Written by: Haotian
Everyone is looking forward to AI+Web3 becoming the catalyst for this round of bull market, as can be seen from the high valuations and heavy bets given by VCs. The question is, what are the current problems in the AI+Web3 integration track? Combined with this systematic report by @web3caff_zh, let me talk about my views:
1) AI training requires large amounts of data, and Web3 is useful for data tracking and the incentive effects derived from it. In the long run, AI will definitely need the help of web3, but it needs to be clarified that web3 can only solve limited problems of AI.
For example, the main driving force of traditional large-scale data training, continuous algorithm optimization, computer vision, speech recognition technology, game AI and other core areas still depends on large-scale centralized computing power and hardware and software adaptation optimization such as chips and algorithms. In the short term, there is no possibility for web3 to gain a foothold in the direction of expanding the boundaries of AI capabilities, such as deep learning convolutional neural networks, reinforcement learning, and brain-like computing models.
2) Generative AI is only a small branch of the AI sector, but it has accelerated the integration of AI and web3. Because generative AI is an AI inclusive technology that is more application-oriented. Ideally, the basic large models are generally completed by large companies using concentrated computing power and adopting open source policies to drive the upper-level application market. The overall AI market will gradually become long-tailed, and the importance of model fine-tuning and reasoning will be highlighted.
However, once companies that control core computing power and model resources change their open source policies, it will have a direct impact on the overall AI market. To avoid this crisis, an infra that relies more on distributed computing power architecture and distributed reasoning collaboration architecture will become necessary.
3) Web3 can play a key role in the construction of AI distributed frameworks. For example, during model training, blockchain can create a unique identifier for the data source and perform data deduplication to improve training efficiency. When computing power is insufficient, blockchain can use the Tokenomics incentive mechanism to build a distributed AI computing power network. During parameter fine-tuning, blockchain can record different versions of the model, track the evolution of the model, and perform refined control.
In the model reasoning stage, ZK, TEE and other technologies can be used to build a decentralized reasoning network to enhance communication and mutual trust between models; in the edge computing and DePIN integration stages, web3 can help build a decentralized edge AI network and drive the integration of AI+DePIN Internet of Things.
4) When Vitalik talked about the integration point of AI+Web3, he stated that AI can be gradually integrated as a participant in the Web3 world, so the integration of AI and web3 will definitely be very slow.
On the one hand, the mainstream web2 world is still focusing on the effectiveness of AI and does not rely too much on the AI behind-the-scenes collaboration framework, which leads to a disconnect with web3. On the other hand, web3 is still in the stage of building basic infra such as distributed computing power networks, distributed reasoning architecture networks, distributed Tokenomics application networks, and distributed AI Agent tool collaboration networks in the field of AI integration, and has not been fully verified and applied by the mainstream web2 demand groups.
In short, in a word, the general trend of AI+Web3 is correct, but the actual implementation and development are not that fast. It may take a cycle or even across cycles to see significant progress, and a little more patience is needed.