OpenLedger 백서

OpenLedger

원문

bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 2쪽 중 1쪽

OpenLedger - AI Blockchain OpenLedger July 1, 2025 Abstract OpenLedger is the AI Blockchain, unlocking liquidity to monetize data, models and agents. The Blockchain records every contribution to the AI lifecycle on-chain, ensuring transparency, traceability, and ac- countability. Proof of Attribution ensures that ownership, credit, and economic rewards are accurately assigned to data providers, model de- velopers, and other participants. OpenLedger creates a collaborative, traceable, and auditable environment for building AI that is open, fair, and owned by all. This paper outlines the design of the AI Blockchain, including its attribution mechanism, incentive structure, and the role of community collaboration in powering a decentralized AI economy. 1 Introduction Artificial intelligence is rapidly evolving, but the infrastructure behind it has not. While large general-purpose models have driven early progress, the future of AI depends on models that are specialized, explainable, and aligned with real-world needs. To enable specialized models, we need specialized data. But without proper attribution, transparency, and provenance, sourcing and sustaining that data becomes difficult. Contributions are hard to trace, incentives are misaligned, and the value created by data providers and model developers often goes unrecognized. There is no reliable way to prove who contributed what or to ensure that participants are rewarded in proportion to their im- pact. 1 OpenLedger introduces the AI Blockchain to solve these problems. It brings data contribution, model refinement, and collaboration on-chain, en- abling transparent attribution and economic alignment throughout the AI lifecycle. With Proof of Attribution embedded into the blockchain, Open- Ledger makes it possible to build AI in a way that is open, verifiable, and collectively owned. 1.1 Why AI Needs a Blockchain AI is built on contributions from a wide range of participants, including data providers, model developers, researchers, and application builders. Today’s AI infrastructure fails to recognize or reward these roles. Contributions are difficult to trace, attribution is often missing, and centralized platforms con- trol access, credit, and monetization. A blockchain designed specifically for AI changes this by introducing three essential capabilities. 1.1.1 Collaboration and Ownership Current AI models are primarily trained on large volumes of internet data. While this works for building general-purpose language models, it lacks the depth and domain specificity required for real-world applications. Specialized AI requires access to high-quality, curated datasets. However, today’s AI stack does not provide a standard way to collect, attribute, or collaborate on this type of data. OpenLedger addresses this gap by enabling open collaboration. Anyone can contribute data, models, or insights, and every contribution is perma- nently linked to its origin. This ensures that ownership is preserved and that contributors receive proper credit for their work. 1.1.2 Monetization and Rewarding Contributions Many contributors in the AI pipeline, including those who provide data, evaluate models, or improve performance, go unrewarded despite the value they create. OpenLedger’s Proof of Attribution ensures that contributors receive economic rewards based on the impact of their work. Anyone can participate in AI development and earn from their contributions without relying on centralized platforms. 2 1.1.3 Transparency and Traceability Most AI systems operate as black boxes. Their internal decision logic, train- ing data, and development history are hidden or unclear. This lack of ex- plainability undermines trust and accountability. OpenLedger records every contribution on the blockchain with verifiable metadata. Data sources, model changes, and downstream interactions are fully traceable and auditable. This enables explainable AI across the full lifecycle. While the need for a blockchain in AI development is clear, not all blockchains are designed to meet this need. Existing general-purpose chains lack the infrastructure to support attribution, coordination, and incentive alignment specific to AI. The next section outlines why these blockchains fall short and how OpenLedger fills the gap. 1.2 Why General-Purpose Blockchains Aren’t Enough Most existing blockchains were designed for financial transactions, asset transfers, or digital collectibles. While these systems have proven valuable for decentralized finance and ownership, they fall short when applied to the needs of artificial intelligence. AI development is not just about transac- tions or token movement. It requires collaboration, provenance, and ongoing contributions across data, models, and evaluation processes. General-purpose blockchains lack native support for attribution, version control of models, structured data flows, and fine-grained