Under the "rigid needs" such as data security, privatization has become the entry threshold for the government and enterprise market. However, under privatization, the limitations of the industry model are not yet known. s solution.
Author|Dou Dou
Editor | Pi Ye
Produced | Industrialist
The field of government and enterprise seems to be becoming another target of AI landing.
Some of the latest developments are: H3C released the private domain large-scale model "Baiye Lingxi" LinSeer for government and enterprise customers; The industry-specific large-scale model "ten thousand machines"...Baidu Smart Cloud, Alibaba Cloud, 360, HKUST Xunfei, SenseTime, etc. have also released related products and solutions for government and enterprise customers.
This trend is also happening overseas. It is reported that Microsoft is using its Azure cloud service to introduce OpenAI's powerful language generation models to government agencies. These models include OpenAI's latest and most advanced large language model (LLM) GPT-4 and GPT-3.
However, judging from the actions of major domestic manufacturers that release large models of government and enterprises, there are not many landing cases. There are many reasons behind it. One is that the product release time is short, and it is impossible to get large feedback in time; the other is that the AI boom caused by large models is still in the outbreak period, and the government and enterprises still need to consider the choice of cooperative manufacturers.
In addition to these factors, several key questions are, are the AI products released by these manufacturers for government and enterprise customers really what government and enterprise need? Or, for government and enterprises, what kind of AI products and services do they need?
1. Government and enterprise market, " wait and see" AI
"The customers we have contacted are all exploring this aspect, and government and enterprises will have demands in this regard." An industry insider told the industry expert, " But for the application of AI large models in the field of government and enterprise, everyone is on the sidelines of waiting and exploring. phase. "
In China, government and enterprises are generally large in size, complex in business, high in customization requirements, and have a heavy burden of IT construction. It is difficult to coordinate between departments and businesses, and it is difficult to break through data barriers. This makes the digital transformation of government and enterprises difficult.
The large model can completely deposit the data of the interaction process between people and people, and between people and machines, so that it can learn and improve independently, and continue to evolve. Put an end to the chimney-style construction in the past, while reducing costs and increasing efficiency, it can also promote the digitalization process of government and enterprises.
This trend of technology application has driven many manufacturers to combine their own TOG products with large models.
However, from an objective point of view, although there are many players who are deeply involved in the AI industry chain, in the field of government and enterprises, due to the consideration of various factors such as computing power, industry professionalism, data security, and localization, it is difficult for manufacturers to provide suitable AI. solution.
As a new digital infrastructure, AI large models still need to be continuously explored and practiced on how to empower the industry. Many companies are also constantly making new attempts.
Zhou Hongyi once posted on Weibo, " Large enterprises or governments want to privatize large models. That is, on the basis of adding public GPT knowledge and capabilities, train a private GPT, only for the enterprise itself or customers to use."
With the same opinion as Zhou Hongyi, there is also Han Xiaoping, director of security product research and development at H3C.
"On the one hand, the data of the government-enterprise large model cannot be out of the domain; on the other hand, it needs to be fed and trained in combination with its business scenario data to generate an actual business scenario app."
He believes that the large-scale model technology of government and enterprises is not the only factor to be considered, but the understanding of user business is more important. For government and enterprise AI services, industry know-how is still the key. Manufacturers without experience in government and enterprise business services can hardly design personalized solutions based on large models.
Both of them aimed at the privatization of the implementation of the government-enterprise AI model.
2. Is privatization a good choice?
The phrase "Completely prohibit the use of ChatGPT" will be mentioned in the hot search of the technology list from time to time. The most recent protagonist is Samsung.
The specific stems from Samsung allowing engineers in the semiconductor department to use ChatGPT to participate in fixing source code issues. But in the process, employees entered confidential data, including the source code body of the new program, the minutes of internal meetings related to the hardware, and other data.
At first, Samsung completely banned the use of ChatGPT, and only partially opened it to the DS department since March 11, but it did not expect that the secrets would be leaked in a short period of time.
Confidential data leakage incidents such as these emerge in endlessly.
In overseas markets, well-known companies such as Apple, Samsung, JPMorgan Chase, and Citibank have made it clear that they prohibit or restrict employees from using large-scale products such as ChatGPT in the workplace.
In fact, the underlying logic of the large model is to label industry data for deep learning and train a vertical model that focuses on a certain industry. This also means that enterprise data must be integrated into the data pool of the general large model.
However, most of the data of government and enterprises involves issues such as commercial secrets and personal privacy, and has very high requirements for data security.
Therefore, the key is to keep the data out of the domain. The private large-scale model is undoubtedly a must for government and enterprises.
