Sahara AI: The Decentralized Blockchain Platform for An

Open, Equitable, and Collaborative AI Economy

Sahara AI Research
research@saharalabs.ai

September 1, 2024

1 Introduction
As Artificial Intelligence (AI) advances, we envision a future where AI is not only capable and
ubiquitous, but also inherently fair, equitable, and accessible to all. In this future, AI will
serve as a universal instrument that enhances lives and fosters global equality of opportunities.
However, at present, control over AI technologies remains concentrated predominantly in the
hands of a few oligopolistic organizations. These centralized AI platforms come with signifi-
cant limitations and risks, including concerns for privacy, deepening economic disparities, and
restricted access to resources. This creates significant barriers that hinder widespread inno-
vation, limit participation in AI development from diverse backgrounds, and restrict access
to AI technologies across different communities.
To empower everyone to own and shape the future of AI, we are building Sahara AI—a
decentralized AI blockchain platform that supports an open, transparent, secure and inclusive
AI ecosystem. At the heart of Sahara AI is the notion of “AI asset”, a novel framework that
defines the ownership and management protocols of private AI resources, such as personal
data and proprietary models.
On our platform, AI developers, data providers and other stakeholders can collaborate to
co-create high-quality AI assets using the platform’s integrated development tools. Through-
out this process, the platform ensures that all contributions are securely recorded and trans-
parently attributed, with clear traceability of each participant’s input. Once created, these
AI assets are made available on the platform, where users can explore and leverage these re-
sources. Users have the flexibility to purchase licenses for access and further development or
even trade shares of the AI assets. Figure 1 presents an overview of this user journey, illustrat-
ing how AI assets move from creation to utilization and user engagement within the Sahara AI
ecosystem. Notably, all transactions within the platform are immutable and traceable, with
ownership protected and the origins of assets recorded. This supports a transparent and fair
revenue-sharing model, ensuring that both developers and data providers receive appropriate
compensation for their valuable contributions whenever revenue is generated.
Sahara AI leverages both blockchain technology and privacy protection methods to develop
a robust provenance infrastructure for AI assets. This infrastructure enables the attribution
of user contributions, protects data privacy, and ensures equitable compensation, all while
emphasizing the importance of user control over their AI assets. Powered by these features,
Sahara AI will deploy a unique, permissionless “copyright” system for all AI assets on its
platform. While “copyright” traditionally implies restrictive control, which may seem contrary
to the open nature of blockchains, this concept is redefined by Sahara AI’s approach—it
represents a tailored framework that ensures contributors retain ownership of their AI assets,
along with fair attribution and compensation for their work—without limiting access and

1
Figure 1: Overview of user journey on the Sahara AI Platform. This diagram demon-
strates the key stages in the AI asset lifecycle within the Sahara AI ecosystem. It highlights
how developers and knowledge providers collaborate to create AI assets, which are then listed
on the AI Marketplace for users to monetize, with all transactions recorded on-chain.

sharing. This platform serves as a one-stop shop for all AI development needs throughout the
entire AI lifecycle—from data collection and labeling, to model training and serving, AI agent
creation and deployment, multi-agent communication, AI asset trading, and crowdsourcing
of AI resources. By democratizing the AI development process and lowering barriers to entry
found in existing systems, Sahara AI provides equal access for individuals, businesses and
communities to co-build the future of AI.
Our platform encourages decentralized governance and community-driven innovation. This
approach ensures that Sahara AI not only adapts to the evolving needs of the AI community
but also leads the charge in setting new standards for ethical and equitable AI practices.
By providing a collaborative environment for developing AI, Sahara AI allows individuals,
small and medium-sized businesses (SMBs), and enterprises to work together, share ideas,
and benefit from the collective intelligence and creativity of a global community.
On the journey to build Sahara AI, we are committed to transforming AI from an in-
strument controlled by the few into a resource that empowers all of humanity. Through the
Sahara AI platform, we seek to break down barriers, foster global innovation, and unlock the
full potential of AI for the betterment of society worldwide.

2 The Sahara AI Platform

The Sahara AI Platform is built on three foundational pillars: Sovereignty and Prove-
nance, AI Utility, and Collaborative Economy. Together, these components create a cohesive
platform where every participant can contribute, collaborate, and benefit.

2
 Figure 2: Three foundational pillars of Sahara AI

Pillar 1: Sovereignty and Provenance

In the evolving landscape of AI, Sahara AI emphasizes Sovereignty and Provenance as key
principle that should define how AI assets and their development processes are owned, man-
aged, and governed. These concepts ensure that all stages of the AI development cycle—from
data collection and labeling, to model deployment and application building—are conducted
in a manner that is decentralized, transparent, and inclusive.
Sovereignty encapsulates the idea that ownership and governance of AI assets should be
decentralized and community-driven. This prevents monopolization and ensures that all stake-
holders have a voice in the AI lifecycle. Provenance, on the other hand, ensures transparency
in attributing contributions and tracking the origin and history of the usage and development
of AI assets. It complements sovereignty by providing a comprehensive, immutable record of
all activities and transactions associated with AI assets.
Sahara AI emphasizes the following critical aspects of sovereignty and provenance:

• Ownership and Attribution: Contributors to AI development (such as data providers,

model trainers, and application developers) have verifiable, on-chain ownership and re-
ceive fair attribution for their contributions.

• Decentralization and Governance: Sahara AI promotes equitable and democratized

control over AI assets. Actions and decisions on AI assets are made transparently
through Sahara Blockchain Protocols and DAOs, ensuring all stakeholders have a say
in the AI development cycle. Additionally, the evolution of AI components within the

3
 platform is governed and driven by the community, allowing for a continuously evolving
framework.

• Trust and Accountability: Detailed record-keeping ensures that every piece of data
and every step in the AI lifecycle is meticulously recorded and traceable on the blockchain,
allowing stakeholders to verify the sources and transformations of data and models.

• Interoperability and Accessibility: AI assets and services are designed to be inter-

operable across different platforms and accessible to a wide range of users, promoting
inclusivity and broad participation in the AI ecosystem.

Pillar 2: AI Utility
Sahara AI empowers users throughout different stages of the AI lifecycle by leveraging a
comprehensive technical infrastructure that delivers seamless AI user experience. We strive
to ensure that every participant within the AI development cycle can efficiently develop,
deploy, and manage AI assets in a trustless, privacy-preserving, and secure environment.
Our platform simplifies operations while embedding strong security measures to defend
against unauthorized access and threats, alongside comprehensive privacy protections focused
on safeguarding user information. These features not only protect user data but also build
trust, enabling users to engage confidently and securely with AI technology.
To deliver maximum utility across all stages of the AI development cycle, Sahara AI
focuses on five key aspects:

• Usability: Sahara AI streamlines the AI development cycle across processes from data
curation, to model development, and agent deployment. This allows participants to
leverage AI to enhance productivity, create high-utility applications and achieve positive
real-world outcomes.

