Does the perfect big model exist?
How can the development of technology establish ethical and governance mechanisms?
Can general artificial intelligence be perfected through reinforcement learning?
......
At this year's World Artificial Intelligence Conference, a number of Turing Award winners gathered together, and these academic experts examined the future development of artificial intelligence from their own perspectives, not only looking up at the sea of stars, but also putting forward specific suggestions on down-to-earth practical problems.
The only way to make AI safe and useful is open source
Yang Likun (Turing Award winner, Chief AI Scientist of Meta AI Basic Artificial Intelligence Research Institute)
The most exciting development of AI is the creation of a technological revolution. I think natural language processing, such as language models, speech models and other symbols or technologies that are essentially related to symbol sequences, is creating a revolution, and what we need to do now is to drive the same generative revolution in image and video.
The best way to get AI to the human level is to make a larger autoregressive LLM (big language model), and making it intelligent is also the way to make it more controllable. Through the idea of goal-driven artificial intelligence, people can set other goals as their safety guardrails, such as so that these systems do not deceive and dominate human behavior, force AI honest goals, submit to human goals, and need to be careful of those goals that make the model curious. Designing these goals to make the system secure will be a formidable engineering challenge.
Some people, fearful of the consequences of artificial intelligence, want artificial intelligence to be strictly regulated as a whole. I completely disagree with this statement. In the long run, the only way to make AI platforms safe, good, and practical is open source.
Imagine a future where everyone will be able to interact with the digital world through AI assistants, which is not a good thing if the technology is controlled by a few companies. In the future, AI systems will become a treasure trove of all human knowledge, and the way they are trained needs to be based on numerous sources. Therefore, we will see more open source LLM and more open source AI systems in the future.
Young Chinese scholars have made breakthroughs in algorithms
Yao Zhizhi (Turing Award winner, academician of the Chinese Academy of Sciences, dean of the Institute of Interdisciplinary Information Sciences, Tsinghua University)
Chinese scientists have made many breakthrough contributions to the development of AI, and Gao Yang, assistant professor at the Institute of Interdisciplinary Information Studies at Tsinghua University, achieved a very important breakthrough in algorithms more than a year ago, which accelerated reinforcement learning by hundreds of times and attracted international attention.
After ChatGPT, the next important goal of AI research is to enable robots with multiple perception capabilities to learn new skills autonomously in different environments. Existing reinforcement learning methods are too slow, and learning new technologies often takes months. Thanks to Gao Yang's research, reinforcement learning can now be completed in just a few hours.
This not only solves practical problems, but also makes theoretical contributions. In the past six or seven years, there has been a "route dispute" at the top of artificial intelligence, that is, whether to take the road of reinforcement learning. The breakthrough of young Chinese scholars has tilted the balance of this debate to the other side. And for the improvement of general artificial intelligence, we still have a long way to go.
As for the application of large models in vertical fields, a lot of paperwork becomes easier after having large language models, and more work can be tried to be done by machines, which will be a direct direction.
Well-developed large models reduce energy consumption several times
David Patterson (Turing Award winner, member of the American Academy of Arts and Sciences, National Academy of Sciences, National Academy of Engineering)
Over the past decade or so, Google has announced five different TPUs (specialized chips tailored for machine learning) with a three-fold difference in performance, and recently developed supercomputers with new technology.
In fact, we've learned a lot when it comes to carbon emissions.
First of all, what we have to do is calculate the energy consumption. Energy consumption only requires a simple formula, and 4 factors - one is the time taken to train the model, as well as the number of processors and the average power of the processor, so that we can get a rough energy consumption data; The last factor is the power of the data center. Multiply these four factors to get energy in kilowatt-hours. To convert to carbon emissions, you need to take a factor, that is, carbon emissions per kWh.
There are many ways we can try to reduce our carbon footprint. The first is to improve technology. Previously, several Google engineers figured out how to change the quality and hardware parameters of the model, reducing energy consumption and carbon emissions by 57 times.
If you are a practitioner or researcher in the field of artificial intelligence, please continue to improve and research on large models, which will reduce energy and emissions by 2 to 4 times. Counting the energy consumption and carbon footprint of a trained model is to allow us to compete better in order to build more effective models, get the right numbers, and help the world continue to roll out efficient hardware.
Next
Let's find out with the editor
What are the wonderful performances in the Pudong branch venue?
01 The multi-functional space witnessed the birth of stars
In the main venue, the 2023 Zhangjiang Intelligent Robot Science and Technology Innovation Competition and Amazon Technologies DeepRacer Youth Challenge were presented. The former is one of the robotics discipline competitions with the greatest influence and the highest comprehensive technical level for college students and graduate students in mainland China, with industrial application as the direction, talent mining and training as the cornerstone, and technical engineering innovation as the path. The latter students came from key universities such as Tsinghua and Fudan, while primary and secondary schools covered many schools in Beijing, Shanghai, Hainan and other places. In this huge area of the new arena, the contestants lead their works to chase and compete against each other, feeling the charm of science and technology, and also showing the charm of technology. Ignite more people's enthusiasm for learning artificial intelligence technology, stimulate their imagination, and explore more possibilities of science and technology.
