"research" 태그가 붙은 하나의 게시물

The demand for deploying large language models (LLMs) on mobile devices is rising, driven by the need for privacy, reduced latency, and efficient bandwidth usage. However, the extensive memory and computational requirements of LLMs pose significant challenges. Enter LinguaLinked, a new system, developed by a group of researchers from UC Irvine, designed to enable decentralized, distributed LLM inference across multiple mobile devices, leveraging their collective capabilities to perform complex tasks efficiently.

The Challenge​

Deploying LLMs like GPT-3 or BLOOM on mobile devices is challenging due to:

• Memory Constraints: LLMs require substantial memory, often exceeding the capacity of individual mobile devices.
• Computational Limitations: Mobile devices typically have limited processing power, making it difficult to run large models.
• Privacy Concerns: Sending data to centralized servers for processing raises privacy issues.

LinguaLinked's Solution​

LinguaLinked addresses these challenges with three key strategies:

1. Optimized Model Assignment:

   • The system segments LLMs into smaller subgraphs using linear optimization to match each segment with a device's capabilities.
   • This ensures efficient use of resources and minimizes inter-device data transmission.

2. Runtime Load Balancing:

   • LinguaLinked actively monitors device performance and redistributes tasks to prevent bottlenecks.
   • This dynamic approach ensures efficient use of all available resources, enhancing overall system responsiveness.

3. Optimized Communication:

   • Efficient data transmission maps guide the flow of information between devices, maintaining the model's structural integrity.
   • This method reduces latency and ensures timely data processing across the network of mobile devices.

A single large language model (LLM) is split into different parts (or segments) and distributed across multiple mobile devices. This approach allows each device to handle only a fraction of the total computation and storage requirements, making it feasible to run complex models even on devices with limited resources. Here's a breakdown of how this works:

Model Segmentation and Distribution​

1. Model Segmentation:
   • The large language model is transformed into a computational graph where each operation within the network is represented as a node.
   • This graph is then partitioned into smaller subgraphs, each capable of functioning independently.
2. Optimized Model Assignment:
   • Using linear optimization, these subgraphs (or model segments) are assigned to different mobile devices.
   • The assignment considers each device's computational and memory capabilities, ensuring efficient resource use and minimizing data transmission overhead between devices.
3. Collaborative Inference Execution:
   • Each mobile device processes its assigned segment of the model.
   • Devices communicate with each other to exchange intermediate results as needed, ensuring the overall inference task is completed correctly.
   • Optimized communication strategies are employed to maintain the integrity of the original model structure and ensure efficient data flow.

Example Scenario​

Imagine a large language model like GPT-3 being split into several parts. One mobile device might handle the initial token embeddings and the first few layers of the model, while another device processes the middle layers, and a third device completes the final layers and generates the output. Throughout this process, devices share intermediate outputs to ensure the complete model inference is executed seamlessly.

Performance and Results​

LinguaLinked's efficacy is demonstrated through extensive testing on various Android devices, both high-end and low-end. Key findings include:

• Inference Speed: Compared to a baseline, LinguaLinked accelerates inference performance by 1.11× to 1.61× in single-threaded settings and 1.73× to 2.65× with multi-threading.
• Load Balancing: The system's runtime load balancing further boosts performance, with an overall acceleration of 1.29× to 1.32×.
• Scalability: Larger models benefit significantly from LinguaLinked's optimized model assignment, showcasing its scalability and effectiveness in handling complex tasks.

Use Cases and Applications​

LinguaLinked is particularly suited for scenarios where privacy and efficiency are paramount. Applications include:

• Text Generation and Summarization: Generating coherent and contextually relevant text locally on mobile devices.
• Sentiment Analysis: Classifying text data efficiently without compromising user privacy.
• Real-time Translation: Providing quick and accurate translations directly on the device.

Future Directions​

LinguaLinked paves the way for further advancements in mobile AI:

• Energy Efficiency: Future iterations will focus on optimizing energy consumption to prevent battery drain and overheating during intensive tasks.
• Enhanced Privacy: Continued improvements in decentralized processing will ensure even greater data privacy.
• Multi-modality Models: Expanding LinguaLinked to support multi-modality models for diverse real-world applications.

