Previously, we outlined Near’s vision to become the hub for User-Owned AI and how having projects that provide training data is a key piece to their AI ecosystem. In this piece, we dive into another important component of Near’s growing AI ecosystem: foundational models and an AI payment infrastructure.
Near chose Pond – a foundational AI model, and Nevermined – a payment infrastructure for AI commerce, to be included in its recent AI cohort. We’ll get into what these projects do and how they are important components of Near’s AI ecosystem.
Pond
Pond is an advanced AI model designed specifically for Web3, leveraging Graph Neural Networks (GNNs) to analyze complex, graph-structured blockchain data. By integrating concepts from GPT, Pond enhances its capabilities, offering accurate insights and predictions while addressing Web3’s unique challenges. Potential applications span behavioral analysis, DeFi security, personalized recommendations, anti-money laundering, market predictions, and more. Pond aims to support a fully automated AGI world, providing an industrial-scale model that acts as the on-chain brain for advanced AI-driven services and applications within the Web3 ecosystem.
Product Features
To address the challenges posed by Web3 data, Pond is utilizing Graph Neural Networks (GNNs) as the foundational model. GNNs are designed to operate on graph-structured data, making them well-suited to the inherent characteristics of blockchain networks.
What are Graph Neural Networks (GNNs)?
GNNs are a class of neural networks designed to process and extract meaningful information from graph-structured data. They leverage the topology of the graph to capture relationships and dependencies between nodes, making them particularly effective for tasks involving interconnected data points, such as Web3 transactions and addresses.
Pond innovatively combines GNNs with Generative Pre-trained Transformers (GPT) to maximize their capabilities. Rather than discarding existing technologies like GPT, Pond adopts a synergistic approach. It integrates attention mechanisms and time/position encoding from language models into GNNs and utilizes GPT for direct text data processing. The GNN model serves as a source of truth for retrieval-augmented generation, alleviating the hallucination issue in GPT and enhancing its knowledge. By combining GPT’s natural language understanding with GNN’s graph analysis proficiency, Pond unlocks hidden insights within Web3 data, creating a comprehensive model that enhances both text and graph data processing.
Pond will develop the foundational model and its associated toolkit, enabling third-party models to be integrated into the large model using a technique pioneered by one of their team members, Liangxi Liu, in his paper “A Bayesian Federated Learning Framework With Online Laplace Approximation,” which was later published in IEEE TPAMI, the top journal in machine learning.
The ‘mega model’ will support third-party use cases and allow them to customize or fine-tune their own models on top of it. In addition to developing their own ecosystem, they have also partnered with third-party distributors such as Ritual, Allora and Phala Network to further expand their developer and application ecosystem.
Another significant revenue stream for Pond comes from the utilization of API calls for their foundation model. Key stakeholders include third-party distribution channels like Ritual and Allora, as well as various models, AI agents, and applications. These APIs enable seamless integration of Pond’s advanced analytical capabilities into a wide range of services, providing essential insights and functionalities. This approach not only broadens Pond’s reach within the Web3 ecosystem but also ensures steady income through API usage fees, enhancing the overall value and applicability of their platform
Nevermined
Nevermined envisions a world where AI agents are the dominant consumers. They believe that AI advancement hinges on agents’ ability to access and transact on needed services and data. However, AI agents lack traditional bank accounts and there is currently a poor UX for payments between agents and humans.
Nevermined is addressing these issues to enable AI agents to transact, generate value, and interact with communities seamlessly. AI agents won’t have bank accounts – instead they will transact using blockchains, digital assets and crypto wallets. This means that the services, models, datasets, workflows, etc., from either 3rd parties or even other agents, will also need to be able to transact with digital assets as well. Nevermined solution is to create a platform that allows AI developers, data publishers and model providers to wrap their APIs in access tokens that enable control over usage. This service is called Smart Subscriptions and allows AI builders and Agents to monetize their AI services.
How does Nevermined Work?
AI builders/creators can monetize their API and set time and price access rules. After the builder/creator sets the subscription terms for API access, a decentralized ID (DID) linked to the creator’s wallet is issued and can be tied to a specific subscription.
The Nevermined App then auto-generates a widget with subscription details for the marketplace for users to discover, but creators can also add this widget to their own sales channels like websites.
Buying access to a subscription means the asset’s smart contract gives the buyer a unique Subscription token (ERC-standard) that can be copied into their app. This Smart Subscription proves ownership of a special access token, letting the user send HTTP requests to the AI service.
Smart Subscriptions
Smart Contracts + Subscription Logic = Smart Subscriptions!
Smart Subscriptions enhance blockchain smart contracts and NFTs, offering greater utility than traditional NFTs. While NFTs typically represent unique assets in a 1-to-1 relationship, Smart Subscriptions allow for 1-to-many relationships, creating asset buckets under one token.
This innovation is particularly useful for AI workflows requiring access to multiple assets like datasets, models, and analysis services. Instead of needing separate NFTs for each component, a single Smart Subscription can represent and orchestrate an entire AI service pipeline.
Nevermined has also added time-based access parameters to Smart Subscriptions. This feature enables token-gated access for various durations, from hours to years, providing flexible and customizable access control for digital assets and services.
How they fit into the Near Ecosystem
Pond
Near intends to work with Pond to incubate a suite of NEAR-native foundational models trained on historical transaction data. This could then be leveraged by other developers in the ecosystem to create a variety of use cases around it.
Nevermined
Near plans to work with Nevermined to create NEAR-native AI Agent Frameworks to integrate across ecosystem apps. This would bring Nevermined’s payment and coordination platform to the Near ecosystem to empower and compensate AI developers and agents.
Wrapping Up
Near is diligently working with various AI projects to sow the seeds of an AI ecosystem. As we touched on, the availability of input data is paramount but so are foundational models and payment and coordination mechanisms such as what Pond and Nevermined are building. With these in place, you can start to see an ecosystem take shape. In our next piece, we will touch on compute and inference providers. Until next time!