• User-Centric Experience: Sahara AI provides an out-of-the-box experience for all

participants in the AI development cycle. Regardless of technical expertise, every par-
ticipant can effortlessly engage with AI technology.

• Security and Privacy: Users benefit from state-of-the-art security measures and pri-
vacy protection. Users can confidently manage their AI assets and computations without
compromising usability.

• High-performance Infrastructure: Sahara AI’s infrastructure supports cutting-edge

AI paradigms and offers a comprehensive toolkit for users to work on advanced AI models
and applications.

• AI-Native Blockchain: Sahara AI is built on the Sahara Blockchain, a Layer 1

blockchain specifically designed with built-in protocol and precompiles for AI trans-
actions across the entire AI lifecycle on the Sahara AI platform.

Pillar 3: Collaborative Economy

The Sahara AI collaborative economy is designed to enable monetization and attribution,
ensuring that all participants are rewarded for their contributions. This means:

• Fair Compensation and Recognition: Users are rewarded in proportion to their

contributions based on provenance of the AI development process, addressing challenges
of global economic disparities.

4
 • Inclusive Participation: The collaborative economy attracts simultaneous partici-
pation from individuals, SMBs, and enterprises, promoting a diverse and vibrant AI
community.

• Trustless Transactions: The Sahara AI platform enables users to monetize their AI

assets through a transparent and efficient process.

3 Design
Building on these three pillars, Sahara AI offers a platform where every participant can
contribute, collaborate, and benefit. The platform is built on a layered architecture designed to
securely and comprehensively support users and developers throughout the entire AI lifecycle.
As illustrated in Figure 3, the Sahara AI platform is structured into four layers:

• The Application Layer acts as the user interface and the major interaction point for
the Sahara AI platform. It provides native built-in applications for users to build and
monetize AI assets.

• The Transaction Layer features the Sahara Blockchain—a Layer 1 blockchain in-
frastructure that manages provenance, access control, attribution and other AI-related
transactions across the AI lifecycle.

• The Data Layer provides abstractions and protocols for data storage, access, and
transfer. It integrates both on-chain and off-chain components to provide seamless data
management across the AI lifecycle.

• The Execution Layer offers essential off-chain infrastructure to support AI utility,

covering a full spectrum of AI functionalities. It provides versatile AI computation
protocols and dynamically allocates computational resources to maximize performance,
ensure scalability, and enhance robustness.

3.1 Application Layer

The Application Layer of the Sahara AI platform serves as the primary interface for partici-
pants of the platform, providing the natively built-in toolkits and applications to enhance user
experience. At its core, the Application Layer is designed with components for participants
to maximize their engagement within the AI ecosystem regardless of technical expertise.

3.1.1 Functional Components

The functional components are the foundational elements that support the robust operations
and security measures of the application layer. These components are designed to ensure the
safe storage, efficient management, and effective deployment of AI.

Sahara ID Sahara ID serves as the cornerstone of identity management within the Sahara
AI platform. It act as the unique identifier for all participants, whether they are AI entities or
human users. This system provides robust identity verification and reputation management,
ensuring secure and transparent interactions across the platform. Through their Sahara ID,
participants gain secure access to AI assets they own or are licensed to, and can track and
manage their contributions and reputation within the ecosystem. Sahara ID also plays a
crucial role in facilitating attributions. It meticulously records each participant’s contributions
to AI projects, ensuring that attribution is clearly tracked and AI “copyrights” are upheld.

5
Figure 3: This layered diagram illustrates the technical architecture of the Sahara blockchain
platform which consists of four inter-related layers. This hybrid infrastructure with both
on-chain and off-chain protocols allows users and developers to effectively contribute to and
benefit from the entire AI development cycle.

6
Sahara Vaults Sahara Vaults are private and secure repositories storing and managing AI
assets, covering both local storage on user nodes and cloud storage on public nodes. These
vaults offer advanced security features, ensuring that all data and assets are protected against
unauthorized access and potential threats. Sahara Vaults strives maintain the privacy, security
and integrity of proprietary AI assets.

Sahara Agent Sahara Agents are AI-driven entities within the Sahara AI platform, each
composed of three integral components—Brain, Perceptor, and Actor. Each component is
specifically designed to perform distinct functions:

• Brain: The strategic core responsible for thought, memory, planning and reasoning. It
processes information and makes informed decisions. Features include:

– Persona Alignment: Tailors the agent’s responses and behaviors to align with spe-
cific user personas, ensuring customized interaction.
– Lifelong Learning: Employs feedback mechanisms and reinforcement learning to
continually enhance its capabilities over time.

• Perceptor: Handles input from various sources, analyzing and interpreting data to in-
form the Brain’s decisions. It features:

– Multimodal Perception: Processes and interprets multiple types of data inputs,

including, but not limited to, visual and auditory.

• Actor: Take the actions as determined by the Brain, based on insights provided by the
Perceptor. It features:

– Tool Utilization: Leverages a wide range of tools and resources to execute actions,
such as searching the web.

3.1.2 Interactive Components

The interactive components of the application layer directly facilitate user interaction with
other users and AI entities, enabling users to actively engage with and utilize AI assets on
the platform. These components make the platform’s capabilities accessible and practical for
a diverse range of applications.

Sahara Toolkits Sahara Toolkits are development and deployment tools designed for par-
ticipants who aim to create and refine AI assets on the Sahara AI platform. These toolkits
cater to a diverse audience by bridging the gap between technical expertise and innovative
execution. For technical users, we provide the Sahara SDK & API, which includes tools
for programming, integration, and customization, allowing for the development of complex
AI functionalities tailored to diverse needs. For users who are less tech-savvy, the Sahara
No-Code/Low-Code toolkits make AI development accessible through intuitive interfaces and
pre-built templates. These platforms empower all users, regardless of their technical skill
level, to participate actively in creating and deploying AI assets.

Sahara AI Marketplace Sahara AI Marketplace is the decentralized hub for publishing,

monetizing and trading of AI assets. It features a comprehensive portfolio of high-quality AI
assets, including proprietary AI agents, models, and datasets. The marketplace integrates
seamlessly with Sahara ID to facilitate ownership protection and access control, leveraging

7
blockchain technology to ensure transparency and security in transactions as well as prove-
nance. Additionally, it offers dynamic licensing and various monetization options, providing
flexible and innovative solutions to meet the diverse needs of users.

