Is it reliable to let a big model serve as a general tutor for children?
Tutoring children in their homework is a "practice" that every parent is destined to experience. "Why can't I solve a question after I have taught it several times?" "I have a bachelor's degree. I thought I could tutor my children at home, but I didn’t expect that I hit a wall in the first grade of junior high school.”I believe everyone must be familiar with the context of these words. After working all day, dragging your tired body to tutor homework in person gives many parents a real headache. In addition, as the children advance in grade, the knowledge points in the courses become more and more abundant, and it becomes more and more difficult for parents to help their children with their homework. Parents can still give you some pointers on some simple questions, but when it comes to "general tutoring", many parents do not fully understand the knowledge and may not be able to answer their children's questions well. In fact, when faced with this problem, parents may wish to seek outside help. Just in the past two days, everyone must have heard that a new star tutor has appeared in the field of education-"Little P Teacher". "Little Teacher P" is not a real person, it is an "AI tutor" who can answer questions in all subjects. It is said that there is no question type that "Little Teacher P" cannot answer in 10 subjects including Chinese, Mathematics, English, Physics, Chemistry, and Students from elementary school to high school. The most valuable thing is that even if it is a mathematical, physical and chemical problem that involves multi-step reasoning and is difficult to understand, "Little P Teacher" can patiently guide you step by step. Formula explanation. In this regard, it is indeed stronger than the mentality of most parents... Children can send text, voice, and picture questions to "Little P Teacher", just like Chatting on WeChat with a real tutor is as natural as chatting with a real tutor online, and the replies are more timely than real tutors, and they are online 24 hours a day. The question is, how to contact "Little P Teacher"? After a preliminary understanding, we know that the "resident" of "Little P Teacher" is the newly released Youdao AI learning machine X20, and its super teaching capabilities are derived from "Ziyue Education Model 2.0". 
Youdao AI Learning Machine The first landing application results. Compared with the first-generation version, Youdao's Ziyue Education Model has been greatly improved in many capabilities such as oral dialogue, educational scene knowledge Q&A, and word processing.
So, how exactly does "Little Teacher P" perform in teaching? Can you afford the title of "AI general teacher"?
#This website experienced Youdao AI learning machine X20 for the first time and conducted an investigation from multiple dimensions. #One-on-one, what is the level of "Little Teacher P"'s lectures?
# As soon as I came up, I asked "Little Teacher P" to tell me a question. "One-to-one lectures" are also the most important aspect for most parents. It affects whether their children can understand clearly, whether they are engaged in learning, and whether their grades can be effectively improved.
Every child’s learning style and blind spots in thinking are different, and unified classroom teaching cannot match everyone’s progress. Therefore, what children need most is not learning materials, but a teacher who knows how to teach students in accordance with their aptitude. It is precisely because of this that parents provide tutoring in person or hire one-to-one tutors to provide targeted tutoring for their children.
Now, "Little P Teacher" may be able to play this role: based on the multi-round interaction capability of large models, "Little P Teacher" can support up to 8 people at a time Take turns to talk, and the children can ask if they don’t understand anything. And when faced with questions, "Little P Teacher" not only gives a brief standard answer, but also provides personalized topic services such as analysis of knowledge points, step-by-step breakdown of ideas, and inferences.
We can try a math application question:
Here, "Little Teacher P" took the Question-based heuristic teaching aims to stimulate children's thinking and initiative. First, the children are reminded that this is a "cyclical repetition" type of question, and then the ideas and methods for dismantling the information are given: first determine how many beads there are in a cycle, and then calculate which cycle the 600th bead falls on. Determine the bead's position in the cycle. 
It doesn’t matter if the child doesn’t understand it yet, “Teacher Little P” will further inform you on how to solve the problem: ## After solving the problem according to the method, "Teacher Little P" will help verify the answer and confirm the thinking process:
Finally, students will be asked about "cyclically repeated questions" ” curiosity, “Little Teacher P” can also introduce background knowledge and summarize key points. 
