Note: This article comes from @ bonnazhu’s Twitter account, and is compiled by Mars Finance as follows:
With the recent release of OpenAI 4o, let me share my views on AI+blockchain:
The generative AI wave led by OpenAI has driven the development of the three sectors of data, storage, and computing. From now on, AI will become their most important customer in the next decade or even decades. The chain of serving AI well and then having AI serve various downstream industry customers and applications is gradually taking shape. AI has become the most important middle layer and engine:
First, AI has driven the demand for upstream infrastructure:
1) Computing: including chip design and production, cloud computing services, data centers, network/power infrastructure, etc.
This link focuses on physics. The training and output of AI results require a lot of computing power, electricity and network resources, and the performance of the chip is the key to efficiency and energy consumption. This determines that chip design companies such as NVIDIA and AMD, wafer foundries such as TSMC and Samsung, and technology giants with cloud computing and data center businesses such as Google, Microsoft, and Amazon are destined to capture the greatest value in this round.
But blockchain is not without its place. Currently, the monopoly of computing power is very obvious. High-performance GPU cards are hard to come by, or you need to pay a high premium to obtain related services from cloud computing vendors. In addition, due to geopolitical reasons, chip bans, etc., the geographical distribution of computing power is also concentrated. The overflow of demand caused by this imbalance makes decentralized computing one of the blockchain sectors that has gained actual benefits in this round of AI wave . There are many projects in this sector, and new projects are constantly emerging. The competition will be fierce, such as @akashnet_ @rendernetwork @gensynai @NodeAIETH @exa_bits @ionet @fluence_project @gpunet @nosana_ai and so on.
However, due to the performance limitations of the blockchain network itself and the contradiction between the high computational complexity of machine learning, complex deep learning must be carried out off-chain and then the results are transmitted to the chain. How to verify whether the computing power provider has performed the training task as required is a difficult point, and the calculation requires calling data and models, which has potential privacy exposure issues. At this time, the power of ZK (zero-knowledge proof) is revealed. At present, there are many projects exploring ZK services for AI, such as @bagel_network @gizatechxyz @ModulusLabs aims to build a machine learning platform, namely ZK machine learning, where developers can deploy AI models and use ZK to verify the AI training and inference process. @ezklxyz focuses on ZKP generators and verifiers for AI services, and @Ingo_zk is delving into ZKP generation hardware acceleration.
In addition, the energy consumption of generative AI (including energy consumption generated by computing and energy consumption caused by heat dissipation) is also quite amazing. It is said that when OpenAI trained GPT-6, it broke down Microsoft's power grid. As the major giants continue to increase their investment in AI data centers (OpenAI plans to work with Microsoft to build a supercomputer called Stargate at a cost of $100 billion), energy consumption will only rise exponentially. However, the construction and renovation cycle of infrastructure such as network/power is very slow, and in countries such as the United States, most of the land is privately owned, and the expansion of power grids and related infrastructure requires private consent. How to motivate private individuals to participate in the expansion of infrastructure, or to reduce their dependence on and burden on the power grid, may be an important topic of #DePin in the future . Of course, in addition to electricity, stable bandwidth is also one of the important infrastructures required by AI. Most data centers tend to be built closer to ISPs (Internet service providers). Where electricity is abundant, network bandwidth resources may not be abundant. How to use #DePin to solve this mismatch problem is also a direction worth looking forward to.
2) Data: including data collection, data labeling/processing, and data transaction/authorization.
Although data is the "food" of AI, most machine learning models can only use processed structured data. Currently, the data sources for machine learning are very wide, and most of them are unstructured and scattered public data, so it takes a lot of time and effort to collect and process these data. This is actually a labor-intensive chore, but it is also a link that blockchain and token economy can well cut into. Currently, the main companies doing this data collection and processing subcontracting business are @getgrass_io @PublicAI_ @AITProtocolThese ones.
However, it should be noted that with the emergence of new machine learning model architectures, the reliance on structured data will change. New technical architectures such as self-supervised learning, GAN, VAE, and pre-trained models can directly use unstructured data for deep learning, bypassing the data processing and cleaning process, which will have a certain impact on the demand for labor-intensive platforms.
In addition, the data that can be publicly captured is only the tip of the iceberg of the world's data. A large amount of data is actually in the hands of private institutions or individual users. Except for some companies that have public APIs to allow calls, most of the data is still not activated. How to allow more data holders to contribute/authorize their own data while protecting privacy well is a key direction. There used to be many platforms for decentralized data transactions, but because they were unable to find parties with data needs, they basically disappeared after several rounds of trials and tribulations. Only a few, such as @oceanprotocol , survived the spring of AI. Their unique Compute-to-data model allows data users to directly calculate on the data set of data sharers without exposing the data, which just solves this privacy pain point.