3.2 Transaction Layer

The Transaction Layer of the Sahara AI Platform features the Sahara Blockchain. The
Sahara Blockchain is a Layer 1 blockchain meticulously designed to meet the platform’s com-
prehensive needs. It comes with protocols for managing ownership, attribution, and a wide
range of AI-related transactions on the platform. The Sahara Blockchain is pivotal in uphold-
ing the sovereignty and provenance of AI assets.

3.2.1 Sahara Blockchain AI Native Features

The Sahara Blockchain distinguishes itself as an AI-Native blockchain through its integration
of specialized Sahara AI-Native Precompiles (SAPs) and Sahara Blockchain Protocols (SBPs)
that support essential tasks throughout AI lifecycle tasks.

Sahara AI-Native Precompile The Sahara Blockchain integrates Sahara AI-Native Pre-
compile (SAP), which are built-in functions that operate at the native level of the blockchain.
These SAPs are precompiled to enable faster execution, lower computational overhead, and
reduced gas costs.

• Training Execution SAPs: This class of SAPs facilitates the invocation, recording,
and verification of off-chain AI training processes. They enable the blockchain to inter-
face seamlessly with off-chain training environments and ensure that training activities
are accurately recorded. Additionally, these SAPs verify the integrity and authenticity
of the training computations performed outside the blockchain. Their aim is to enhance
the credibility and reliability of AI models developed within the Sahara AI Platform.

• Inference Execution SAPs: Inference Execution SAPs support the invoking, record-
ing, and verification of AI inference results generated off-chain. These SAPs invoke
off-chain AI computations required for inference and provide a robust mechanism to
verify that the AI inferences, which may include predictions, classifications, or other
outputs, are derived from legitimate computations. By integrating these functionalities,
the Sahara Blockchain ensures that all AI inference activities are transparent, verifiable,
and accountable.

Sahara Blockchain Protocols The Sahara Blockchain also implements AI-specific pro-
tocols through smart contracts, collectively known as Sahara Blockchain Protocols (SBP).
These SBPs provide a structured and secure framework for managing various aspects of the
AI lifecycle. They ensure that AI assets and computation results are handled transparently
and reliably.

• AI Asset Registry SBPs: These protocols manage the initial registration and track-
ing of AI assets on the blockchain. They establish a comprehensive ledger that uniquely
identifies AI models, datasets, agents, and other AI-related assets, ensuring their prove-
nance is verifiable. The AI Asset Registry SBPs focus on the static aspects of AI asset
management, such as recording the creation, identity, and origin of assets.

• AI Licensing SBPs: The AI Licensing SBPs decide on-chain rights to access or uti-
lize AI assets. It enforces access control with different types of on-chain licenses (see

8
 section 3.5.3 for details) ensuring that only authorized entities can use or access specific
AI assets. This protocol helps maintain the security and proper usage of AI assets,
facilitating compliant and controlled distribution of AI capabilities.

• AI Ownership SBPs: These protocols maintain clear, non-transferable, and non-

fungible ownership records of AI assets. It ensures that the ownership details of AI
models and datasets are securely stored on-chain, providing indisputable proof of own-
ership. This helps in promoting the “copyright” of AI and managing asset ownership
transparently.

• AI Attribution SBPs: The AI Attribution SBPs track ongoing contributions through-

out the AI lifecycle and manage the distribution of rewards based on these contributions.
These protocols ensure that all inputs to the development and evolution of AI assets
are recorded and that any revenue generated is fairly allocated to contributors.

3.2.2 Sahara Blockchain Design

The Sahara Blockchain employs a Proof-of-Stake consensus mechanism and focuses on pro-
viding a suite of AI native features to support application layer and work seamlessly with
the execution layer to achieve strong AI utility. We utilize the Tendermint algorithm [1] for
Byzantine Fault-Tolerant consensus. This ensures a high degree of fault tolerance and enables
the network to reach consensus even in the presence of malicious nodes. Additionally, the
Sahara Blockchain follows a modular design to offer flexibility in administrative domains and
supports scalable solutions. Our design choice aims to address the following chain features:

• Efficiency: Building on top of the Tendermint algorithm, the Sahara Blockchain not
only enjoys the inherent benefits of high-performance properties such as fast block con-
firmation time, quick average block time, and near-instant finality, but also enhances
them further by introducing additional optimizations tailored for real-time data pro-
cessing and high-speed transactions.

• Scalability: Our modular design supports horizontal scalability and off-chain scaling
solutions, such as Layer 2 solutions. We plan to support rollup solutions to alleviate the
load on the main chain to ensure efficient scalability as demand grows. Rollups aggre-
gate multiple transactions off-chain and submit them to the main blockchain, thereby
significantly enhancing throughput and lowering transaction costs. Our approach aim
to maintain high levels of security and decentralization while providing a robust solution
for increased usage in the future.

• Interoperability: We design Sahara Cross-chain Communication (SCC) Protocol to

facilitate seamless interactions with other blockchains. Our protocol allows for the se-
cure and permissionless transfer of any type of data encoded in bytes between different
blockchains without relying on third-party intermediaries. By leveraging SCC, we en-
sure trustless and permissionless interactions, where any party can operate a relayer to
transfer information between blockchains. Additionally, it supports cross-chain bridges
to enable the seamless transfer of assets between different blockchain networks.

• EVM-Compatibility: The Sahara Blockchain’s built-in virtual machine is fully com-

patible with the Ethereum Virtual Machine (EVM). This compatibility allows developers
to leverage the extensive ecosystem of Ethereum tools, resources, and community sup-
port, including frameworks like Solidity. Developers can effortlessly write and deploy
smart contracts on the Sahara Blockchain, ensuring they can also operate across other

9
 EVM-compatible blockchains with minimal to no code modifications. We provide de-
velopers with a familiar and powerful environment to build decentralized applications
efficiently on Sahara.

• Low Gas Fees: The Sahara Blockchain implements a highly efficient fee structure to
minimize transaction costs. By optimizing transaction batching and utilizing dynamic
fee mechanisms, we aim to keep gas fees economically viable, even as network demand
grows. We are striving to achieve cost-efficiency on our blockchain to make our platform
more accessible and attractive for developers and users in an effort to foster greater
participation and engagement within the ecosystem.

