Ein Jahr für Menschen entspricht einem Tag für KI. Die Substitution ist unvermeidlich. Sutton, Wang Xingxing und andere beantworten die vier letzten Fragen der KI.

On the eve of the 2025 Inclusion·Bund Summit, the world's top ten AIs posed rhetorical questions to humanity: What jobs should always be reserved for humans? Should AI be held accountable if it makes wrong decisions and causes damage? If AI can help you optimize your quality of life, how much privacy are you willing to sacrifice? Will you actively avoid some AI assistance to prevent your own abilities from deteriorating? Are you worried that AI hallucinations and information cocoons will lead humanity astray? Is it "raising a tiger to cause trouble" for humans to make AI increasingly powerful...

Rather than saying these are questions from AI to humans, it's more accurate to say they are humanity's self - reflections. In the current era where "one human year equals one AI day", what humanity most urgently needs are answers - answers about the future.

Ma Yi, the dean of the School of Computing and Data Science at the University of Hong Kong, said that his phone has been "bombarded" by parents. The most frequently asked question is, What should my child study to avoid being replaced by AI upon graduation?

AI not only drags ordinary people into a state of both excitement and anxiety but also brings anxiety to major political and economic entities around the world. On September 11, Richard Sutton, the winner of the 2024 Turing Award and the "father of reinforcement learning", said in his keynote speech at the Bund Summit that the outside world's fears of AI - induced biases, unemployment, and even human extinction are exaggerated and have been incited by certain organizations and individuals seeking to profit from them.

Where does the source of all the good things in the world lie? Sutton's answer is that artificial intelligence and human prosperity will come from decentralized collaboration.

Question 1: How well can AI perform tasks?

Wang Xingxing: AI's task - performing ability is on the verge of an explosion

Those already in the industry have less anxiety because they have a deeper understanding of "what AI cannot do".

"Currently, AI can write articles and create paintings better than 99.99% of people. But when it comes to really making AI perform tasks, it's still a desert." This was the first statement made by Wang Xingxing, the founder and CEO of Unitree Robotics, a hot "star" in the field of humanoid robots, after Unitree Robotics officially announced its IPO.

Ma Yi also put forward a similar view. He believes that currently, AI, represented by large models, is still in the most primitive stage of 'phylogenetic intelligence'. Relying on a large number of parameters and pre - trained data, it not only consumes a large amount of resources and has low efficiency but also lacks individual memory and self - awareness. Ma Yi reviewed the four stages of intelligent evolution: from the phylogenetic genetic intelligence represented by DNA, to the ontogenetic intelligence formed when biological individuals develop brains and perception systems, then to the group intelligence achieved through language, and finally to artificial intelligence in the true sense.

Even looking back from the most primitive stage of AI, Wang Xingxing admitted that the thing he regrets the most is not studying AI more than a decade ago: "In 2011, I was also very interested in AI, but at that time, AI was very niche. After reading a few books, I thought there was limited scope for development, so I later switched to working on robots."

In fact, the integrated development of AI and robots is giving birth to a brand - new embodied intelligence industry, which means enabling robots to have AGI capabilities and be able to perceive, plan, and act autonomously like humans. At this year's Bund Summit, many of the displayed robots can already perform basic tasks in some work fields, such as cooking, rescuing people, and placing detonators.

A robotic dog rescues a baby from the ruins

However, Wang Xingxing still believes that there are considerable challenges in the current development of embodied intelligence. Firstly, at the data level, the problems of data collection and quality are still prominent, and the utilization rate of data needs to be improved. Secondly, at the model level, the integration of multi - modal data is not ideal, and aligning the model with the robot's control modality is also a difficult point. For example, when trying to make a robot learn to do housework based on a generated video, although the video generation itself may be quite good, it is still very challenging to align the video generation with the robot's control modality.

As the saying goes, the simplest principle leads to the greatest truth. No matter what era, entrepreneurs will encounter common problems. Wang Xingxing admitted that as the company's personnel scale grows, the efficiency of collaboration may decrease. He needs to spend time exploring more efficient organizational management methods and can't be a "hands - off boss".

"When it comes to really making AI perform tasks on a large scale, we are still on the verge of a massive explosive growth." Wang Xingxing said that the AI era is very fair. As long as one is smart and willing to work, towering trees will eventually grow in the desert. For the young generation aspiring to innovation and entrepreneurship, he suggested "forgetting past experiences, learning the latest knowledge, and whole - heartedly embracing the new era".

Question 2

What is the key variable in AI competition?

Wang Jian: Calculation is indispensable on the journey to Mars

"At this moment of open - source, OpenAI has taken the wrong side of history." At the end of January this year, with the open - sourcing of Chinese large models such as DeepSeek, Sam Altman, the CEO of OpenAI, made a shocking statement.

In 1998, Netscape, the best and most open browser in the industry at that time, went open - source. This was a watershed event in the Internet era and a key variable in the development of the Internet. In the era of artificial intelligence, the choice between open - source models and closed - source models has become the key variable in AI competition.

Wang Jian, the founder of Alibaba Cloud and the director of the Zhijiang Laboratory, believes that simply open - sourcing code can no longer solve the problem. Opening up data and computing resources is the only way to push AI forward.

Space has always been the greatest resource. "Today, we can't just use AI on mobile phones and computers. AI should not be absent from space. In addition to communication satellites, navigation satellites, and remote - sensing satellites, AI will give rise to a fourth type of satellite - the computing satellite." The Zhijiang Laboratory where Wang Jian works has sent large AI models into space via satellites.

