Entering 2023, ChatGPT will push the world into a new era - the era of large models. It not only triggers the overall upgrading of the AI industry, but also allows various large models to emerge in endlessly. The key reason behind it is that large models can generally improve productivity , and many companies in the industry are actively looking for opportunities to apply large models and generative AI, hoping to make a difference in the industry.
Indeed, the reason why ChatGPT is called the "iPhone moment" of AI is that the generative AI represented by ChatGPT can make it possible for everyone to order computers to solve problems. It plays the role of assisting people, serving people and even surpassing people. With this revolutionary technological breakthrough, ChatGPT took the lead in setting off an application boom in search engines and various tool software, and aroused industry users' interest in ChatGPT related technologies. attention and learning. At the same time, a large number of downstream applications have also captured new technology and industrial opportunities. It is hoped that through various large-scale models and engineering capabilities, the capabilities of ChatGPT-like products can be transferred to the original applications, so as to better empower the digital intelligence of enterprises. transformation.
But we must also see that it is not a "one-off" process to make a large model from "existing" to "usable". More companies usually face limited data resources, difficulty in computing power investment, poor model generalization ability, and The development bottleneck of scarcity of high-level talents. It is precisely because of this that for more enterprises, a more "pragmatic" approach in the future is to choose a large-scale model service provider that suits them, "stand on the shoulders of giants", and better "use" large-scale models.
It is precisely because of this urgent market demand that, as a practitioner and enabler of digital transformation, iSoftStone has always insisted on creating value for industry customers with full-stack digital technology in recent years, especially in the field of large models. Power is also actively exploring and practicing, hoping to provide industry customers with professional services for large-scale implementation of large-scale models and open up large The "last mile" of the model landing will better accelerate the new era of embracing large-scale models in thousands of industries.
Four Challenges of Large Model Landing
Undoubtedly, few doubt the disruptive impact of big models on the future. But the reality is that domestic exploration of large models is still in the early stages. No matter in the research and development, iteration or use stages, large models are a "luxury" that consumes a lot of resources and is not cheap to use. In addition to the high cost, enterprises want to deploy and use large models in actual business scenarios, and they also face many implementation difficulties such as data, parameter tuning, and talents.
In this regard, Huo Yu, vice president of iSoftStone's digital innovation service line , said that since this year, iSoftStone has actively explored and practiced in the field of large-scale models, and found that large-scale models face many challenges in implementing them in industries and enterprises. Observe from several dimensions:
First, from the perspective of computing power, training large-scale models requires a lot of computing resources and capital investment, which is a huge challenge for many enterprises. "The company's investment in the basic computing power of the large model is expected to be nearly 100 million yuan this year, and it will continue in the future. In this field, if you need to form a competitive product or solution, tens of millions of capital investment can only be regarded as the starting threshold , It can be said that the computing power and training costs of large models are extremely high, including large industry models and large vertical field models that are often mentioned now, so computing power is a high threshold that cannot be avoided.” Huo Yu said.
Second, from the perspective of data, in the case of generative AI, the language model trained needs to have higher richness and complexity in order to better understand and generate various language expressions. For example, different languages have different characteristics such as syntactic structure, vocabulary usage rules, and semantic relations, so the training model needs to include more language knowledge and rules to adapt to these differences. At the same time, different languages may also contain some special expressions, cultural background, etc. The training model also needs to consider these factors to better simulate and generate various language expressions. Not only that, in order to improve the richness and ability of the model, it is also necessary to use a larger, diverse, and real language dataset to improve the generalization ability and richness of the model, so that it can better adapt to various language scenarios and application requirements.
In Huo Yu's view, the corpus problem of large model training is essentially the problem of enterprise data governance. On the one hand, enterprises must obtain a large amount of available and credible data; on the other hand, data also needs to be governed to form a structure. And standardized data, so as to better train the required large model service.
Third, from the perspective of algorithms, the development of AI models began with the release of AlexNet, and most of the subsequent research focused on the exploration of the depth and breadth of the model. So far, typical models such as BERT and GPT have appeared, and pre-training models have also been launched. At present, the parameters of large models in China have climbed to hundreds of billions and trillions, and their accuracy is constantly refreshing SOTA.
But at the same time, for enterprises, how to choose a suitable pre-training model, how to fine-tune based on a mature large model in specific scenarios and tasks can quickly produce accurate results, and whether to choose a commercial or open source large model For model services, how to balance cost and training effect is also a very "headache" problem.
Fourth, from the perspective of talents, as more and more companies begin to widely apply large-scale models, related talent needs have become a new challenge. Taking AI trainers who perform database management, algorithm parameter setting, human-computer interaction design, performance test tracking and other auxiliary tasks in the actual use of artificial intelligence products as an example, relevant data show that there are currently a large number of such talents. The gap” needs to be filled urgently.
It is not difficult to see that although we have witnessed the deep integration of large models with scenes and industries and achieved good results, it has been verified that large models have not only been applied in technology companies, but have also taken steps towards thousands of industries. The implementation of large models is not a simple matter, and new challenges brought about by computing power, data, algorithms, and talents still need to be solved.
