What do you have to do to "tame" a large model that is not controlled?

Author | Huawei Wing

Interview guest|Wang Wenguang, Vice President of Daguan Data

More than a year after GPT exploded, the argument that "almost all fields need to be reconstructed with large models" has been deeply rooted in the hearts of the people, both at home and abroad. More than 200 manufacturers in China have set off a "100-model war", an endless stream of 100 billion and trillion large-parameter models, and rapid iterations of performance effects and application directions, all of which show the upsurge of large models being embraced by all walks of life. However, at a time when more industries are eager to try large models, there are also many practical landing problems emerging, and controllability is one of them.

At the upcoming AICon Global Artificial Intelligence Development and Application Conference and Large Model Application Ecology Exhibition on May 17th, InfoQ invited Wenguang Wang, Vice President of Data, to give a speech on how to improve the interpretability, operability, and controllability of large models by using knowledge graphs, RAG, and large model fusion technology from the perspective of large model related technologies and illusions. Before the conference, InfoQ interviewed Wenguang Wang about the uncontrollability of large models and the application of technical paths.

The following is an edited transcript of the interview.

Where is the "uncontrollable" large model?

InfoQ: Speaking of controllability, where is the output of large models "uncontrolled" right now?

Wang Wenguang: The content of the output of the large model is generated according to the prompt input by the user, which is determined by the ability of the model itself, and it is impossible to control the output of the model from the details. That is, large models are inherently uncontrollable. In practical application, the uncontrollability of the large model can be said from two aspects: first, the output content is consistent with the user's expectation is credible, and the inconsistency with the expectation is the so-called illusion; The second is that controllability may not necessarily be needed when using, such as writing novels and scripts, etc., even if it is imaginative, there is no big problem.

In addition, for Chinese users, there are also scenarios where controllability is highly demanding. For example, sometimes it is required that the output must be given verbatim. But controllability and illusion are different concepts, illusion is inconsistent with facts, and controllability is consistent with expectations.

InfoQ: Is the issue of controllability the biggest obstacle to the implementation of large models today? What is the effect of the existing large-scale model products in the industry?

Wang Wenguang: It can't be said to be an obstacle completely, but it should be divided into scenarios, but it is an obstacle in some scenarios with high controllability requirements, such as applications in the manufacturing and financial fields. In other words, the higher the requirements for the accuracy of the output results, the greater the impact on controllability.

I don't think the goal of a large model is controllability, but the ability of the model itself. Its level of intelligence is not strongly related to controllability, the stronger the large model, the better the controllability, but the controllability can be done in other ways.

InfoQ: From a security and compliance perspective, how can the industry work together to drive the control of large models?

Wang Wenguang: This problem is mainly solved by the provider of the large model, and it is necessary to ensure that the output content adapts to local regulations, customs, privacy and moral requirements.

The three main ways to deal with it

InfoQ: What aspects of the big model need to be addressed to solve the problem of controllability?

Wang Wenguang: There are many methods for this, and the most commonly used is RAG (Retrieval Enhanced Generation) technology, which retrieves what is needed and then inputs it into the model through the method of prompt words. Others will use the method of analyzing the activation link in the neural network, which is more difficult and very expensive, so it may not really be used too much.

InfoQ: What are the current approaches to solving the problem of controllability in the industry?

Wang Wenguang: RAG is commonly used, especially in applications, but RAG itself will also have several subdivisions. The first is the search engine, which uses this method to find the approximate range of the answer, and then inputs it into the large model through prompt words, so that it can give the answer; The second is the vector database, which uses the vector method to retrieve content, but compared with the search engine, it may also have problems such as retrieval efficiency and accuracy. Because the starting point of search engines is quite high, it is not easy to do a good job of a search engine.

In addition, the knowledge graph is used more in the industry, and its advantage is that it has a lot of predefined structures for the business, which can more easily find accurate answers, and then use large models to generate answers into a reasonable text to answer.

Mainstreamly, these three methods are used: search engine retrieval, vector retrieval, and knowledge graph augmentation. In terms of application, there are more of the first two in the general field, and the knowledge graph in the professional field is better.

InfoQ: How much can knowledge graphs improve the controllability of large models? How effective is it applied to Cao Zhi's large model?

Wang Wenguang: Knowledge graph and large model are a complementary relationship. In principle, large models essentially call them the results of inductive reasoning, while knowledge graphs are more deductive reasoning. From a practical point of view, the large model is a probabilistic output, which cannot be accurately controlled, and at the same time, even if there is an error, it cannot be edited. The disadvantages of knowledge graphs are high construction costs, there are many structured costs, and logical reasoning requires the ability to understand the business, which is what large models are good at, for example, large models can be used to build knowledge graphs and understand language. The combination of the two can just realize a highly intelligent system that can be applied on the ground. The integration of knowledge graph and Cao Zhi's large model is very good in terms of effect, and is accepted by customers in a wide range of industries such as finance, manufacturing, and energy.

