Original title: State of Crypto+AI 2024
Original author: MagnetAI
Original source:mirror
Compiled by: TechFlow
Summary of key points
We conducted an in-depth analysis of 67 Crypto+AI projects and classified them from the perspective of Generative AI (GenAI). Our classification covers:
GPU DePIN
Decentralized computing (training + inference)
Validation (ZKML+OPML)
Encrypted Large Language Model (LLM)
Data (general + AI specific)
AI Creator Apps
AI Consumer Applications
AI Standards (Tokens + Proxy)
AI Economy
Why write this article?
The Crypto+AI narrative has attracted a lot of attention. Many reports on Crypto+AI are emerging, but they either cover only part of the AI story or explain AI only from the perspective of cryptocurrency. This article will explore this topic from the perspective of AI, explore how cryptocurrency supports AI, and how AI can bring benefits to cryptocurrency to better understand the current Crypto+AI industry landscape.
Part 1: Decoding the Generative AI Landscape
Let's explore the entire Generative AI (GenAI) landscape starting with the AI products we use every day. These products are usually composed of two main components: a large language model (LLM) and a user interface (UI). For large models, there are two key processes: model creation and model utilization, commonly referred to as training and inference. As for the user interface, it comes in many forms, including conversation-based (such as GPT), vision-based (such as LumaAI), and many other forms that integrate inference APIs into existing product interfaces.
calculate
Digging deeper, computing is fundamental for both training and inference, and relies heavily on underlying GPU computing. While the physical connections of GPUs in training and inference may be different, GPUs are common as infrastructure components of AI products. On top of this, we have orchestration of GPU clusters, called clouds. These clouds can be divided into traditional multi-purpose clouds and vertical clouds, with vertical clouds focusing more on and optimizing for AI computing scenarios.
storage
Regarding storage, AI data storage can be divided into traditional storage solutions, such as AWS S3 and Azure Blob Storage, and storage solutions optimized specifically for AI datasets. These specialized storage solutions, such as Google Cloud's Filestore, are designed to improve data access speed in specific scenarios.
train
Continuing the discussion of AI infrastructure, it is crucial to distinguish between training and inference as they are significantly different. In addition to general computing, both also involve a lot of AI-specific business logic.
For training, the infrastructure can be roughly divided into:
Platform: Designed specifically for training, it helps AI developers efficiently train large language models and provides software acceleration solutions such as MosaicML.
Base model providers: This category includes platforms like Hugging Face that provide base models that users can further train or fine-tune.
Frameworks: Finally, there are various basic training frameworks built from scratch, such as PyTorch and TensorFlow.
reasoning
For reasoning, it can be roughly divided into:
Optimizers: Specialize a range of optimizations for specific use cases, such as algorithmic enhancements to support parallel processing or media generation. An example is fal.ai, which optimizes inference for the text-to-image process, increasing diffusion speed by 50% over general methods.
Deployment platform: Provides general model inference cloud services, such as Amazon SageMaker, to facilitate the deployment and expansion of AI models in different environments.
application
While AI applications are countless, they can be roughly divided into two categories based on user groups: creators and consumers.
AI Consumers: This group mainly uses AI products and is willing to pay for the value these products bring. A typical example is ChatGPT.
AI Creators: On the other hand, AI creator applications are more about inviting AI creators to their platforms to create agents, share knowledge, and then share profits with them. GPT Market is one of the most famous examples.
These two categories cover almost all AI applications. Although more detailed classifications exist, this article will focus on these broader categories.
Part 2: How Cryptocurrency Can Help AI
Before answering this question, let’s summarize the main advantages that cryptocurrency can bring to AI: monetization, inclusivity, transparency, data ownership, cost reduction, etc.
From the vitalik.eth blog: A high-level summary of the intersection of crypto + AI
These key synergies help the current landscape primarily in the following ways:
Monetization: Unique cryptographic mechanisms such as tokenization, monetization, and incentives enable disruptive innovation in AI creator applications, ensuring an open and fair AI economy.
Inclusiveness: Cryptocurrencies allow permissionless participation, breaking down the limitations imposed by closed, centralized AI companies today. This enables AI to be truly open and free.
Transparency: Cryptocurrencies can use ZKML/OPML technology to make AI completely open source, putting the entire training and reasoning process of LLM on the chain, ensuring the openness and permissionlessness of AI.
Data ownership: By enabling on-chain transactions to establish account (user) data ownership, users can truly own their AI data. This is particularly beneficial at the application layer, helping users effectively protect their AI data rights.
Cost reduction: Through token incentives, the future value of computing power can be realized and the current GPU cost can be significantly reduced. This approach greatly reduces the cost of AI at the computational level.
Part 3: Exploring the Crypto+AI Landscape
Applying the benefits of cryptocurrency to different categories in the AI landscape creates a new perspective on the AI landscape from a crypto perspective.
Large language model layer
GPU DePIN
We continue to outline the AI+Crypto blueprint based on the AI landscape. Starting with large language models and starting with GPUs at the base layer, a long-term narrative in cryptocurrency is cost reduction.
With blockchain incentives, we can significantly reduce costs by rewarding GPU providers. This narrative is currently called GPU DePIN. While GPUs are used not only in AI but also in gaming, AR, and other scenarios, the GPU DePIN track typically covers these areas.
Those focused on the AI track include Aethir and Aioz Network, while those working on visual rendering include io.net, render network, and others.
Decentralized computing
Decentralized computing is a narrative that has been around since the inception of blockchain and has evolved significantly over time. However, due to the complexity of computing tasks (compared to decentralized storage), it usually requires limiting computing scenarios.
