xinhuanet financial observation | big model application: difficulties and breakthroughs
2024-09-18
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
xinhuanet beijing september 18 title: observation on the application of large models: difficulties and breakthroughs
reporters ran xiaoning, chen tingyu, ling jiwei, zhao qiuyue
there are too many big models but too few applications. big models need to be run and used... how to complete the "last mile" of big model application has become the focus of discussion both inside and outside the industry recently.
the "hundred model war" has begun, and it is not an exaggeration. at present, there are more than 190 large models registered with the state internet information office, with more than 600 million registered users. this year, all large models have been reduced in price, and some have even been reduced to free use. "don't go for models, go for applications", a well-known company has expressed this view many times, reflecting that the competition in the large model track has entered a new stage of ecology and "customer acquisition".
from the flourishing of a hundred flowers to the implementation of scenes, what challenges does the application of big models face, and how to solve the problems and activate the ecosystem? recently, the reporter interviewed big model development companies, design and research and development institutions, manufacturing companies and terminal manufacturers to explore the stories behind how big models promote research and development, improve production quality and efficiency, and enable c-end to improve user experience.
big model reconstructs the relationship between scientific research and production
we are currently at a critical stage where ai is reshaping productivity and production relations for scientific research. since it was first proposed in 2018, ai for science has reached a consensus in the academic community as a new scientific research paradigm, and ai has brought revolutionary impacts to the field of scientific research. the empowerment of new technologies has improved scientific research efficiency, promoted original scientific research innovation, and demonstrated the great value that artificial intelligence has brought to scientific research. ai large models have led to fruitful results for scientific research teams.
tian zhongqun, an academician of the chinese academy of sciences and honorary director of the tan kah kee innovation laboratory, said that ai has created possibilities for theoretical calculations. the improvement of algorithms and the increase in computing power have greatly improved computing efficiency. he used an analogy to say, "doing basic scientific research is like being on an isolated island with only cliffs and no paths. now, ai provides a tool that can help researchers have the opportunity to ride the waves and climb the cliffs."
in the field of electrochemistry, scientific research needs to solve the problem of new energy industrialization. for example, if a new energy storage power station, especially a large-scale energy storage power station, has a safety problem, the consequences will be very serious. ai technology for electrochemistry provides more guarantees for energy security and can better detect, control, feedback and control. for another example, in a battery energy storage system, the collection, processing and feedback of parameters involve a huge amount of data. relying on traditional manual processing methods, the fastest time is calculated in "days". but for ai, a few hours, minutes or even seconds may be enough.
"ai helps scientists discover problems, analyze problems, provide feedback on problems more quickly, and conduct proactive control to form a closed loop, effectively improving safety and efficiency," said tanaka gun.
protein is the material basis of all life activities and can be called the "crown jewel" of molecular biology. based on the uni-mol molecular conformation model released by deepin, a variety of general capabilities such as molecule generation and property prediction can be realized. in the field of drug discovery, the virtual dynamics molecule generation method vd-gen developed based on uni-mol can directly generate molecules with high binding affinity in the cavity of protein targets.
sun weijie, co-founder and ceo of deepin, introduced that in recent years, deepin and its collaborators have launched a series of scientific big models and underlying engines, including the dpa molecular simulation big model, the uni-mol molecular conformation big model, the uni-fold protein folding big model, the uni-rna gene sequence big model, the uni-dock high-performance molecular docking engine, and the uni-smart scientific literature multimodal big model.
deepin science's large-scale scientific model system (photo provided by deepin science and technology)
there are about 200 million known proteins on earth, and each protein has a unique spatial structure. after a long process of life evolution in nature, protein molecules can spontaneously complete the entire folding process in an instant. however, if scientists want to predict their folding method by calculating the interactions between amino acid molecules, they have to exhaust all possible protein configurations, which will take more time than the age of the entire universe.
"now, by using ai technology, the three-dimensional conformation of proteins can be accurately calculated in a very short time. scientists are also further exploring the use of ai to design proteins that do not exist in nature or to modify existing proteins according to specific functional requirements." professor jinbo xu, computational biologist, founder and chief scientist of molecular heart, said, "predicting protein structure through ai has greatly improved people's understanding of proteins, understanding how proteins perform their biological functions, and recognizing the interactions between proteins and non-proteins, so that people can better understand the molecular processes of life, which has an important impact on fields such as biology, medicine and pharmacy. for example, through ai protein structure prediction, accurate protein targets can be found more quickly, helping drug developers design more effective drug molecules."
