WAIC 2025: AI will definitely continue to grow, and it's a difficult problem for humanity to prevent it from becoming the "ultimate villain".

The World Artificial Intelligence Conference (WAIC) in Shanghai has regained its popularity!

On July 26, 2025, this year's WAIC attracted more than 1,200 guests from over 30 countries and regions. Among them were 12 laureates of top awards such as the Turing Award and the Nobel Prize, over 80 Chinese and foreign academicians, and representatives from multiple world - class laboratories.

At the opening event on July 26, Geoffrey Hinton, the 2024 Nobel laureate and a master in the AI field, gave a speech. This well - known scholar, who has always adhered to the "AI threat theory", once again reiterated the dangers of the unregulated development of AI and called on the global community to establish a "safety network" for AI research.

In addition, Yan Junjie, the founder and CEO of MiniMax, a rising domestic AI startup, said in his speech that the growth of increasingly powerful AI has almost no limits. With the decline in training costs, "AI in the future will be more accessible to the public."

Meanwhile, Peng Zhihui (also known as Zhihui Jun), the co - founder and CTO of Zhipu Robotics, a domestic robotics startup that has recently been rumored to "backdoor list", brought the company's robot product Lingxi X2 on stage to perform a "cross - talk", making the "partnership" between robots and humans more tangible.

01 Both Lions and Babies Need "Supervision"

Even though his disciples and followers occupy half of the Silicon Valley AI technology circle, Geoffrey Hinton, the "grandmaster" in the AI field, has always gone against the grain of his industry peers and consistently adhered to the "AI threat theory".

At the WAIC conference on July 26, Hinton once again reiterated his concerns about the rapid development of AI in his speech.

When briefly describing the development of AI technology over the past 30 years up to the current large - model stage, Hinton believes that the way current large models understand language is the same as that of humans.

"Humans might just be large language models and will also produce hallucinations like large language models, creating a lot of hallucinatory language." It has to be said that this insight from Hinton fits the so - called "human - machine" meme in current social media very well.

However, the problem is that compared with the human "carbon - based brain", the "silicon - based brain" of AI has the inherent advantages of being storable, replicable, and capable of "instant transmission". This means that with technological development, the industry generally believes that it is only a matter of time before AI becomes smarter than humans. As a form of existence, these "intelligent agents" represented by AI will surely demand "survival" and "control".

Hinton believes that current AI may be like a "three - year - old child" and is still easily manipulated by humans, but the future situation is uncertain. He also compares current AI to a young lion, and there are only two possibilities for keeping a lion: "Either train it not to attack you or eliminate it."

Judging from the current global AI progress, no single country can truly "eliminate AI" by halting technological development. There is only one way left: the world needs an AI safety agency to train AI to "do good".

"How to train an AI that doesn't want to rule humans is the ultimate problem humanity faces," Hinton said at the end.

02 Some Call for Regulation, While Others Seek Deregulation

The master's speech at WAIC is thought - provoking, but in North America, where Hinton is based, the situation is a bit "tone - deaf". OpenAI and Anthropic, founded by Hinton's students, are already companies valued at hundreds of billions of dollars. Not to mention that Silicon Valley venture capitalists have placed huge bets on AI startups in the past two years.

A prominent manifestation is that as AI companies increase their lobbying expenses in the US political arena, the US regulatory authorities have officially loosened control over AI development.

Also this week, on July 23 local time in the US, then - President Donald Trump released the Artificial Intelligence Action Plan. In this document, the US regulatory authorities determined to ensure the leading position of the US in AI from aspects such as data, standards, and talent:

1. Invest in R&D

Significantly increase long - term federal investment in basic and applied AI research, especially in areas such as next - generation AI, AI safety, and trustworthy AI.

2. Unleash AI Data Resources

Promote the safe opening of the vast amount of data sets owned by the federal government to AI researchers and the public, providing high - quality "fuel" for model training.

3. Set AI Technical Standards

Under the leadership of the government, in cooperation with the industry and academia, formulate global benchmarks, standards, and norms for AI technology to ensure that AI systems are safe, reliable, interpretable, and fair.

4. Cultivate an AI - Ready Workforce

Reform STEM education, promote apprenticeship and retraining programs, attract and retain top global AI talent, and reserve human resources for the AI economy.

5. Strengthen International Collaboration

Establish an AI alliance with allies and partner countries, jointly formulate rules, counter the abuse of AI by "authoritarian countries", and promote the open and democratic application of AI.

6. Protect Key Technologies

Strengthen the protection of key US AI technologies, algorithms, and hardware (especially semiconductors). Use means such as export controls and investment reviews to prevent technology from flowing to strategic competitors.

It can be seen that the US is giving the green light to the development of AI technology and using geopolitical issues to prevent competitors from starting on an equal footing.

03 The "Experience Era" of Robots Is Coming

At this year's WAIC conference, humanoid robots are undoubtedly the most attractive highlight, without a doubt.

During the main forum session, Peng Zhihui (Zhihui Jun), the co - founder of Zhipu Robotics, brought the company's product Lingxi X2 on stage to perform a "human - robot cross - talk".

Cross - talk emphasizes speaking, imitating, joking, and singing, while for robots and humans, "understanding is everything". Lingxi X2 stated on stage that the collaboration between humans and robots should be based on "consensus". But how to establish a consensus between robots and humans and break through the key to human - robot collaboration? Zhihui Jun said that this is the area the company will focus on, and it is important to move forward with more peers.

Therefore, Zhihui Jun also announced the open - source plan for "Zhipu Lingqu OS" on site, hoping to work with more people to promote the integration of the current robot system ecosystem and the breakthrough of new embodied intelligence technologies.

As if to endorse embodied intelligence, Richard Sutton, the 2024 Turing laureate and a professor in the Department of Computer Science at the University of Alberta, Canada, who participated in the conference via video link, believes that the data currently used for training large models is almost exhausted. But there's no need to be discouraged because this means that the next era of AI - the Era of Experience - is about to arrive.

Different from the past when AI was trained with "static" data, in the future, AI can be made to obtain knowledge and improve abilities by "experiencing" the external environment and objects, just like human infants. Although this goal is still a bit far off, currently, a large number of robotics startups are indeed constantly training and learning in the "physical world".

This is why top scholars in the industry, including Fei - Fei Li, have shifted from "AI" to "physical AI", emphasizing that for artificial intelligence to truly enter the real world, it must understand and learn the whole world from a three - dimensional perspective.

This article is from the WeChat official account "GeekPark" (ID: geekpark), written by Jing Yu, and published by 36Kr with permission.