In the process of personalized teaching interaction with "Little P Teacher", the child's learning effect has been maximized. In this process, we can also feel the difference in style between "Little P Teacher" and other large conversational models. It understands the process of young people's learning knowledge and cognitive ability development better, and is willing to use a simple and easy-to-understand language to communicate with children. "Communication" problem-solving ideas. In the past, most applications of AI technology at the course tutoring level only provided adaptive learning based on rules and knowledge points, while generative large models brought intention understanding, natural language Breakthroughs at the communication level are expected to achieve true "teaching students in accordance with their aptitude." On the one hand, large models can better understand open conversations. For example, in the process of asking questions, children's expressions may not be so clear. It may be difficult for previous AI to understand such open expressions, but today's large models have the ability to accurately understand the incomplete expressions of teenagers during the learning process, and The obtained information is matched with the content in the knowledge document, and targeted responses are given. On the other hand, with the breakthrough of context window technology, large models can understand context more comprehensively and handle more complex tasks and requirements. In the interaction with children, the longer context window of Ziyue Education Model 2.0 helps "Little P Teacher" better process the contextual information of the conversation and generate more coherent, accurate, and comprehensive answers. The obscure "Li Sao", "Teacher Little P" actually made me understand itWe know that for children who are accustomed to the modern Chinese language system, it is often difficult to fully understand ancient works. There must also be many difficulties in the process of answering classical Chinese questions: sentence segmentation problems, common knowledge background of ancient culture, and the confusion of Tongjia characters. , translation skills, etc. There have been researchers who have conducted evaluations on multiple large models, among which ancient Chinese is one of the most challenging tasks. Although ancient Chinese is difficult to understand, it is not an insurmountable problem for large models who are already good at "language". Specifically, by introducing high-quality proprietary data sets, customizing the basic language model through supervised fine-tuning, and then improving it with the help of human feedback reinforcement learning methods, we can quickly make up for the model's lack of ability in ancient Chinese. The advantage of "Little P Teacher" is that it can not only translate it into modern Chinese, but also provide explanations of cultural background knowledge and analysis of knowledge points. It is said that "if you drink hard and read "Li Sao" well, you can be called a famous person", but this long lyrical poem from the pre-Qin period is too difficult to understand, so I will Let "Teacher Little P" help me interpret it: Not only ancient Chinese, but also any long text content, large models Their understanding ability is stronger than that of ordinary human teachers. This is because large models are good at understanding and processing long pieces of text, large-scale data sets, or large knowledge bases. Even if the content of the teaching materials is constantly updated and the background knowledge involved changes, the large model can achieve continuous learning. Multi-modal input question, is it possible? As we all know, the question forms of various subjects are very diverse. For example, math questions include formulas and geometric shapes, physics questions include force analysis diagrams, and chemistry questions include Molecular structure, etc. This requires "AI tutors" to have the same multi-modal content recognition capabilities as human teachers. Over the past year, large model technology has made many breakthroughs in multi-modal tasks, but there are still many forms of multi-modal recognition in real scenes. Challenges include text not necessarily appearing in standardized print, and factors such as the “deformation” and “shadow” of the input image making it more difficult to identify and understand. For another example, "multi-image intertwined dialogue" may appear in test questions from different disciplines. The model needs to compare multiple input images and understand the questions related to the images, and generate answers and explanations based on this. One of the highlights of "Little P Teacher" is that it supports full input scene recognition, including inputting questions in text, pictures, audio and other forms. For example, if you send a blurry picture to "Teacher Little P", you don't need to ask additional questions. It can immediately analyze the meaning of the question and the solution. 
This function stems from the progress in multi-modal capabilities of "Ziyue Education Model 2.0". Especially for tasks involving complex image and text input, "Little P Teacher" performs equally well. #The reason why large models can demonstrate powerful image and text understanding capabilities is closely related to their own multi-modal context learning (in-context learning) capabilities. Compared with traditional visual reasoning models, multi-modal large models have learned rich open world knowledge during large-scale pre-training and can generalize to unseen objects or concepts with zero or few samples. . Specifically, large models are able to transfer tasks to unseen problems as long as they are given a few image-text pairs as examples during training. For example, when faced with math, physics and chemistry test questions that contain images, "Little P Teacher" can not only identify the text in the image, but also accurately capture the semantic association between the image and the text. . It seems that parents can finally be liberated. Of course, with the all-powerful "Little P Teacher", we do not recommend that parents completely let go of their children's learning. After all, this is also an important part of the closer parent-child relationship (dog head). #Is this the time when AI takes over the teaching of all subjects? #So far, we have basically understood the teaching skills of "Little Teacher P". However, there is another difficult challenge in the child's learning process: "universal curriculum tutoring". If any AI solution wants to be better implemented in the field of