Wind up AIGC
AIGC has become the new business watershed of our time, leading the Cambrian explosion of content and creativity. Whether it is the frontier technology industry or the entire economic and social pattern, it will produce huge changes. In order to allow entrepreneurs to better embrace this era, the dark horse of entrepreneurship will take China's AIGC No. 1 service platform as its own responsibility, and conduct a series of reports from the perspective of software and hardware platforms, investors, industry applications and other dimensions by interviewing a number of well-known entrepreneurs, investors, scholars and entrepreneurial practitioners in the industry, so as to discuss new opportunities in the era of large models and show new industrial strength. This is the eighth part.
The guest of honor in this issue is Sun Bin, President and COO of Bamboo Intelligence. Founded in 2015 by Jian Renxian, former vice president of Microsoft (Asia) Internet Engineering Institute, Bamboo Intelligence is a leading domestic artificial intelligence enterprise in the era of cognitive AI with NLP driven by large and small models as the core technology, through dialogue, knowledge, training and generation of four product platforms, to achieve cross-industry, cross-scenario large-scale landing, has served more than 500 large and medium-sized enterprises. When the LLM (Large Language Model) technology represented by ChatGPT set off a global wave, Bamboo has upgraded all product lines and launched mature AIGC enterprise-level products using a variety of LLM technologies to help knowledge producers such as white-collar workers improve productivity and help enterprises complete digital and intelligent transformation.
In the second season of the dark horse AIGC theme series live broadcast on June 8, Sun Bin shared the theme of "The Landing and Exploration of Big Language Model in the To B Field", and showed the business insights, layout and important achievements of Bamboo Intelligence in the implementation of generative AI technology.
The following is the transcript of this live broadcast:
Today I will share the development of large language models that I have personally experienced as an NLP (natural language processing) practitioner team, and the future we see.
I believe that many people have recently learned about Chat GPT on the Internet, and have analyzed many characteristics, advantages or shortcomings of big language models and generative AI. But in fact, big language models, including our NLP (natural language processing) industry, have been developing for a long time. The Bamboo Intelligence team has been focusing on NLP for 8 years, and these 8 years are also the period when NLP has truly been industrialized. During this period, we have been in the NLP track, studying how to commercialize NLP services, how to market, how to provide services to enterprises, how to achieve growth on the C-end, and have accumulated some experience and customers.
01
Industry changes and four major challenges
In the field of artificial intelligence, the processing of images, speech, and semantics are some important tracks, and NLP is one of the parts on natural semantic understanding.
In fact, in terms of the semantic understanding of artificial intelligence, it can be divided into two parts: one is the very familiar human-machine question and answer, we call it short text, which is used for interaction; The other is long text, which allows machines to read, understand, extract, and extract existing unstructured documents and articles. These two fields are both strong areas of NLP, and they are precisely the fields of industrialization. The big language model we are familiar with today shows you high understanding and high dialogue ability in dialogue, but what makes everyone more amazing is the generated articles, the code written, and the language translated.
In these two areas (short text and long text), various NLP teams have achieved a lot of results in the past, but in the process of their implementation, they mainly face four major challenges, and now the emergence of large language models has greatly changed the situation.
First, there is deep semantic understanding and reasoning. Before the outbreak of large language models, all language models were relatively lightweight, for some specific scenarios, the first challenge at that time was how to accurately understand the meaning of the customer, and everyone was constantly using new model optimization algorithms to understand the customer's intention, understand the customer's context and their emotions, but now the big language model has greatly improved this ability.
Second, intelligent question and answer based on knowledge graph. This is also the most criticized place for large models now, because the knowledge of large models is not very accurate. In the past, the best practice was to use the NLP model to understand the customer's dialogue intention, and then use the knowledge and answers given by the customer to achieve accurate matching, which can help the dialogue robot give the correct response.
The third challenge is how to achieve accurate matching of questions and answers through semantic training. For example, in the financial industry, there are actually standard answers to the service content of credit cards, but the questions of customers are very different, and how we use language models to understand customer intentions is a difficult problem. In the field of To B services, the accuracy of the answers is required to be very high, otherwise it will mislead or cause damage to customers, and this challenge continues into the era of large language models.
The fourth challenge is the flow of tasks. As you know, any big language model has a strong chat ability, but it is difficult to get it to do a job, such as getting a credit card, reissuing a ticket, or querying a ticket. This multi-round process-based dialogue can be easily imitated by large language models, but to accurately provide this data, another layer of technical control is required.
