news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

text｜zhao yanqiu and zhou xiangyue

edited by niu hui

the implementation of the big model in central state-owned enterprises is entering a new stage. at a conference in september, a senior industry insider told digital intelligence front that the implementation of the big model is no longer limited to a certain industry, but has blossomed in all walks of life. the depth and breadth of its implementation within each enterprise have also taken a big step forward.

industry observations show that some leading central soes have already completed the initial pilot implementation of large models and are gradually moving towards more core scenarios.since may this year, as the price war of large models continued to ferment, the implementation of large models has been further accelerated. a number of scenarios have begun to be explored and replicated on a large scale, and at the financial level, roi has turned positive."xie guangjun, vice president of baidu, told digital intelligence frontline.

in the second half of this year, the implementation of the large model will move forward again.in just two and a half months, the number of publicly available large-scale model-related winning bids in the market has exceeded the total number of winning bids in the first six months of the year.. the number of projects awarded in a single month has also entered the hundreds of stages. big models are accelerating their full bloom in all walks of life. among them, energy, finance, education, internet and other industries are particularly active.

under the wave of rapidly changing technology, industry exchanges are accelerating. in recent months, various ai conferences of all sizes have emerged in various places, and baidu is reportedly going to hold the 2024 baidu cloud intelligence conference on september 25. more exchanges and collisions of technologies are coming.

01

central state-owned enterprises have taken another big step forward

in 2024,china southern power grid has completed the procurement of several large-scale model-related projects, such as "key technology project of ai basic training facilities for power system of southern power grid general dispatching-research and application of nlp large model technology for power system", "design research and component development of safety training and reasoning function set for large models in power industry by southern power grid research institute in 2024". relevant teams are working with baidu smart cloud for joint innovation.

in the financial industry, a large state-owned bank only carried out a small-scale pilot with a few thousand people in the customer service assistant and counter assistant scenarios last year. this year, these functions have been officially launched in several major customer service centers across the country and promoted to front-line counter employees across the country. the daily active users of the large model have reached tens of thousands.

in the automotive field, a person from geely research institute admitted at a recent conference that "nowadays, if car companies go out to sell cars without large models on them, they are embarrassed to sell their products." large models are used in smart cockpits and autonomous driving, and also enable digital marketing, user operations and other scenarios. recently, they have begun to empower all companies under geely.

in the government affairs market, local governments have made large-scale model planning in combination with local industries.by the end of this year or the first quarter of next year, the intelligent computing center will be put into operation on a large scale, and then the combination of large models and local industries will be launched.. "a government official talked about the progress of the market. science, education, culture and health have begun to be applied. "some tertiary hospitals have even invested tens of millions to try out scenarios." a medical professional said that the directors are concerned about case generation, management of human, financial and material resources, and scientific research, which are closely related to service levels, management and scientific research levels.

ports, a major channel for import and export trade, have also been explored and laid out. "we have seen that many customers have written big models into their it plans for the next three years." hu wei, general manager of baidu intelligent cloud logistics and transportation solutions, told digital intelligence frontier that unlike the past few rounds of technological changes, this round "fewer people are watching and more are trying it out."

for example, shandong port, which ranks first in the world in cargo throughput, has completed pilot exploration of big models in intelligent question and answer and intelligent question and number, and these two functions have been launched to all employees.

in the chemical industry, sinochem information and baidu smart cloud are exploring the use of large models for the research and development of new materials. the knowledge assistant "hua xiaoyi" can retrieve and answer professional knowledge such as molecular characteristics and molecular synthesis routes through natural language questions. at present, the search efficiency of specific molecules has increased by more than 5 times, and the efficiency of research and development work has been greatly improved.

this enthusiasm is also reflected in the bidding market. public data showsfrom january to august this year, the number of domestic large-scale model projects has reached five times the number for the whole year of 2023, and the amount of the winning bid has reached twice that of the whole year of last year.among them, the leading large-scale model manufacturers still dominate in terms of winning orders. baidu ranks first in four key indicators: the number of large-scale model projects, the amount of winning bids, the industries covered, and the number of central state-owned enterprises covered.

according to incomplete statistics from digital intelligence frontier,from july 1 to september 15, in just two and a half months, the total number of large model-related winning bids was at least 286., successfully surpassing the total number of projects in the first six months of the year.

among the purchasers, operators, energy, education, government affairs, finance, etc. still placed the most orders. an obvious sign is that some companies have put forward more and more segmented demands, and purchases for data collection and governance, large model security, talent training, etc. have increased significantly.

for example, china southern power grid completed the bidding for at least 15 large-model related projects in the two and a half months from july to mid-september. state grid also completed bidding for multiple large-model related projects in q3.

li chao, general manager of the energy and power industry of baidu smart cloud, told digital intelligence frontline that china southern power grid released its own controllable large model, "big watt", as early as september last year. among them, baidu smart cloud qianfan large model platform provides technical support for power dispatching scenarios. in q3 this year, china southern power grid completed the selection of large language models for its core business department, china southern power grid general dispatching, and baidu was successfully selected as the technical service provider.

