On July 7, the 2023 World Artificial Intelligence Conference-Generative AI and Big Model: Change and Innovation Forum, hosted by Qiming Venture Capital, was successfully held at the Shanghai World Expo Exhibition & Convention Center. Renowned experts and scholars, top investors and leading entrepreneurs in the frontier fields of artificial intelligence and other technologies gathered together to explore the driving force of China's scientific and technological innovation.
Based on the perspective of entrepreneurship and investment, this forum shared the latest generative AI cutting-edge information and practical exploration, and released the industry's first generative AI insight report based on in-depth research on more than 100 enterprises at home and abroad, comprehensively interpreting the industrial transformation of generative AI and the ten prospects for the future.
The big models that lead the new generation of AI are exerting great influence, moving from research labs to industry and gradually into our lives.
As one of the earliest and most active venture capital institutions in the field of AI in China, Qiming Venture Capital joined the World Artificial Intelligence Conference for the first time to host this forum. From AI 1.0 to AI 2.0, after more than ten years of deep cultivation and cutting-edge insights, Qiming Ventures' investment companies are all over the AI industry ecological chain - from chips, underlying basic software, basic models to downstream vertical applications, many companies have become leaders in their fields.
In the opening speech, Zhou Zhifeng, Partner of Qiming Venture Capital, delivered a speech entitled "Generative AI and Big Model Lead Industrial Transformation and Innovation". "Big models and generative AI are leading the biggest wave of technology today, with clear signals from technologies and products such as Stable Diffusion, ChatGPT, Midjourney, and Copilot showing explosive user adoption and historically high per capita output from the companies behind them," he said. The 'first principles' of the big model are to massively compress all the digital information in the human world. The accumulation and development of computing power, algorithms, and data over the past 40 years have driven the explosion of generative AI today, and enterprises at the infrastructure, model, and application layers are all given the opportunity to grow rapidly. The new wave of AI brings new product forms, and models are not only technical infrastructure, but products that reach users - users + models = applications. Companies with +AI and AI+ models from the perspective of venture capital are ushering in rapid development, but in the face of the boundary expansion of industrial giants and large-model companies, startups need to find their own golden channel. Qiming Venture Capital identified the signal that AI 2.0 has crossed the technical singularity, and laid out a number of large model companies in advance before the market tipping point, and will continue to witness the rapid development of generative AI with you in the future. ”
Pre-trained large model is the infrastructure of a new generation of artificial intelligence applications, and training high-precision 100 billion bilingual dense models in Chinese and English is of great significance to the research and application of large models. Professor Tang Jie of Tsinghua University talked about the development process of AGI in his speech, saying that since 2019, the large model of ultra-large-scale knowledge graph and ultra-big data has provided core support for the future of cognitive intelligence. After continuous research and development, the 100 billion pedestal model trained in 2022 provides a solid foundation for subsequent research and development work. Among them, the ChatGLM-6B model has been downloaded more than 3 million times since it was open source, and it has been rated as the first place in the open source influence of large models by the "Chinese Intelligent Large Model Map Research Report". In the future, we will continue to explore the frontiers of cognitive large models, and gradually unlock the capabilities of In-context, environmental interaction and self-optimization, and self-reflective learning.
Large models and multi-modalities significantly drive the demand for computing power infrastructure. Guantai Ye, Partner of Qiming Venture Capital, and Dr. Gang Liang, Partner of Wall Cheng Technology, discussed how high-performance general-purpose GPUs can support the training and inference of large models in China, and demonstrated the excellent strength of China's core training and inference.
