Editorial Department Organized from meet2024 qubit | qbitai
Whether the facts of a legal case are clear and how to determine the relationship between it and the law are all logical. β
At the meet2024 intelligent future conference, Li Dahai, co-founder and CEO of Facewall Intelligence, emphasized the importance of logical reasoning ability when talking about the landing of large models in the legal industry.
He believes that the AGI revolution is the fourth technological change alongside the steam revolution, the electric revolution and the information revolution, and as the core technical route of this change, the most important thing is the logical reasoning ability that the large model can be truly applied to the production process.
As the earliest team to make large models in China, Facewall Intelligence has done a lot of meticulous work on logical reasoning in the process of model training, and has divided it into multiple dimensions including induction, deduction, time, and space, and has specially overcome and improved them one by one.
According to reports, the 100 billion multi-modal large model CPM-Cricket of Facing Wall Intelligence can be benchmarked against GPT-35 level, and at the same time the logical reasoning ability is very outstanding. In the logical reasoning test of the public examination, the total accuracy of the CPM reached 6376%, even more than GPT-4's 6188%γ
However, as all industries and enterprises see the application prospect and value of large models, how can we unleash the greater potential of large models and promote the development and transformation of productivity?
The answer given by Li Dahai is "large model + agent".
He likened the large model to the engine of the car, which provided the power for the car. But if you want to build a car, you also need a steering system, you need a car chassis, and all other components including the interior to really provide users with a complete car product.
Li Dahai believes that on the basis of the large model engine, a series of upper-layer technologies need to be superimposed, including memory ability, use of tools, etc., so as to bring more and more extensive applications and imagination space, and AI agent (agent) is the carrier of this series of technical capabilities.
In order to fully reflect Li Dahai's thinking on the "large model + agent" track, Qubit edited and sorted out the content of his speech on the basis of not changing the original meaning.
About Meet Smart Future Conference: Meet Conference is a top business summit in the field of intelligent technology hosted by qubits, dedicated to the implementation and industry application of cutting-edge technology technology. This year, a total of dozens of mainstream ** and live broadcast platforms reported and broadcast the MEET2024 conference live, attracting more than 3 million industry users to participate online, and the total number of ** on the whole network exceeded 20 million.
Presentation Takeaways. The large model should be able to be really used in the production environment, the most important thing is the logical reasoning ability, the large model is a new technological revolution, which can be compared with the industrial revolution, the power revolution, and the information revolution, the large model is the underlying engine, and if you want to do a good job as an agent, you need the engine to provide strong power output, and the most fundamental change brought by the large model is the change in the relationship between man and machine, and the two will become more equal in the future, and the future world will be connected by agents The world of agents (the following is the full text of Li Dahai's speech).
Large models should have logical reasoning as their core competency.
Facewall Intelligence is a large model company incubated by Tsinghua University's Thunlp Laboratory, and our team began to cultivate the field of large models before the company was officially established. At the same time, we also launched the open-source community of OpenBMB, in which we have made a lot of open-source technologies and tools for large models.
Together with Tsinghua University and the OpenBMB open source community, we have built a "one body, two wings" industry-university-research large model ecology.
Face Wall Intelligence is the earliest team to make large models in China, and the world's first Chinese pre-trained large model was launched by our team in December 2020, called CPM-1. In November 2023, we released the latest 100 billion multimodal large model CPM-Cricket, which can be benchmarked against GPT-35 levels.
The core capability of the model lies in logical reasoning.
From the very beginning of its establishment, we have realized that the most important thing for a large model to be truly applied to the production environment is the ability to reason logically. Therefore, in the process of model training, we have done a lot of detailed work on logical reasoning, which is divided into multiple dimensions including induction, deduction, time, space, etc., and specially tackled and improved one by one.
In order to test the logical reasoning performance of the model in the real scene, we simulated the test questions of the public examination in the past three years for the large model, and the results showed that the overall accuracy of the CPM reached 6376%, even more than GPT-4's 6188%γIn the English GMAT test, the score of the large model of face-to-wall intelligence is 93% of GPT-4, which is very close, and some question types even exceed GPT-4.
Recently, we have seen in our work with clients in the legal industry that use cases are very focused on and dependent on the logical reasoning capabilities of the model. Whether the facts of the case are clear or not, and how to determine the relationship between the facts and the legal provisions, are all about logic. After the evaluation of industry customers, the logical reasoning performance of the large model of Facewall Intelligence is the most prominent, which is exactly in line with the actual needs of customers.
Today, I believe that the technical route of the large model has formed a consensus in the entire industry, but everyone must still think about whether the large model is a technology wave like web3 or a ten-year industrial change.
In my opinion, the AGI revolution with large models as the core is the fourth major technological change, which can be compared with the steam revolution, the electric revolution, and the information revolution, and will last for at least 20-30 years. In a few years, the production and life of the entire human society will be turned upside down because of the evolution of the AGI revolution.
Large model + agent" to create more imagination space.
The large model is like the engine of the car, powering the car. But if you want to build a car, you also need a steering system, you need a car chassis, and all other components including the interior to really provide users with a complete car product.
Therefore, on the basis of the large model engine, it is also necessary to superimpose a series of upper-layer technologies, including memory ability, use of tools, etc., so as to bring more and more extensive applications and imagination space, and AI agent is the carrier of this series of technical capabilities.
The agent has the typical characteristics of six dimensions: personality, IQ, emotional intelligence, perception, values and growth, so as to adapt to various application scenarios. At the same time, in order for individual agents to exert more powerful capabilities, they also need to be connected and coordinated to handle and complete more complex tasks.