xu jinbo believes that ai big models are particularly suitable for solving problems in life sciences. "the most successful example of the combination of computing and biology is ai protein structure prediction, which is the greatest contribution of ai to biology and even to the entire scientific community so far. but this is just the beginning and it is far from over."
neworigin (darwin), an ai protein generation model independently developed by molecular heart, is an ai protein basic model that integrates sequence, structure, function and evolution. it has learned a large amount of highly professional and complex multimodal data, and can "customize" functional proteins on demand according to industrial application needs. at present, the neworigin model has been widely used in innovative drug research and development, materials, food, chemical industry, agriculture and other fields, and has made breakthroughs in many types of difficult industrial tasks such as macromolecular drug design, protein stability optimization under extreme environments, enzyme activity optimization, enzyme-specific substrate docking, protein de novo design, etc.
ai big models are helping scientific research from earth to the distant deep space, playing an important role in deep space exploration. ouyang ziyuan, an academician of the chinese academy of sciences and the first chief scientist of china's lunar exploration project, said that with the rapid advancement of human deep space exploration activities, exploration data has grown explosively. in terms of data management, my country has already gained a first-mover advantage; in terms of data application, my country must give full play to its existing artificial intelligence technology advantages.
in addition to the samples returned by detection, human research on the geological evolution of the moon mainly relies on the research of lunar geological objects such as crater identification. the size, depth, shape and other characteristics of the crater are important bases for studying the geological evolution of the moon. according to statistics, there are currently more than 1 million lunar craters with a diameter of more than one kilometer on the moon, and the number of craters with a diameter of less than one kilometer has not been determined so far. if it is completely dependent on manual labor, it is "almost impossible" to complete the identification of all lunar craters.
at the 2024 big data expo, the institute of geochemistry of the chinese academy of sciences and alibaba cloud jointly released the world's first "lunar science multimodal professional large model" (abbreviated as "lunar professional large model"). the large model is built based on alibaba cloud's tongyi series of models. at present, the accuracy rate of lunar crater age and morphology has reached more than 80%. the application of the lunar professional large model has greatly improved scientific research efficiency: researchers only need to input lunar crater images and related questions, and the lunar professional large model can call tongyi vision and multimodal models to determine the modal type corresponding to the image from 17 types of multimodal data (including spectrum, elevation, gravity and other data).
the implementation of large scientific research models requires understanding of the underlying scientific laws
although ai big models have brought huge value to scientific research, they still face a series of related scenario challenges in the implementation of scientific research big models in downstream industries.
xu jinbo believes that in addition to the necessary basic conditions such as algorithms, computing power, and data, the development of large protein generation models also requires two major professional advanced capabilities: first, the integration of multiple disciplines such as computer science, biology, and physics, familiarity with various methods such as ai, molecular dynamics, and quantum computing, and the ability to consider the cross-domain integration of sequence and structure, main chain and side chain, evolution and omics in parallel in practice. the second is the ability to step out of the laboratory and sink into the real industrial environment, and be close to the real industrial needs in terms of demand, verification, and implementation. at present, talent teams with these capabilities and conditions are very scarce.
sun weijie believes that only by allowing ai to first learn the scientific laws and mathematical distribution of the microscopic particle universe can we try to solve important problems in the microscopic world.
"the existing algorithm system and model system bring an opportunity to truly reconstruct the world from the atom. the explosion of ai has brought scientific research a systematic opportunity to reconstruct production tools, productivity and production relations. starting from atoms, software, data, representation and the final manufacturing link will all be reconstructed. although the prosperity of the digital world has promoted the development of large models, what we need to pay more attention to is the long-term outcome, and the outcome of ai's reconstruction of scientific research productivity will be to move towards the era of 'intelligent atomic manufacturing'," he said.
"the physical world is made up of microscopic particles, and what we are concerned about are molecular structures, protein gene sequences, molecular simulations, etc. at the microscopic level. these new modalities cannot be covered and processed in the classical large models, and the large models of the past have not been able to truly understand the underlying scientific laws of all things in the universe." sun weijie said.
he further elaborated that it is necessary to let the ai big model understand the universe of microscopic particles and try to solve the problems of the microscopic world, but there is often a lack of very effective data because the microscopic world of particles is invisible and intangible. "in the microscopic world, the best and proven method at present is ai for science. ai for science is opening up a new scientific research paradigm, closely combining artificial intelligence with basic scientific research, and giving ai the ability to understand the microscopic world. the three pillars of the big model of the physical world, the big model of the digital world, and the big model of embodied intelligence constitute the three main tracks for work and entrepreneurship in the field of ai today. ai for science is one of the three pillars of ai and the only way to agi." sun weijie said.
