Sentient 백서

Sentient

원문

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

and adoption of open-source AI artifacts across developers, enterprise teams and end-user products. Because the protocol is open and composable, artifacts can be registered, composed and reused across workflows, turning distribution into a cooperative network effect. 3.3.1 Distribution goals Ensure useful artifacts reach the right audiences. Usage drives both revenue and token emissions, incentivizing builders to design broadly applicable components, including for other artifacts. 3.3.2 Leveraging the network of artifacts Artifacts are both services and consumers; they can be invoked by or call other components. Pipelines (e.g., research agent + search + summarizer + translation + evaluator) propagate usage and revenue upstream. Sentient surfaces these compositions via Chat, dashboards and partner APIs, compounding network effects. 3.3.3 Segmented distribution paths Developer-centric: SDKs, templates, CLIs, package manager publishing with license metadata, portals and forums. Composability means upstream components accrue usage even when end users never call them directly. Enterprise: Private deployments via containers or TEEs, enterprise-grade licenses with SLAs, on-prem data handling and audit trails; subscription and usage-based terms that route revenue to creators. Product: API endpoints and UI components for consumer apps (e.g., summarization, recommen- dations). Multi-artifact experiences cascade usage credit and revenue up the chain. 3.3.4 Matching and recommendation A matching engine (inspired by Kaito) connects artifacts to users based on semantic tags (domain, modality, compliance, license) and signals (usage, ratings, stake). High-quality artifacts rise as adoption grows. 3.3.5 Community ownership and distribution incentives Broader stakeholder ownership amplifies distribution. Temporary incentives (e.g., emission multi- pliers for milestones) accelerate early adoption and decay over time. Adoption by other artifacts 8 counts toward usage, encouraging reusable components and joint ventures. 4 Verifiable Trust Primitives Sentient’s protocol is supported by technology primitives that ensure proper accounting, security and verifiability. These primitives are developed in the open and shared with the ecosystem. 4.1 Fingerprinting Fingerprinting is an AI-native cryptographic primitive that helps prove ownership and detect unauthorized copies of models. Unlike digital signatures on static files, fingerprinting works for generative models whose behavior is defined by billions of parameters. Sentient’s fingerprinting library embeds secret query–response pairs into the model during fine-tuning or adapter training. A fingerprint consists of a sequence of triggers and corresponding responses. Triggers are inputs drawn from a distribution unknown to the public; responses are outputs with certain lexical or statistical properties. 4.1.1 Embedding fingerprints Techniques include: • Specialized fine-tuning– train on a small dataset containing secret trigger pairs, preserving normal performance. • Parameter expansion– use low-rank adapters or prompt-tuning vectors to encode fingerprints without altering base weights. •Benign data mixing– mix triggers with benign data to prevent overfitting and detection. • Model mixing– weighted averaging to bundle multiple fingerprint sets while preserving fidelity. Fingerprint parameters (key length, number of triggers, redundancy) tune robustness vs. detection time. 4.1.2 Detection and enforcement 1.Challenge generation– owner/auditor generates secret triggers. 2. Model interrogation– send triggers; fingerprinted models return matching responses with high probability. 3.Evidence recording– record detections and verify on-chain against commitments. 4. Remediation– redirect revenue, slash offending hosts, suspend licenses; repeat offenders lose hosting rights. Further resources: • Fingerprintinglibrary(OML1.0): https://github.com/sentient-agi/OML-1.0-Fingerprinting •Research paper (AlphaXiv):https://www.alphaxiv.org/abs/2502.07760 9 4.2 Trusted Execution Environments (TEEs) — General Overview TEEs provide hardware-enforced isolation so code and data remain confidential even on untrusted hosts. Enclaves attest the exact code binary and AI model hash used, ensuring proper execution. 4.2.1 Secure Execution Code, weights and inputs load into an enclave (isolated memory). The host OS/hypervisor cannot read or modify it. After execution, the enclave emits an attestation document; critical actions (payments, key handling, governance) can be gated on valid attestation. TEEs vs. other verifiable compute • ZK proofs: full cryptographic verification; currently high latency/cost for large models (improv- ing rapidly). •MPC: splits secrets across nodes; reduces single-node trust with protocol/latency trade-offs. • TEEs: near-native speed, general-purpose; trust vendor root-of-trust; be mindful of side-channels. NVIDIA supports GPU TEEs on several datacenter cards (e.g., Phala). Sentient invests in TEEs for hardware-independent guarantees and negligible overhead, while encouraging exploration of TEE/MPC/ZK options across partners. 