Amazon Web DeepRacer Junior Challenge
At the same time, the three-day 2023 RoboCup Robot World Cup China Competition and Zhangjiang Intelligent Robot Science and Technology Innovation Exhibition was also held at the same time, with a total of 20 universities participating, including 26 participating teams and a total of 249 participants.
2023 RoboCup Robot World Cup China Competition and Zhangjiang Intelligent Robot Science and Technology Innovation Exhibition Site
In addition, the Metaverse Core Achievements Exhibition and the AI Innovation and Application Exhibition focus on the phased achievements of the construction of metaverse application scenarios in the fields of medical and health, classroom education, business operations, digital cultural tourism, and intelligent manufacturing, and also allow more participants to witness and participate in the metaverse world where virtual and real are integrated and virtual and strong.
Metaverse Core Achievements Exhibition and AI Innovation and Application Exhibition
02 Large public spaces to make the "immersive" future experience unimpeded
Let users discover and expand features, let participants experience and immerse themselves in the scene, and let ramblers think and explore the future. During WAIC 2023, a series of cutting-edge technology scenarios such as the "meta-universe" will continue to enter the "reality" at the Zhangjiang Science Hall.
Many of the companies exhibited were "repeat customers" of the last World Artificial Intelligence Conference, and this time they have brought "upgraded" AI products. Rich interactive application experience scenarios around the themes of meta-universe, artificial intelligence + art, intelligent robots and other themes, each of which is eye-opening. (For more metaverse "fresh" experiences, you can read more about today's two.) ）
Interactive app experience scenarios
Diversified integration and innovation activities such as global innovation project roadshows, open source bazaars, and talent bazaars provide important spaces to help cutting-edge smart sparks collide and multi-party resources continue to converge and link.
Global Innovation Project Roadshow/Open Source Bazaar/Talent Bazaar
The AI Book Club joined hands with three top publishing houses, People's Posts and Telecommunications Press, Tsinghua University Press, and China Machinery Industry Press, to bring a feast of ideas to the participants. In an inclusive and multi-dimensional open public space, nearly 1,000 latest books gather here, focusing on the latest trends and innovations in the field of artificial intelligence, and authors of a series of best-selling books also gather here to bring wonderful sharing, and there are on-site signing activities to promote reader exchanges.
03 Focus on cutting-edge dialogue, multi-space to gather industry strong voices
A top-level ideological feast where wisdom collides and inspires new ideas - the National Society Chairman Forum was held here.
National Society Presidents Forum
Focusing on the theme of "Digital Transformation of Integrated Cities in the Yangtze River Delta, the Latest Frontier Trends of Artificial Intelligence", the forum specially invited the presidents and vice presidents of national societies in the field of artificial intelligence and its interdisciplinary fields, as well as top academicians and experts in cross-fields, including nearly 20 academicians of the two academies and a number of senior experts and scholars. They come from different fields such as artificial intelligence, information, control, computers, mathematics, operations research, urban planning, transportation, civil engineering, etc., and represent the most cutting-edge ideas and ideas.
Artificial intelligence is at the forefront of scientific and technological development in today's world, and AI chips, as one of the core technical components to realize artificial intelligence, are constantly breeding the power of innovation. A chip theme forum called "From 'End' to 'Cloud', Bravely Climbing the Peak of 'Core'", brought together many corporate guests engaged in AI large chips, expressed their opinions on market expansion, scenario application, standard establishment and other fields, and jointly created the "core" movement. In addition, there are sub-forum activities such as "ITU Metaverse Focus Group Working Group Meeting", "AI Unbounded, Intelligent Chain Global" and "AI Empowering the Ocean - Intelligence: Endless Future and 2023 China Intelligent Ship Innovation Alliance High-end Academic Forum".
Chip theme forum
As the bearing space of the Zhangjiang branch of this conference, and also the main venue of Zhangjiang Science City's science and technology exchange activities, nearly 30 forums in Zhangjiang Science Hall were held at the same time during the conference. Including one Pudong Forum with the metaverse as the main axis and highlighting the achievements of artificial intelligence applications, driving industrial development with the construction of major application scenarios, and launching the "unveiling" of major application scenarios of the metaverse in 2023, creating a benchmark scene with strong influence; One session of the "Second Plenary Session of the Metaverse Focus Group" hosted by the International Telecommunication Union ITU, where celebrities from all walks of life shared their insights and experiences on the accelerated integration of the metaverse and the real economy, and discussed important topics such as international standards, key technologies, application implementation, and industrial ecology. The two thematic forums focused on technological innovation and application scenarios in the two frontier fields of "AI chip" and "intelligent driving", respectively, to show the world the AI R&D strength of science and technology powers. There will also be 20+ ecological forums, integrating a variety of technological elements, where guests will share their latest research results and practical experience, and discuss the most cutting-edge technology trends and innovation opportunities around the themes of 9 major application scenarios of the metaverse, AI-enabled urban digital transformation, developer vitality, and 365 days of open source.
Related links
Focus | World Artificial Intelligence Conference 2023: From the laboratory to the application field, AI shows new skills
The above content has been comprehensively sorted out by the small editor, for reference only, please refer to the latest information released by the official and the on-site situation.
end.
Article and picture source: Wen Wei Po APP, Zhangjiang Science Hall public account
If there is any infringement, please contact Xiaobian in the background to delete