big model empowerment manufacturing needs to deepen integrated application
ai big models are gradually penetrating into all aspects of the manufacturing industry, becoming one of the core technologies for the manufacturing industry to become intelligent, flexible and automated, bringing new opportunities to the manufacturing industry.
the government is actively promoting the specific application and innovation of big models in the field of production and manufacturing. in april this year, the science and technology department of the ministry of industry and information technology proposed to promote the widespread use of artificial intelligence in the production and manufacturing links, and emphasized "taking the deep integration of artificial intelligence and manufacturing as the main line, laying out general big models and industry big models, and accelerating artificial intelligence to empower new industrialization."
guided by this trend, many industries such as home appliances, automobiles, and chemicals have explored the practical application of large ai models.
injection molding is an important process in the production of washing machines. its production process seems to be nothing more than the opening and closing of the mold, but behind it are complex processes and parameters such as temperature, pressure, molding cycle, mold health, energy consumption, etc. in the past, it could only be debugged by manual experience. today, the "black box" problem of the injection molding process has been solved.
walking into haier tianjin washing machine interconnect factory, you can see that the tianzhi industrial big model independently developed by cosmoplat transforms the industrial experience of injection molding masters into quantifiable data and indicators. the relevant person in charge said that through the reasonable adaptation of the injection molding big model and the expert model, the overall energy consumption of the injection molding machine is optimized and reduced by 6%-10%, and the production cycle is increased by 5%-12%. it is understood that the tianzhi industrial big model can read industrial language, understand industrial processes and mechanisms, generate industrial execution instructions and execute industrial machinery control. it has been applied in haier's washing machine interconnect factories in tianjin and foshan.
the picture shows the injection molding equipment of haier tianjin washing machine interconnect factory (photo provided by kaos)
"the industrial scenarios and processes in automobile manufacturing generally have pain points such as high glue utilization rate, high labor costs, the need to optimize production scheduling, and the difficulty in manually achieving optimal process scheduling in process design." at the 2024 "generative artificial intelligence + automobile" supply and demand matching and results transformation event held by the china industrial internet research institute, many automobile companies mentioned the series of pain points faced by the industry.
"generative ai technology helps to automatically arrange welding processes and program assembly engineering, thereby obtaining the optimal load balance and improving the efficiency of process documentation," said zheng chunqi, deputy general manager of gac aion new energy automobile co., ltd. ye tongsheng, automotive industry director of cloudwalk technology group co., ltd., further added that the analytical capabilities of large models can monitor the quality of vehicle production throughout the entire process.
in terms of intelligent petrochemical industry, intelligent industrial equipment and efficient experimental research, the intelligent prediction technology of catalytic cracking settler coking solves the coking problem of catalytic cracking settler. according to deng chun, a professor at the national key laboratory of heavy oil of china university of petroleum, generative artificial intelligence promotes the application of intelligent operation of industrial equipment in the chemical industry through material perception, reaction mechanism, core equipment, process optimization and system optimization.
however, the implementation of ai big models in manufacturing is not smooth sailing. the survey found that the amount of data in manufacturing companies is huge and fragmented, making it difficult to effectively aggregate and maximize the value of data. in addition, the computing power cost, deployment cost, and trial and error cost of big models are high, and there is a particular shortage of industrial ai technical talents.
"the implementation of big models in manufacturing enterprises requires the joint efforts of algorithm engineers, data engineers, and front-line managers of the enterprise, especially investing a lot of energy and time in data collection and labeling, model fine-tuning, and process optimization." sun linjun, vice chairman of the hainan artificial intelligence association, member of the digital committee of the zhejiang federation of industry and commerce, and founder of ai technology company real intelligence, said that in the face of the complex supply chain of manufacturing enterprises, the real agent intelligent body can use rpa to integrate and optimize data resources and low-cost alternative interfaces; with the help of big models, the threshold for data use can be lowered, complex business operations can be decomposed through reasoning, and rpa can be scheduled to automatically complete business operations, and then by reasonably controlling the model deployment cost and computing power resources, it can be applied to various business links and business processes such as raw material procurement, inventory management, production planning, and logistics distribution, and play a good role in reducing costs and increasing efficiency.
despite facing many challenges, the vast majority of the companies interviewed firmly believe that the application prospects of ai big models in manufacturing are still broad.
xu xiaolan, president of the china electronics society, suggested that we should give full play to my country's advantages such as a complete industrial system, large industrial scale, rich application scenarios, and abundant engineering talents, focus on consolidating the foundation of ai technology, deepening the integrated application of ai, and improving the industrial development ecosystem, deepen the empowerment of artificial intelligence technology, accelerate the development of new quality productivity, and promote new industrialization.
