"Only when a technology can be commercialized can it achieve a positive cycle."
Source: Know
Author|Xu Jingjing
Rarely, Chalk held a grand launch event for a new product. Zhang Xiaolong, the founder and CEO of Chalk, couldn't even hold back the spoilers for the new product a month in advance, which shows the high level of internal attention and expectations for AI products.
At today's press conference, Chalk officially launched the first self-developed vertical domain model focusing on the vocational education industry, and announced that it will officially launch the C-end AI product - Chalk AI teacher "Chalk Head" on August 1.
Moreover, Chalk believes that this chalk AI teacher can "afford to pay the fee". Zhang Xiaolong is convinced: "Only when a technology can be commercialized can a positive cycle be realized." ”
01
"Challenges remain, but the process of reshaping education is unstoppable"
According to Chen Jianhua, CTO of Chalk, the Chalk AI teacher will accompany the students throughout the process of preparing for the exam based on the ten service scenarios (from the entrance test to theoretical learning, to intensive exercises, simulated papers, and finally to the final sprint stage).
Chen Jianhua reviewed the process of exploring AI inside Chalk.
At the end of 2022, when ChatGPT was released, the team has been thinking about three questions about how to inject vitality into business development with large model technology:
1. Is the importance of prompts underestimated or overestimated?
Chalk believes that more often than not, the importance of prompt prompts is underestimated, and a high-quality prompt can better exert the capabilities of large models. And a high-quality prompt requires not only a lot of skill and design, but also a deep understanding and abstraction of the problem domain.
2. Will AGI general artificial intelligence be realized in the short term?
Chalk's understanding of AGI is based on existing large models, and it also needs the ability to superimpose long-term memory, better logical reasoning and self-evolution. For the realization of AGI, Chalk remains optimistic for a long time, but believes that it will face greater challenges in the short term.
3. Is there a need for a vertical model?
There has always been a view in the industry that with the general enhancement of the capabilities of general large models, there is no need for models in vertical domains, and Chalk does not agree with this. Chalk believes that based on unique data, long-term accumulation of teaching and research, and in-depth insight into users, the vertical model will definitely be able to make better results than the general large model in the field of vocational education.
With a preliminary judgment, Chalk will begin to explore the combination of large models and education from the first quarter of 2023, and there are limitations in the performance of general large models in vertical scenarios such as education:
In some scenarios, the performance of the general model is not as expected. For example, in the answering scenario, the accuracy of the general large model is very low when answering objective questions such as line tests. In the proposition scenario, the general large model can only imitate the form of the question, and cannot meet the needs of the difficulty, content, and test center setting of the question.
Of course, the general model also has good points, such as interview reviews.
In summary, Chalk then has preliminary conclusions and results:
First, the large model is easy to take the lead in improving the internal efficiency of the B-side. For example, when it is applied to the interview review scenario, the content of the review is given through the large model, and then the tutor checks and outputs it to the student, which can achieve very good results.
Chen Jianhua revealed: "Interview reviews are just needed for students to prepare for the exam, which was mainly done manually by teachers before, with a large workload, low efficiency and high unit price. With the help of AI-assisted teachers to conduct interview comments, the review time of real teachers can be shortened from 20 minutes to less than 5 minutes. According to our reviews, the availability rate of AI interview reviews can also be above 90%. ”
Second, in small scenarios such as reviews, the general large model can achieve very good results and is controllable.
Third, in the education scenario with high accuracy requirements, the landing of C-end applications will face very big challenges, and the biggest challenge is the "illusion" of large models. And the "hallucination" problem is very fatal for the education industry, which has a very low tolerance for error.
From this, Chalk also further clarified the next way to go:
First of all, it is extremely necessary to develop a large vertical model.
"The vocational education examination has a comprehensive and unique examination system. In the face of this kind of investigation system, the performance of the general large model is not satisfactory. Developing a large vertical model means that we can achieve better results with smaller models and lower costs. This is very necessary for chalk and is practical. ”
Second, hallucinations can be overcome with the help of RAG retrieval aids.
