BasedAI is a decentralized peer-to-peer (P2P) network designed for Zero-Knowledge Large Language Models (ZK-LLMs) and smart contracts, founded by Sean Wellington. It introduces an infrastructure that integrates Fully Homomorphic Encryption (FHE) with large language models (LLMs), facilitating secure and private AI computation without compromising data privacy.^[1][2][3][4]

Overview

BasedAI is built on a distributed network of machines that enable the transformation of traditional LLMs into encrypted zero-knowledge LLMs (ZK-LLMs) through a mechanism known as "Cerberus Squeezing." This framework optimizes performance and data privacy by allowing miners to process user interactions with LLMs without decrypting the inputs or outputs. This innovation addresses the performance degradation often seen in FHE (Fully Homomorphic Encryption)-compliant environments by streamlining interactions between users, miners, and validators.

The platform also integrates various blockchain technologies, drawing on features from Polkadot's Substrate, Ethereum's EVM, and GPU configurations from RNDR. BasedAI offers a comprehensive framework that allows for launching smart contracts, managing assets, and interacting with the network’s consensus mechanisms.^{[1][2][3][4][7]}

Technology

BasedAI leverages Fully Homomorphic Encryption (FHE) to ensure that data remains encrypted throughout its lifecycle - storage, transmission, and computation. This approach enables computations on encrypted data without requiring decryption, ensuring full privacy in AI operations. The platform’s mining process incorporates zk-LLMs to support this secure, encrypted AI environment.

Central to BasedAI’s ecosystem are processing nodes called "Brains," which execute computational tasks associated with LLMs. These Brains form the backbone of the decentralized infrastructure, enabling distributed and privacy-focused AI processing.^[1][2][3][5]

Testnets and Development

BasedAI’s development roadmap includes two primary testnets: Prometheus and Cyan. The Prometheus testnet was completed, and the Cyan testnet, launched in June 2024, serves as a release candidate aimed at testing the integration of Brain NFTs and their associated tokens. This phase focuses on emissions, rewards, and the administration of these tokenized assets. BasedAI is expected to proceed to a mainnet launch following these testnet phases.^{[3][4][5][6][7]}

BasedAI Ecosystem

The platform operates on a decentralized blockchain network, emphasizing data privacy and security in AI applications. By integrating zk-LLMs with homomorphic encryption, BasedAI aims to provide a secure and transparent AI environment, while also decentralizing ownership of AI models. Users interact with the network using the Based Command Line Interface (Based CLI), which supports a wide range of functions, from wallet management to interacting with computational nodes.

Key Features

• Cerberus Squeezing: The mechanism that converts traditional LLMs into zk-LLMs, enabling secure, encrypted AI processing.
• Fully Homomorphic Encryption (FHE): Ensures data remains encrypted throughout its lifecycle while supporting computations.
• Based CLI: A command line interface that facilitates user interaction with the platform, allowing users to manage wallets, interact with Brains, and participate in consensus mechanisms.
• Testnet Phases: BasedAI’s development is marked by its testnets, Prometheus and Cyan, which are essential for testing its functionality before full deployment.

Applications

BasedAI’s primary applications revolve around enhancing privacy in AI processing, democratizing AI model ownership through Brain NFTs, and creating a secure environment for deploying smart contracts. Its integration of AI and blockchain technologies also supports advanced use cases such as decentralized AI-powered tools, secure data processing, and privacy-preserving computations.

BasedAI represents a novel fusion of decentralized blockchain technology and artificial intelligence. By leveraging homomorphic encryption and zk-LLMs, it offers a framework that prioritizes data privacy and security while enabling the decentralized ownership and operation of AI models.^[2][3][4][5]

Tokenomics

Rewarding Participation: Incentives and Token Emissions

BasedAI incentivizes participation in its ecosystem through the issuance of $BASED tokens. Every 10 seconds, 10 $BASED tokens are distributed among active Brains, which are key nodes responsible for computational tasks. The rewards are allocated based on two primary factors: the amount of $BASED staked in each Brain and the operational performance of the validators and miners linked to those Brains. The highest-performing 30% of Brains receive additional bonuses, creating a competitive environment that rewards efficiency and reliability. All $BASED tokens are two-way bridgeable with the wrapped BasedAI ERC-20 token on Ethereum.