These four major challenges in the past NLP industry still exist, but to a large extent, it can be solved more easily with large language models, which is the benefit of technological change to the industry. In this regard, we have made some cases using large models and gained customer recognition. For example, the customer service center of a securities customer originally used a traditional customer service robot, which exposed many problems in use, such as business knowledge and idle chat mixed management is not easy to maintain, a large number of knowledge materials content and group mismatch, answer accuracy rate is too low, etc., 40%~50% of the phone finally have to be transferred to manual service. However, with the blessing of technology, the proportion of robot services can be increased by 30 percentage points, which can greatly save manpower and improve efficiency.
Therefore, in the application process of today's AI, especially in the To B industry, it is not as everyone thinks, one model can solve all problems. It's actually a platform logic like the following image:
The bottom layer is the basic technology, the AI module above, and the secondary development through some APIs, and finally access the customer's business scenario. Recently, everyone has focused on the writing ability and dialogue ability of large language models, but from the perspective of practitioners serving enterprises, although the underlying model capabilities have made great breakthroughs, there are still many capabilities that need to be improved in order to better realize the application.
On the occasion of the outbreak of today's large-language model industry, we call on all professional teams with experience in the track to control the large model, so that the large model can be implemented in multiple scenarios to achieve industrial development. Not only the big model itself should develop rapidly, but the innovative business brought by the big model should also achieve rapid development.
02
The future of industry
Next, let me share the analysis from the perspective of practitioners, what changes will the big language model bring to the industry?
First of all, think about it, what is the essence of the Chat GPT phenomenon, and what is the experience it brings to everyone? I believe that all people who have had conversations with it are not just for fun, everyone will feel that "I am talking to an agent". What does this mean? It means that humans can communicate with language models. It can understand you, it can answer you, it can help you execute. The best thing about the big language model is that it brings enough knowledge in a "violent" way, and then it can understand people's emotions and talk to people, which just solves the bottleneck problem of previous human-machine interaction.
From this point of view, my first judgment is that the big model will change our current software paradigm. I believe that colleagues who do IT have a personal experience of this, when the PC appears, when the Internet arrives, when mobile applications begin to explode, the software paradigm has changed, from industrial software to PC software, to websites and mobile APP, and then to the current big model boom, the software paradigm will be rewritten again.
So, after the big language model comes, what will be the paradigm of software applications? We can mobilize a wide variety of applications through dialogue. If we were application-oriented in the past, in the future, we can very clearly imagine that people will communicate with mobile phones or smart hardware, and then it will complete the corresponding software operations, the shield between all applications will be broken, the ability of the application will be called, and the large language model will directly mobilize the application capability to form a new interaction mode, which will be a new operating system, a new "iPhone moment".
In the next 2~3 years, a large number of our C-end applications, will change due to the change of dialogue mode, it is likely that it is no longer touch input, but voice input, many actions will also break the boundaries of applications, may have an AI assistant in each mobile phone, it can schedule multiple application capabilities through dialogue, order, ride, purchase can be completed by voice, and then thousands of applications based on ChatGPT models will appear.
The above is the change on the C side, so how will the B side change? My second judgment is that on the B side, the private domain knowledge of enterprises/industries will become crucial, resulting in a large number of enterprise Chat GPT and industry Chat GPT.
At present, the characteristics of the large language model are that it can highly understand human semantics and can do some deep work, such as writing, reasoning, analysis, etc., but its knowledge cannot be relied upon. This is because the Internet data used for training is unreliable, and today's big language model is actually a conversational model, not a question answering model. It is created for dialogue, so it compromises, it admits mistakes, and in order for the conversation to continue, it changes the content according to the preferences of the person with whom it is interlocuted, but it is not a model with the right knowledge.
But for enterprise customers, our industry directors, customer service, marketing personnel, and policy consultants must not give customers inaccurate knowledge. Therefore, on the B side, it will definitely develop into such a paradigm: the big language model does communication and understanding, plus private domain knowledge - accurate private domain knowledge - and then drives the application of the industry/enterprise, and the application of the industry/enterprise will also change because of this change.
So, how to construct the private domain knowledge of enterprises? We can review the development process of enterprise digitalization, at the earliest, we called equipment networking digital, which is the first generation of production digitalization; In the second step, we use ERP, including manufacturing systems, to achieve full business process IT, which is the second generation of asset digitization.
