AI’s Three Questions

As the 2025 World Artificial Intelligence Conference (WAIC2025) approaches, three core questions about AI development have been raised: the mathematical question, the scientific question, and the model question. These inquiries aim to delve into how the new wave of AI revolution will influence the evolution of human civilization.

However, at the conference, these questions were distilled into a more straightforward and intuitive query: Can large models generate a better world?

Understanding Data

The pace of development for large models is evidently much faster than human evolution. Xu Li, Chairman and CEO of SenseTime, recalls that in 2012, when AI pioneer Geoffrey Hinton’s team first won the championship at ImageNet, the scale of machine learning was roughly equivalent to transferring ten years of human knowledge to AI. Fast forward to the era of generative AI, when ChatGPT processes 750 billion tokens, this is akin to a natural language creator writing for 100,000 years.

For large models that develop at lightning speed, the issue of “data hunger” is pressing. These models have nearly covered all publicly available data. It is estimated that by 2027 to 2028, the natural language data available on the internet may be exhausted. In reality, the speed of language generation has not kept pace with the growth of computational power, creating a “backward gap” for large models.

How can we better provide the “oil” of data to support development? The industry has proposed several paths, interestingly all requiring the involvement of large models. One approach is to mine hidden “oil” by sourcing data from the real world. “We find that traditional enterprises have the desire to embrace large models, but their data assets are not structured,” said Tan Bin, CMO of StarRing Technology. He likens large models to supercars, while companies only possess oil fields, emphasizing the urgency of converting these fields into high-quality fuel.

Another bolder idea is for large models to generate data. However, this data must be generated based on an understanding of the real world; otherwise, it risks producing hallucinations and misrepresenting reality. Recently, SenseTime released the “KAIWU” world model, primarily used in intelligent driving scenarios. Xu Li illustrated this with the example of merging in traffic. Collecting data from the real world would be a time-consuming project, but now the world model can generate videos of merging from seven camera angles, adjusting details like weather, vehicle types, road structures, and speeds to create various possible data scenarios.

“As models become more capable and our understanding of the world deepens, the unity of understanding and generation allows for more interactive possibilities,” Xu Li believes. This shift in addressing data issues is moving from passive to active, aiding progress across many industries and providing more opportunities for exploring the real world.

Exploring Efficiency

What is the greatest help that large models provide? The overwhelming answer is efficiency. Over the past few years at the World Artificial Intelligence Conference, the rapid enhancement of efficiency brought by large models has been exhilarating.

During this year’s conference, numerous groundbreaking cases were showcased. At Baidu’s booth, a product called “MiaoDa” was demonstrated, with the slogan “Create an application in one sentence.” A reporter input a natural language command: “Please help me design a professional website for Shanghai tourist attractions.” The large model quickly processed the request, breaking it down into four components: searching for attractions, key sections, design requirements, and functional needs. It then automatically assembled a virtual development team with diverse skills. The reporter observed as the webpage was built in real-time, with roles like architect, development engineer, and UI designer coming online sequentially. Without seeing a single line of code, the webpage was completed in three minutes, and “MiaoDa” even named it “Exploring the Magical Shanghai.” The webpage included all classic attractions and set up a message board and booking interface.

An engineer observing the product noted that building such a website in the past would have taken about 40 days with a team of architects, operations, product managers, backend developers, and testers. The current efficiency left him in awe. Baidu’s MiaoDa brand leader, Zhu Guangxiang, explained that the underlying technology combines “multi-agent collaboration + multi-tool invocation,” utilizing models like Wenxin to mobilize different domain experts based on user commands, achieving astonishing efficiency.

Similar cases were abundant at the conference. For instance, Tongyi Qianwen showcased the AI programming model Qwen3-Coder, which excels in coding capabilities and intelligent agent invocation. A novice programmer could complete a week’s work in just one day with its assistance. The AI verification system from Dewu, which won the highest honor at the World Artificial Intelligence Conference, has already penetrated the industry, demonstrating the ability to generate an authentication report for a pair of sneakers in just five seconds.

The Safety Question

When discussing the benevolence of large models, safety is an unavoidable topic. Each year, the World Artificial Intelligence Conference prioritizes AI safety governance as a top-tier issue, as it concerns the future of humanity.

Just before the conference, an international consensus on AI safety was released, signed by over 20 industry experts and scholars, including Geoffrey Hinton and Yao Qizhi, calling for increased global investment in AI safety. The signatories generally believe that humanity is at a critical turning point—AI systems are rapidly approaching and may soon surpass human intelligence levels.

Implementing comprehensive AI safety education is urgent. At the conference, a reporter experienced a “face-swapping” scenario: standing in front of a screen, their face was scanned, and the system generated a highly realistic “digital mask” that replicated the facial expressions and movements of the real person. However, using the AI face verification model from Hehe Information, all the “fake faces” were accurately identified.

“The ‘fake faces’ generated in our interactive exhibit can be produced by current mainstream general large models,” a team member from Hehe Information explained. These AI safety issues should not be underestimated, as they are applicable in scenarios like bank identity verification, remote account opening, and large transaction validation. “We hope to remind people of the importance of AI safety, allowing large models to ‘generate’ a better world rather than pose threats.”

To address challenges such as agent overreach and excessive delegation, Tsinghua University, in collaboration with Ant Group, has upgraded the large model safety solution “Ant Tianjian.” This solution is based on the security philosophy of “attack as a means of defense,” constructing a full-process protection system through a technology stack of “alignment-scanning-defense.” This solution will be gradually open-sourced and collaboratively built with the industry to establish a trustworthy AI ecosystem.