Entering the large-scale model circuit as a service
Based on this, since the beginning of this year, iSoftStone has actively deployed large-scale model tracks, not only investing in the construction of AI computing power infrastructure, but also from the perspective of service and implementation, hoping to rely on its own AI talent resources, AI platform tools , AI cooperation ecology and the experience and methodology accumulated in "practice", and industry customers to create large-scale model services that can be implemented, and better help companies welcome the arrival of the era of large-scale models. Specifically:
First of all, in terms of AI talent resources, this is the unique advantage of iSoftStone in the large-scale model track. Relying on more than ten years of technology accumulation and industry experience, the company has not only been able to provide industry customers with "full-stack" digital technology services after horizontally pulling and integrating its own service capabilities, but also accumulated a large number of high-level Quality digital talent.
"Our first entry point is the direction of AI talents. We invest in a dedicated computing power platform as an engineering practice environment for training engineers to get started and familiarize themselves with large models. This will enable engineers and architects of related development languages to play with large models faster and more proficiently. , hope that through investment in computing power and talents, they can open up the chain between large models from products to landing applications, and become experts in the field of large models. In the future, the capabilities of these expert resources can not only be passed on to more On the other hand, for enterprises, they can also directly obtain the help of expert resources, directly carry out the large-scale model development required by the enterprise or provide related services such as parameter tuning." Huo Yu said.
Secondly, in terms of AI platform tools, iSoftStone has also created iSoftStone Tianxuan 2.0 MaaS platform. According to reports, based on industry service needs, iSoftStone Tianxuan 2.0 MaaS platform is integrating iSS Model Ops Platform, iSS Model Dev Tools, iSS Scenario Plugin Application Service Platform (iSS Model Plugin Store) and other products, it can provide customers with large model data processing, large model one-stop operation services, continuous training, tuning, deployment, reasoning and digital asset management, data security and other services.
It is worth mentioning that iSoftStone's newly upgraded integrated platform for training and pushing is based on the Ascend hardware base, adopts iSoftStone G420K training platform and iSoftStone G210K reasoning platform, integrates Euler operating system and other components, and is equipped with its own AI middle platform , can provide users with a variety of interactive AI models, deeply adapt to different AI application scenarios, and can be applied in many industries and fields such as central state-owned enterprises, education and scientific research, and finance.
Third, in terms of AI cooperation, iSoftStone has established ecological cooperation with industry leaders and mainstream large-scale model manufacturers. Not only is it the first to connect to Microsoft Azure cloud GPT4, but it is also an ecological partner of Baidu Wenxin Yiyan, Huawei Cloud Pangu Model, Ali Tongyi Qianwen, Yuanchengxiang ChatImg2.0, and is also actively researching such as ChatGLM, Open source large models such as DeepSpeed Chat, OpenAssistant, Alpaca, and LLaMA.
Huo Yu believes that the greatest value of iSoftStone's comprehensive "layout" in AI cooperation ecology lies in the fact that the team has accumulated a lot of experience and methodologies through the use of these third-party commercial large-scale model services and open-source large models. These practices and experiences are also of great reference value and reference significance for industry customers to implement large-scale models, which can prevent enterprises from taking "detours" in the process of implementing large-scale models. This is also a relatively "safe" way of implementation.
Finally, in terms of the implementation of large-scale models, iSoftStone is also working with industry customers to better promote enterprises to embrace the era of large-scale models through the form of "co-creation".
"Nowadays, many industry customers have found iSoftStone. They think that it is difficult for enterprises to build a large-scale model. The advantage of these industry customers is that they have a lot of industry data, but they don't know how to make these data Better realization of corpus, or how to enable large models to achieve training or reasoning faster on the basis of saving computing power, etc. In addition, building large industry models, early planning and selection are also a threshold , and these are exactly the jobs that iSoftStone is good at, so the two parties can jointly promote the implementation of large-scale models through co-creation. At present, the company has launched relevant cooperation with many customers in the banking and insurance industries." Huo Yu said .
As mentioned in IDC's latest "AI Large-scale Model Technical Capability Assessment Report, 2023": "For industry users, while paying attention to the completeness of the manufacturer's large-scale model technology stack, they should focus on examining the manufacturer's industrial application experience. The main focus should be on the application layer, applying technology to actual business scenarios, planning in advance, and accumulating industry and scenario experience and data, so that we can 'stand on the shoulders of giants' to create differentiated competitive advantages." From this From a perspective, the four advantages accumulated by iSoftStone in the field of large-scale model services can undoubtedly better help enterprises embrace large-scale models and accelerate their digital and intelligent transformation.
Get through the "last mile" of landing
As a matter of fact, iSoftStone’s service mode of “slotting” a large model track is itself the result of careful consideration and careful consideration.