Conclusion​

LinguaLinked represents a significant leap forward in deploying LLMs on mobile devices. By distributing the computational load and optimizing resource use, it makes advanced AI accessible and efficient on a wide range of devices. This innovation not only enhances performance but also ensures data privacy, setting the stage for more personalized and secure mobile AI applications.

The rise of large language models (LLMs) and decentralized computing has introduced significant challenges, especially regarding the verification and integrity of AI computations across distributed systems. The 6079 Proof of Inference Protocol (PoIP) addresses these challenges by establishing a robust framework for decentralized AI inference, ensuring reliable and secure computations.

The Challenge: Security in Decentralized AI Inference​

Decentralized AI inference faces the unique problem of ensuring the integrity and correctness of computations performed across a network of distributed nodes. Traditional methods of verification fall short due to the non-deterministic nature of many AI models. Without a robust protocol, it's challenging to guarantee that the distributed hardware returns accurate inference results.

Introducing Proof of Inference Protocol (PoIP)​

6079 Proof of Inference Protocol (PoIP) provides a groundbreaking solution for securing decentralized AI inference. It uses a combination of cryptoeconomic security mechanisms, cryptographic proofs, and game-theoretic approaches to incentivize correct behavior and penalize malicious activity within the network.

Core Components of PoIP​

Inference Engine Standard​

The Inference Engine Standard sets the compute patterns and standards for executing AI inference tasks across decentralized networks. This standardization ensures consistent and reliable performance of AI models on distributed hardware.

Proof of Inference Protocol​

The protocol operates across multiple layers:

1. Service Layer: Executes model inference on physical hardware.
2. Control Layer: Manages API endpoints, coordinates load balancing, and handles diagnostics.
3. Transaction Layer: Uses a distributed hash table (DHT) to track transaction metadata.
4. Probabilistic Proof Layer: Validates transactions through cryptographic and economic mechanisms.
5. Economic Layer: Handles payment, staking, slashing, security, governance, and public funding.

Ensuring Integrity and Security​

PoIP employs several mechanisms to ensure the integrity of AI inference computations:

• Merkle Tree Validation: Ensures that input data reaches GPUs unaltered.
• Distributed Hash Table (DHT): Synchronizes transaction data across nodes to detect discrepancies.
• Diagnostic Tests: Evaluate hardware capabilities and ensure compliance with network standards.

Economic Incentives and Game Theory​

The protocol uses economic incentives to encourage desirable behavior among nodes:

• Staking: Nodes stake tokens to demonstrate commitment and increase their credibility.
• Reputation Building: Successful tasks enhance a node's reputation, making it more attractive for future tasks.
• Competitive Game Mechanisms: Nodes compete to provide the best service, ensuring continuous improvement and adherence to standards.

FAQs​

What is the Proof of Inference Protocol?​

The Proof of Inference Protocol (PoIP) is a system designed to secure and verify AI inference computations across decentralized networks. It ensures that distributed hardware nodes return accurate and trustworthy results.

How does PoIP ensure the integrity of AI computations?​

PoIP uses mechanisms like Merkle tree validation, distributed hash tables (DHT), and diagnostic tests to verify the integrity of AI computations. These tools help detect discrepancies and ensure the correctness of data processed across the network.

What role do economic incentives play in PoIP?​

Economic incentives in PoIP encourage desirable behavior among nodes. Nodes stake tokens to demonstrate commitment, build reputation through successful tasks, and compete to provide the best service. This system ensures continuous improvement and adherence to network standards.

What are the main layers of the PoIP?​

The PoIP operates across five main layers: Service Layer, Control Layer, Transaction Layer, Probabilistic Proof Layer, and Economic Layer. Each layer plays a crucial role in ensuring the security, integrity, and efficiency of AI inference on decentralized networks.

