Top Ten Buzzwords of 2023 Announced, AI Large Models Selected


Domestic AI large models have made breakthroughs, with the "Emu3" model from the Zhiyuan Institute being published in the journal Nature, marking the first international recognition for a Chinese large model in the field of academic excellence, indicating significant progress in underlying architecture innovation.
Although AI large models perform well on reasoning tasks, their core deficiency lies in their inability to learn from mistakes. OpenAI researcher Jerry Tworek pointed out that the lack of an effective correction mechanism after model failure has become a key obstacle in achieving artificial general intelligence.
According to the SuperCLUE-VLM multimodal evaluation, Google's Gemini-3-pro scored 83.64 points to win first place, leading comprehensively in three key dimensions: basic cognition, visual reasoning, and application. The performance of domestic models has also attracted attention.
Six departments in Beijing jointly issued measures to promote the upgrading of the medical device industry through data circulation and AI large models. Key initiatives include building high-quality medical data sets and improving data circulation policies to promote safe and compliant data applications, meeting the needs of enterprises and research institutions.
Tsinghua University published a study in "Nature Machine Intelligence", introducing the new concept of "ability density", challenging traditional AI evaluation standards. The research emphasizes that attention should not only be paid to the number of model parameters, but also to the level of intelligence within each parameter, questioning the scale rule that larger models are necessarily more capable.