3) Storage: including database, data backup/storage system
Most of the data used in training and inference of deep learning models are retrieved from databases or data storage backup systems. Databases and backup/storage systems can be thought of as "refrigerators", but they are actually quite different. The former focuses on management and needs to support frequent reading and writing, as well as complex queries (such as SQL) and retrieval, while the latter focuses on large-scale, long-term backup and archiving, and needs to ensure privacy, security, and non-tamperability.
Database and storage complement each other and serve AI deep learning together. A typical scenario is: data is extracted from the database, preprocessed and cleaned, and converted into a format suitable for model training. The processed data can be stored in decentralized storage to ensure data security. During the model training phase, training data is read from decentralized storage for model training. The intermediate data and model parameters generated during the training process can be stored in the database for quick access, fine-tuning, and updating.
This sector is where blockchain excels, @ArweaveEco @Filecoin @storj @Sia__Foundation is in this field, and even @dfinity can be included. However, I increasingly feel that @ArweaveEco is the most suitable solution for serving AI: its one-time payment and permanent storage model, supplemented by many database projects in the ecosystem and the newly released parallel architecture AO computing network, are perfectly adapted to the needs of multi-threaded tasks in deep learning, which enables it to well support the deployment of machine learning.
Second, AI performance determines the upper limit of downstream applications:
Although AI has been applied in industry and agriculture (2B) to some extent, the breakthroughs we have seen this round are mainly 2C applications based on large language models (LLM). We can divide these applications into two categories:
The first category is actually just the embodiment of a large language model, such as some AIGC platforms, which generate the results that users want based on user instructions. However, the performance of this type of application mainly depends on the AI model used, and the main LLM model is monopolized by giants, so the differences between applications are often small and the moat is relatively narrow; the other category is to use AI models to improve the functions and user experience of existing products, such as search engines and games with added AI capabilities, including @_kaitoai @ScopeProtocol @EchelonFND
In addition, the generative AI wave has also boosted a new application ecosystem - AI Agent, that is, an intelligent robot that has the ability to independently perform tasks and make decisions based on user intentions. AI Agent is essentially based on the LLM model, adding more complex execution and processing logic to enable it to serve different application scenarios. In fact, the prototype of this agent already exists in the field of cryptocurrency, such as the liquidation bot of the DeFi lending protocol and the arbitrage bot of the decentralized trading platform. Although these DeFi Bots have some characteristics of intelligent robots, they are purely on-chain and do not support off-chain behavior. Because they are based on smart contracts, they require external triggers to start.
In the absence of AI, an external keeper network is currently used to connect the off-chain and on-chain. For example, the price oracle is such a typical example, and @thekeep3r is also an example. The emergence of AI Agents has given a new idea, that is, intelligent robots can complete the work by themselves and realize automation. The main on-chain AI Agents are: @autonolas @MorpheusAIs ; Other more general AI Agents include @chainml_ @Fetch_ai ; and the AI Agents that focus on companionship and human-computer interaction include @myshell_ai @virtuals_io @The_Delysium , this type of agent is characterized by anthropomorphism, providing emotional value, and having the imagination space to be applied to various games and metaverse.
Third, write at the end:
AI is actually a fusion narrative. Its emergence has connected several crypto sectors that were originally isolated or even unable to find market fit. Currently, AI is still in the era of large infrastructure investment. Upstream sectors such as data, storage, and computing are the most direct and continuous beneficiaries. They are more sensitive to the development of AI and have higher certainty.
However, for investors in this industry, the risk is that most of the dividends may not be in the cryptocurrency market. The current AI effect in the currency market is more of a spillover effect from the imbalance of supply and demand in the traditional market , or pure speculation. As for downstream applications, the performance ceiling depends on the AI model, and the AI model is still in the process of continuous iteration, and the combination of AI and products is still being explored, and the market fit has yet to be verified. This makes the future of downstream applications still relatively uncertain, and the certainty is not as high as that of the upstream sector.
Of course, there are projects like @bittensor_ and @ritualnet , which I think should be called AI ecological platform projects. They do not simply focus on a certain business in the upstream or downstream, but through the design of architecture and economic mechanisms, enable various providers of upstream and downstream businesses to access and deploy on their platforms or chains to achieve the so-called artificial intelligence collaboration. These projects have a grand vision, but the demand capture problems currently faced by the upstream and downstream of blockchain AI will also be reflected in them, and their valuations are relatively high. However, compared to betting on a specific project, the risk of betting on these platforms will be relatively small.
In the short term, whether blockchain can continue to benefit from AI dividends may still depend on the imbalance of supply and demand in the upstream sector, especially the continued shortage of supply. However, in the medium and long term, the verifiability, immutability and token incentives of blockchain can indeed bring new possibilities to AI. Among them, zero-knowledge proof is a powerful tool that can protect privacy and achieve trusted verification , perfectly solving the problem of blockchain serving the high computing requirements of AI deep learning under performance limitations.