Conclusion​

The 6079 Proof of Inference Protocol represents an interesting advancement in the field of decentralized AI. By ensuring the security and reliability of AI computations across distributed networks, PoIP indicates a new way for broader adoption and innovation in decentralized AI applications. As we move towards a more decentralized future, protocols like PoIP will be useful in maintaining trust and integrity in AI-powered systems.

• source: https://cuckoo.network/blog/2024/07/03/proof-of-inference-protocol-decentralized-ai
• telegram: https://cuckoo.network/tg
• discord: https://cuckoo.network/dc

In decentralized AI, ensuring the integrity and reliability of GPU providers is crucial. The Proof of Sampling (PoSP) protocol, as outlined in recent research from Holistic AI, provides a sophisticated mechanism to incentivize good actors while slashing bad ones. Let's see how this protocol works, its economic incentives, penalties, and its application to decentralized AI inference.

Incentives for Honest Behavior​

Economic Rewards​

At the heart of the PoSP protocol are economic incentives designed to encourage honest participation. Nodes, acting as asserters and validators, are rewarded based on their contributions:

• Asserters: Receive a reward (RA) if their computed output is correct and unchallenged.
• Validators: Share the reward (RV/n) if their results align with the asserter's and are verified as correct.

Unique Nash Equilibrium​

The PoSP protocol is designed to reach a unique Nash Equilibrium in pure strategies, where all nodes are motivated to act honestly. By aligning individual profit with system security, the protocol ensures that honesty is the most profitable strategy for participants.

Penalties for Dishonest Behavior​

Slashing Mechanism​

To deter dishonest behavior, the PoSP protocol employs a slashing mechanism. If an asserter or validator is caught being dishonest, they face significant economic penalties (S). This ensures that the cost of dishonesty far outweighs any potential short-term gains.

Challenge Mechanism​

Random challenges further secure the system. With a predetermined probability (p), the protocol triggers a challenge where multiple validators re-compute the asserter's output. If discrepancies are found, dishonest actors are penalized. This random selection process makes it difficult for bad actors to collude and cheat undetected.

Steps of the PoSP Protocol​

1. Asserter Selection: A node is randomly selected to act as an asserter, computing and outputting a value.

2. Challenge Probability:

   The system may trigger a challenge based on a predetermined probability.

   • No Challenge: The asserter is rewarded if no challenge is triggered.
   • Challenge Triggered: A set number (n) of validators are randomly selected to verify the asserter's output.

3. Validation:

   Each validator independently computes the result and compares it with the asserter's output.

   • Match: If all results match, both the asserter and validators are rewarded.
   • Mismatch: An arbitration process determines the honesty of the asserter and validators.

4. Penalties: Dishonest nodes are penalized, while honest validators receive their reward share.

SpML​

The spML (sampling-based Machine Learning) protocol is an implementation of the Proof of Sampling (PoSP) protocol within a decentralized AI inference network.

Key Steps​

1. User Input: The user sends their input to a randomly selected server (asserter) along with their digital signature.
2. Server Output: The server computes the output and sends it back to the user along with a hash of the result.
3. Challenge Mechanism:
   • With a predetermined probability (p), the system triggers a challenge where another server (validator) is randomly selected to verify the result.
   • If no challenge is triggered, the asserter receives a reward (R) and the process concludes.
4. Verification:
   • If a challenge is triggered, the user sends the same input to the validator.
   • The validator computes the result and sends it back to the user along with a hash.
5. Comparison:
   • The user compares the hashes of the asserter's and validator's outputs.
   • If the hashes match, both the asserter and validator are rewarded, and the user receives a discount on the base fee.
   • If the hashes do not match, the user broadcasts both hashes to the network.
6. Arbitration:
   • The network votes to determine the honesty of the asserter and validator based on the discrepancies.
   • Honest nodes are rewarded, while dishonest ones are penalized (slashed).

Key Components and Mechanisms​

• Deterministic ML Execution: Uses fixed-point arithmetic and software-based floating-point libraries to ensure consistent, reproducible results.
• Stateless Design: Treats each query as independent, maintaining statelessness throughout the ML process.
• Permissionless Participation: Allows anyone to join the network and contribute by running an AI server.
• Off-chain Operations: AI inferences are computed off-chain to reduce the load on the blockchain, with results and digital signatures relayed directly to users.
• On-chain Operations: Critical functions, such as balance calculations and challenge mechanisms, are handled on-chain to ensure transparency and security.