In fact, there are typical cases of swarm intelligence in both human society and nature. Just as we need teams and organizations to bring individuals together, bee colonies, ant colonies, and fish colonies in nature also exhibit higher levels of intelligence than individual individuals.
Based on these thoughts, Facewall Intelligence has begun to lay out the technical route and landing direction of "large model + agent" since June 2023, and in the past few months, it has released a series of AI agent application frameworks driven by large models, which we call the "AI agent troika" of Facewall Intelligence.
The first is the agentverse agent general platform.
It constructs a rich virtual space in which a large number of agent experts are defined, with different personas and professional capabilities.
When a user makes a request, the agents immediately start the teaming process. This is a strategic recruitment process to determine which experts should be committed to a particular task. Once these experts form a team, they begin to negotiate with each other on the details of the task and clarify the division of labor. After the negotiation is completed, it moves to the execution phase, where each agent completes the corresponding work according to its role, and then integrates it.
There is also a strategic planner throughout the process to ensure that all agents work together to form a final product, which is compared with user needs, and iteratively improved if there are large deviations. The universal nature of this framework allows us to build on it and carry out a wide range of work.
The second is the Xagent super agent application framework.
It is a superintelligence that can disassemble complex tasks and perform task distribution based on dynamic instructions. It acts as an agent expert, planning according to the needs of the person and accomplishing the goals proposed by the user.
On the basis of this plan, if the user does not enter enough information, it interacts with the user and gathers the necessary information.
At the end of each step of the plan, Xagent also evaluates whether additional work is needed after each step is executed, and the whole process is a dynamic structure. In benchmarks, Xagent's capabilities have surpassed AUOTGPT across the board.
For example, when you send a command to xagent, "I have friends visiting on the weekend, please recommend a few restaurants for me", the superagent will not immediately list a long list of restaurants, but will first ask you about your preferences, asking if you prefer a quiet environment or a specific type of diet to understand your needs.
Its first step is to interact with you, not to perform tasks immediately; Next, conduct a restaurant search based on your response; Then, collate the search results and come up with several scenarios with pros and cons analysis. When the scenario is ready, it's presented as a visualization for you to choose from. Once you've made your choice, it will book the restaurant directly for you via a connected API.
This is different from the one-step question and answer model that we are usually familiar with, and it shows a better quality experience provided by agents.
The third is the chatdev multi-agent collaborative development framework.
It can help us build a virtual AI software company, set up agents with different roles such as CEO, CTO, product manager, programmer, designer, etc., and connect them through a communication network called a "conversation chain".
The interaction process of these roles aligns with the waterfall model of software development, including software design, system testing, and documentation.
We let these AI agents cooperate according to a clear division of labor, and communicate and interact through natural language, with an average of less than 3 minutes and an average cost of less than 3 yuan, and a simple software development can be completed "less than the time and money of a glass of Coke".
In this work, we also made a function called HAI (Human-Agent-Interaction), which allows people to interact with the agent through natural language, and in this way, the agent can get enough data feedback to make their work better, which is an exploration of the growth of the agent we just mentioned.
In just over two months, the number of ChatDev stars on GitHub has skyrocketed, exceeding 180,000, ranking first in the trending ranking for many consecutive days.
The coupling between the large model and the agent is very important.
Facewall intelligence needs to do both the large model and the agent at the same time
Because the large model is the underlying engine, if you want to be a good agent, you need the engine to provide strong power output.
In practical applications, if you find that the agent may not work well enough in which directions, you need to have real-time feedback at the bottom layer to be able to modify and optimize it in time. There are many limitations to relying solely on external large models, and the coupling between the two is very important.
At present, we have many scenarios on the technical route of "large model + agent".
In the financial scenario, Facewall Intelligence has carried out in-depth cooperation with domestic leading bank customers, and provides customers with strong language dialogue and logical reasoning capabilities for their "intelligent wealth assistant" products through large models, and answers users' professional questions in financial consulting and other business consulting.
In the legal scenario, Facewall Intelligence creates a large legal model, combined with agent technology to provide legal personnel with a powerful assistant, and assists in extracting key points of the case, clarifying the facts, sorting out legal terms, etc., so as to greatly shorten the time of case processing and improve efficiency.
There are about 30 million cases in China every year, but the number of grassroots legal personnel is very limited. After interviews, we found that on average, each grassroots legal staff has to deal with 3-4 cases per day, but there are many complex cases, and the processing time will be very long, and a case that needs to be litigated may be scheduled.
Three or four months later. In this case, a legal model is needed to better help the legal department greatly improve the efficiency of case handling.
The most fundamental change brought about by the large model is the "change in the relationship between man and machine". It enables machines to interact with each other in the same way as humans through natural language, and to think logically and handle complex tasks.
The emergence of large models will make the relationship between machines and humans more equal, everyone can have an AI partner, and "large model + agent" can help people create a more intimate and understanding of your life intelligent partner.
We believe that the future world will be a world linked by agents, that is, a new era of "Internet of Agents" proposed by Facewall Intelligence.
The big model is the fourth technological revolution. Every technological change takes a very long time: the first industrial revolution took hundreds of years, the information revolution took about 50 years, and I believe that the AGI revolution will not be short.
2023 is only the first year, and I believe that there are still many problems to be solved and breakthroughs, including task planning, multimodality, higher-level cognition, and smaller models.
I hope that all partners in the industry will work together to overcome and solve these problems, so that the whole society and people's lives will be better.