3.3 Data Layer

The Data Layer in the Sahara AI platform is an abstraction designed to optimize data man-
agement throughout the AI lifecycle. It acts as a vital interface that connects the execution
layer to diverse data management mechanisms and seamlessly integrating both on-chain and
off-chain data sources. This conceptual layer not only ensures efficient access to data but also
enhances system integrity and performance.

3.3.1 Data Components

On-chain Data On-chain data includes, but is not limited to, critical AI asset metadata,
attributions, commitments, and proofs. This ensures all contributions and interactions within
the Sahara AI platform are transparent and accountable.

Off-chain Data Significant datasets, AI models, and supplemental information are stored
off-chain due to storage limitations and cost considerations associated with on-chain data
management. By utilizing off-chain storage solutions, we can handle vast amounts of data
without compromising performance while remaining efficient and cost-effective.

3.3.2 Data Management

Security At Sahara AI, we prioritize security above all to protect our platform and user
data. We implement a comprehensive suite of security measures to ensure robust protection
across all interactions:

• Advanced Encryption: We secure all data using the latest encryption practices, which
ensures that sensitive information remains protected both in transit and at rest. This
level of security is crucial for maintaining confidentiality and integrity within our ecosys-
tem.

• Access Controls: In synergy with AI licensing SBPs, our platform implements on-chain
licenses that serve as stringent access control systems. These controls strictly limit data
access to authorized personnel only, significantly enhancing the security and compliance
of our platform. Furthermore, this approach leverages the inherent properties of the
Sahara Blockchain, such as transparency and immutability, to ensure that these controls
are not only effective but also verifiable and secure.

• Private Domain Storage: We enable users to store their data in private domains, which
offer enhanced security features while allowing seamless interaction with our platform.
This capability ensures that users retain control over their sensitive data, while still
benefiting from the robust functionality and connectivity of our platform.

10
Data Availability At the Sahara AI platform, we proactively tackle the Data Availability
(DA) problem by implementing off-the-shelf solutions that ensure all block data is verifiably
accessible to all network participants. This strategy is critical not just for maintaining net-
work integrity and trust, but also for enhancing our blockchain’s scalability. By enabling
efficient and reliable data verification, particularly for light clients who do not store the entire
blockchain, we significantly reduce the network’s dependency on full nodes. This approach
minimizes bandwidth and storage demands, and facilitates smoother and more scalable net-
work operations. Integrating these DA solutions into our Data Layer boosts overall system
performance and enables the network to handle increased volumes of transactions and data
without compromising speed or security.

Indexing In the Sahara AI platform, we actively enhance on-chain data management with
advanced indexing techniques tailored to our blockchain architecture. These methods sig-
nificantly improve data retrieval speeds and query efficiency. We leverage off-the-shelf de-
centralized indexing solutions alongside our proprietary solutions to boost scalability and
performance. This integration within our Data Layer ensures seamless interactions between
execution layers and data sources. Our approach enables the platform to manage increasing
data volumes and complex queries efficiently.

Storage Our storage strategy for off-chain data employs a hybrid model that combines
the strengths of both decentralized and centralized systems to optimize cost-efficiency and
scalability. We use both decentralized storage solutions like IPFS for critical data that benefits
from immutability and distributed hosting, while leveraging more traditional cloud storage
solutions for large volumes of data where speed and availability are paramount. This dual
approach allows us to effectively manage data storage costs while ensuring high availability
and rapid access.

3.4 Execution Layer

The Execution Layer is the off-chain AI infrastructure of the Sahara AI platform that interacts
seamlessly with the Transaction Layer and Data Layer to execute and manage protocols
related to AI computation and functionality. Based on the execution task, it securely pulls
data from the Data Layer and dynamically allocates computational resources for optimal
performance. During execution, the Execution Layer leverages versatile protocols that are
efficient, private, and integrity-preserving. It interacts with the Sahara Blockchain to record
all execution activities and proofs, ensuring provenance and trust. Furthermore, our AI
infrastructure supports high performance, meaning that it is expedient, elastic, and resilient.

3.4.1 High Performance Infrastructure

Sahara AI’s Execution Layer’s underlying infrastructure is designed to support high-performance
AI computation with the following properties:

• Expedient: The Execution Layer is building to ensure rapid and reliable performance
by efficiently coordinating AI computations across various contributors and participants
within our infrastructure.

• Elastic: To handle varying levels of traffic, the Execution Layer is building robust
autoscaling mechanisms. This feature will ensure our infrastructure remains highly
available under high traffic conditions.

11
 • Resilient: The Execution Layer is built with fault tolerance to ensure system stability
and reliability. Further complemented by the Sahara Blockchain, our infrastructure is
partition tolerant. In the event of failures, the system can quickly recover to maintain
the integrity of the processing workflow and minimizing downtime.

3.4.2 Abstractions
Abstractions are foundational to the implementation of various AI assets on the Sahara AI
platform. They provide the necessary conceptual framework for managing datasets, AI mod-
els, computation resources, vaults, and AI agents.

Core Abstractions Core abstractions are essential components that form the basis of AI
operations on the Sahara AI platform.

1 Datasets The datasets abstraction represents the curated data from our participants
that fuel AI training and augment inference processes.

2 AI Models The AI Model abstraction encapsulates the models and architectures for a
variety of machine learning tasks. Our primary models of interest are generative models
[2, 3, 4, 5], with an emphasis on large language models (LLMs) [6, 7, 8]. Sahara AI
optimizes specifically for transformer-based models [9, 10].

3 Computation The computation abstraction encompasses the resources that supports

the execution of AI tasks. Sahara AI’s infrastructure dynamically allocates these re-
sources for optimal performance and cost-efficiency. This includes not only cloud-based
GPUs, but also decentralized contributions from network participants.

High-Level Abstractions High-level abstractions build upon core abstractions and pro-
vides higher-level functionalities and integrations on the Sahara AI platform.

1 Vaults Vaults abstractions are the execution interfaces behind Sahara Vault. They are
structured to manage the accessibility and effective use of AI assets and are crucial for
operational processes.

2 AI Agents AI agents abstractions are the exeuction interfaces behind Sahara Agent.
Specifically, we are interested in LLM-based agents [11, 12, 13] where LLMs are the
brains. These abstractions enable the agents to perform complex reasoning, engage in
natural language interactions, and execute a range of cognitive and decision-making
tasks efficiently [14, 15, 16, 17].