Advantages of spML​

• High Security: Achieves security through economic incentives, ensuring nodes act honestly due to the potential penalties for dishonesty.
• Low Computational Overhead: Validators only need to compare hashes in most cases, reducing computational load during verification.
• Scalability: Can handle extensive network activity without significant performance degradation.
• Simplicity: Maintains simplicity in implementation, enhancing ease of integration and maintenance.

Comparison with Other Protocols​

• Optimistic Fraud Proof (opML):
  • Relies on economic disincentives for fraudulent behavior and a dispute resolution mechanism.
  • Vulnerable to fraudulent activity if not enough validators are honest.
• Zero Knowledge Proof (zkML):
  • Ensures high security through cryptographic proofs.
  • Faces challenges in scalability and efficiency due to high computational overhead.
• spML:
  • Combines high security through economic incentives, low computational overhead, and high scalability.
  • Simplifies the verification process by focusing on hash comparisons, reducing the need for complex computations during challenges.

Summary​

The Proof of Sampling (PoSP) protocol effectively balances the need to incentivize good actors and deter bad ones, ensuring the overall security and reliability of decentralized systems. By combining economic rewards with stringent penalties, PoSP fosters an environment where honest behavior is not only encouraged but necessary for success. As decentralized AI continues to grow, protocols like PoSP will be essential in maintaining the integrity and trustworthiness of these advanced systems.

• source: https://cuckoo.network/blog/2024/06/17/proof-of-sampling-posp-protocol-decentralized-ai
• telegram: https://cuckoo.network/tg
• discord: https://cuckoo.network/dc

Arbitrum Nitro, developed by Offchain Labs, is a second-generation Layer 2 blockchain protocol designed to improve throughput, finality, and dispute resolution. It builds on the original Arbitrum protocol, bringing significant enhancements that cater to modern blockchain needs.

Key Properties of Arbitrum Nitro​

Arbitrum Nitro operates as a Layer 2 solution on top of Ethereum, supporting the execution of smart contracts using Ethereum Virtual Machine (EVM) code. This ensures compatibility with existing Ethereum applications and tools. The protocol guarantees both safety and progress, assuming the underlying Ethereum chain remains safe and live, and at least one participant in the Nitro protocol behaves honestly.

Design Approach​

Nitro's architecture is built on four core principles:

• Sequencing Followed by Deterministic Execution: Transactions are first sequenced, then processed deterministically. This two-phase approach ensures a consistent and reliable execution environment.
• Geth at the Core: Nitro utilizes the go-ethereum (geth) package for core execution and state maintenance, ensuring high compatibility with Ethereum.
• Separate Execution from Proving: The state transition function is compiled for both native execution and web assembly (wasm) to facilitate efficient execution and structured, machine-independent proving.
• Optimistic Rollup with Interactive Fraud Proofs: Building on Arbitrum’s original design, Nitro employs an improved optimistic rollup protocol with a sophisticated fraud proof mechanism.

Sequencing and Execution​

The processing of transactions in Nitro involves two key components: the Sequencer and the State Transition Function (STF).

• The Sequencer: Orders incoming transactions and commits to this order. It ensures that the transaction sequence is known and reliable, posting it both as a real-time feed and as compressed data batches on the Ethereum Layer 1 chain. This dual approach enhances reliability and prevents censorship.
• Deterministic Execution: The STF processes the sequenced transactions, updating the chain state and producing new blocks. This process is deterministic, meaning the outcome depends only on the transaction data and the previous state, ensuring consistency across the network.

Software Architecture: Geth at the Core​

Nitro’s software architecture is structured in three layers:

• Base Layer (Geth Core): This layer handles the execution of EVM contracts and maintains the Ethereum state data structures.
• Middle Layer (ArbOS): Custom software that provides Layer 2 functionality, including decompressing sequencer batches, managing gas costs, and supporting cross-chain functionalities.
• Top Layer: Drawn from geth, this layer handles connections, incoming RPC requests, and other top-level node functions.

