The Current Status and Challenges of Web3 AI Development

NVIDIA's stock price has reached a new high, and the progress of multimodal models further consolidates the technological advantages of Web2 AI. From semantic alignment to visual understanding, and from high-dimensional embeddings to feature fusion, complex models are integrating various modalities of expression at an unprecedented speed, building an increasingly closed AI stronghold. Meanwhile, there has been a small bull market for cryptocurrency and AI-related stocks.

However, this wave seems to have nothing to do with the cryptocurrency sector. Recent attempts of Web3 AI in the Agent direction show a deviation in orientation. Trying to assemble a Web2-style multimodal modular system using a decentralized structure is essentially a misalignment of technology and thinking. In today's context, where module coupling is strong, feature distribution is unstable, and computing power requirements are centralized, multimodal modularity is difficult to establish in Web3.

The future of Web3 AI should not be limited to imitation, but should adopt a strategically circuitous approach. From semantic alignment in high-dimensional spaces, to information bottlenecks in attention mechanisms, and to feature alignment under heterogeneous computing power, Web3 AI needs to take "surrounding the cities from the countryside" as its tactical program.

Challenges Facing Web3 AI

Semantic alignment difficulties

Web3 AI is based on a flat multi-modal model, making it difficult to achieve high-dimensional embedding space. This leads to misalignment of semantics and poor performance. High-dimensional embedding space is crucial for understanding and comparing different modal signals, but the Web3 Agent protocol struggles to achieve this.

Most Web3 Agents are merely simple wrappers around existing APIs, lacking a unified central embedding space and cross-module attention mechanisms. This results in information being unable to interact between modules from multiple angles and levels, constrained to a linear pipeline, exhibiting a single function and failing to form a complete closed-loop optimization.

Attention mechanism design is limited.

Low-dimensional space limits the precise design of attention mechanisms. High-level multimodal models require precise attention mechanisms, which need high-dimensional space as a foundation.

Web3 AI's modular design makes it difficult to achieve unified attention scheduling. The lack of a common vector representation, parallel weighting, and aggregation capabilities prevents the construction of a "unified attention scheduling" capability like that of Transformer.

Feature fusion stays at a shallow level

The discrete modular assembly leads to feature fusion remaining at a superficial static stitching stage. Web3 AI often adopts the approach of discrete module assembly, lacking a unified training objective and cross-module gradient flow.

Compared to the complex feature fusion methods of Web2 AI, the fusion strategies of Web3 AI are too simplistic, making it difficult to capture deep and complex cross-modal associations.

Barriers to Entry and Development Directions in the AI Industry

The technological barriers in the AI industry are deepening, but the pain points for Web3 AI have not yet fully emerged. Web2 AI has made significant investments in the development of multimodal systems, creating strong industry barriers.

The development of Web3 AI should follow the strategy of "surrounding the city from the countryside." Small-scale trials should be conducted in edge scenarios to ensure a solid foundation before waiting for the emergence of core scenarios. Suitable directions include lightweight structures, easily parallelizable and incentivizable tasks, such as LoRA fine-tuning, behavior alignment post-training tasks, crowdsourced data training and annotation, small foundational model training, and cooperative training on edge devices.

Before the Web2 AI dividends completely disappear, Web3 AI needs to carefully choose its entry points, focusing on projects that can penetrate from the edges, combine points and surfaces, advance in circular ways, and remain flexible and agile. Only in this way can it find a foothold in the future AI competition.