For more than a year, big model technology and applications have been developing rapidly.Whether it isOpenai, Google, and Baidu, Tencent, bytes, etc. The big model products have been iterated many times, and the application is constantly expanding.And next, what development trend will the big model present is still the focus of the industry's attention.A few days ago, at Taihu Dialogue · Artificial Intelligence+Forum, many academic and industry people deeply discussed related topics.
The five development directions of AI models
At the meeting, the Chinese Academy of Engineering, a professor of lecturer at Tsinghua University, and Intelligent Industry Research Institute (AIR) Dean Zhang Yaqin talked about the five directions of the development of AI models.The first is multi -mode intelligence, that is, combining language, text, video, laser radar point cloud, 3D structure information, 4D space -time information, and even biological information to achieve multi -scale and cross -modal intelligence perception, decision -making and generation.
The second is independent intelligence, which is about to be used as a tool. Development can independently plan tasks, write code, mobilize tools, and optimize paths to achieve high self -iteration, upgrade and optimization, and realize autonomous intelligence.Here we need to consider how to call each other between the models, how to use the tools, and how to learn between the basic models.
For example, the simulation hospital of the independent evolution of medical intelligence, they first builtLLM Agent's simulation hospital includes two types of roles and eight key processes of doctors and patients.Proposal MEDAGENT-ZERO medical intelligence independent evolution strategies that do not rely on artificially label data.There are two ways of self -evolution: accumulating samples from correctly, and summarizing experience from the wrong answer.After diagnosis of tens of thousands of patients, the medical intelligence body diagnosis is equivalent to doctors for two years for two years. Medicine can only achieve the best on the Medqa subset.
The third is marginal intelligence, deploying large models to the edge device end, such asAI PC, AI mobile phones, AI TVs, etc., to achieve high -efficiency, low power consumption, low cost, low delay treatment and response, thereby achieving marginal intelligence.
AIR and Yaxin have both cooperative cases in this regard, such as video analysis and optimization of edge all -in -one video, using motion vector, incremental movement identification box, significantly improve the target tracking performance; large model reasoning performance optimization, based on runtime parameter cache mechanism, Intelligent and efficiently managing the MOE model Experts for replacement, reducing memory demand & gt; 50%; 5G dedicated network computing power internal technology, using the base station BBU idle computing power to provide services. Based on the process of seizing the process, the BBU calculation can be effectively improvedRecycling efficiency of power resources.
The fourth is physics (body) intelligence: big models are being used in unmanned cars, robotics, factories, transportation, Communication, grids, power stations, and other physical infrastructure improve their level of automation and intelligence, thereby achieving intelligence.Zhang Yaqin believes that when it comes to physical intelligence, robots will be the key landing direction of large models. Unmanned driving is actually another form of robots.Driving and so on.
The fifth is biological intelligence. It applies large models to human brains, life, and organisms, to realize biological intelligence with large models and biological connections, and finally realize the integration of information intelligence, physical intelligence, and biological intelligence.
likeTypical cases of AI+Life Health, Brainco Brainwood Interface Technology, incubated in the Harvard University Innovation Labs, which are creative by the Chinese team, including biomedical engineering, material science, brain science, brain diseases, and AI interdisciplinary teams.Entry -type brain machine interface mass production products appear on Time Weekly NASA official website.It can be applied to autism, sleep disorders, closed -loop mindfulness meditation, depression, Alzheimer, etc.
In addition, in the roundtable forum, many industries have made detailed discussions on the development trend of large models.Zhou Jingren, the chief technical officer of Alibaba Cloud Intelligent, said that a trend of a very obvious model is multi -modal. Now there are more and more knowledgeable bodys.It is the key to blending all kinds of knowledge together.
In his opinion, what needs to be considered is how to make large models more toward human intelligence, and also requires memory, how to combine long -term memory and short -term memory.In this way, in the future, big models can develop into more individual and distinctive products, and can have different thoughts in different fields.At the same time, the big model has more interaction with the real world, and the relevant information will be fed back to the model, which allows the big model to develop the ability of self -renewal and optimization.
However, as far as the current maturity of large models is concerned, it remains to be studied.Zhou Jingren talked about itIn the IT era, the tentacles of information can be available. After the big model, the tentacles of knowledge can be within reach.Now, a very clear point is that as an assistant, the big model can provide various counseling opinions, but the final decision is still people. Now this big model of the assistant has been done very well. This is a big model. This is a big model.The first step in development.
The second step is to make the big model more accurate. There are still many questions about the answers of the big model. How to make it do ten times and ten times is accurate. Zhou Jingren believes that this still has room for optimization.Furthermore is to allow the model to learn and evolve yourself. Of course, this step is still under research.
The dean of the Macau Industrial Technology Research Institute and the partner of the capital capital management told Jiang Xu that the multi -mode and large model isAGI feasible road.He believes that how to influence the big model and change the physical world is a trend, such as making intelligence (robot, etc.) possible, which will also make future pension more convenient.
Lecturer of Huayan University of Tsinghua University,Ma Weiying, chief scientist of Air, talked about a very new angle.He believes that the focus that everyone is focusing on now is that the big models can be seen, listened, can be written, etc. The opportunities in these directions are mainly in Internet factories, such as Google, Microsoft.
So, do you have new opportunities, not just to generate articles and pictures.What Ma Weiying is considering is that he hopes that it is not just the key to imitating human beings. He believes that there are many opportunities to explore biological and natural fields, such as biology, chemistry, materials, etc.In his opinion, AI will have huge opportunities in these areas in the next ten years.
Today, as a tool assistant, the big model has very powerful functions and has a very mature application.Now the big model has more intelligence, such asGPT-4O can already express feelings in human interaction.And further, to achieve independent learning and self -performers, that is, independent intelligence, it still needs continuous research.In terms of the opportunities brought by the big model, in addition to the text, pictures, video generation and other aspects of the major Internet factories, for new enterprises, exploring from the fields of nature, such as biology, chemistry, materials, etc., may see that you may watch it.New opportunities.
For more electronic components requirements,pls see:https://www.megasourceel.com/products.html