Cross-Chain Interaction​

Arbitrum Nitro supports secure cross-chain interactions through mechanisms like the Outbox, Inbox, and Retryable Tickets.

• The Outbox: Enables contract calls from Layer 2 to Layer 1, ensuring that messages are securely transferred and executed on Ethereum.
• The Inbox: Manages transactions sent to Nitro from Ethereum, ensuring they are included in the correct order.
• Retryable Tickets: Allows for the resubmission of failed transactions, ensuring reliability and reducing the risk of lost transactions.

Gas and Fees​

Nitro employs a sophisticated gas metering and pricing mechanism to manage transaction costs:

• L2 Gas Metering and Pricing: Tracks gas usage and adjusts the base fee algorithmically to balance demand and capacity.
• L1 Data Metering and Pricing: Ensures costs associated with Layer 1 interactions are covered, using an adaptive pricing algorithm to apportion these costs accurately among transactions.

Conclusion​

Cuckoo Network is confident in investing in Arbitrum's development. Arbitrum Nitro's advanced Layer 2 solutions offer unmatched scalability, faster finality, and efficient dispute resolution. Its compatibility with Ethereum ensures a secure, efficient environment for our decentralized applications, aligning with our commitment to innovation and performance.

• source: https://cuckoo.network/blog/2024/06/06/introduction-to-arbitrum-nitro
• telegram: https://cuckoo.network/tg
• discord: https://cuckoo.network/dc

The combination of blockchain and AI is gaining significant market attention. With ChatGPT amassing hundreds of millions of users quickly and Nvidia’s stock soaring eightfold in 2023, AI has firmly established itself as a dominant trend. This influence is spilling into adjacent sectors like blockchain, where AI applications are being explored.

Currently, crypto plays a complementary role in AI, offering significant potential for growth. Most organizations are still in the exploratory phase, focusing on the tokenization of computing power (cloud and marketplace), models (AI agents), and data storage.

Decentralized crypto technology doesn’t directly boost efficiency or reduce costs in AI training but facilitates asset trading, attracting previously unused computing power. This is profitable in today’s compute-scarce environment. Tokenizing models enables decentralized community ownership or usage, reducing barriers and offering an alternative to centralized AI. However, decentralized data remains challenging to tokenize, requiring further exploration.

While the market hasn’t reached a consensus on AI and crypto, the ecosystem is shaping up. Here are a few categories we will review today: Infrastructure-as-a-Service Cloud, computing marketplaces, model tokenization and training, AI agents, data tokenization, ZKML, and AI applications.

Infrastructure-as-a-Service Cloud​

As the AI market grows, GPU cloud computing projects and marketplaces are among the first to benefit. They aim to incorporate unused GPU resources into centralized networks, lowering computing costs compared to traditional services.

These cloud services aren't considered decentralized solutions but are integral parts of the web3 + AI ecosystem. The idea is that GPUs are scarce resources and hold intrinsic value.

Key Projects:

• Akash Network: Decentralized cloud computing marketplace based on Cosmos SDK, using Kubernetes for orchestration and reverse auction pricing for cost reduction. Focuses on CPU and GPU computing.
• Ritual: AI infrastructure network integrating AI models into blockchain protocols. Its Infernet platform enables smart contracts to access models directly.
• Render Network: Decentralized GPU rendering platform focusing on both rendering and AI computing. Moved to Solana for better performance and cost.
• NetMind.AI: AI ecosystem providing a marketplace for computing resources, chatbot, and life assistant services. Supports a wide range of GPU models and integrates Google Colab.
• CUDOS: Blockchain compute network similar to Akash, focusing on GPU computing via Cosmos SDK.
• Nuco.cloud: Decentralized computing cloud service based on Ethereum and Telos, offering a range of solutions.
• Dynex: Blockchain for neuromorphic computing, using Proof-of-Useful-Work for efficiency.
• OctaSpace: Decentralized computing cloud, operating on its own blockchain for AI and image processing.
• AIOZ Network: Layer 1 decentralized compute platform for AI, storage, and streaming.
• Phoenix: Web3 blockchain infrastructure for AI computing and data-driven networks.
• Aethir: Cloud infrastructure for gaming and AI, based on Arbitrum.
• Iagon: Decentralized storage and computing marketplace on Cardano.
• OpFlow: Cloud platform focusing on AI and rendering, using NVIDIA GPUs.
• OpSec: Emerging decentralized cloud platform aiming to build the next-generation supercomputer.