02

big companies focus on "going with the flow"

the earliest companies to make the leap to large models in this round almost all had a foundation in small models. for example, icbc had already launched a project around 2021, hoping to use a "larger model" to build business applications such as ocr. "after chatgpt came, they took the opportunity to set up the large model project," a financial person told digital intelligence frontline.

the production safety market is subject to policy supervision, which is almost a bottom line for central state-owned enterprises. previously, longyuan power, the largest secondary wind power company of the state energy group, adopted a traditional small model safety solution to manage more than 200 wind farms and more than 10,000 wind turbines under its umbrella, forming a closed loop from safety monitoring, analysis, alarm to processing at the production site. in the second quarter of this year, baidu won the bid for the upgraded version of the project, which will combine the generalization and more accurate recognition characteristics of the cv large model and the ability of the large language model to make the previously large amount of collected information truly useful.

what are the things that the small models could not do before but the big models can now play a role in? hu wei gave an example, the dispatching of the port, the staff scheduling, the berths, the yards, etc., were all done with small models in the past, and the results were quite good, such as the yard planning.in fact, it is a mathematical algorithm problem, and the small model is very suitablehowever, when combining these into “integrated scheduling”, the small model cannot achieve large-scale collaborative computing and coordinated planning across data sets and devices. “we particularly hope that the big model can solve the problem. the challenge is that the data aggregation in the port takes time.”

in the industrial sector, the recognition rate of traditional cv small models is relatively low when encountering small samples. however, with the help of cv large models, the recognition rate of small sample data can be improved.

many enterprises also lack knowledge management, accumulation and application platforms. now, with the help of big models and knowledge retrieval, enterprises can form a knowledge management platform capability to complete the tasks of making implicit knowledge explicit, structuring explicit knowledge, associating structural knowledge, and mobilizing associative knowledge.

many companies hold regular production and operation meetings, and aligning data is a very difficult task. with the help of large language models, results can be summarized and extracted faster and more accurately to complete preliminary analysis, which is of great value to users.

03

to enter the core business, we need to find the "hands and feet"

li chao observed that in the past two years, the application of big models was still in its infancy, concentrated in assistant scenarios such as office and customer service. this is completely inconsistent with the expectations of the management of central state-owned enterprises with investments of tens of millions, hundreds of millions or even tens of billions. now, big models need to penetrate into the core business of enterprises.

“in the energy industry,we are investing in one direction - simulation optimization. "li chao said. in the industry, there are a large number of work tasks related to mechanisms and scientific computing, which cannot be done by large language models or large cv models. "if these problems are not solved, the future we envision - taking the large model as the capability center and providing a complete agent service - will not be realized." li chao said that in the future, large models must be combined with small models in professional fields to go deep into customers' core business scenarios.

"in the past few years, we have laid some foundations in this direction." li chao said that baidu provides simulation optimization engines and works with industry partners to enter power grid dispatching, petroleum and petrochemical refining, and oil and gas exploration scenarios to explore the implementation of practical projects related to mechanism models, such as steady-state analysis of power grid flow in power grid dispatching, optimization of air energy island operation in petroleum and petrochemicals, and desulfurization and denitrification. in september this year, baidu also participated in the bidding for the national pipeline network intelligent dispatching.

“these are the core production directions that customers are very concerned about.these directions are the essential "hands and feet" when building some intelligent services as a whole with large models as the scheduling center in the future.li chao said that they will put special efforts into planning these "hands and feet". the big model is the brain, and only with these "hands and feet (professional apis)" can the core scene services be truly realized.

"at the port, we first let customers see that the big model is reliable through questions and answers and data, and then gradually cut into the core business system." hu wei said, "we are already planning the next phase at shandong port group, and will go deeper into business scenarios."the port is a complex transportation hub, and the core is the dispatching of goods, people, trucks, gantry cranes, etc.previously, the tos system (terminal operating system) introduced by the port gradually added ai algorithms. the big model can further solve the problem of more metadata access.

as large models move toward production core systems,one of the core competencies of the global big model competition is logical reasoning"in the tests we conducted on customer sites, we found that there was a big gap between the logical reasoning of different large models for complex problems." in addition, the video resources generated by the large number of cameras deployed in the port were actually not used. one of the current directions is to use these existing hardware and image resources.use multimodal large models for integration to achieve better global coordination"this is a direction we are exploring."

04

74% of ai workloads are in the cloud

the implementation of a large model is a complex systematic project. as enterprises enter the deep water zone, some doorways and paths for implementation are accelerating to the surface.