3.4.3 Protocols for the Execution Layer

The Execution Layer of the Sahara AI platform orchestrates complex AI operations through a
suite of specialized protocols designed to facilitate efficient interactions among various abstrac-
tions. These protocols manage a wide range of activities, including AI access, training and
execution. They work seamlessly with Sahara Blockchain Protocols (SBPs) to meticulously
record execution details. This includes, but not limited to, tracking who performed specific
actions, what resources were accessed, and the contributions made by different entities.

12
Abstraction Execution Protocols Abstraction execution protocols provide the necessary
framework to ensure that all high-level abstractions operate efficiently and securely.

• Vault Execution Protocols The Vault Execution Protocols standardize interactions

with the vaults within the Sahara AI platform’s Execution Layer, detailing how vaults
are accessed and utilized for various AI processes.

– Direct Access Protocol: This protocol allows users to perform targeted queries
on the vault to retrieve specific pieces of information quickly and securely.
– Downstream Model Training Protocols: These protocols facilitates the use
of stored data within the vault for training downstream AI models. It supports a
diverse array of training paradigms [10, 6, 18, 19] from the use of both unstructured
data for pre-training to structured data for subsequent fine-tuning.
– Retrieval-Augmented Generation (RAG) Protocol: The RAG [20, 21] pro-
tocol harnesses relevant data retrieved from the vault to enhance the outputs of
generative models.

• Agent Execution Protocols The Agent Framework Protocols manage the interactions
and coordination of AI agents within the Execution Layer. These protocols include:

– Communication Protocols: Sahara AI supports a range of communication

strategies to suit different coordination needs:
∗ Hierarchical Communication: In this strategy, one agent acts as a leader,
with other agents reporting to it [22, 23, 24].
∗ Peer-to-Peer Communication: In this strategy, all agents interact as equals,
sharing information and making decisions collaboratively [12, 25, 26].
– Multi-Agent Coordination Protocols: Effective coordination and orchestra-
tion of multiple agents are essential for achieving complex objectives. Within the
Sahara AI platform, these protocols involves sophisticated task allocation, synchro-
nization, evolution, and collaboration strategies.

Collaborative Computation Protocols The Execution Layer introduces Collaborative

Computation Protocols to facilitate joint AI model development and deployment among multi-
ple participants. Under this collaborative framework, our solution includes privacy-preserving
compute modules and computation fraud proof mechanisms to ensure privacy and integrity
throughout the AI development process, as well as Parameter-Efficient Fine-Tuning (PEFT)
Modules for efficient tuning of AI models.

• Collaborative Model Training Protocols Collaborative model training is a key

component that allows multiple users or systems to jointly build AI models. This
includes:

– Decentralized Training: Sahara AI is building to support decentralized train-

ing, which allows participants to contribute computational resources [27] and train
models collectively as a distributed training system [28].
– Model Aggregation: Beyond decentralized training, Sahara AI is also building to
support model aggregation with cutting-edge model merging techniques [29, 30, 31]
to allow multiple participants to combine their proprietary models into a single,
collective model.

13
 • Collaborative Model Serving Protocols Collaborative model serving allows multi-
ple users or systems to jointly serve AI models for inference. This includes:

– Decentralized Serving: Similar to decentralized training, participants will be

able to host segments of models for serving [32, 33, 27]. We are building this to be
scalable and efficient to foster seamless collaboration among participants.

• Add-on Modules

– PEFT Modules Sahara AI is building to support various PEFT techniques [34,

35, 36], such as Low-Rank Adaptation (LoRA) [37], to efficiently and portably
customize AI models. These PEFT modules enhance model performance with
minimal computational overhead and can be easily integrated as plug-and-play
components for large models as LLMs.
– Privacy Preserving Compute Modules Privacy-preserving compute modules
within the Execution Layer ensure that proprietary data and model parameters
remain secure and confidential throughout the processing lifecycle. To cater to
a wide range of use cases and scenarios, we offer a suite of privacy-preserving
techniques:
∗ Differential Privacy (DP): Employed during both inference and training to
add noise to data or computations [38, 39, 40, 41]. This ensures that individual
data points cannot be distinguished.
∗ Homomorphic Encryption (HE): Used for scenarios involving two-party
computations, lightweight computation tasks, and preprocessing computations
such as random number generation for larger models [42].
∗ Secret Sharing (SS): Utilized for general transformer structure inference
tasks [43, 44]. SS divides data into parts and distributes them across different
nodes, preventing any single node from accessing the complete data.
– Computation Fraud Proof Modules We are also building to support fraud
proofs generation [45, 46] for computation results from our collaborative computa-
tion protocol. These proofs will be verified on-chain to protect against malicious
or faulty computations.

3.4.4 Integrations
The Execution Layer integrates seamlessly with other layers of the Sahara AI platform to
provide a cohesive and efficient AI infrastructure.

• Transaction Layer The Execution Layer collaborates with the Transaction Layer to
manage the sovereignty and provenance of AI assets. All execution, contribution, and
usage activities involving AI assets—such as model access, data usage, and computation
results—that occur on the execution layer are logged and sent to the Sahara Blockchain
for recording. This is facilitated by SAPs and SBPs.

• Data Layer The Execution Layer utilizes the vault abstraction and corresponding
protocols to interact with the Data Layer, securely accessing data from vaults for training
and RAG. This integration leverages mechanisms from both the Data Layer and the
Execution Layer’s computation protocols to ensure privacy, security, and integrity.

14
3.5 Economic System
The economic system of Sahara AI is designed to create a collaborative, fair, and rewarding
environment where all participants can benefit from their contributions. Whether it’s operat-
ing a node, contributing knowledge and data, fine-tuning a model, or building an autonomous
agent, Sahara AI ensures fair compensation, transparent transactions, and inclusive partici-
pation. By leveraging Sahara Blockchain and its SAPs, Sahara AI provides a robust system
that upholds these principles, fostering trust and equitable opportunities for all.

AI asset Investment Model Sahara AI further enhances the collaborative economy by

adopting an innovative investment model. This model allows users to invest resources such as
capital, data, compute, and technical AI expertise in exchange for stake in AI models, agents,
and applications. This mirrors the strategies of major GPU/cloud giants like Microsoft and
Amazon, who invest in foundational model companies such as OpenAI and Anthropic. These
giants offer GPU resources in exchange for stake, which in the future secures their vested
interest in AI advancements. This approach not only provides AI developers with essential
computational power but also aligns the interests of hardware suppliers with the success
of foundational model companies. As a result, it creates a mutually beneficial ecosystem
that accelerates AI innovation while potentially reshaping industry dynamics. By extending
this model across the entire AI lifecycle—from data to compute, model development to final
application—Sahara AI enriches the collaborative economy, making it more dynamic and
inclusive.