Computing-resource marketplaces​

Decentralized computing-resource marketplaces utilize user-provided GPU and CPU resources for AI tasks, training, and inference. These marketplaces mobilize unused computing power, rewarding participants while reducing barriers to entry.

These GPU computing marketplaces often focus on the narrative of decentralization rather than service utility. Projects like io.net and Nosana, leveraging Solana and DePin concepts, show a huge growth potential. Investing early in GPU markets during peak demand phases can offer high returns through incentives and ROI.

Key Projects:

• Cuckoo AI: A decentralized marketplace that rewards GPU miners serving AI models with daily ERC20 payments. It uses blockchain smart contracts and focuses on transparency, privacy, and modularity.
• Clore.ai: A GPU rental platform using PoW. Users can rent GPUs for AI training, rendering, and mining tasks. Rewards are tied to the amount of their token held.
• Nosana: An open-source, Solana-based GPU cloud computing provider. Focuses on AI inference and is developing connectors for PyTorch, HuggingFace, TensorFlow, and community libraries.
• io.net: An AI cloud computing network leveraging Solana blockchain technology. Offers cost-efficient GPU resources, supporting batch inference and parallel training.
• Gensyn: An L1 protocol for deep learning model training. Aims to improve training efficiency through a trustless, distributed system. Focused on reducing training costs and increasing accessibility.
• Nimble: A decentralized AI ecosystem that combines data, computing power, and developers. Aims to make AI training more accessible and has a decentralized, composable framework.
• Morpheus AI: A decentralized computing marketplace built on Arbitrum. Helps users create AI agents to interact with smart contracts.
• Kuzco: A distributed GPU cluster for LLM inference on Solana. Offers efficient local model hosting and rewards contributors with KZO points.
• Golem: An Ethereum-based CPU computing marketplace that’s expanded to GPUs. One of the earliest peer-to-peer computing networks.
• Node AI: A GPU cloud marketplace offering affordable GPU rentals through EyePerformance.
• GPU.Net: A decentralized GPU network providing infrastructure for generative AI, Web3, and high-end graphics rendering.
• GamerHash: A platform utilizing gamers' spare computing power for crypto mining while introducing a play-to-earn model for low-end devices.
• NodeSynapse: A GPU marketplace offering Web3 infrastructure, GPU computing, server hosting, and a unique revenue-sharing model for token holders.

Model tokenization and training​

Model tokenization and training involve converting AI models into valuable assets and integrating them into blockchain networks. This approach allows for decentralized ownership, data sharing, and decision-making. It promises improved transparency, security, and monetization opportunities while creating new investment channels.

The crucial factor is recognizing projects with real innovation and technical challenges. Simply trading ownership or usage rights of AI models isn't true innovation. The real advances come from verifying model outputs effectively and ensuring secure, decentralized model operation.

Key Projects:

• SaharaLabs: Focuses on privacy and data sharing with tools like Knowledge Agent and Data Marketplace, helping secure data operations and draw clients such as MIT and Microsoft.
• Bittensor: Builds a decentralized protocol for AI models to exchange value. It uses validators and miners to rank responses and enhance the overall quality of AI-driven applications.
• iExec RLC: A decentralized cloud computing platform that ensures resource security via Proof-of-Contribution consensus while managing short-term computational tasks.
• Allora: Rewards AI agents for accurate market predictions, utilizing consensus mechanisms to validate agent forecasts in a decentralized network.
• lPAAL AI: Provides a platform for creating personalized AI models that can handle market intelligence, trading strategies, and other professional tasks.
• MyShell: Offers a flexible AI platform for chatbot development and third-party model integration, incentivizing developers through native tokens.
• Qubic: Leverages proof-of-work consensus for AI training, with the Aigarth software layer facilitating neural network creation.