"when we meet with customers, the first thing we do is to help them understand the boundaries of the big model." hu wei told digital intelligence frontier that in order to avoid inconsistent cognition that leads to failure to deliver or a situation that is too far from customer expectations, they now basically have each project,all of them will go through the complete process of "light consultation + on-site implementation".

in building intelligent computing power, idc data shows that74% of ai workloads are in the cloudbut today, the technical paradigm of intelligent computing has changed. in the cpu cloud era, people are most concerned about elasticity and extreme cost-effectiveness, while gpu cloud is about whether it can exert the computing power of a large cluster. such a cluster is not only expensive, but also has certain technical barriers. therefore, cloud vendors still play a major role.

in addition to computing power, the importance of data is becoming increasingly apparentaccording to incomplete statistics from digital intelligence frontier, in q3 2024, purchases related to data purchase, collection, and governance are increasing significantly.

at the data level, for example, the data of port customers are stored in the servers of various terminal companies. hu wei recommends that customers set up special data teams.

when it comes to data, there are deeper issues.”nowadays, when the industry discusses big models, it must talk about data, but frankly speaking, a lot of it is just empty talk."an industry insider admitted.

"in the era of big models,how to prepare data, how to manage it, and how to apply different data to different stages of large model training, most people are still groping in the dark.li chao said that baidu also encountered many pitfalls in this process.

for example, when training an industry model based on a general model, there are technical tricks to how to proportion the data. if you feed too little to the big model, the effect will not be obvious; if you feed too much, the model will not converge easily, and it may even lead to a decrease in the general ability of the model... these pitfalls were later settled and formed into a set of tools and methodologies, which were exported through the qianfan platform and a dedicated technical service team.

there is also a tendency for a gap to appear between large model technology and applicationsmany "user units do not understand ai, and ai units do not understand the industry." li chao introduced that in order to solve the gap, in addition to equipping solution architects with industry backgrounds, they will also focus on selecting partners. "for energy industry partners, we basically only select two types, one is the client industry unit, and the other is the company that provides human resources outsourcing services in the client industry company." li chao told digital intelligence frontline that their common feature is that they are deeply involved in and understand the client's business, and understand the it industry.

in addition, when the big model is put into practice in a specific scenario,there is still a lot of engineering work to be donefor example, the big model is used to output answers in the form of "general, specific, and specific", and the answers are different each time, but some customers are more accustomed to the "specification first, then generalization" method, and hope that the answers can be "reproducible". "even if there are only minor changes in word order and grammar, some port leaders cannot accept it," hu wei said, giving an example, which requires them to do a lot of engineering work to ensure that the output of the big model meets customer expectations.

from the overall situation of enterprise implementation, customer demand is 360 degrees. interestingly, all cloud companies are currently shifting from cloud partners to ai transformation strategic partners, upgrading their overall technical capabilities around ai infrastructure, algorithm models, data, deployment optimization, and customization.

05

the human factor

a cio of a technology company encountered considerable resistance in the construction and application promotion of a large model: front-line employees and departments did not cooperate, and even the leaders’ advice was useless.after the company used the code assistant, the overall efficiency increased by 1/3"although we rarely talk about reducing staff, it does mean that people who write code will have to move to the front-end or back-end of development." front-line employees are worried about being laid off, and some business departments have a sense of territory. these are all realistic problems.

in many traditional large enterprises,the biggest resistance comes from work inertiamany front-line employees are experienced workers. if they are asked to use equipment maintenance assistants, they will say, "if there is something, i might as well try it myself, or call lao zhang. the three of us will work together for the rest of our lives."

“the large model is a one-man project"it needs to be pushed forward from top to bottom," a financial industry insider observed. senior leaders publicly supported the benchmark project. "sometimes not opposing it is also a form of support."

"we pulled the business departments in at the beginning to participate in the construction and final adoption and promotion," said jin jianhua, founder and ceo of ianalysis. for example, for a maintenance assistant, the group may have its own ideas, the subsidiaries may have their own ideas, and the frontline employees may have their own ideas. how can we align everyone's profit expectations?it’s a question of balancefor example, some knowledge graph construction is done by senior experts from secondary companies, so that everyone can advance towards a common goal.

companies also need to regularly disclose the operation status and business benefits of benchmark projects and publicly provide incentives. "for example, we can give employees some points and link them to some systems and exchange them for some gifts," hu wei observed. "we will directly link them to performance and bonuses," said a cio of a technology company. "although it is a bit simple and crude, it does work." this requires the cooperation of management and the mobilization of hr resources.

the retention rate of first-time users is also very important. we need to find ways to improve the first-time login experience of employees after the revision and iteration. for example, if the entrance is clear enough and what was not accurate before is now accurate, corporate employees will continue to use it.

from the perspective of the business department, designing stories that are appealing and visual, and letting some veterans share their experiences, may be contagious. the project's daily and monthly active data are equally important. these arecultivating an ai culture in your company。

the wave of big model implementation is still surging. under the torrent, all players in the industry chain are continuously increasing their efforts to promote more companies to achieve digital transformation.

baidu will also make a big move in the near future and will hold a2024 baidu cloud intelligence conferenceby then, baidu smart cloud qianfan platform will usher in a new upgrade in application development tools, large models, and tool chains. the three major ai application products - baidu smart cloud keyue, wenxin kuaima baidu comate, and baidu xiling digital human will also be completely upgraded. at the same time, multiple sub-forums such as smart finance, smart industry, smart transportation, smart government affairs, smart cars, and embodied intelligence will be held to achieve more technical exchanges and collisions.