education, developing general cognitive abilities is a key point that must be broken through. The "AI general tutor" must have a massive knowledge reserve and be well integrated in order to be able to deal with problems in various subjects. Since this year, GPT-4 has taken the lead in proving that large models can master multi-disciplinary knowledge and have achieved good results in various examinations such as bar qualification, SAT and GRE. results. Nowadays, we know that this "all-rounder" can be used not only to pass exams, but also to play a huge role in the field of education. Just like "Little P Teacher", it can provide students with personalized learning support, answer questions, and provide explanations and guidance for any subject in primary and secondary schools. Whether it is GPT-4 or "Little P Teacher", the reason why it is capable of all subject tasks is essentially inseparable from the characteristics of large models that are "good at learning knowledge" open. Human cognition is limited by the nature of biological intelligence, resulting in a ceiling in cognitive ability. For ordinary people, it is not easy to be proficient in teaching in a certain field. The large model is a rather huge knowledge container, which can refine and absorb massive amounts of multidisciplinary knowledge and acquire very powerful cognitive abilities. At least it can become a "generalist" in the field of primary and secondary school curriculum teaching. There are currently some effective methods for making large models "universal": one is to use a specified data set to test large models Make fine-tuning to gain general ability in a short time; there is also a more popular method recently, which uses RAG (Retrieval Augmented Generation) technology and information from proprietary data sources to supplement text generation, allowing large models to quickly absorb plug-ins. Knowledge base to keep knowledge updated and iterated. Thinking a step further, this universal ability of large models also qualifies them to play a "grading" role in teaching similar to that of human teachers. As long as standardized evaluation criteria and a step-by-step evaluation process are written out, the large model can quickly adapt to the evaluation process, help children correct mistakes, and develop a better way of thinking. #In the long run, "AI full-subject tutoring" supported by large-scale model technology will definitely bring profound changes to the education industry. Having said that, the learning machine is an electronic product after all, and parents should also consider it from the hardware perspective. Youdao AI learning machine X20 has also made efforts in all aspects: in order to simulate the learning experience of paper materials, X20 uses a customized ink mode with a high-definition matte-plated eye protection screen, superimposed with six eye protection functions. X20 is also equipped with a 4096-level pressure-sensitive smart writing pen, which fully protects children's vision while they are immersed in solving questions. #X20 also pioneered the "three-in-one" model. It is not only a learning machine, but also a training machine, and it is also a child's exclusive computer. In learning machine mode, X20 provides "5.3" strategic cooperation resources, and has built-in massive genuine test questions from the same source for the high school and college entrance examinations to choose from; in computer mode, you can use the original dedicated intelligent keyboard and built-in keyboard input practice software to standardize typing fingering. In addition to "Little P Teacher", the new generation of virtual oral private tutor Hi Echo is also a large-model native application built into the X20, and has completed the virtual Comprehensive upgrade of human images, scenes, dialogue capabilities, and teaching modes: Hi Echo version 2.0 adds a new "speaking grading" function, and has new virtual human images for users to flexibly choose.In addition, taking into account students' learning progress and English proficiency, Hi Echo 2.0 designs exercises with more diversified and in-depth conversation scenarios. At the same time, based on the technical support of the "Ziyue Education Model", Youdao AI learning machine X20 also provides intensive grammar teaching for all question types, including reading comprehension, composition correction and other functions. 
In this year when large models were surging, there were Many researchers have pointed out that the education field is one of the best scenarios for the implementation of large models. Now, we have seen the future form of large-scale model implementation in the field of education from "Little P Teacher". In terms of teaching process, it changes the traditional way of "simply distributing learning resources" and provides a way to guide children to develop independent learning abilities through discussions and questions and answers. It can also "prescribe the right medicine to the case" based on the questions asked and push relevant knowledge in a targeted manner, achieving a leap from "one person has one face" to "one thousand people have one face". This is the educational concept we have been pursuing for thousands of years. In terms of teaching mode, because large models have strong knowledge transfer and combination innovation capabilities, they can achieve comprehensive understanding and draw inferences about different subjects and skills, making the relationship between AI and education The combination is no longer limited to a certain discipline or link. The generalization ability and generation ability emerging from the general large model have greatly reduced the labor costs required for teaching work, and will also give birth to a new form of education. At the same time, we can see that generative large models continue to make breakthroughs in more fields, such as text-based pictures, text/pictures-generated videos, text/pictures-generated videos, and text/pictures-generated videos. Generate 3D and other directions. These new technologies are also expected to be integrated into the capabilities of "AI tutors" and derive new multi-modal communication methods, making AI application scenarios in the education field more diverse and exerting more value. With the continued progress of generative AI technology, a comprehensive revolution in the educational paradigm is bound to occur. The above is the detailed content of The comprehensively upgraded "General AI Tutor" is officially put into use, ushering in a revolution in family education. For more information, please follow other related articles on the PHP Chinese website!