In the scenario of student exam preparation, Chalk has accumulated a large amount of high-quality knowledge base data. Through RAG, the large model's understanding of the questions can be more suitable for the needs of the students, and the answers will be more accurate.
Based on this, in order to provide more professional and accurate education solutions, Chalbi is determined to invest in the development of self-developed vertical large models.
Of course, in the process of developing the vertical model, Chalk inevitably encountered some challenges.
The first challenge is intent recognition.
Chen Jianhua said: "In a relatively open environment, when the vertical model really plays the role of a teacher and answers students' questions, the way students ask questions will be much more complicated than we imagined. For example, for a question, users can not only ask questions, but also ask options, and can also extend the questions based on the knowledge points corresponding to the question. And the way the student has to ask questions variously, he may ask, 'Teacher, this question is too difficult, can you tell me again?' This involves the explanation of the topic. He may ask, 'Teacher, this question is too difficult, what should I do about the central comprehension question?' This involves answering questions about knowledge points. He may also ask, 'Teacher, this question is too difficult, how should I learn the test?' 'This is the need for learning planning. ”
Chalk's solution is to continuously clarify the boundaries of the scene and optimize the algorithm strategy after more than ten rounds of data annotation. According to reports, chalk teaching assistants and teachers have provided about 500 million dialogue templates, the chalk question bank has been practiced a total of 3.748 billion times, and users have done 61.3 billion questions. On this basis, Chalk subdivides user dialogue scenarios into 11 categories and 42 subcategories, processing more than 300,000 conversation templates, improving the ability of AI teachers to deal with complex conversations. Chen Jianhua said that after continuous data labeling and algorithm optimization, the accuracy of chalk large model intention recognition can reach more than 98%, which significantly improves the problem of large model illusion.
The second challenge is to give a precise answer based on how to avoid hallucinations.
On the one hand, based on Chalk's 10 years of teaching and research accumulation and unique data, the RAG system was built, and at the same time, CoT and the chain of thought were used to let the large model reason step by step. On the other hand, the divide and conquer method is used to strengthen the understanding of the problem by the large model, and the complex problem is split into simple enough tasks, which is far better than directly handing over the complex tasks to the model for processing.
In terms of effect feedback, Chen Jianhua revealed that according to the internal test data, taking the question and answer scene as an example, under the effect of "Chalk Vertical Large Model + RAG", Chalk AI teachers are better than mainstream general large model products on the market in terms of speech, data, common sense, judgment, and quantity.
(Image from chalk)
It is revealed that in the next step, Chalk is developing richer AI scenarios: it is expected that this year will be launched to apply for AI teachers, who will use heuristic Q&A to guide students to review and sort out key points, answer, and at the same time can be corrected and explained. In addition, Chalk is also preparing for interviews with AI teachers, AI teachers from public institutions and teacher programs.
"Challenges remain, but the process of reshaping education is unstoppable." Chen Jianhua said.
02
Behind the Birth of Chalk AI Teacher: How Does the Team Collaborate?
At the press conference, Liu Shuai, vice president of Chalk, talked about the birth process of Chalk AI teacher from the perspective of team building and teaching and research.
In terms of team building, "in order to enable the rapid implementation of AI teachers, we first built a very professional team, including teachers with rich teaching experience, tutors who provide users with guidance and Q&A work on the front line all year round, as well as teaching and research staff who have been deeply engaged in teaching and research all year round, operation colleagues who know how to communicate with users very well, and a large number of customer service colleagues who provide consultation for users." ”
As an example of solving the user's intent recognition mentioned above, Liu Shuai said: "Our tutors provide users with Q&A work every day, and they are very clear about the user's expression and understand the user's intention. At our peak, we had 3,013 tutors working on both tutoring and Q&A. After collecting the questions and intentions of so many students, they will be synchronized to the teaching and research and the teachers of the big course, and they will give professional classification and professional answers. ”
Next, the team was tapping how AI could be smarter and more content-sensitive.
