Generative AI and the Emergence of Big Model Entrepreneurship; General artificial intelligence is in the ascendant.

Trainee reporter Chen Yukang

The "Generative AI and Big Model: Forum for Change and Innovation" of the 2023 World Artificial Intelligence Conference hosted by Qiming Venture Capital was held in Shanghai on July 7. This forum is the only forum hosted by venture capital institutions at this World Artificial Intelligence Conference.

The reporter learned from the forum that with the explosive growth of computing power and data, artificial intelligence has entered the era of 2.0: large-scale pre-training learning based on massive data has greatly reduced the basic research and development costs of application threshold and AI industrialization, while generative AI can automatically generate original information content, synthesize data and physical object models, which enables AI to assist human work in a large number of actual scenarios and has the opportunity to generate a wide range of social values. Generative AI has also become the theme of the times that innovative entrepreneurs, investors and government departments pay close attention to.

Computing infrastructure is in an explosive period of growth.

Zhou Zhifeng, a partner of Qiming Venture Capital, delivered a speech entitled "Generative AI and Big Model Leading Industry Transformation and Innovation". He said that the big model and generative AI are leading the biggest new wave of science and technology, and the clear signals come from technologies and products such as Stable Diffusion, ChatGPT, Midjourney and Copilot, which show explosive user adoption and unprecedented high per capita output of the enterprises behind them. The’ first principle’ of the big model is to compress all the digital information in the human world on a large scale. The accumulation and development of computing power, algorithms and data in the past 40 years have promoted the outbreak of today’s generative AI, and enterprises in infrastructure layer, model layer and application layer have been given opportunities for rapid development.

"The new wave of AI brings new product forms. The model is not only a technical infrastructure, but a product that reaches users-user+model = application. Both +AI and AI+ companies are experiencing rapid development from the perspective of venture capital, but facing the boundary expansion of industrial giants and large model companies, start-ups need to find their own golden channels. Qiming Venture Capital recognized the signal that AI 2.0 has crossed the technological singularity, and laid out a number of large model enterprises in advance before the market tipping point, and will continue to witness the rapid development of generative AI with you in the future. " Zhou Zhifeng said.

Large-scale model and multi-modal have significantly boosted the demand for computing power infrastructure, and "China Core" has outstanding strength in training and reasoning.

Ye Guantai, a partner of Qiming Venture Capital, said that in the big model era, it is essential to train models with hundreds of billions of parameters and trillions of parameters, and the rapid growth of large model parameters has continuously put forward higher requirements for GPU capabilities. GPU with large computing power needs stability, expansibility, delay control, cost performance, etc. to support various models, showing a typical’ barrel theory’. The promotion of large computing power chips needs the ecological support of the whole upstream and downstream. The era of AI has arrived, and the close cooperation between chip companies and big model companies will surely usher in a brilliant future.

Liang Gang, a partner of Biqi Technology, introduced the company’s chip products. He said that the BF16 computing power of the BR100 chip of Biwa Technology reaches 512 TFLOPS, which is leading in the industry and supports various data precisions such as TF32, BF16 and INT8. Among them, the computing power of INT8 can reach 1024TOPS. In addition, in terms of bandwidth, BLINK, which is independently developed by Bijie Technology, can support single-machine 8-card interconnection, and can realize multi-machine multi-card high-speed interconnection through IB network card. The software ecosystem of SUPA has been built, which has supported several large models and training and reasoning frameworks, and adapted several large model algorithms.

After the introduction of ChatGPT, the paradigm of artificial intelligence is changing. According to Liang Jiaen, chairman and CTO of Yunzhisheng, the big model has brought 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 models with big data will become a power plant in the AI2.0 era and will drive the upgrading of core elements of industries such as talents, technology and market.

Humanoid robot welcomes "iPhone" moment.

The era of generative AI and large model also provides new opportunities for the upgrade of humanoid robots.

Xiong Youjun, co-founder, chief technology officer and executive director of Ubisoft Technology, said that humanoid robots are welcoming the "iPhone" moment. Humanoid robots are more suitable for human environment and provide more natural human-computer interaction to meet people’s emotional and companionship needs. The evolution and development of AI technology is the key to promote the intelligence of humanoid robot. The combination of humanoid robot and ChatGPT pushes AI to a new height of embodied intelligence.

"ChatGPT empowers humanoid robots to bid farewell to developed limbs and simple minds; The humanoid robot will extend ChatGPT’s capabilities, help it transcend text processing and information space, and have the ability to interact with the physical world. The humanoid robot market is now in full bloom and will be more prosperous in the future. " He said that Ubisoft developed a small humanoid robot in 2012, and invested in developing a large humanoid robot Walker in 2016. So far, Walker has gone through four iterations, making it the first large humanoid robot in China to be commercialized.

The development of robot industry in China has multiple foundations. Shao Tianlan, founder and CEO of Mecarmand Robot, said that human beings have been trying to make robots smarter and have adopted many challenging technologies. Multi-modal large model can organically integrate the basic capabilities of robots to achieve a higher degree of intelligence, and give robots the possibility of combining more and more complex actions, while reducing the development and programming costs. The big model has brought new impetus to the further promotion of China’s industry, and many new products and applications can be combined under the existing conditions. China has more advantages in the development of a new wave of robot industry, including talent base, industrial chain base and customer base.

The big model of subdividing the track has emerged.

During this forum, Qiming Venture Capital and Unfinished Research jointly released the "Generative AI" report. The report mentioned that in 2023, the big model was pushed to a peak, marked by the release of GPT-4, and the generative artificial intelligence moved towards general artificial intelligence and entered the stage of innovative application.

A big model of subdividing the track has emerged. According to the report, unlike startups in the direction of productivity tools that emerged rapidly in 2022, in 2023, a larger proportion of new companies focused on the innovation of underlying technologies. Large-scale model startups began to differentiate. While the general large-scale model startups were in the ascendant, many vertical large-scale model companies oriented to specific directions such as medical care, e-commerce, scientific research, industry, autonomous driving and robotics began to appear.

For the business opportunities that may emerge in the future, the report mentioned that within three years, the core driving force of subversive AI applications comes from the innovation of the underlying model, and the two cannot be decoupled, and the role of the model will be greater than that of product design. At present, the generative AI market is in the early stage of technology dominance, and there are opportunities for platform enterprises with a market value of 100 billion US dollars.

Computing power is still the scarcest resource at present, and it is also the most profitable resource. According to the report, computing power is the largest part of the cost structure of the big model, and the performance of GPU determines the pace of this emerging industry. With the progress of computing power and models, more start-ups are pouring in. They have grabbed the dividend of time, but they are also facing competition and possible giant crush-this is the blue ocean of start-ups and the reefs under the waterway.

In terms of regulatory trends, the report believes that China is encouraging the development of general artificial intelligence while rapidly introducing regulatory measures for generative artificial intelligence and soliciting opinions. Beijing, Shanghai and Shenzhen are the most ambitious first echelons, all of which have put forward more ambitious artificial intelligence research, innovation and industrial goals. The EU continues to take the lead in supervision and legislation, taking the lead in introducing GDPR (General Data Protection Regulation) five years ago. The United States is more concerned about the leading position of artificial intelligence technology and is forming a regulatory framework based on risk management.