Now, we have a big language model, with the private domain knowledge of the industry/enterprise, and really let the HR department, administrative department, sales department, customer service department of the enterprise... All knowledge is expressed in the form of dialogue or reading, which really makes the enterprise intelligent.
After the development of artificial intelligence to a certain stage, knowledge can be used, and evolution has been achieved, and we will evolve from the digital age to the digital intelligence era.
In the following time, we will see more and more enterprises turn the knowledge of the department, the knowledge of the enterprise and even the knowledge of the industry into a knowledge base using artificial intelligence technology, into knowledge that can be built and called by AI, forming a knowledge flow, and then turning this knowledge into digital people, serving our enterprises and serving our customers.
We can predict that in the coming year, these contents of enterprise services will grow exponentially, at least 10 times more. Entrepreneurs in the To B industry must be prepared for this.
03
Four new challenges for the future
So, what new challenges will the development of large models face in the future?
In addition to the four challenges mentioned earlier, as a technology practitioner, I would like to share with you a few of the new challenges we see:
The first challenge is large model reading or knowledge graph pre-building? In the past we did a lot of knowledge graphs, but today's large language models can read documents, can read those unstructured data. So, is it still necessary to build a knowledge graph? In other words, is it necessary to preset the answer today to facilitate the question and answer and query, or do you want the big language model to read the content by itself and then give you the answer?
In fact, these two practice paths can complete a lot of content queries, but in the end, which effect is better and which has high accuracy, I believe that different scenarios should use different modes. One might ask if it would be better to combine the two? The answer is worth looking forward to, and I hope that our team will use engineering capabilities to give results.
The second new challenge is "Prompt? Embedding？ Fine-tuning? "All three words are particularly hot words right now. Prompt is the prompt word, Embedding refers to the embedded interface, and Fine-tuning refers to model fine-tuning, all of which are the work to be done to train large models. But at present, not many teams can adjust the model well, and it may become worse and worse in the process of fine-tuning. So my advice to you today is not to be obsessed with fine-tuning large language models, and ultimately to take controllable results and high-quality goals as the standard. The ability to use large language models plus its own engineering capabilities, such as knowledge graphs, calls to customer data, and then the ability to use data to ultimately meet customer needs is king.
The third new challenge is whether to make a large language model or a professional model? My point of view is that the general large language model has its advantages, the professional model also has its scenarios, and each model actually has its own ability characteristics. We believe that the general large language model is suitable for dialogue and training at the To C end, and the professional model is suitable for acquiring professional knowledge and completing professional tasks in the industry.
The fourth new challenge: Should large models be invoked in the cloud or deployed in the private domain? 10 years ago, everyone was talking about whether public cloud or private cloud was better, but we see that today the two coexist. There are similarities between cloud computing and AIGC industry development, and the general model is suitable for various small and medium-sized enterprises, which is universal and flexible; The professional model has high security, data can be controlled, it has to serve the enterprise, the data must be accurate, and different tasks must be completed. Therefore, we can predict that in the future, there will be several leading enterprises to provide the best big language models to serve everyone, but at the same time, there will also be thousands of industry private clouds, enterprise private clouds, thousands of industry models and enterprise models.
In the future, the public-owned large model will definitely become stronger and stronger, will be led by several leading enterprises, and the private model of the industry will surely bloom, which is also the business opportunity of many of our To B enterprises.
We believe that the ChatGPT phenomenon will bring us a huge AIGC dividend. Careers in writing, painting, and creation will be greatly effective. A big model is ultimately a tool, and those who can use tools will eliminate those who can't.
The smart home industry will have great development, in the past, the environment of each family was too complex, so there was no way to preset various conversations in the home environment, and now after the application of the large language model, it will greatly promote the smart home industry.
There are also personal assistant apps, which I think will grow tremendously. AI can help you book flights, order meals, change dates, make reservations, and even buy things, and we expect a lot of personal assistant applications to explode in the second half of this year.
In addition, there are also the "IP crowdsourcing" model of the meta-universe, emotional companionship products, etc., which will usher in an outbreak.
Bamboo Intelligence is an NLP professional company established in 2015, our core capability is to combine short text NLP and long text NLP processing capabilities, combined with our professional models, as well as existing large language models, to achieve different application scenarios, make different digital employees, to provide services for enterprises and individuals. We also hope to be able to encourage all players in the industry to enjoy the industrial explosion brought by AIGC and jointly do a good job in the implementation of AI in various industries.