4.2.2 Data Privacy and Compliance Examples Consumer PII: decrypt only inside the enclave, process, then re-encrypt; attestations prove which audited binary/model handled the PII. Enterprise/regulatory: protect data “in use” for healthcare/finance/legal; enforce TEE-only access with attestable proofs. Verifiable AI evaluation: audit pipelines by attesting code, model and dataset hashes; enable trustless scoring and benchmark execution. 4.2.3 Sentient Enclaves Framework (SEF) SEF is an end-to-end stack for confidential AI on AWS Nitro Enclaves (CPU TEEs). Nitro protects user data from the cloud host; Enclaves protect execution from the machine host. SEF adds: •Reproducible buildsproducing EIFs with deterministic hashes •Remote Attestation Web Serverinside the enclave •File-system monitorfor unconditional attestation of data ingress/egress •Transparent VSock proxieswith optional kill-switch semantics SEF emphasizes security, ergonomics and reproducibility for verifiable computation. SEF repo:https://github.com/sentient-agi/Sentient-Enclaves-Framework 10 4.2.4 Final Considerations TEE vendor trust (silicon/firmware) is a shared assumption across AWS/Intel/AMD. SEF is open- source and portable to AMD/Intel using EIF on QEMU. This flexibility supports broad protocol applications, from agent hosting to evaluations. 5 Tokenomics Sentient’s network channels economic rights to those who create, maintain and curate its intelligence. Emissions are distributed among constituencies for alignment over time. Significant portions are reserved for builders (developers, researchers, data curators, evaluators, maintainers), early backers (with extended vesting), and the Sentient Foundation (non-profit steward for grants, safety, community and protocol development). An initial airdrop seeds a broad base of stakeholders. Token Utility • Payment for AI artifacts: users pay for models, agents, datasets and evaluation; contracts route revenue to creators, hosts, evaluators; splits are artifact-governed. • Staking and delegation: holders stake on artifacts or delegate to representatives (Reps) who vote on emissions/policy; delegators share Rep rewards. • Governance: token-weighted voting for emissions, gauges, funding and upgrades; proposals via holders or delegates. • Liquidityandcomposability: pairwithartifact-specifictokens/derivatives; bondingcurves/liquidity pools to crowdfund artifacts. Token Emission Mechanism Figure 3: Token Emission Mechanism. 11 Sentient mints continuously to fund growth and reward participation. DAO controls the annual emission rate (fraction of current supply per epoch) and allocation policy across pools. 1. User incentives– emissions to users of Sentient interfaces (chat, APIs, dashboards, partner apps) to bootstrap network effects. 2. Representative and staker rewards– base stipends to Reps and stake-weighted distribution to stakers. Stakers receive the majority of their Rep’s yield. Define: StakeWeight (SW) = Token_staked× ( 1 + 3× LockDuration (years)MaxLockDuration (4 years) ). 3. Artifact emissions– the largest allocation for AI artifacts and contributors. A gauge blends three signals: V∈(0,1)(stake-weighted normalized Rep votes) SWS∈(0,1)(artifact’s stake weight share) RS∈(0,1)(artifact revenue share in rolling period) W=λV+x·SWS + (1−x−λ)RS, withλ+x≤1, λ,x∈[0,1]. Governance tunesλandxto emphasize metrics over time. This mechanism creates a loop: useful artifacts attract users; engagement generates revenue and signals; emissions reward contributors and stakers; governance rebalances as the network evolves. Governance Holders govern the protocol via direct democracy and representative decision-making. The DAO controls economic levers (emission rate, gauge parameters, treasury). Reps are DAO-elected validators or trusted community members who lock stake to assign relative emission weights. Any holder can initiate proposals and vote or delegate. Delegators can migrate stake among Reps; the DAO can appoint or deprecate Reps. Checks and balances aim for robustness and adaptability. Summary The Sentient tokenomics design remains true to foundational goals—monetize open-source AI, coordinatecontributorsanddistributevalue—whileincorporatingnewprotocolfeatures. Distribution ensures broad participation and long-term alignment. Utility turns the token into both fuel and glue. A dynamic emission mechanism funds growth, rewards stakeholders and adapts via gauge parameters and bootstrapping budgets. Layered governance empowers the community to steer the network’s evolution with transparency and accountability, creating a resilient, extensible economic framework for a community-owned intelligence network. 12
프로젝트로 돌아가기