To make answers more credible, graph technology will become the next step in the development of large language models (English: Large Language Model, abbreviated LLM).
To look back at the accelerated development of large models, we need to push the timeline back to December last year, when ChatGPT became a phenomenon on a global scale. The user side is driving the market side to explode rapidly. As of now, nearly a hundred major language models have appeared around the world.
When dozens of large domestic models appeared, ChatGPT’s attention was quickly divided. According to SimilarWeb data, the growth rate of ChatGPT’s visits in the early stage was astonishing, with a month-on-month growth rate of 131.6% in January. In the later period, the growth rate gradually slowed down as time went by. Global desktop and mobile traffic to the ChatGPT website fell 9.7% in June compared with May, while unique visits to the ChatGPT website fell 5.7%. Sameweb data shows that the time visitors spend on the site also dropped by 8.5%.
The usage rate of ChatGPT is gradually decreasing, which is inseparable from its difficult to guarantee accuracy. Many users have reported on social platforms that during the current AI chat process, they sometimes receive answers that seem correct but are ridiculous when read carefully. Or, fine-tuning the way they ask questions will result in completely different answers, and some even provide Chaotic information is difficult to discover intuitively and requires certain industry knowledge to discover.
In fact, correctness has become an unavoidable issue in the development of all large language models. Especially when large models need to be applied to business scenarios in industries such as medical care, health, insurance, banking, finance, and industrial manufacturing, the requirement for fault tolerance is close to zero. Because once a large model provides wrong or even biased information, it may cause huge legal liability or public relations crisis.
Therefore, from the perspective of actual development, the battle on the battlefield of large language models will surely enter the next stage - allowing the answers to get rid of serious nonsense and become 100% credible.
Talking nonsense seriously has become one of the biggest loopholes in the big model
After consulting with people in the large model industry, the author learned that the currently popular large models are mainly GPT-type models. Most of these models are pure decoder structures, using the n-gram method and based on neural network modeling with parameters Θ. According to The first k words are used to predict the k+1th word. During the pre-training process on massive corpus, the model can learn the language rules, effective information, etc. in the corpus, so that it can output reasonable relevant information according to the given prompts in the generation stage.
However, in the process of practical application, the "illusion" phenomenon of large models has become a key reason that hinders its widespread application.
The phenomenon of "hallucination" means that the model outputs wrong or untrustworthy results with high confidence. When the model has "hallucination", the training data it uses cannot prove the rationality of the output. Behind this, there may be a connection with the corpus of large models that are complex, diverse and full of errors. Gartner, a well-known international analysis agency, pointed out in the latest research report: "The use of ChatGPT cannot list citation sources, and its reliability is mostly based on source information. However, these sources themselves may be wrong and inconsistent." At the same time, Tencent Group Senior Executive Vice President Tang Daosheng also said frankly in a public speech: "Currently, general large models are generally trained based on a wide range of public documents and online information. The information on the Internet may have errors, rumors, and biases, and a lot of professional knowledge Insufficient accumulation of industry data results in insufficient industry pertinence and accuracy of the model, and high data noise.”
It can be seen that solving the "illusion" phenomenon has become a key threshold for the development of large models, especially for the industrial application of large models. As long as there is a 1% deviation, it will plant the seeds of risk for the enterprise and become the threshold for implementation.
In this context, graph technology has the opportunity to break the deadlock and give its own correct answer.
Let connection be the key to breaking the situation
If large models are an important "alarm bell" that wakes up data, allowing enterprises to release business value from their silent data through analysis, then graph technology is the key pointer on the alarm bell, using structured and orderly data associations to allow Silent data releases value rationalization, empowering large language models from the bottom to become more efficient, more accurate, and more intelligent.
In this article, we first clarify the concept of graph technology. Graph technology uses "point-edge" as the data structure to intuitively and concisely describe the complex relationships between business entities. It takes the storage, query, and computing performance optimization of complex relationships as the first design principle and can mine data from graph-structured data. Valuable knowledge or rules to guide business decisions.
In the process of large model development and application, graph technology can reduce the occurrence of hallucinations.
As mentioned above, the application principles of large models mostly revolve around Transformer technology, which uses training data to condense into generalized knowledge and then precipitate the next character. But large models fail to recall facts and generate representations with incorrect facts.
Graph technology is based on graph theory, a mathematical theory used to study graphs (data structures composed of vertices and edges) and their properties, through structured linked data for calculation and analysis. When graph technology is deeply integrated with large models, the generalized knowledge in large language models can be abstracted and used to create knowledge graphs, and the relationships between entities can be effectively captured with the help of data association, allowing for deeper reasoning, retrieval and analysis of data. The logical relationship can effectively improve the knowledge understanding, accurate output and logical reasoning capabilities of large models, fill the logical loopholes of large models, and achieve the accuracy, interpretability and transparency of large language models and other generative AI results. .
Objectively speaking, with the help of graph technology, the current fierce competition for large models will reach the next milestone.
Graphics technology will become the ticket to the next round of large model competition
The technological competition may seem long, but in fact it changes rapidly, and those entering the game need to grasp more possibilities. Looking at the current international market where large models and graphics technology are deeply integrated, we can get a glimpse of the future.
Paying attention to foreign countries, we can find that mainstream large models overseas have begun to use graph technology to manage Context and Prompt, and deeply integrate the concept of graph into the pre-training process, so that traditional large models have better reasoning, logic and Explainability. It is understood that Neo4j, the leading graph database manufacturer, has announced a cooperation with Google Cloud's large language model Vertex AI. Enterprise customers can use the Neo4j product in Google Cloud Platform to build knowledge graphs to obtain more accurate, transparent and explainable generative formulas. AI insights. In the academic field, foreign research on papers related to large models and industry knowledge graphs is also at the forefront of the world.
Compared with the first implementation of foreign graphics technology, domestic large model manufacturers have not yet made large-scale investments in the field of graphics technology. However, there are also graphics technology manufacturers that are actively exploring the combination of technologies and application implementation between the two.
Chuanglin Technology, a quasi-unicorn in the graph database, took the lead in leveraging its industry-leading technical concepts and practical experience. In early 2023, it announced for the first time that it would “formally access Baidu Wenxin Yiyan’s capabilities to create a full range of graph analysis artificial intelligence products to achieve Graph Intelligent Analysis” . And recently, Chuanglin Technology has once again cooperated with Huawei's newly released Pangu large model to jointly promote the joint development of graph technology and large models. It is understood that Chuanglin Technology will further integrate graph technology and various large models in the future, give full play to the high performance, easy expansion, real-time and other advantages of the domestic Galaxybase graph database, and break the gap between domestic graph technology and general large models and industry large models. Landing bottleneck.
Summarize
For large models, overcoming the "illusion" phenomenon has become a must, and graph technology will surely become a key factor in overcoming this challenge. In the arena of large models, the combination of graphics technology will also become a new ticket for large model manufacturers.
If you want to know more about the practice and achievements of Chuanglin Technology's Galaxybase graph technology in the field of large models, please follow Chuanglin Technology's Galaxybase official account and leave a message for consultation. Let's work together to find new growth points for large models.