May 14, 2025, was an exciting day for Wang Jian. The "Three - Body Computing Constellation" composed of 12 computing satellites sent an 8B AI model identical to that on the ground into space for the first time. This means that data processing can be completed anywhere in space, and inter - satellite connectivity can also be achieved. He explained that after sending satellites into the solar orbit, it is almost impossible to transmit data back to the ground for processing. Only by sending AI and computing power into space can humans truly leave the Earth.

Wang Jian hopes that numerous entities will jointly complete this "Three - Body Computing Constellation" and share space under the principle of openness. In the future, every satellite will be open to anyone in the world.

On the journey to Mars, humans cannot do without calculation and AI. This is the most exciting prospect for the next decade, or even two decades.

Question 3

Where is the end - point of AI?

Sun Xuan: Nuclear fusion

The end - point of AI is energy, and the end - point of energy is nuclear fusion. Sun Xuan, a professor at the School of Nuclear Science and Technology of the University of Science and Technology of China, reiterated the industry's consensus: Nuclear fusion is the key technology to unlock the next - generation civilization.

The power consumption of one million GPUs is equivalent to one - eighth of Beijing's electricity consumption. Currently, AI accounts for 1.5% of the world's electricity consumption. If we compare AI to the "brain of the Earth", the human brain accounts for 20% of the body's energy consumption. Therefore, some people predict that AI's electricity consumption will also account for more than 20% of the world's total electricity consumption. These figures mean that the AI field alone will create a huge energy gap.

Since 2020, capital investment in nuclear fusion has increased significantly. Leading international technology companies such as NVIDIA, Google, and OpenAI have all entered the nuclear fusion field. Betting on this ultimate energy source has become an industry consensus, and the nuclear fusion field is regarded as being on the verge of commercialization.

According to a report released by the Fusion Industry Association (FIA) in July 2024, the total cumulative investment in global nuclear fusion commercial companies has reached $7.1 billion, a year - on - year increase of $900 million. Capital market financing has repeatedly hit new highs. Among the 35 surveyed companies, 89% are optimistic about achieving grid - connected power generation before the end of the 2030s.

However, the ultimate energy source comes with ultimate challenges. There are still technical difficulties on the road to the "artificial sun". Sun Xuan explained that the core scientific difficulty in achieving nuclear fusion lies in confining a plasma with a temperature as high as hundreds of millions of degrees. "It's a bit like trying to trap a temperamental beast in a cage. It's very difficult."

Currently, the mainstream technologies for pursuing controllable nuclear fusion are divided into two major directions: laser inertial confinement and magnetic confinement. The implementation conditions for both require extremely high - level engineering construction. Whether it's developing extremely precise giant lasers or building something as large as the ITER (International Thermonuclear Experimental Reactor), there are problems of high cost and long construction periods.

At the Bund Summit, a reporter from IT Times saw the legendary "artificial sun", which is the latest controllable nuclear fusion model developed by Hefei Xingneng Xuanguang Technology.

The hybrid path of "magnetic inertial confinement" can significantly reduce costs, shorten construction time, and improve iteration efficiency, becoming a way to "tame the beast". In addition, Sun Xuan also put forward a breakthrough prospect: "Can we create a fusion reactor that can learn and design itself, without relying on our existing experimental data but exploring based on physical rules, just like the Go software AlphaZero in the past?"

Question 4

What is the definite direction of AI?

Richard Sutton: AI has entered the 'experience era'

Richard Sutton believes that the human data dividend is approaching its limit, and artificial intelligence is moving from the 'human data era' to the 'experience era', with far greater potential than before.

The goal of most machine - learning today is to transfer existing human knowledge to static AIs lacking autonomous learning abilities. "We are gradually reaching the limit of human data. Existing methods cannot generate new knowledge and are not suitable for continuous learning, which is crucial for the effectiveness of intelligence." He believes that we are entering the "experience era" and need a new data source generated through the direct interaction between intelligent agents and the world. This is how humans and other animals learn, the "37th move" in AlphaGo's self - play, and also the path through which AlphaProof recently won a silver medal in the International Mathematical Olympiad.

Sutton explained that "experience" refers to observation, action, and reward, and these three signals are transmitted back and forth between the intelligent agent and the world. "Knowledge and experience can be learned from experience. The intelligence level of an intelligent agent depends on its ability to predict and control its input signals. Experience is the core and foundation of all intelligence." He also pointed out that reinforcement learning has led us into a new experience era, but to unleash its full potential, two currently immature technologies are still needed - continual learning and meta - learning technologies.

"Humanity's most remarkable superpower is its ability to collaborate better than any other animal. Humanity's greatest success lies in collaboration itself - the economy, the market, and the government are all products of successful collaboration," Sutton said. Artificial intelligence and human prosperity come from decentralized collaboration. 'Collaboration doesn't always happen, but it is the source of all good things in the world. We must seek collaboration, support it, and strive to institutionalize it.'

Looking forward to the future of artificial intelligence, he proposed four practical "prediction principles": First, there is no consensus on how the world should work, and no single view can override the others. Second, humans will truly understand intelligence and create it with the help of technology. Third, the current human intelligence level will soon be surpassed by super - artificial intelligence or humans enhanced by super - intelligence. Fourth, power and resources will flow to the smartest intelligent agents. Therefore, in the process of human development, the replacement by artificial intelligence is inevitable.

Looking at the history of the universe, Sutton divides it into four eras: the particle era, the stellar era, the replicator era, and the design era. He believes that the uniqueness of humans lies in "pushing design to the extreme" and creating things that can design themselves, which is also the goal pursued through artificial intelligence today. Humans are at least the catalyst, the midwife,