InfoQ: How much does RAG improve the control of large models? How effective is it applied to Cao Zhi's large model?

Wang Wenguang: The biggest improvement direction is that using this method to improve the large model is equivalent to turning open-ended questions into multiple-choice questions. In the implementation of Cao Zhi's large model, a large number of methods of integration with knowledge graphs were used.

InfoQ: How can you avoid the limitations of RAG itself in the application practice of controllable large models?

Wang Wenguang: To implement large models, RAG technology is inevitable to encounter, and using other technical methods will only be more difficult or the effect will not meet expectations. The specific limitations depend on the approach, and each of the three directions of RAG has its own difficulties. The limitation of a search engine is complexity, and a search engine is a vast and complex system; Vector retrieval is very simple at first glance, but the controllability is very poor, and it is impossible to change it when encountering problems. Knowledge graphs, like search engines, are very complex knowledge systems, which are complex to learn, and a knowledge graph is often done for different businesses, and it is difficult to build a comprehensive knowledge graph.

Our current approach is to use all three methods in one system, and each method has weaknesses, so use other methods to supplement them. If you only have one of these methods, you will get 60 points at most, which is actually quite difficult to do well.

Relying on large models alone will never meet expectations

InfoQ: There is a lot of overlap between large models and knowledge graphs, will they replace each other?

Wang Wenguang: I don't think they can replace each other. For example, humans are already smart, but they still need to consult encyclopedias when they need precise expertise. It's the same for large models, and it's impossible for them to remember everything, especially in professional fields, so I often say that knowledge graphs are encyclopedias for large models; In addition, large models also need to be updated, and larger models will be slower to update, and training will take time. So large models always need some way to supplement information, and a knowledge base is a good choice. Therefore, I often say that books are the ladder of human progress, and knowledge graphs are the ladder of progress of large models (artificial intelligence), haha.

InfoQ: Can large models feed back into the construction and development of knowledge graphs? Can the knowledge graph based on large models be unified?

Wang Wenguang: The most direct impact is that now that there is a large model, some research directions of knowledge graphs are no longer done, such as Q&A. Because the large models do a good job in these aspects, they can be combined with each other to do it. The ensuing impact is that everyone can have more energy to do other directions of the knowledge graph, such as reasoning, which may also be a research direction that the knowledge graph will integrate with large models in the future.

InfoQ: What techniques are available now and in the future to help improve the controllability of large models?

Wang Wenguang: At present, I think the main ones are the three methods just mentioned, and the enhancement of the ability of the large model itself, such as training a technology for a specific field, which can be used but the cost is relatively high, and it seems that everyone does not do too much in the language model, and the effect may not be so good.

InfoQ: How long do you think it will take for large models to be generally accepted by the industry and the general public in terms of controllability?

Wang Wenguang: I think it is difficult to rely on large models alone, and it may never meet everyone's expectations, so we must combine the methods mentioned above. Because no matter how good the large model is, if there is no relevant content in the corpus (such as what just happened), it will definitely not be a good answer. As an example, asking about Chang'e-6 without retrieval enhancement now (May 8, 2024) is all serious nonsense.

InfoQ: What are you going to share with the audience at the upcoming AI Con?

Wang Wenguang: I will mainly talk about two parts, which are also aspects that everyone is more concerned about. First, how to solve the controllability, we will mainly combine search engines, knowledge graphs and vector databases; The second is the case we are actually doing, because the biggest problem with the large model now is how to land.

Speaker Introduction:

Wang Wenguang, currently serving as the vice president of Daguan Data, the title of senior engineer, and the elite talent of the "Pearl Plan" in Pudong New Area, has won the second prize of Guangdong Science and Technology Progress Award, the second prize of Shanghai Computer Federation Science and Technology Progress Award, and the second prize of Shanghai Pudong New Area Science and Technology Progress Award. Expert in the compilation of artificial intelligence standards, author of "Knowledge Graph: Theory and Practice of Cognitive Intelligence", participated in the compilation of "Intelligent Text Processing Practice", consultant expert and article author of "New Programmer * New Decade of Artificial Intelligence", focusing on knowledge graph, general artificial intelligence AGI, large model, AI big engineering, NLP, cognitive intelligence, reinforcement learning, deep learning and other artificial intelligence directions.

Original link: What do you need to do to "tame" the uncontrolled large model? | Interview with Wang Wenguang, Vice President of Daguan Data_Machine Learning/Deep Learning_Huawei _InfoQ Selected Articles