HKUST's Xu Jialong: The Moat of AI Agents Is Not Yet Solidified, Model Differences Reflect Efficiency Rather Than Disruptive Breakthroughs

On April 21, during the roundtable discussion “Decoding Web 4.0: When AI Agents Take Over On-Chain Permissions,” Xu Jialong, Associate Vice President of Hong Kong University of Science and Technology, commented on the “moat of AI agents.” He noted that there are differences in the model training paths and technical systems behind different AI agents, leading to significant variations in actual user experience. Recently, some new models and tools have shown superior performance in generation quality and execution efficiency, even demonstrating higher ceilings in development output. However, he pointed out that at the current stage, these differences have not yet formed a decisive generational gap, being closer to “efficiency improvement” rather than a “paradigm breakthrough.” In other words, competition among agents is still in a rapid evolution phase, with no stable and insurmountable technical barriers emerging yet. The iteration pace of current AI agents and large models is extremely fast, with new products or capabilities appearing almost weekly, driving the industry forward. However, from the perspective of practical use and business decision-making, whether there is a need to continuously and frequently follow these changes still requires careful evaluation.