As the latest computing scenario, AI has naturally spawned a series of decentralized computing projects. Compared with GPU DePIN, these decentralized computing platforms not only provide cost reduction, but also meet more specific computing scenarios: training and inference. They are orchestrated across the WAN, significantly enhancing scalability.
Achieving scale and cost-efficiency with gensyn.ai
For example, platforms that focus on training include AI Arena, Gensyn, DIN, and Flock.io; platforms that focus on inference include Allora, Ritual, and Justu.ai; and platforms that handle both aspects include Bittensor, 0G, Sentient, Akash, Phala, Ankr, and Oasis.
verify
Verification is a unique category in Crypto+AI, primarily because it ensures that the entire AI computing process, whether training or inference, can be verified on-chain.
This is essential to keep the process completely decentralized and transparent. In addition, technologies like ZKML protect data privacy and security, enabling users to own 100% of their personal data.
According to the algorithm and verification process, it can be divided into ZKML and OPML. ZKML uses zero-knowledge (ZK) technology to convert AI training/inference into ZK circuits, making the process verifiable on the chain, as shown in platforms such as EZKL, Modulus Labs, Succinct, and Giza. On the other hand, OPML uses off-chain oracles to submit proofs to the blockchain, as shown in Ora and Spectral.
Encryption basic model
Unlike general-purpose large language models such as ChatGPT or Claude, encrypted base models are retrained with large amounts of encrypted data, giving them a specialized knowledge base for cryptocurrencies.
These base models can provide powerful AI capabilities for crypto-native applications such as DeFi, NFT, and GamingFi. Currently, examples of such base models include Pond and Chainbase.
data
Data is a key component in the field of AI. In AI training, datasets play a vital role, while in the reasoning process, a large number of user prompts and knowledge bases also require a large amount of storage.
Decentralized data storage not only significantly reduces storage costs, but more importantly ensures data traceability and ownership.
Traditional decentralized storage solutions such as Filecoin, Arweave, and Storj can store large amounts of AI data at very low costs.
Meanwhile, newer AI-specific data storage solutions are optimized for the unique characteristics of AI data. For example, Space and Time and OpenDB optimize data preprocessing and querying, while Masa, Grass, Nuklai, and KIP Protocol focus on the monetization of AI data. Bagel Network focuses on user data privacy.
These solutions leverage the unique strengths of cryptocurrency to innovate in data management in areas of AI that have previously received less attention.
Application Layer
1. Creators
In the Crypto+AI application layer, creator applications are particularly noteworthy. Given the inherent monetization capabilities of cryptocurrency, it is natural to incentivize AI creators.
For AI creators, the focus is divided into low/no code users and developers. Low/no code users, such as bot creators, use these platforms to create bots and monetize them through tokens/NFTs. They can quickly raise funds through ICOs or NFT Mints, and then reward long-term token holders through shared ownership (such as revenue sharing). This fully opens up their AI products to be jointly owned by the community, thus completing the AI economic life cycle.
Additionally, as Crypto AI creator platforms, they solve early to mid-term funding and long-term profitability issues for AI creators by leveraging the tokenization advantages inherent in cryptocurrencies, and provide services at a fraction of Web2’s typical commission rates — demonstrating the zero operating cost advantage brought by the decentralization of cryptocurrencies.
In this field, platforms such as MagnetAI, Olas, Myshell, Fetch.ai, Virtual Protocol and Spectral provide agent creation platforms for low/no code users. For AI model developers, MagnetAI and Ora provide model developer platforms. In addition, for other categories such as AI+ social creators, there are platforms such as Story Protocol and CreatorBid that are tailored for them, while SaharaAI focuses on the monetization of knowledge bases.
2. Consumers
Consumer refers to AI that directly serves cryptocurrency users. Currently, there are fewer projects on this track, but the existing ones are irreplaceable and unique, such as Worldcoin and ChainGPT.
3. Standards
Standards is a unique track in Crypto that features the development of independent blockchains, protocols, or improvements to create AI dApp blockchains or enable existing infrastructure (such as Ethereum) to support AI applications.
These standards enable AI dApps to embody the strengths of cryptocurrency, such as transparency and decentralization, providing essential support for creators and consumer products.
For example, Ora extends ERC-20 to provide revenue sharing, and 7007.ai extends ERC-721 to tokenize model inference assets. In addition, platforms such as Talus, Theoriq, Alethea, and Morpheus are creating on-chain virtual machines (VMs) to provide execution environments for AI agents, while Sentient provides a comprehensive standard for AI dApps.
4. AI Economy
AI Economy is a major innovation in the field of Crypto+AI, emphasizing the use of tokenization, monetization, and incentive mechanisms of cryptocurrency to achieve the democratization of AI.
The AI Economic Lifecycle Developed by MagnetAI
It highlights the AI sharing economy, community co-ownership, and shared ownership. These innovations have greatly promoted the further prosperity and development of AI.
Among them, Theoriq and Fetch.ai focus on agent monetization; Olas emphasizes tokenization; Mind Network provides re-staking benefits; MagnetAI integrates tokenization, monetization and incentive mechanisms into a unified platform.
in conclusion
AI and cryptocurrencies are natural partners. Cryptocurrencies help make AI more open, transparent, and irreplaceably support its further prosperity.
In turn, AI expands the boundaries of cryptocurrency, attracting more users and attention. As a universal narrative for all of humanity, AI also introduces an unprecedented mass adoption narrative to the crypto world.