3.5.1 Economic Roles

In the economic system, the Sahara AI platform includes the following important roles:
• Developer: Developers create AI models, tools, and applications on the Sahara AI
platform. Their incentives include revenue through usage fees, licensing, and sales of
their creations.
• Knowledge Provider: Knowledge providers consist of annotators and reviewers. An-
notators are responsible for providing personalized data and professional knowledge,
while reviewers ensure the quality of the collected data. They are compensated based
on the quality and utility of the data they provide.
• Consumer: Consumers are end-users and businesses that utilize AI models, tools, and
applications from the platform. They pay for services based on usage, which ensures a
sustainable economic cycle. Consumers benefit from access to cutting-edge AI solutions
that can enhance their operations and drive innovation.
• Investor: Investors provide capital and resources (GPUs, Cloud servers, RPC nodes,
etc.) to support the development and deployment of AI assets. In return, they receive
the stake of the AI assets, earn revenue generated by AI assets and potential income
from the price appreciation of the AI assets.
• Operators: Operators are responsible for contributing storage and computing power
and maintaining the network infrastructure, both on-chain and off-chain. They ensure
the smooth operation and security of the platform. Operators are incentivized through
rewards for their role in maintaining network stability and performance.
• Validators: Validators maintain the integrity and security of the Sahara Blockchain.
They verify transactions and secure the network. Validators earn rewards for their
critical role in ensuring the trustworthiness and reliability of the blockchain.

15
3.5.2 Growth Flywheel

Figure 4: Sahara Dual Growth Flywheel

Overall, as illustrated in Figure 4, Sahara AI employs the dual growth flywheel model to
drive the sustainable and scalable growth in Web3 ecosystem and AI ecosystem. This dual
approach ensures that both Web3 ecosystem and AI ecosystem seamlessly work in synergy
under a self-sustaining loop.

AI Ecosystem The AI ecosystem growth flywheel starts with developers who release and
engage in AI relevant tasks. These AI relevant tasks attract knowledge providers to co-operate
with developers creating AI assets within the ecosystem. The AI assets created then attract
consumers, who contribute to an increase in revenue from AI assets. This increase in revenue
captures the interest of investors, leading to shares appreciation. This appreciation triggers
a rise in developer’s net profit, which in turn, draws more developers into the ecosystem.
This continuous influx of developers leads to more innovation, further driving the creation of
valuable AI assets and perpetuating the growth cycle of the AI ecosystem.

Web3 Ecosystem The Web3 ecosystem growth flywheel begins with developers who drive
product innovation by building applications on the Sahara Blockchain. This innovation ac-
celerates user engagement, attracting more users to the platform. Increased user engagement
helps to promote a flourishing ecosystem, leading to an increase in the number of transactions.
As transactions grow, the revenue of basic services improves. This enhancement in revenue
attracts more operators and validators to the ecosystem, which further enhances infrastruc-
ture stability. A stable infrastructure, in turn, attracts even more developers, fueling further
product innovation and continuing the cycle of growth.

Synergy The AI and Web3 ecosystems are closely interconnected, with developers at the
core of both. In the AI ecosystem, developers engage in AI relevant tasks to create AI
assets that drive innovation and attract investment, fueling further growth of developers.
More developers with these AI advancements are then leveraged within the Web3 ecosystem,

16
enhancing product innovation and user engagement. The increased engagement boosts trans-
actions and revenue in the Web3 ecosystem, which in turn provides a stable platform for more
AI developments, attracting more developers. This synergy ensures that as each ecosystem
grows, it supports and accelerates the growth of the other, creating a powerful cycle of mutual
reinforcement.

3.5.3 Capitalization of AI Assets

The capitalization of AI assets within the Sahara AI platform is structured into three distinct
instruments: Receipts, Shares, and Licenses.

Receipt Receipt is an on-chain, non-transferable and non-fungible digital proof for AI assets
ownership. Some specific parameters of the receipts, such as eligibility, allocation ratios, or
the mechanism followed, will be finalized before developers and knowledge providers co-create
AI assets. The receipts of high-quality AI assets will significantly benefit AI developers and
relevant contributors in building their on-chain reputation. A higher on-chain reputation can
lead to wider visibility within the Sahara AI ecosystem, increased opportunities to sell shares
or licenses of AI assets, additional access to features in the Sahara AI ecosystem, and so forth.
As such, AI developers are highly motivated to create high-quality AI assets.

Share Share is an on-chain digital proof that represents the right to revenue sharing of AI
assets. The percentage of revenue sharing depends on the number of shares held. Some specific
parameters of the shares, such as eligibility, allocation ratios, or the mechanism followed, will
be finalized before developers and knowledge providers co-create AI assets. The revenues from
shares are generally derived from the revenue generated from the sales of licenses for AI assets
or commission fees from trading shares and licenses.
Owners can either retain their shares to earn part of the revenue generated by AI assets or
sell their shares on the secondary market for immediate arbitrage. High-quality AI assets will
ensure a continuous flow of revenue for the holders of shares, which will also lead to an increase
in the price of shares in the secondary market. This will further motivate AI developers to
create high-quality AI assets and incentivize consumers to buy and trade the different types
of licenses.

License License is an on-chain digital proof that represents permission to access or utilize
AI assets. Licenses have different rights to access or utilize the AI assets:

• Partnership License: Access permissions and payment methods are concluded in a cus-
tomized agreement, typically involving a long-term profit-sharing model based on the
user’s benefits.

• API License: Access is granted through a secure authentication process and requires
approval based on predefined usage criteria with fixed payment for each call.

• Full-access License: Full access to the AI assets with all internal parameters with one-
time payment.

• Long-term License: Unlimited calls to AI assets within a specific time duration with
one-time payment.

Some specific parameters of the license, such as eligibility, allocation ratios, or the mecha-
nism followed, will be finalized before developers and knowledge providers co-create AI assets.

17
4 Governance
The governance of Sahara AI Platform emphasizes decentralized and community-driven inno-
vation and decision-making. Actions and decisions are made transparently through the Sahara
DAO, which empowers the users who made significant contributions to propose, discuss, and
vote on key initiatives. The Sahara Foundation supports the initial setup of the Sahara DAO,
guiding the community towards a fully decentralized governance. Our governance model
ensures broad participation in the evolution of Sahara AI Platform.