Liang Gang pointed out: "The BF16 computing power of the BR100 chip of Walltech reaches 512 TFLOPS, which is leading in the industry, and supports TF32, BF16, INT8 and other data accuracy, of which the computing power of INT8 can reach 1024TOPS. In addition, in terms of bandwidth, the BLINK independently developed by Walltech can support single-machine 8-card interconnection, and realize multi-machine multi-card high-speed interconnection through IB network card. The software ecosystem of SUPA, which has supported several large models and training and inference frameworks, and adapted to multiple large model algorithms. There are no shortcuts in the semiconductor industry, so Walltech will continue to focus on product development and cooperation with users, do things steadfastly, and finally speak with products. Listening to customer needs is the top priority, in terms of hardware, Walltech will focus on computing power, video memory, bandwidth, cluster interconnection and security needs; In terms of software, Walltech will form a 'application-optimization-feedback-iteration' cycle with partners and customers to continuously improve itself. ”
Ye Guantai pointed out: "In the era of large models, computing power is indispensable for training models with hundreds of billions of parameters and trillions of parameters, and the rapid growth of large model parameter scale has continuously put forward higher requirements for GPU capabilities. The GPU with large computing power needs to support a variety of models, need to have stability, scalability, delay control, cost performance, etc., showing a typical 'barrel theory', and the promotion of large computing power chips requires the entire upstream and downstream ecological support. The era of AI has arrived, and the close cooperation between chip companies and large model companies will surely usher in a brilliant future. ”
Subsequently, on behalf of Qiming Venture Capital, Zhou Zhifeng released the "Generative AI" report on the cooperation between Qiming Venture Capital and Unfinished Research, and conducted in-depth interpretation. The report comprehensively presents a series of topics that have attracted much attention, such as the overall framework of mature AI, the industry map, the pricing model of generative AI and large-model enterprises, the global AI talent distribution, advantageous fields, and the must-read papers jointly recommended by the world's top AI practitioners, as well as the top ten prospects in big language model, multimodal model, and business development.
The report pointed out that 2023 pushed the large model to a peak, marked by the release of GPT-4, generative artificial intelligence in the direction of general artificial intelligence, into the stage of innovative application. The infrastructure, model, and application layers will form the new architecture for generative AI. Currently, generative AI is in the early stages of the market, and the most important feature of this stage is that research, application and regulation work together to open up the innovation path of generative AI. (Please pay attention to the official account of Qiming Venture Capital and enter "generative AI report" in the background to get the full report download method.) You can also visit the official website of Qiming Ventures and download it through this link: https://www.qimingvc.com/sites/default/files/State_of_Generative_AI_2023.pdf)
As a leader in the field of cognitive intelligence, Zhipu AI has injected the capabilities of large models into all walks of life by practicing the MaaS concept. Dr. Zhang Peng, CEO of Zhipu AI, delivered a keynote speech on "Cognitive Big Model and Application Preliminary Study", saying that GPT-3 released in 2020 has brought large models into the "usable stage", and Zhipu AI is also fully engaged in the research and development of large models in the same year. In June this year, Zhipu AI upgraded the 100 billion model ChatGLM to the second generation, launched models with more parameter scale, and redesigned the API product matrix to adapt to the needs of a wider range of application scenarios. At present, Zhipu AI provides services such as model private deployment, API call, and model customization, which can provide full-stack solutions for ultra-large enterprises, medium and large enterprises, and small enterprises. In addition, Zhipu AI has also launched an integrated platform for training and promotion to help users quickly use private data to use models, and provides full-link AI application production tools to help users complete the development of exclusive AI applications in one stop.
In the era of generative AI and large models, what new possibilities will humanoid robots create? Dr. Youjun Xiong, co-founder, chief technology officer and executive director of UBTECH, introduced that humanoid robots are ushering in the "iPhone" moment in his keynote speech on "The Development and Future of Humanoid Robots". Humanoid robots are more suitable for human environments and provide more natural human-computer interaction to meet human emotional and companionship needs. The evolution and development of AI technology is the key to promoting the intelligence of humanoid robots, and the combination of humanoid robots and ChatGPT has pushed AI to a new height of embodied intelligence. ChatGPT empowers humanoid robots, making them say goodbye to developed limbs and simple minds; Humanoid robots will extend ChatGPT's capabilities to help it transcend text processing and information space and interact with the physical world. The humanoid robot market is currently blooming and will be more prosperous in the future. UBTECH developed a small humanoid robot in 2012 to a large-scale humanoid robot Walker in 2016, and Walker has undergone 4 iterations so far, making it the first large-scale humanoid robot in China to achieve commercial landing. The company is also one of the very few companies in the world with humanoid robot full-stack technology, and relies on the core technology of humanoid robot to realize the commercial application of multiple scenarios.
