures

Product

Product

Ritual Superchain: and AI coprocessor for all blockchains

• a state of sovereign modular execution layers, each containing specialized Stateful precompiles
• a GMP layer to facilitate interop between existing blockchains and Ritual superchain
• Ritual's AI VM contains not only SPCs, but also base layer infra to facilitate optimized execution, including inference engine binaries and vector dabases

Infernet

Flow

1. Smart contracts use interface to initiate requests including model ID, input features, etc.
2. Infernet nodes listen for calls and collect models from storage
3. Infernet computes inference and delivers output (alongside optional proofs)
4. Coordinator delivers the response directly to the contract

SDK

The SDK closely resembles that of off-chain oracle systems (and cross chain messaging protocols)

Server Contracts

Coordinator.sol

Events

• SubscriptionCreated
• SubscriptionCancelled
• SubscriptionFulfilled

Functions

• createSubscription()
• cancelSubscription()
• getSubscriptionInterval()
• deliverModelOutput()

EIP712Coordinator.sol

Functions

• createSubscriptionDelegate()
• deliverModelOutputDelegate()

Functions

• _receiveModelOutput
• rawReceiveModelOutput: Callable only by address(COORDINATOR)

Client/Developer Contracts

Developers may inherit two interfaces:

• Callback Consumer: one-time computation consumption
• SubscriptionConsumer: recurring computation consumption

both inherit the base consumer

BaseConsumer.sol

Handles receiving model output from coordinator.

Functions

• rawReceiveModelOutput(): public entrypoint, relays to
  • same as below
• _receiveModelOutput()
  • subscriptionId
  • interval
  • redundancy
  • node
  • response
  • proof

CallbackConsumer.sol

• technical reference

Functions

• _requestModelOutput() returns uint32 ID
  • containerId
  • features
  • maxGasPrice
  • MaxGasLimit
  • redundancy

SubscriptionConsumer.sol

• technical reference

Functions

• getModelFeatures()
  • subscriptionId
  • interval
  • timestamp
  • caller
• _createModelOutputSubscription() returns uint32 ID
  • containerId
  • maxGasPrice
  • MaxGasLimit
  • frequency
  • period
  • redundancy
• _cancelModelOutputSubscription()
  • ID

Delegator.sol

Exposes a signer address allowing authorized EOA to sign off on actions on behalf of contract.

Allows developers to create Coordinator subscriptions off-chain, on behalf of a contract, by signing a DelegateSubscription from signer and submitting to EIP712Coordinator.createSubscriptionDelegatee()

Types

Subscription

• owner: on-chain address (smart contract), which receives subscriptions and can cancel
• containerId: unique identifier of compute container you want an Infernet node to run. Subscriptions' features are passed into this container and output directly returned.
• features: two ways to expose input data/features. Encode data in this parameter or leave empty and expose a getModelFeatures() function in consumer contract.
• frequency: 1, 2, etc, or UINT256_MAX implying perpetual
• period: in seconds, 0 if once
• redundancy: how many unique nodes do I want a response from each time.
• activeAt: start time
• maxGasPrice: max price of response tx
• maxGasLimit: contracts callback function gas _ DELIVERY_OVERHEAD_WEI (fixed surcharge)

Nodes

People can run nodes permissionlessly

• doc reference

Workflows

Two main workflows: training and inference

Uses MLFlow