Huo Yu told me: "When the wave of large-scale models hit, the first thing we rejected was to build a general-purpose large-scale model platform by ourselves, and we still considered starting from the track of large-scale models in the industry. The data and corpus in the vertical industry are also very important to make a large-scale industry model. Therefore, as a service-oriented company, it is most suitable for iSoftStone to enter this market in a service mode. We can use expert resources, platform tools, As well as experience and methodology, coupled with the data resources in the hands of industry customers, the cooperation between the two parties in the form of co-creation can realize the landing of large models in a relatively fast way.”
Similarly, entering the large-scale model track with the "service" model also enables iSoftStone to observe and look at the entire large-scale model market from a more "overall" perspective, and has accumulated and summarized many large-scale model implementations that are worth paying attention to. key matters.
The first is about the selection of open source and commercial large models. In this regard, Huo Yu believes that based on business needs, Chinese enterprises will also choose to directly access commercial large-scale models, or choose to deploy localized services based on open-source large-scale models, just like public cloud and private cloud deployment.
"Leaving aside the problem of computing power resources, if we only look at the size of the model parameters, the capacity of the open source large model is close to that of the commercial model on a relatively small scale, and gradually has the momentum to catch up. In addition, the relative pursuit of large and comprehensive general commercial models For large models, in terms of model training in vertical fields, the number of open source models has exceeded that of commercial models, so companies don’t have to worry too much about open source large models lagging behind mainstream commercial large model platforms. In the field of industry large models, it should be said that the entire market is It is not yet mature, and is still in the stage of exploration and practice, which requires the entire industry chain to promote the gradual maturity of the industry's large model through co-creation, and finally get through the last mile of landing." Huo Yu said.
Second, in addition to computing power and model parameters, enterprises should focus on the precipitation of data and corpus. In Huo Yu's view, corpus is indeed a scarce resource at present, but many companies do not realize this. If companies want to make large-scale models in the future, data and corpus are unavoidable problems. It is precisely because of this that the quality of data and corpus has become particularly important, and a series of tasks such as cleaning, labeling, and governance of the underlying data have become more urgent.
"We are currently helping some industry customers with data governance. We judge that the work of data and corpus will continue for a long time, but many tasks such as data labeling and data collection will become 'tools + manual'. At the same time, these data and corpus will pay more attention to industry attributes. Not only that, but in the specific practice process, we have also summarized and precipitated how to form a standardized methodology for corpus or data, and how many parameters need to be 'fed'. The corpus can achieve a more appropriate price-performance ratio, which is what iSoftStone is doing. On this basis, we are also responsible for the setting of business rules, mathematical modeling, parameter tuning, and follow-up long-term optimization. It can provide relevant services for industry customers." He said.
Thirdly, the vector database is also an important direction of iSoftStone's attention. Vector databases are mainly used in the fields of AI and machine learning. In these fields, data is usually presented in the form of vectors, which can effectively solve the problems of storing and querying unstructured data such as text, pictures, audio, and video; and in the era of large models, the current management and processing of unstructured data The mainstream method is to convert the semantic content of unstructured data into multi-dimensional vectors by using Embedding Models such as RNN or Transformer, and directly store and process these vectors, which can be used for training and learning of large models. Provide better data support.
Huo Yu believes that the vector database will play three roles and values in the large model, including that it has become the core technology to promote the iteration and evolution of the large model, and at the same time, it will also have a qualitative impact on the customization needs of enterprises. Impact; more importantly, the vector database will also bring about relatively large changes in the delivery of future data-based projects.
"My understanding is that the vector database is another 'digitization' of data science. It can be simply compared with the original database with a table structure, which is represented by a two-dimensional or multi-dimensional structure. Then the vector database uses a Therefore, with the development of vector database, computing power and AGI technology, the future data delivery mode will change, which will also produce a new business model, and we are also highly concerned about this.” Huo Yu said .
Looking back, iSoftStone's pragmatic style of "actions speak louder than words" has allowed it to build its own unique competitive advantage in the large-scale model service track. The investment of real money in AI computing power resources, for a service-oriented enterprise, is behind the demonstration of determination and the grasp of the general trend of the market. This enables iSoftStone to provide sufficient computing power resources for engineers and developers to get started and accumulate practical experience; The new value brought by the model.
"In the future, we will continue to reserve more expert resources, continue to invest in platforms and tools, and strengthen the quality of data and corpus according to the customer's business application scenarios, and ultimately improve the efficiency of delivery, with a better innovative service model Get through the last mile of the large model landing." Huo Yu said.
Between "practicing by yourself" and "watching from the sidelines", iSoftStone chose the former, which also gave it a deeper understanding and insight into the large-scale model market. At the same time, its bold practice in the past period also made It has more accumulation and precipitation in terms of talents, tools, ecology and even methodology, which is why iSoftStone has the ability and confidence to allow more industry customers to accelerate the embrace of the era of large-scale models.
Shenyao's Science and Technology Observation was founded by a senior technology media person, Shenski, who has 20 years of experience in dissemination of enterprise-level technology content. He has long focused on the observation and thinking of industrial Internet, enterprise digitalization, ICT infrastructure, and automotive technology.
