SAI大家说 | AI主题学术讲座(总第三十期)

School of Artificial Intelligence

The rapid scaling of artificial intelligence has catalyzed unprecedented advancements, yet the prevailing “more data, more compute” paradigm is increasingly hitting unsustainable economic and environmental ceilings. As we approach the limits of data availability and energy efficiency, the next frontier in AI is not merely larger models, but smarter, more efficient, and more adaptable systems. This talk introduces a research agenda focused on building intelligent systems that achieve high performance through deep data insights rather than brute-force scaling. 

The speaker will propose a new paradigm for sustainable multimodal intelligence that addresses two critical bottlenecks: data efficiency and open-world adaptation. First, he will address data efficiency through a principled data curation framework, validated by his work on the SenseNova LLM series, which achieves state-of-the-art performance using only 18% of the data volume required by flagship open-source models. Second, he will tackle open-world adaptation by presenting robust, training-free methods for out-of-distribution detection and open-vocabulary segmentation. These innovations bridge the gap between foundational research and large-scale industrial deployment, with proven impact in complex domains such as manufacturing and remote sensing. The speaker will conclude by discussing his future research trajectory toward self-evolving, unified multimodal agents—a necessary evolution for AI to become truly reliable, scalable, and capable of navigating real-world complexity.

All of you are warmly welcome!