First of all, it is necessary to transform the professional knowledge in the field of public examination into a learning language that AI can recognize, build a knowledge framework for AI teachers, and provide it with learning nourishment. Chalk has collected a large number of books, question banks, courses, knowledge bases, exams and other content. Secondly, in order to allow AI teachers to give targeted learning suggestions and tutoring for users with different portraits, Chalk defined the ability value for each test point. In addition, starting from the foundation of the students and the situation of the students themselves, the students are finally divided into nine categories.
Next, the chalk needs to be divided into different stages for different products. Taking the system class as an example, it includes four stages: theory, strengthening, brushing questions, and sprinting. Each stage has different learning objectives and learning tasks. How to push his learning tasks and goals to users at each stage, and what kind of supervision plan to provide, has become the focus of Chalk's work. "We will eventually take into account the completion of the students' learning and the acceptance of the students, which is another difficulty in our work."
At present, Chalk AI Teacher has launched seven major functions, and Liu Shuai introduced three of the key functions:
About the course context function. In order to realize the context of the course, it is necessary to have a clear transition between knowledge points and knowledge points in the teaching process, and a clear cut between topics and topics. Only in this way can the AI teacher quickly identify and locate the location of knowledge points and questions, and accurately push them when students raise doubts.
About the question answering function. When Chalk builds the question bank, all the easy-to-mistake questions and easy-to-make options have been marked, which can facilitate the AI teacher to quickly locate the wrong points and accurately provide users with questions and answers.
About the Q&A function. Q&A is actually the most personalized part of the service. Chalk Job Assistant collects the application data in recent years, and then combines the user's own situation and position positioning, which can provide users with personalized and targeted guidance and suggestions for the application.
Liu Shuai said that Chalk AI Teacher will also be one of the highest-end products of the mainstream of chalk in the future.
In terms of internal efficiency, Liu Shuai gave an example, in February 2021, a chalk tutor could only serve 40 students at the same time, and in February 2024, a tutor can serve 120 students at the same time, and at the peak, one teacher can serve 150 students.
"We hope that in the near future, AI teachers will not only be able to provide Q&A tutoring, but they will also be able to appear in the classroom as a virtual teacher like a real teacher and present us with the teaching of various courses." Liu Shuai looked forward.
At the press conference, Liu Shuai also sorted out the evolution of the public examination training industry: from the initial offline 1.0 stage, to the 2.0 stage of online large classes, to the 3.0 stage of online and offline combination, and then to the 4.0 stage supported by AI technology.
In this evolutionary process, Chalk has gone through the 2.0 stage of the system class, and in 2019, the 3.0 stage of the OMO boutique class was launched. "Next, with the application and upgrading of AI technology, chalk will gradually enter the 4.0 stage in 2024."
03
Make an AI product that "can afford to pay for": "Only when a technology can be commercialized can a positive cycle be realized"
Zhang Xiaolong's view has always been that a modern enterprise should be systematic and product-oriented (not particularly personal). Over the years, Chalk has also been moving along this line of thought: from the early system classes to today's Chalk AI teachers, they are all typical products that combine new technology and high-quality content.
He believes: "Science and technology liberate people and serve people. After it liberates people, it makes people's lives more interesting and meaningful, so that people can do more creative things. New technologies may be painful in the short term, but in the medium to long term, they will create more different types of jobs and jobs. ”
At the end of the press conference, Zhang Xiaolong emphasized that Chalk AI Teacher, as a C-end AI product, is "charged" and "can afford to pay".
"We believe that this product can afford to pay a fee, and only when the business model is established can the company continue to invest in technology. Only when a technology can be commercialized can a positive cycle be realized. China has very broad application prospects, and commercialization, in turn, can accelerate technological progress. Zhang Xiaolong said. It is hoped that the positive cycle of technology in specific fields and vertical fields will be realized as soon as possible, so that various industries and fields can be better and faster development, especially in the application field. ”
"Friends in China's science and technology field, don't all focus on the general model, we can have more specific services for specific fields, realize commercialization as soon as possible, and make contributions to China's social economy and employment." Zhang Xiaolong finally suggested.
END
The author of this article is Xu Jingjing
"Edtech This Year · 2022" + "Training Industry This Year · 2021" + "EdTech Industry Atlas 2022-2023", on sale!