# BasedAI: A decentralized P2P network for Zero Knowledge Large Language Models (ZK-LLMs)

**Authors:** Sean Wellington  
**arXiv:** [2403.01008](https://arxiv.org/abs/2403.01008)  
**Date:** 2024-03-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.01008v1/

## Key points

- BasedAI is a decentralized peer-to-peer network integrating Fully Homomorphic Encryption (FHE) with large language models (LLMs) through its innovative framework.
- The network introduces "Cerberus Squeezing" into the mining process to convert standard LLMs into encrypted zero-knowledge LLMs (ZK-LLMs).
- BasedAI miners can efficiently process user prompts without decrypting queries or responses, enhancing data privacy and security by leveraging generative adversarial networks.
- An initial allocation scheme includes owner adjustments and stake-weighted incentives to incentivize participation and ensure fair resource distribution.
- Validator diversification incentive is implemented to promote decentralization and mitigate centralization risks in the network.
- BasedAI demonstrates commitment to sustainability and inclusivity through fiscal yield calculations, emission functions rewarding intelligence, and fostering a sustainable ecosystem for all participants.
- Temporal Fusion Transformers (TFT) play a crucial role in BasedAI's consensus mechanism by enforcing rules during transition periods and building trust among peers over time credits.
- The network secures BASED block rewards through robust mechanisms, ensuring the utility and value of the native token ($BASED).
- By expanding ZK-LLM applications and emphasizing privacy, BasedAI sets new standards for decentralized infrastructure and data protection in the era of advanced technologies.

## Summary

BasedAI is a decentralized peer-to-peer network that revolutionizes the integration of Fully Homomorphic Encryption (FHE) with large language models (LLMs) through its innovative framework. This distributed system introduces a groundbreaking mechanism known as "Cerberus Squeezing" into the mining process, enabling the conversion of standard LLMs into encrypted zero-knowledge LLMs (ZK-LLMs). By leveraging insights from generative adversarial networks, BasedAI miners can efficiently process user prompts without decrypting queries or responses, significantly enhancing data privacy and security. Moreover, BasedAI implements an initial allocation scheme that includes owner adjustments and stake-weighted incentives to incentivize participation and ensure fair distribution of resources. The network also incorporates a validator diversification incentive to promote decentralization and mitigate centralization risks. Through examples of fiscal yield calculations and emission functions rewarding intelligence, BasedAI demonstrates its commitment to fostering a sustainable and inclusive ecosystem for all participants. Furthermore, Temporal Fusion Transformers (TFT) play a crucial role in BasedAI's consensus mechanism by enforcing rules during transition periods and building trust among peers over time credits. The network secures BASED block rewards through robust mechanisms, ensuring the utility and value of the native token ($BASED). By expanding ZK-LLM applications and emphasizing the importance of privacy in the digital age, BasedAI sets new standards for decentralized infrastructure and data protection in the era of advanced technologies.