reward systems. They cannot represent the full lifecycle of an AI system, from raw data to deployed model. Attempts to retrofit these blockchains for AI often result in workarounds that compromise scalability, traceability, and contributor in- centives. OpenLedger is built specifically to address these gaps. It is not a blockchain for general applications that happens to support AI. It is the AI Blockchain, designed from the ground up to handle attribution, model tracking, contri- bution history, and collaborative ownership. OpenLedger does not adapt blockchain to AI. It redefines what a blockchain can be by aligning it directly with how AI is created, shared, and sustained. With a clear understanding of what general-purpose blockchains cannot offer, it becomes important to revisit the current direction of the AI indus- try itself. The next section examines the limitations of general-purpose AI models and the growing importance of specialized, fine-tuned intelligence. 3 Feature General Blockchain OpenLedger (AI Blockchain) Purpose Built for DeFi, NFTs, etc. Built only for AI and model workflows Data Attribution No native tracking Every contribution is tracked on-chain Rewards For validators and min- ers For data and model con- tributors Security Focus on transactions Protects AI contribu- tions and usage Provenance Limited Full history of models and datasets Governance Votes on protocol up- grades Votes on model quality and improvement rules Table 1: Comparison of General Blockchains and OpenLedger 1.3 The Shift from General Models to Specialized AI AI research is shifting from the pursuit of ever-larger, general-purpose models to the development of highly optimized, domain-specific intelligence . While foundational models are trained on broad internet data, they often lack applicability in specialized contexts. As a result, the industry now prioritizes adaptability, efficiency, and application-specific intelligence , which requires: • Fine-tuning models for specialized applications in sectors like fi- nance, healthcare, legal, and cybersecurity. • Reducing computational costs by leveraging smaller, optimized models rather than running expensive, general-purpose LLMs. • Enhancing explainability through specialized models that pro- vide interpretable, domain-specific justifications . The idea is not to replace foundational models, but to coexist and utilize the existing foundational models to make them even 4 more intelligent. Instead of competing with large-scale AI models, Open- Ledger enables fine-tuned, specialized AI models to work in tandem with foundational AI, unlocking greater efficiency, accuracy, and real-world applicability. To support this transition, OpenLedger provides aframework for model attribution, decentralized fine-tuning, and governance, ensuring that AI builders and contributors receive fair recognition and financial incentives for improving models. The shift toward specialized AI models signals not just a technical change but a broader economic one. As AI systems become more autonomous and capable, they are redefining how value is created and exchanged in digital environments. The following section explores this economic transition and its implications. 1.4 Economic Shift from the Internet to AI: The Need for AI-Native Platforms AI is not just a technological shift, it is an economic transformation . Traditional internet-based revenue models, such as advertising, SEO, and centralized data monetization, are being disrupted by AI-driven au- tomation. This shift is causing fundamental changes in how digital economies function: • Search engines and SEO-based businesses are losing value as AI-driven assistants replace traditional search interactions . • Content creation is increasingly AI-dominated , reducing tradi- tional monetization opportunities for human creators. • The legacy internet economy (advertising, centralized data ownership) is collapsing , necessitating a new system for AI- driven economic transactions . OpenLedger introduces AI-native economic infrastructure, ensuring that AI models and agents operate within a sustainable, decentralized economy where contributors, developers, and liquidity providers are directly incentivized through tokenized AI models . A robust economic foundation requires clear roles and responsibilities. OpenLedger defines a set of key stakeholders who contribute to and benefit 5 from the AI Blockchain. The next section outlines these roles and how they interact within the ecosystem. 1.5 Key Stakeholders in the OpenLedger Blockchain The OpenLedger blockchain is built around a collaborative model, where multiple participants contribute to AI model creation, validation, and adop- tion: • AI Model Developers – Build, train, and optimize AI models for deployment. • Data Contributors – Provide domain-specific data with verifiable attribution, ensuring transparent model improvements. • Validators – Secure the network, validate AI model performance, and prevent misuse or low-quality contributions. • Applications and AI Agents – Consume AI models for real-world automation, integrating them into decentralized ecosystems. • Protocol Governors – Stake OPEN tokens to earn voting power and guide the future of AI model development. They evaluate proposals, vote on their progression, and ensure that only high-quality models backed by the community advance through the lifecycle. 