With the launch of ChatGPT, the paradigm of artificial intelligence is shifting. As a leader in the AI 1.0 era, Yunzhisheng recently officially released the Shanhai model with 60 billion parameters, handing over its first answer sheet in the AGI field. Liang Jiaen, Chairman and CTO of Yunzhisheng, shared the power of generative AI to empower the future in his speech. Liang Jiaen pointed out that large models have ushered in a new era of AI and opened a new paradigm of "data dynamics": pre-training, instruction fine-tuning and feedback reinforcement. The integration of big data into big models will become a power plant in the AI 2.0 era, and will drive the upgrading of core elements of the industry such as talents, technology, and markets. Yunzhisheng's self-developed Shanhai model has strong general capabilities, and focuses on the construction of related capabilities such as plug-in expansion and domain enhancement, and empowers customers around industries such as smart IoT and smart healthcare. In the future, Yunzhisheng will continue to explore the limits of intelligence and lead industrial upgrading.
The influence and value of the open source community in the AI 2.0 era is self-evident. This forum specially invited Wang Tiezhen, head of Hugging Face's China region, and Zhou Jiangong, founder of Weijian Research, to have an in-depth discussion on how the power of open source can promote the development of generative AI. Wang believes that the new wave of AI 2.0 (generative AI) models shows great versatility compared to AI 1.0 models, allowing every user to use them, not just AI tools used by engineers. Although closed-source models such as ChatGPT and GPT-4 are recognized as the best models, they have problems such as data security, deployment cost and controllability, and Hugging Face's 100 billion parameter open source big model BLOOM provides users with the opportunity to solve the above problems and make large models and AI universal. In addition, Hugging Face thrives on an open source community that provides developers around the world with ideas and techniques to share. Zhou Jiangong expressed surprise that the global model open source community such as Hugging Face has only 200 employees, and said that the industry needs to brainstorm how AI can benefit more people and industries, and how to ensure the credibility and security of large models in the development process.
As the world's leading artificial intelligence company dedicated to linking consumers and goods, Zhiyuan Technology has developed its own AI large model, tens of billions of parameter scale and unique training to make the large model better at understanding people and goods on the basis of universal capabilities. Dr. Zhou Bowen, founder of Zhiyuan Technology and Huiyan Chair Professor of Tsinghua University, shared how to empower enterprises with a unified large model base in their 5D innovation capabilities. Zhou Bowen pointed out that generative AI is the only way for AGI, and ChatGPT has opened a new inflection point in artificial intelligence, and AI has changed from "competing with people" to "collaborative interaction" to help humans solve problems. Zhou Bowen said that the deepening of the application of large models will return to the essence of business, the integration of AI and consumer industry will develop into a future trend, and a unified large model will drive enterprises to transform from digitalization to intelligence. ProductGPT, the large model base created by Zhiyuan Technology, will reconstruct the binary relationship between products and users, empower the whole life cycle of enterprise 5D (opportunity insight, explosive product definition, scheme design, drive R&D, marketing transformation), and provide employees with productivity tools in different links such as knowledge management, market insight, and customer delivery.
Driven by generative AI, robot technology innovation is also accelerating industrial digitalization. The landing of large models in the manufacturing field needs to interact with the physical world, and what difficulties and breakthroughs will arise in this process. Chen Nan, Executive Director of Qiming Venture Capital, and Shao Tianlan, Founder and CEO of Mecamand Robotics, discussed the theme of "Embodied AI Embodied Intelligence Promotes Industrial Digital Upgrade". Shao Tianlan pointed out that humans have been trying to make robots smarter and have adopted many challenging technologies. Multimodal large models have the opportunity to integrate the basic capabilities of robots more intelligently to achieve a higher degree of intelligence and give robots the possibility to combine more and more complex actions, while reducing development and programming costs. The large model has brought new impetus to the further improvement of China's industry, and can combine a lot of new products and new applications under the existing conditions. China has more advantages in the development of the new wave of robot industry, including talent base, industrial chain foundation and customer base.
Chen Nan said that when Mecamand Robot was established in 2016, it was determined to use AI + 3D vision to promote the ubiquity of intelligent robots, and after several years of efforts, it can be said that the achievements have been remarkable. With the popularity of large models, the concept of Embodied AI embodied intelligence has once again received great attention. Embodied intelligence emphasizes end-to-end intelligence through strong AI capabilities and strong coupling with machine control, capable of handling more complex and ambiguous tasks. The new AI+3D technology empowers the automation of the heavy industry industry, helping China's industrial digital upgrade, and even the level of the entire industrial field.
— End —
Qubits QbitAI · Headline number signed
Follow us and be the first to know the latest scientific and technological trends