However, privatization means that government and enterprise customers need to collect data for training themselves. In general, the timeliness and accuracy of AI decision-making depend on data quality, quantity, and computing power. This means that the more dimensions, quantity, and quality of data input to the pre-training model, the higher the true value of the large model.
A controversial point is whether the private models created by professional manufacturers can meet the needs of government and enterprises compared with the capabilities of general-purpose large models.
On March 30, the founder of Bloomberg, Bloomberg released a paper, BloombergGPT: A Large Language Model for Finance.
In the paper, it introduces its financial large language model trained on the financial field data set of 363 billion tokens and the general data set of 345 billion tokens. And conduct financial field assessment and general field assessment.
The results show that the model has the best comprehensive performance in financial domain tasks; in general tasks, the model's comprehensive score is also better than other models of the same parameter magnitude, and the score on some tasks is higher than the parameter quantity Larger models.
This means that a large model trained based on professional domain corpus has a better understanding in the domain than a general-purpose large model.
One implication is that in other specific domains, dedicated large language models can also be developed, and the effect may be better than general large models.
However, the conclusions of this evaluation experiment still cannot represent all industries, and some problems still need to be addressed. For example, there is actually a big gap in the dimension and quality of financial data between China and foreign countries. As the first industry model, this financial model replaces different dimensions and quality . Whether the data can maintain the accuracy of its conclusions is yet to be known.
In addition, in addition to data-sensitive government and enterprise scenarios such as government affairs systems, city brains, and finance, and other terminal scenarios, the corpus data of the general-purpose large model still has great advantages.
Therefore, for government and enterprises, the AI services they need are capable of integrating general-purpose large models as a capability supplement while ensuring data security.
3. Large-scale models of government and enterprises also need "group battles"
There are specializations in the art industry, and the same is true for the ability of large models.
In terms of the application of large AI models in government affairs, Zhou Hongyi once said bluntly that he predicted that China will not have only one large model in the future, and that each city and each government department will have its own proprietary large model.
Under the wave of innovation and innovation, the localization of the full technology stack has become the entry threshold for serving government and enterprise customers. Therefore, for government and enterprises, Xinchuang's ecological capabilities have become an important criterion for them to choose manufacturers.
At present, there are only a handful of AI large-scale model manufacturers that can realize the localization of the entire technology stack. Whether it can attract more Xinchuang partners becomes the key, learn from each other's strengths, or achieve a win-win situation.
In addition, on the basis of localization, different scenario applications also require cooperation with different manufacturers, such as the advantages of Internet cloud manufacturers in terminal software applications; the network advantages of operators; Huawei's advantages in chips and systems.
In addition to cooperation based on customer needs, there is also pressure from the scarcity of computing power and other resources.
With the arrival of a new round of AI boom, servers need to be equipped with a large number of chips such as GPU/NPU/FPGA/ASIC to support high computing power, but at present, computing power is not only expensive but also resource-intensive.
At present, 95% of the chips used for machine learning, including the ChatGPT large model, are Nvidia's A100 (or the domestically used alternative product A800), and the unit price of this chip exceeds 10,000 US dollars. In addition, the performance of Nvidia's latest generation H100 chip far exceeds that of A100, but the price is also higher, about 250,000 yuan.
Under the industrial ecology of AI large-scale models that learn from each other's strengths, it will greatly eliminate the repeated construction of models, realize the reuse of data, and greatly save the utilization rate of computing power resources. At the same time, it can also reduce the AI deployment cost of government and enterprise customers.
Therefore, for manufacturers who deploy large government and enterprise models, on the premise of privatization, through open source, API interfaces, etc., use the general large model and customer or their own integrated data to help customers adjust the model and form a customized AI Products, or will promote the rapid landing of large government and enterprise models.
Some manufacturers are also verifying the feasibility of this path.
For example, China Electronics Technology Co., Ltd. provides a new paradigm of industry application in four stages of "general intelligent model large cycle + industry intelligent model small cycle" double cycle and "model training + test evaluation + scene fine-tuning + credibility enhancement" for the party, government and enterprises.
For another example, LinSeer, released by H3C for government and enterprises, not only supports the model of "private domain large model + H3C ICT infrastructure" to ensure data security, but also supports the model of "customer-selected large model + H3C ICT infrastructure" to meet Various needs.
The trend of government and enterprise AI services is gradually clear, that is, government and enterprise may need a "buffet" service model. Customers can choose chips, model technology parties, data integration parties, software vendors, etc. that suit their needs.
Whether AI, which has experienced many cold winters, can seize the opportunity of this era still needs time to verify.