AI Agent​

AI agents, or intelligent agents, are entities capable of autonomous understanding, memory, decision-making, tool use, and performing complex tasks. These agents not only guide users on "how to" perform tasks but actively assist in completing them. Specifically, this refers to AI agents that interact with blockchain technology for activities like trading, offering investment advice, operating bots, enhancing decentralized finance (DeFi) functionalities, and performing on-chain data analysis.

Such AI agents are closely integrated with blockchain technology, enabling them to generate revenue directly, introduce new trading scenarios, and enhance the blockchain user experience. This integration represents an advanced narrative in DeFi, creating profit through trading activities, attracting capital investment, and fostering hype, which in turn drives a Ponzi scheme-like cycle of investment.

Key Projects:

• Morpheus: A decentralized AI computing marketplace built on Arbitrum that enables the creation of agent AIs that operate smart contracts. The project is led by David Johnston, who has a background in investment and executive leadership. Morpheus focuses on community engagement with a fair launch, security audited staking code, and active development, although updates on AI agent codes are slow and core module progress is unclear.
• QnA3.AI: Provides comprehensive services for information management, asset management, and rights management throughout their lifecycle. Using Retrieval-Augmented Generation (RAG) technology, QnA3 enhances information retrieval and generation. The project has quickly grown since its inception in 2023, with significant increases in user engagement and application.
• Autonolas: An open market for creating and using decentralized AI agents, offering tools for developers to build AI agents capable of connecting to multiple blockchains. Led by David Minarsch, a Cambridge-educated economist specializing in multi-agent services.
• SingularityNET: An open, decentralized AI service network that aims to democratize and decentralize general-purpose AI. The network allows developers to monetize their AI services using the native AGIX token and was founded by Dr. Ben Goertzel and Dr. David Hanson, known for their work on the humanoid robot Sophia.
• Fetch.AI: One of the earliest AI agent protocols, which has developed an ecosystem for deploying agents using the FET token. The team is composed of experts from prestigious universities and top companies, focusing on AI and algorithmic solutions.
• Humans.ai: An AI blockchain platform that brings together stakeholders involved in AI-driven creation within a creative studio suite, allowing individuals to create and own their digital likenesses for the creation of digital assets.
• Metatrust: A Crypto-enabled AI agent network that offers a comprehensive Web3 security solution covering the entire software development lifecycle, founded by a globally recognized research team from Nanyang Technological University.
• AgentLayer: Developed by the Metatrust team, this decentralized agent network utilizes OP Stack and EigenDA to enhance data efficiency and overall performance and security.
• DAIN: Building an agent-to-agent economy on Solana, focusing on enabling seamless interactions between agents from different enterprises through a universal API, emphasizing integration with both web2 and web3 products.
• ChainGPT: An AI model designed for blockchain and crypto, featuring products like an AI NFT generator, AI-powered news generator, trading assistant, smart contract generator, and auditor. ChainGPT won the BNB Ecosystem Catalyst Award in September 2023.

Data tokenization​

The intersection of AI and cryptography in the data sector holds significant potential, as data and computing power are foundational resources for AI. While decentralized computing can sometimes reduce efficiency, decentralizing data is reasonable because data production is inherently decentralized. Thus, the combination of AI and blockchain in the data sector offers substantial growth potential.

A major challenge in this field is the lack of a mature data marketplace, making effective valuation and standardization of data difficult. Without a reliable valuation mechanism, projects struggle to attract capital through token incentives, potentially breaking the "flywheel effect" even for high-potential projects.