the introduction of large models into terminals promotes the large-scale popularization of ai
if the cloud-based big model demonstrates the powerful technical capabilities of ai, then edge-side ai is the vehicle for accelerating the popularization of the benefits of ai technology.
why is edge ai so popular? cicc released a research report stating that edge ai is the next stage of ai development. by empowering terminal hardware with large models, it is expected to start a wave of ai applications.
judging from the layout of terminal manufacturers, oppo proposed to equip about 50 million users' mobile phones with ai functions this year, lenovo is fully promoting the aipc "one-in-one multi-terminal" smart terminal strategy, xiaomi su7 is equipped with an ai big model, and changhong's multiple series of tvs are equipped with the changhong yunfan ai big model... it can be seen that the entry of big models into the terminal side has already shown an accelerating trend.
the picture shows the distribution and positioning of large models on the terminal side (source: iresearch)
putting large models into small terminals has unlimited prospects but also challenges. in the view of zhang peng, ceo of zhipu ai, how to make the model smarter based on a lower transmission volume and how to provide unique resources on the mobile terminal to support the operation of large models are major challenges encountered when large models enter terminals.
zhipu ai, which was established only five years ago, has now become a rising star in the field of model development. when talking about how to create space for the implementation of large models, zhang peng believes that cooperation is the key. "on the one hand, we need to do a good job of self-developing large models. on the other hand, hardware manufacturers, model algorithm manufacturers, operating systems and other ecological technology manufacturers need to work together." zhipu ai has collaborated with intel, qualcomm and others to enable large models to run on various terminals such as pcs, mobile phones, and cars.
"turn off the front camera", "mute the computer volume", "generate a summary based on my reading habits"... consumers have found that it is easier to use a brand's new computer. a simple conversation with the computer can replace the previous complicated operations. according to the manufacturer, with the iteration of future versions, it will even be able to help users realize more difficult functions such as editing and sending emails, making personalized posters, and understanding the meaning of a picture.
"in 2022, lenovo started planning to place the big model locally," said zheng aiguo, general manager of lenovo group's global sme products and solutions. at that time, the parameters of the big model were too large to be carried locally. until last year, when big model development companies focused on 6-7b, lenovo began to realize that the implementation of the big model local plan was easier. in may last year, lenovo combined the end-side ai big model with the pc project.
wang bin, director of xiaomi group's ai lab, believes that on the one hand, models are getting smaller, while on the other hand, computing power is increasing, hardware capabilities are becoming stronger, and coupled with various demands, it is expected that some killer ai applications will appear on the terminal side.
the reporter sorted out the ai mobile phone products sold on an e-commerce platform and found that voice, image and ai assistant are the three most concentrated functional points of ai mobile phones. at present, mainstream mobile phone manufacturers are actively deploying lightweight large models to make ai unconscious and vigorously explore ai application scenarios. market research reports predict that ai mobile phones will show great value prospects in personal intelligent assistants and improving office capabilities. it is becoming a reality that each terminal has its own built-in ai assistant, which is seen as an ai future that everyone can enjoy.
ai image recognition, voice wake-up, computational photography and other edge ai use cases may seem simple, but they actually have very strong requirements on the chip's computing power and ddr bandwidth.
the rapid growth of ai use cases on smartphones highlights the importance of qualcomm. "qualcomm incorporated the concept of ai into the entire soc from the very beginning of the design," said wan weixing, head of ai product technology in china at qualcomm. in response to the challenge of large models to ddr bandwidth, qualcomm has developed quantization and compression technologies to reduce the size of the model. in response to the high computing power requirements of large models, qualcomm has made very professional designs on npu to meet the diverse needs of different use cases.
in addition, hundreds of models are in full bloom, and the large models on the end side are further subdivided into many categories. however, the communication capabilities between models on different terminals have not yet been established. therefore, sun mingjun, director of the zhongguancun zhiyong artificial intelligence research institute, believes that "in the future, the unification of technical standards and interoperability will be a very large technical problem that needs to be solved by all parties."
"application is king", this concept is particularly important in the development and implementation of large model technology. during the survey and interview, industry insiders generally agreed that the real value of large models lies in solving practical problems and creating real value for users.
as the "ai+" action advances in depth, we expect that with the collaboration of multiple parties, big model technology will ride on the "flywheel effect", empowering thousands of industries and accelerating the formation of new quality productivity, while also feeding back technology iteration and performance improvement in continuous implementation, opening up a new realm for big model applications.
