On September 5, the 2024 Inclusion Bund Conference officially opened in Shanghai. At the main forum of the opening ceremony, focusing on the topic of large-scale model landing industry, MiniMax founder and CEO Yan Junjie, Baichuan Intelligence founder and CEO Wang Xiaochuan, Ant Group President and CFO Han Xinyi and Geek Park founder and president Zhang Peng On-site conversations were held, and the latest practices and thinking in the industry were shared. They believe that application breakthroughs are the key to the development of China's general artificial intelligence technology (AGI).
18 months is usually an important milestone in the business field, and AI technology has also made breakthrough progress during such a cycle. In Yan Junjie’s view, there have been two main changes in model capabilities in the past two years—improvement in effectiveness and reduction in cost.
From GPT3 to GPT4, the intelligence level of the model has changed by an order of magnitude, and the cost of inference has also been reduced. The cost from GP4 to GP4O has dropped by more than ten times. This makes many applications possible, and the industry is paying attention to when a super APP will be born.
According to Wang Xiaochuan's recollection, when Baichuan Intelligent was established in April last year, he was very excited and thought it was the beginning of a new era. Baichuan was fortunate to be able to participate in the new technological changes as the first echelon. He does not agree with the "AGI bubble" statement and judges that the development of the large model is in line with industry expectations in terms of results.
"Last year, the industry was mainly building model foundations, and China was also following up to catch up with the huge market expectations to prove itself." Wang Xiaochuan said that he was still a little uncomfortable at the time, feeling that it was too fast, and felt like he was being pushed back, because every day When a new paper comes out, it is inevitable that there will be a huge amount of anxiety.
"Now I feel that in a normal and rhythmic breakthrough, we can continue to solve problems. This year, the industry began to explore reinforcement learning, and the technical field has taken a solid step." Wang Xiaochuan said.
However, in addition to technology, Wang Xiaochuan believes that if large models want to move to the next stage of development, the key is to achieve two major integrations: one is the combination of knowledge search and large models, and the other is the combination of large model enhancement. Currently, he feels that the combination of search and models is being explored at home and abroad, and the global AI development is lower than expected.
In this regard, Yan Junjie believes that in the future, the error rate of large models will be reduced by orders of magnitude, and infinite-length memory and multi-modality will be realized. These are the three most essential technical issues. If the model error rate can be reduced to single digits, it means that the reliability of future models will increase by an order of magnitude, and AI will have the opportunity to transform from an "assistant" status into a replacement or improvement of productivity, and the penetration rate will also increase.
The AGI in Yan Junjie’s mind is “Intelligence With Everyone”. When MiniMax was established more than two years ago, it had already produced its earliest models and products. However, the entire industry had not yet become popular at that time, and the model capabilities were not strong enough to solve the productivity problem, so we started with entertainment. As model capabilities improve, some productivity attributes begin to emerge. He believes that the industry may usher in greater changes in the future.
Han Xinyi observed from an industrial perspective that technology is now ahead, and applications have not yet fully caught up. In the future, the industry will need to keep up with technology at the application level so that the technology can be used. Secondly, create more technical problems in use to make the direction of technological development more focused and certain. “For AI to be well applied, it must be easy to use, useful and affordable.”
Ant will release three new AI products at this conference, including the AI independent App "Zhi Xiaobao" launched by Alipay. What Ant wants to do is to introduce artificial intelligence into people's daily life services and business environments, so that the AI ecosystem can operate sustainably in the long term.
In Wang Xiaochuan's view, when talking about AGI, we can refer to the "human" standard, that is, the ability to think about knowledge and provide services like a human assistant. Medical care is a market he focuses on. Baichuan Intelligent not only firmly bets on this market, but also regards "doctors" as a measuring stick to verify AGI. It believes that only after bringing in concrete scenarios, expectations for technology will become clearer and users will be able to have more experience. Efficiency and experience upgrades.
To sum up, the industry is optimistic about the future development of AI technology and looks forward to seeing more technological innovations applied in practice, thereby promoting the overall progress of business and society.