Key Projects:

• Synesis One: A crowdsourcing platform on Solana where users earn tokens by completing micro-tasks for AI training. Collaborating with Mind AI, the platform supports various data types and robot process automation. Mind AI has agreements with GM and the Indian government, raising $9.5 million in funding.
• Grass.io: A decentralized bandwidth marketplace enabling users to sell surplus bandwidth to AI companies. With over 2 million IP addresses, only active users earn rewards. They have raised $3.5 million in seed funding led by Polychain and Tribe Capital.
• GagaNode: A next-generation decentralized bandwidth market addressing IPv4 shortages using Web 3.0 technology. It is compatible with multiple platforms, including mobile, and the code is open-source.
• Ocean Protocol: Allows users to tokenize and trade their data on Ocean Market, creating data NFTs and data tokens. Founder Bruce Pon, formerly with Mercedes-Benz, is a lecturer on blockchain and decentralization technologies, supported by a global advisory team.

AI Application​

Combining AI capabilities with current crypto businesses opens up new avenues for enhancing efficiency and functionality across various sectors like DeFi, gaming, NFTs, education, and system management.

DeFi​

• inSure DeFi: A decentralized insurance protocol where users buy SURE tokens to insure their crypto assets. It leverages Chainlink for dynamic pricing.
• Hera Finance: An AI-driven cross-chain DEX aggregator integrated with multiple chains, such as Ethereum, BNB, and Arbitrum.
• SingularityDAO: An AI-powered DeFi protocol offering investment management. Partnering with SingularityNET, they secured $25M to scale their AI tools and ecosystem.
• Arc: Offers an AI-driven DeFi ecosystem via Reactor and DApp. Acquired Lychee AI, joined Google Cloud's AI Startup Program, and released AI-powered ARC Swaps.
• AQTIS: An AI-supported liquidity protocol aiming to establish a sustainable ecosystem with $AQTIS akin to gas on Ethereum.
• Jarvis Network: Utilizes AI algorithms to optimize trading strategies for crypto and other assets. Their native token JRT circulates actively.
• LeverFi: A leveraged trading protocol developing an AI DeFi solution with Microsoft. Secured $2M from DWF Labs for AI investment management.
• Mozaic: An auto-farming protocol blending AI concepts and LayerZero technology.

Gaming​

• Sleepless AI: Uses AI blockchain for a virtual companion game. The first game, HIM, features unique SBT characters on-chain. Binance launched Sleepless AI's token via Launchpool.
• Phantasma: A gaming-focused layer 1 blockchain offering smartNFTs and AI smart contract encoders.
• Delysium: An AI-driven Web3 gaming publisher with an open-world framework. They offer Lucy, an AI Web3 OS, and enable AI-Twins creation for interactive gaming.
• Mars4.me: An interactive 3D metaverse project supported by NASA data. Secured long-term funding from DWF Labs.
• GamerHash: Leverages excess computing power during high-end gaming for crypto mining. Their Play&Earn feature provides tasks for lower-spec computers.
• Gaimin: Founded by an eSports team, it combines cloud computing with gaming to build a dApp that provides extra GPU rewards.
• Cerebrum Tech: A generative AI, gaming, and Web3 solution provider. Recently raised $1.8M for gaming and AI expansion.
• Ultiverse: A metaverse gaming platform that raised funds from Binance Labs to launch its AI-driven open metaverse protocol, Bodhi.

NFT​

• NFPrompt: An AI-driven platform where users can generate NFT art. Binance's Launchpool supports NFP staking for rewards.
• Vertex Labs: A Web3 and AI infrastructure provider with blockchain, AI computation, and Web3 brands like Hape Prime.

Education​

• Hooked Protocol: Offers social educational games and has Hooked Academy, an AI-driven educational tool backed by ChatGPT.

System​

• Terminus OS: A Web3 OS based on blockchain-edge-client architecture, designed for AI-era proof of intelligence. ByteTrade secured $50M in funding for its development.

Conclusion​

The fusion of AI and cryptocurrency is opening up groundbreaking opportunities across various sectors, from cloud computing and AI applications to data tokenization and zero-knowledge machine learning. This combination of technologies is already showcasing its transformative power. With more innovative projects on the horizon, the future of AI and crypto will be diverse, intelligent, and secure.

We're eager to see new collaborations and cutting-edge technologies that will expand the reach of AI and blockchain into even broader applications.