4.1 The Sahara DAO

The Sahara DAO is dedicated to complete democratization, minimize governance to essential
functions, promoting autonomy and innovation. Users who made significant contributions to
the Sahara ecosystem play a crucial role in shaping the platform, proposing, discussing, and
casting votes on various governance proposals. They have the option to either vote personally
or assign their voting rights to a trusted representative.
The Sahara DAO is also designed to ensure the platform’s independence and transparency,
reflect the broader community’s interests, prioritize fair compensation for contributions to
the ecosystem and strategically manage resources to balance short-term needs with long-term
goals.

4.2 The Sahara Foundation

The Sahara Foundation represents a vital element within the expansive, dynamic network of
enterprises, collectives, and individuals dedicated to realizing the Sahara Vision. Its function
is to facilitate the creation of the Sahara DAO, foster the growth of the Sahara ecosystem,
and advance the underlying technology. During the formative phase of the Sahara DAO, the
Foundation supports the Sahara DAO to identify effective strategies on long-term development
and significantly enhance its prospects for a decentralized, transparent and community-driven
governance.
The Foundation also supports the SBP as an open-source initiative, fostering community
engagement to expand its technical infrastructure and sponsoring research on scalability,
security, and decentralization.

5 The Future
In the future enabled by Sahara AI, artificial intelligence is no longer controlled by a select few
but is instead a shared resource, accessible to all. This new world of AI is guided by principles
of transparency, inclusivity, fairness, and, most importantly, user privacy and control.
In this future, AI enthusiasts can log into Sahara AI to explore a wide array of AI modules,
fine-tune these models with their proprietary datasets, and share their improvements with the
community. They have full control over their AI assets, ensuring that their contributions are
protected while still earning rewards.
Data scientists, once constrained by centralized barriers, can now leverage Sahara AI’s
vast pool of data and models to advance their research. With the platform’s robust privacy
features, they can refine models and deploy solutions while maintaining control over their
data and intellectual property.
For business leaders, Sahara AI provides a strategic advantage with tailored AI solutions
that align with their company’s specific needs. They can drive innovation within their or-
ganizations, secure in the knowledge that their data is protected and their AI assets remain
under their control.

18
 Meanwhile, individuals who may not be AI experts can also find a welcoming environment
within Sahara AI. They can engage with cutting-edge projects, contribute their own expertise,
participate in governance, and support ethical AI initiatives, all while maintaining full control
over their contributions and earning rewards that reflect their involvement.
As the Sahara AI ecosystem evolves, we seek to build a community where everyone—regardless
of background or expertise—has a meaningful role in shaping the future of AI. This is not
just about building a platform; it’s about creating a decentralized, open environment where
the benefits of AI are shared equitably, and where privacy and control over AI assets are
paramount. Together, we will transform AI into a force that empowers individuals and com-
munities alike, guiding us toward a more connected, secure, and fair world.

References
[1] Ethan Buchman, Jae Kwon, and Zarko Milosevic. The latest gossip on bft consensus,
2019.

[2] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint
arXiv:1312.6114, 2013.

[3] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil
Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in
neural information processing systems, 27, 2014.

[4] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models,
2020.

[5] Alec Radford and Karthik Narasimhan. Improving language understanding by generative
pre-training. 2018.

[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla
Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan-
guage models are few-shot learners. Advances in neural information processing systems,
33:1877–1901, 2020.

[7] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux,
Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien
Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and
efficient foundation language models. ArXiv, abs/2302.13971, 2023.

[8] Albert Qiaochu Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Deven-
dra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume
Lample, Lucile Saulnier, L’elio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock,
Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed.
Mistral 7b. ArXiv, abs/2310.06825, 2023.

[9] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in
neural information processing systems, 30, 2017.

[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-
training of deep bidirectional transformers for language understanding. arXiv preprint
arXiv:1810.04805, 2018.

19
[11] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu,
Linxi (Jim) Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent
with large language models. ArXiv, abs/2305.16291, 2023.

[12] Sirui Hong, Xiawu Zheng, Jonathan P. Chen, Yuheng Cheng, Ceyao Zhang, Zili Wang,
Steven Ka Shing Yau, Zi Hen Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, and
Chenglin Wu. Metagpt: Meta programming for multi-agent collaborative framework.
ArXiv, abs/2308.00352, 2023.

[13] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with
dynamic memory and self-reflection. ArXiv, abs/2303.11366, 2023.

[14] Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. Dynamic llm-agent net-
work: An llm-agent collaboration framework with agent team optimization. ArXiv,
abs/2310.02170, 2023.

[15] Murray Shanahan, Kyle McDonell, and Laria Reynolds. Role play with large language
models. Nature, 623:493–498, 2023.

[16] Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji.
Unleashing the emergent cognitive synergy in large language models: A task-solving agent
through multi-persona self-collaboration. In North American Chapter of the Association
for Computational Linguistics, 2023.

[17] Wenlong Huang, P. Abbeel, Deepak Pathak, and Igor Mordatch. Language models
as zero-shot planners: Extracting actionable knowledge for embodied agents. ArXiv,
abs/2201.07207, 2022.

[18] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela
Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman,
Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welin-
der, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow
instructions with human feedback, 2022.

[19] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning,
and Chelsea Finn. Direct preference optimization: Your language model is secretly a
reward model, 2023.

[20] Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin,
Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian
Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp
tasks. ArXiv, abs/2005.11401, 2020.

[21] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei
Sun, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language
models: A survey, 2024.

[22] Yongchao Chen, Jacob Arkin, Yang Zhang, Nicholas Roy, and Chuchu Fan. Scalable
multi-robot collaboration with large language models: Centralized or decentralized sys-
tems? arXiv preprint arXiv:2309.15943, 2023.

[23] Rui Hao, Linmei Hu, Weijian Qi, Qingliu Wu, Yirui Zhang, and Liqiang Nie. Chatllm
network: More brains, more intelligence. ArXiv, abs/2304.12998, 2023.

20
[24] Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu,
Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn
to cooperate in organized teams. arXiv preprint arXiv:2403.12482, 2024.

[25] Yashar Talebirad and Amirhossein Nadiri. Multi-agent collaboration: Harnessing the
power of intelligent llm agents. ArXiv, abs/2306.03314, 2023.

[26] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang,
Zhaopeng Tu, and Shuming Shi. Encouraging divergent thinking in large language models
through multi-agent debate. ArXiv, abs/2305.19118, 2023.

[27] Alexander Borzunov, Dmitry Baranchuk, Tim Dettmers, Max Ryabinin, Younes Belkada,
Artem Chumachenko, Pavel Samygin, and Colin Raffel. Petals: Collaborative inference
and fine-tuning of large models, 2023.