Emission Halving Schedule

To mitigate inflation and encourage long-term participation, BasedAI follows an annual halving schedule, which progressively reduces the block reward. This halving is designed to balance the supply of $BASED and maintain incentives for miners and validators. Starting in April 2024, the block reward will halve from 10 $BASED tokens to 5, with subsequent halvings every year.^[2][3][4][5]

Governance

Staking and Validator Incentives

BasedAI implements a staking mechanism that allows users to stake $BASED tokens to any Brain or specific validator. Validators play a crucial role in maintaining the network’s functionality by processing computational tasks for Brains. Stakeholders earn rewards from the emissions allocated to the validators they are staked to. Additionally, to become a validator or miner, a user must perform a permanent memory operation that associates their address with a Brain.

Allocation of Rewards

75% of emissions are allocated to nodes—validators and miners—performing essential tasks, while 25% goes to the Brain owners or specific validators managing the Brain.

Owner Adjustments and Stake-Weighted Rewards

Brain owners have the flexibility to adjust how rewards are distributed within their respective nodes, enabling a competitive landscape where validators are rewarded based on performance. The system also includes stake-weighted incentives, which provide larger rewards to participants with more significant investments in specific Brains.^{[2][3][4][5][7]}

Utility

Brain NFTs and Validator Diversification

BasedAI promotes validator diversification by limiting rewards to the top 70% of validators, incentivizing participants to spread their stakes across multiple Brains rather than concentrating on a few. This dynamic encourages broader network participation and mitigates centralization risks. Validators cannot duplicate stakes across Brains, fostering a system where each Brain operates independently with distinct staked validators.

Mitigation of Centralization

To further decentralize the network, BasedAI enforces a cap on the maximum stake per Brain, preventing any single entity from controlling more than 0.5% of the network's total stake in $BASED. This ensures a more equitable distribution of rewards and reduces the risk of oligopolistic control over the network.

Temporal Fusion Transformer (TFT) Enforcer

One of BasedAI’s unique features is the Temporal Fusion Transformer (TFT) Enforcer, which is designed to optimize reward distribution by analyzing network data over time. The TFT model evaluates various factors such as transaction volumes, block times, and emission rewards, dynamically adjusting incentives to validators and Brains. This automated system helps ensure that rewards are distributed fairly, minimizing opportunities for manipulation or collusion.^{[2][3][4][5][6][7]}

Sean Wellington

Sean Wellington is a researcher at Based Labs and holds an educational background from UC Berkeley. He has contributed to the development of BasedAI’s decentralized infrastructure, with a focus on integrating Fully Homomorphic Encryption (FHE) with large language models (LLMs) to maintain data privacy during interactions without the need for decryption. A key aspect of his work includes the development of the "Cerberus Squeezing" mechanism, designed to optimize performance in FHE-compliant environments. This system enables BasedAI miners to process and respond to encrypted LLM queries, drawing on techniques from generative adversarial networks (GANs) to support privacy.

Wellington’s work seeks to address performance challenges often encountered in FHE-based computing environments, improving the communication between users, miners, and validators within the network. His research and development contributions extend beyond BasedAI, as reflected in a range of patents and academic publications within the areas of cryptography, decentralized systems, and artificial intelligence. These include patents such as Electronic Chain of Custody Method and System (2006) and Methods and Systems for Managing Distributed Concurrent Data Updates of Business Objects (2016), which highlight his involvement in innovations related to distributed data management and document compression.

Throughout his career, Wellington has focused on exploring decentralized technologies, particularly in the areas of artificial intelligence and cryptography, in order to address ongoing challenges related to data privacy and security in digital systems.