2 Architecture The OpenLedger architecture[fig 1] is structured to provide an efficient, ver- ifiable, and economically sustainable framework for decentralized specialized model development. It consists of two primary layers: the blockchain layer and the specialized model layer. Each of these layers plays a distinct role in ensuring that specialized models are secure, interpretable, and capable of interacting with external environments. 2.1 EVM Compatible Blockchain The foundational layer of OpenLedger is built on an EVM-compatible Blockchain. This ensures low-latency transactions while leveraging the ro- bust security and liquidity provided by ethereum. The blockchain layer plays 6 Figure 1: Openledger - AI Blockchain, unlocking liquidity to monetize data, models and agents a crucial role in maintaining an immutable record of specialized models, own- ership, incentives, data points, and proof of attribution. Transactions are op- timized using rollups to ensure scalability while maintaining cryptographic integrity. For a transaction set T , the state transition function σ is defined as: σ(s, T) → s′, (1) where s and s′ represent the system state before and after processing T , ensuring verifiable and deterministic execution of smart contract logic for model registration, staking, and governance. 2.2 Proof of Attribution Ensuring transparent and verifiable data attribution is fundamental to Openledger. By leveraging Proof of attribution techniques , we enable model creators and data contributors to receive fair compensation while maintaining an auditable record of how data influences specialized models. This section details the full pipeline of data attribution, mathematical for- mulations, and mechanisms to ensure trust, accountability, and incentives within the OpenLedger ecosystem. 2.2.1 Why Data Attribution is Necessary As AI models evolve, ensuring proper attribution of data contributions is essential for: • Tracking how individual training data points influence model behavior. 7 • Rewarding data contributors based on their contribution’s impact. • Reducing bias and misinformation by penalizing low-quality data. • Enabling verifiable proof-of-attribution for AI-generated outputs. In Openledger, proof of attribution ensures that each data source is cryptographically linked to model outputs, providing an immutable, decentralized ledger of contribution . 2.2.2 Mathematical Framework for Influence-Based Attribution To measure the influence of a given data point di on the model’s output y, we define the influence function: I(di, y) = ∂L(y, θ)∂θ · ∂θ∂di (2) where: • L(y, θ) is the loss function for generating output y. • θ represents the model parameters. • ∂L∂θ measures how model parameters affect the loss. • ∂θ∂di measures how training data influences model parameters. 2.2.3 Efficient Influence Computation Computing full Hessian-based attributions is computationally expensive. In- stead, Openledger adopts [1], an efficient approximation for real-time at- tribution: IDataInf(di, y) = LX l=1 1 λl 1 n nX j=1 Ll,ij λl + Ll,ii − Ll,i ! (3) where: • Ll,ij = ∇θlℓTi ∇θlℓj represents pairwise loss impact. • λl is a stabilization parameter. 8 This allows efficient batch computation of attribution scores while en- suring scalability [1]. Here’s the updated OpenLedger Data Attribution Pipeline section with the correct math, current influence model (inference-based), and fee-reward flow—all formatted to match your system and phrased cleanly: 2.2.4 OpenLedger Data Attribution Pipeline Step 1: Data Contribution • Data contributors submit structured, domain-specific datasets intended to power Datanets. • Each data submission is uniquely attributed, enabling traceable and verifiable contributions. Step 2: Influence Attribution During Inference When a model serves an inference, we compute how much each data point di contributed to the final output y. We define an inference-level influence score: I(di, y) = α · F (di, y) Where: • α is a constant weight, • F (di, y) measures the direct impact of data point di on the model’s output y. Only points with I(di, y) > 0 qualify for reward distribution. Step 3: Inference Fee Calculation and Contributor Rewards Each inference request incurs a fee to cover computation and platform costs: Feeinference =  Tin 1000 · Rin  +  Tout 1000 · Rout  + Fplatform Where: • Tin: number of input tokens 9 • Tout: number of output tokens • Rin: rate per 1000 input tokens (in OPN) • Rout: rate per 1000 output tokens (in OPN) • Fplatform: flat platform fee (in OPN) If the platform fee is percentage-based: Fplatform = η ·  Tin 1000 · Rin + Tout1000 · Rout  After retaining the platform fee: Fnet = Feeinference − Fplatform This is split into: Fnet = Fmodel + Fstakers + Fcontributors With: Fmodel = β · Fnet, F stakers = γ · Fnet, F contributors = δ · Fnet Each contributor ci is rewarded based on their proportional influence: wi = I(di, y)Pn j=1 I(dj, y) ⇒ Rewardci = wi · Fcontributors Example Say: • Tin = 800, Rin = 0.2 • Tout = 1200, Rout = 0.4 • Fplatform = 0.5 OPN 10 Feeinference =  800 1000 · 0.2  +  1200 1000 · 0.4  +0.5 = 0.16+0.48+0.5 = 1.14 OPN Fnet = 1.14 − 0.5 = 0.64 OPN If split as: Fmodel = 0.448, F stakers = 0.064, F contributors = 0.128 Then, for a contributor with wi = 0.25: Rewardci = 0.25 · 0.128 = 0.032 OPN This system enables provable, real-time attribution , ensuring con- tributors are rewarded proportionally to their data’s influence on model inference. It discourages low-quality inputs and promotes the curation of high-impact, verifiable datasets. 2.3 Openledger AI Studio The Openledger AI Studio is responsible for hosting, training, and evaluating explainable, domain-specific specialized models. Unlike generic foundation models, specialized models in OpenLedger are optimized for niche domains while ensuring complete transparency in their attribution and training pro- cesses. 2.3.1 Datanets Specialized models rely on high-quality, domain-specific data to function op- timally. The Datanets acts as an on-chain data aggregation and attribution mechanism, ensuring that every contributor is appropriately rewarded. Data points D submitted to the reservoir are assigned credibility scores based on staking weights wi: C(D) = nX i=1 wi · f (xi, yi), (4) where xi and yi represent input-label pairs, and f (xi, yi) is a function deter- mining data quality and reliability. 11 2.3.2 Supervised Fine-Tuning Once a sufficient dataset has been aggregated, specialized models undergo supervised fine-tuning . The model parameters θ are optimized via Em- pirical Risk Minimization (ERM) : θ∗ = arg min θ 1 n nX i=1 L(yi, fθ(xi)), (5) where L represents the loss function, and fθ denotes the specialized model function. Once a sufficient dataset has been aggregated, specialized models un- dergo supervised fine-tuning . The model parameters θ are optimized via Empirical Risk Minimization (ERM) : θ∗ = arg min θ 1 n nX i=1 L(yi, fθ(xi)), (6) where L represents the loss function, and fθ denotes the specialized model function. To improve efficiency, Openledger employs fine-tuning, where the spe- cialized data collected is turned into intelligence, which can inturn power up the more purposeful and powerful AI Agents. 2.3.3 Reinforcement Learning with Human Feedback(RLHF) To enhance model reasoning and adaptability, Openledger integrates rein- forcement learning (RL) with human feedback. The reward function R(θ) for model updates is defined as: R(θ) = nX i=1 wi · (V (yi, fθ(xi)) − α · L(yi, fθ(xi))), (7) where V (yi, fθ(xi)) represents validator-assigned scores based on correctness and interpretability, and α is a regularization parameter to prevent overfit- ting. Users who provide high-quality feedback are rewarded with stake in- centives, while those who attempt to manipulate the system face stake slash- ing. 12 2.3.4 Model Evaluation and Deployment: OpenLoRA OpenLoRA is a multi-tenant LoRA model serving framework that enables scalable, low-latency inference for specialized models in Openledger. It optimizes GPU utilization and inference latency through: 2.3.5 Multi-Tenant GPU Infrastructure for LoRA Model Serving OpenLoRA allows multiple LoRA models to share a single pre-trained back- bone model. This reduces GPU memory overhead and optimizes computa- tional efficiency. 2.3.6 Segmented Gather Matrix-Vector Multiplication (SGMV) for LoRA Model Execution SGMV enables batched execution of LoRA adapters while maintaining effi- cient memory access patterns. Given input tensor x, LoRA transformations are computed as: y = xW + xAB, (8) where W is the pre-trained model weight, andA, Bare LoRA-specific adapters. 2.3.7 On-Demand Model Loading and Fast Switching Between LoRA Models OpenLoRA dynamically loads LoRA adapters, reducing cold-start times by keeping backbone models in GPU memory and swapping in only the adapter weights. 2.3.8 GPU Resource Scheduling and Load Balancing Requests are dynamically assigned to GPUs based on batch size constraints and memory availability. A new request Rnew is assigned to GPU k∗ where: k∗ = arg max k {Rk | Rk < Bmax, Mk > Mrequired} (9) This ensures optimal throughput and load balancing. 13 2.3.9 Request Migration and Memory Optimization for KvCache Management KvCache storage is dynamically managed to prevent memory overflow, with request migration occurring when a GPU reaches capacity. Requests are reassigned while preserving previous inference states to maintain seamless execution. OpenLedger will host models and integrate with leading agent frame- works, ensuring seamless availability while establishing an infrastructure to develop more
프로젝트로 돌아가기