Zhang Yaqin: Transformer will be gradually reconstructed within five years, and AGI will be achieved within 15-20 years | Tsinghua AIR Wuxi Innovation Center established

Bai Jiao sent from Aofei Temple
Quantum Bit | Public Account QbitAI

In the next five years, there will be major breakthroughs in AI technology architecture, and Transformer will be gradually reconstructed.

Achieve general artificial intelligence (AGI) within 15-20 years and pass the "New Turing Test" .

At the "Taihu Dialogue: Artificial Intelligence+" and Tsinghua University AIR Wuxi Innovation Establishment Ceremony, Academician Zhang Yaqin made such a trend judgment and named five important development directions worthy of attention, including multimodality, embodied intelligence, biological intelligence, agent intelligence, and edge intelligence.

Among them, driverless cars are the most representative of embodied intelligence. He even said: It will be the biggest embodied intelligence application in the next five years .

2025 is the "ChatGPT moment" for autonomous driving. In 2030, autonomous driving will become mainstream, and 10% of new cars will have L4 capabilities.

This event included the main forum and three roundtable forums. Many experts, including Academician Zhang Yaqin, shared many views worthy of reference on topics such as smart industry, industrial applications, and autonomous driving.

Zhang Yaqin: Achieving general artificial intelligence within 15-20 years

At the main forum, Academician Zhang Yaqin shared some views on the development of artificial intelligence, mainly including the core development trends of current large models, the evolution process of AI, the development prospects of driverless cars, and specific governance recommendations.

He first named the five development directions of the AI ​​big model , including multimodal intelligence, autonomous intelligence, edge intelligence, embodied intelligence, and biological intelligence.

Taking embodied intelligence as an example, Zhang Yaqin talked about the application of driverless cars in detail. He believes that driverless cars are the biggest embodied intelligence application in the next five years and the first embodied intelligence to pass the "New Turing Test". Large models/generative AI accelerate L4 generalization capabilities, combined with vehicle-road-cloud integrated collaboration, improve driving safety and traffic efficiency, and help realize the ChatGPT moment of driverless cars.

In 2025, the "ChatGPT moment" of driverless cars will be realized ; in 2030, driverless cars will become the mainstream of the market, and it is expected that 10% of new cars will have L4 level autonomous driving capabilities.

Currently, the development of large AI models is at a critical stage.

Zhang Yaqin predicts that in the next 10 years, large models and generative AI will become mainstream technologies and industrial routes; and that general artificial intelligence (AGI) will be achieved within 15-20 years and pass the "New Turing Test" .

Further splitting is done around various intelligent directions.

• 0-5 years: Information intelligence . Within 0 to 5 years, in the field of information intelligence, the understanding and generation of language, images, sounds and videos must pass the new Turing test.

• 0-10 years: Physical intelligence (embodied intelligence) . Within 0 to 10 years, in the field of physical intelligence, we will achieve the ability of large models to understand and operate in physical environments and pass the new Turing test.

• 0-20 years: Biointelligence . In the field of biointelligence, we will focus on the human body, brain-computer interface, organisms, pharmaceuticals and life sciences, realize biointelligence that connects large models with organisms, and pass the Turing test.

Specifically in terms of industrial ecology, technology, algorithms and other dimensions, these trends and directions were mentioned.

The basic big model will be the technical foundation of the artificial intelligence era, and will form a new industrial ecology together with the vertical industry model and edge model. The scale of such an ecology will be 100 times larger than that of the PC era and more than 10 times larger than that of the mobile Internet era.

At present, the core elements of the big model are mainly Token-based (unified representation) Scaling Law.

In addition, he also talked about the fact that large models now require new algorithm systems, and Transformer, Diffusion, and AR will be gradually reconstructed within 5 years.

Compared with the human brain, existing algorithms have problems of low efficiency and high energy consumption. It is necessary to develop a new algorithm system, including world models, DNA memory, agents, reinforcement learning (RL), probabilistic systems and decision systems, to achieve a 100-fold efficiency improvement.

In the next five years, there will be a major breakthrough in AI technology architecture. The current mainstream AI technology framework Transformer/Diffusion/AR may be gradually reconstructed by new technologies in the next five years.

Finally, Academician Zhang Yaqin talked about suggestions on the governance of AI development.

AI technology can only usher in a new leap forward by establishing a scientific classification system, ID entity mapping, increasing investment, setting red lines and boundaries, and strengthening international cooperation.

There is also a roundtable forum

Regarding the driverless driving scenario that Academician Zhang Yaqin highlighted, Gu Weihao, co-founder and CEO of Momenta, who is a player in the field of autonomous driving, further shared his practical experience and future prospects.

He first talked about the biggest challenges of autonomous driving, which mainly lie in four aspects:

Data level : The way to obtain and process high-quality data is crucial. Data is not only a technical issue, but also related to the design of products and business models.

Visual model application : Currently, the 1 billion parameter visual model performs well in applications, but there is limited room for improvement. The problem may lie in token selection, encoding method or data quality.

Edge configuration : Autonomous driving technology must ultimately be implemented at the edge, and how to configure large cloud models at the edge with limited resources is a problem we are currently solving.

Verification problem : After developing large models and solutions, how to verify their effectiveness requires exploring new verification methods and data sets.

In this case, what level of autonomous driving can users achieve before they are willing to pay for it? Gu Weihao directly pointed out that it can save effort, worry and money .

Effort saving It refers to the driver's decision-making. Users may hope that this car can help me relieve energy in the future.

Peace of mind The key is how AI can express the chain of thinking, so that users can understand and handle the scenario with peace of mind.

Saving money It is more intuitive. He gave a specific example. For example, using L4 to solve the problem of the last mile of terminal delivery, it costs an average of 30 cents per express delivery, which reduces the cost of the express delivery outlet owner by more than 10 cents.

Talking about the future autonomous driving ChatGPT moment , he said that by then all understanding of the physical world will be applicable to robot development.

Each robot has a specific purpose, but their common foundation is the understanding of the physical world and the construction of world knowledge. This is an important foundation for the future development of all robots.

The main forum of "Taihu Dialogue: Artificial Intelligence +" focused on the development trend and application prospects of the intelligent industry. Academician Curtis Carlson took "Influential Innovation" as the core topic and emphasized, "We must improve the way of value creation. Artificial intelligence will completely change the model of value creation." Academician Zhang Hongjiang deeply analyzed the current path of large model development and looked forward to the future of Auto-pilot, which will change the future structure of human organizations and employment.

It is worth mentioning that the Tsinghua AIR Wuxi Innovation Center was officially established, led by Academician Zhang Yaqin. It can be said that Tsinghua and Wuxi were connected for the first time because of AI.

Turing Award winner Academician Yao Qizhi also delivered a speech online. He said that the establishment of AIR Wuxi Innovation Center is not only another important milestone in the development of AIR, but also can provide solid technical support for Wuxi's artificial intelligence industry, inject strong innovation momentum, and further promote Wuxi to achieve new breakthroughs in emerging science and technology fields.

In the future, Wuxi will begin to add new footnotes to the map of artificial intelligence industry development.

Full text of Zhang Yaqin's latest AI views: https://mp.weixin.qq.com/s/8ep4croIlpckijXI2cp8BQ

Click here ???? Follow me, remember to mark the star~