[28] Ziheng Jiang, Haibin Lin, Yinmin Zhong, Qi Huang, Yangrui Chen, Zhi Zhang, Yanghua
Peng, Xiang Li, Cong Xie, Shibiao Nong, Yulu Jia, Sun He, Hongmin Chen, Zhihao
Bai, Qi Hou, Shipeng Yan, Ding Zhou, Yiyao Sheng, Zhuo Jiang, Haohan Xu, Haoran
Wei, Zhang Zhang, Pengfei Nie, Leqi Zou, Sida Zhao, Liang Xiang, Zherui Liu, Zhe Li,
Xiaoying Jia, Jianxi Ye, Xin Jin, and Xin Liu. Megascale: Scaling large language model
training to more than 10,000 gpus, 2024.

[29] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge
fusion by merging weights of language models, 2023.

[30] Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael
Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon
Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned
models improves accuracy without increasing inference time, 2022.

[31] Michael Matena and Colin Raffel. Merging models with fisher-weighted averaging, 2022.

[32] Zhuoran Zhao, Kamyar Mirzazad Barijough, and Andreas Gerstlauer. Deepthings: Dis-
tributed adaptive deep learning inference on resource-constrained iot edge clusters. IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems, 37:2348–
2359, 2018.

[33] Bingyang Wu, Yinmin Zhong, Zili Zhang, Gang Huang, Xuanzhe Liu, and Xin Jin. Fast
distributed inference serving for large language models. ArXiv, abs/2305.05920, 2023.

[34] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient
fine-tuning for large models: A comprehensive survey, 2024.

[35] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and
singular vectors adaptation of large language models, 2024.

[36] Yixiao Li, Yifan Yu, Chen Liang, Pengcheng He, Nikos Karampatziakis, Weizhu Chen,
and Tuo Zhao. Loftq: Lora-fine-tuning-aware quantization for large language models,
2023.

[37] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang,
Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021.

21
[38] Martin Abadi, Andy Chu, Ian Goodfellow, H. Brendan McMahan, Ilya Mironov, Kunal
Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016
ACM SIGSAC Conference on Computer and Communications Security, CCS’16. ACM,
October 2016.

[39] Björn Bebensee. Local differential privacy: a tutorial, 2019.

[40] Tong Wu, Ashwinee Panda, Jiachen T. Wang, and Prateek Mittal. Privacy-preserving
in-context learning for large language models, 2023.

[41] Da Yu, Saurabh Naik, Arturs Backurs, Sivakanth Gopi, Huseyin A. Inan, Gautam
Kamath, Janardhan Kulkarni, Yin Tat Lee, Andre Manoel, Lukas Wutschitz, Sergey
Yekhanin, and Huishuai Zhang. Differentially private fine-tuning of language models,
2022.

[42] Tianyu Chen, Hangbo Bao, Shaohan Huang, Li Dong, Binxing Jiao, Daxin Jiang, Haoyi
Zhou, Jianxin Li, and Furu Wei. The-x: Privacy-preserving transformer inference with
homomorphic encryption, 2022.

[43] Kanav Gupta, Neha Jawalkar, Ananta Mukherjee, Nishanth Chandran, Divya Gupta,
Ashish Panwar, and Rahul Sharma. Sigma: secure gpt inference with function secret
sharing. Cryptology ePrint Archive, 2023.

[44] Théo Ryffel, Pierre Tholoniat, David Pointcheval, and Francis R. Bach. Ariann: Low-
interaction privacy-preserving deep learning via function secret sharing. Proceedings on
Privacy Enhancing Technologies, 2022:291 – 316, 2020.

[45] KD Conway, Cathie So, Xiaohang Yu, and Kartin Wong. opml: Optimistic machine
learning on blockchain, 2024.

[46] Haochen Sun, Jason Li, and Hongyang Zhang. zkllm: Zero knowledge proofs for large
language models, 2024.

22
A Glossaries
• AI Models: AI models are algorithms designed to simulate human intelligence by
learning from data. These models can recognize patterns, make decisions, and even
generate content based on what they’ve learned. AI models range from narrow AI,
which specializes in specific tasks, to more general models that handle a broader range
of functions. Examples include neural networks, decision trees, and regression models.

• Generative Models: These AI models are designed to create new content, such as
text, images, music, or video. Generative models work by learning patterns from large
datasets and using this knowledge to generate new data that resembles the original.
Common types of generative models include transformers.

• Transformers: A type of deep learning model architecture that excels in processing

sequential data, like language. Transformers use self-attention mechanisms to weigh the
importance of different parts of the input data simultaneously, making them effective
for tasks like translation, text generation, and more. Transformers are the foundation
for many modern large language models.

• Large Language Models (LLMs): These are a subset of transformers trained on

massive amounts of text data. LLMs are capable of generating human-like text by
predicting the next word in a sequence based on the context provided. They are versatile
and can perform various tasks such as writing, translation, and summarization.

• AI Agent: An AI agent is an autonomous program that interacts with its environment

to perform tasks, make decisions, and achieve specific goals. AI agents can be simple,
like a virtual assistant that schedules appointments, or complex, like a self-driving car
that navigates through traffic.

• Parameter Efficient Fine-Tuning (PEFT): PEFT refers to techniques that enable

fine-tuning large pre-trained models with fewer parameters, making the process more
resource-efficient. PEFT methods like LoRA or adapters focus on updating only a small
part of the model, reducing computational cost while maintaining performance.

• LoRA (Low-Rank Adaptation): LoRA is a PEFT method that adds small, trainable
matrices to the layers of a large language model. Instead of updating the entire model
during fine-tuning, LoRA modifies these low-rank matrices, which represent a smaller
subset of the model’s parameters. This approach significantly reduces the number of
parameters that need adjustment, speeding up the process and reducing resource usage
while maintaining or enhancing performance.

• Retrieval Augmented Generation (RAG): RAG is a hybrid approach that com-

bines retrieval-based methods with generative models. In RAG, the model first retrieves
relevant information from a database or document set and then uses that information
to generate a response. This method allows the model to provide more accurate and
contextually relevant answers, especially in cases where up-to-date or specialized infor-
mation is required.

• Differential Privacy: Differential privacy is a technique used to protect individual

privacy in datasets by adding random noise to the data. This ensures that the data can
be used for analysis without revealing sensitive information about individuals, making
it useful in contexts like sharing private records or training AI models with personal
data.

23
• Model Merging: The process of combining two or more AI models to create a new
model that benefits from the strengths of each original model. Model merging can
be used to improve performance, incorporate diverse knowledge, or combine different
capabilities into a single, more powerful model.

24