Smart stuff
Author | Cheng Qian
Edit | Shadow of indifferenceZhidong reported on June 10 that yesterday, iFLYTEK Xinghuo Model V1.5 was unveiled, upgrading the three comprehensive capabilities of open-ended knowledge question and answer, logical reasoning and mathematical ability, and multi-round dialogue, and released the Spark APP that supports pure voice input and multimodal input. At the same time, iFLYTEK also announced the new progress of the application of the Xinghuo model in the office, education, medical and industrial fields.
The text generation, language understanding, knowledge question answering, logical reasoning, mathematical ability, and code ability of the Spark model are all continuously improving, among which the knowledge question answering ability is increased by up to 24%, and the logical reasoning ability is improved by 10%.
In fact, the results of the combination of the Spark model with education and office scenes have already appeared. On May 6, iFLYTEK Spark model was released and integrated into education and office scenarios, iFLYTEK President Wu Xiaoru revealed that from June 1 to 8 this year, the number of activations of iFLYTEK AI learning machines with functions such as oral sparring, writing assistant, and composition correction increased by 214% year-on-year, and the number of user activations increased by 176% and 205% in the user activation of iFLYTEK office books and iFLYTEK hearings equipped with functions such as regular discourse and one-click drafting.
It can be seen that the new user experience is greatly activating the needs of users.
It is worth mentioning that on June 9, the last day of the national college entrance examination, Liu Cong, president of iFLYTEK Research Institute, also demonstrated the ability of the Xinghuo model to answer the college entrance examination papers, and did mathematical function problems and language reading comprehension questions, which can give a logical and clear analysis process.
On the occasion of the release of the Spark model, Liu Qingfeng, chairman of iFLYTEK, announced the three upgrade points in the year of Spark, and will overtake ChatGPT at the end of October. This is also the only domestic large model manufacturer that clearly gives the upgrade time. Now is the key point for its iterative upgrade, Liu Qingfeng said, because the dream of long-term ism needs to be realized one milestone after another.
First, the three comprehensive capabilities are upgraded, and users can customize 200+ "personal" assistants
iFLYTEK Spark Model V1.5 has been upgraded for the three comprehensive capabilities of open-ended knowledge Q&A, logical reasoning and mathematical ability, and multi-round dialogue that users need the most.
At the same time, iFLYTEK released the Spark APP and Mini Program, which supports full-voice dialogue, multi-modal input, etc., and simultaneously launched the "Spark Assistant Creation Center", allowing users to create large and small intelligent assistants based on their own needs.
Users who want to build an assistant from 0 to 1 need to enter the corresponding assistant instructions in the background. If you want to enrich the content of the text, users can also add keywords to the assistant instructions, such as "cheerful humor", "quoting scriptures", etc. And the newly generated assistant will also be synchronized to the PC and mobile device.
At present, the Spark Assistant Creation Center has released more than 200 assistants.
1. Open-ended knowledge Q&A, the text content can cite scriptures and supplement analytical insights
At present, large models are more difficult to solve the update of new knowledge, and the phenomenon of Zhang Guan and Li Dai is prone to occur when answering some factual questions. Based on its powerful language understanding ability, the large model can deeply understand the needs of users, extract knowledge from real-time databases and information databases, and then provide answers to users through summary expression capabilities, so as to ensure the timeliness and accuracy of results.
Wu Xiaoru said that in fact, open-ended knowledge Q&A ultimately spells the natural language understanding ability of the big model.
June 9 is the last day of the national college entrance examination, taking this opportunity, the president of iFLYTEK Research Institute asked the model of Xinghuo on the spot, "What is the essay question of the first volume of the national college entrance examination in 2023?" And analyze the meaning it conveys". The Spark model not only gives the essay topic, but also describes the content behind the question.
In terms of long text generation capabilities, the Spark large model has also been further improved. When asked, "The college entrance examination has just ended, the child is about to start a new beginning, please write a warm letter to your child", the text frame generated by Xinghuo is clear, and the text is also quoted to enrich the article.
There is also the situation of the general artificial intelligence industry, "what are the new trends in China on general artificial intelligence, and analyze China's foundation and advantages." After listing the new trends in China, Xinghuo further analyzed the advantages of talent, data, policies and other aspects, and finally added the current challenges. Liu Cong said that Xinghuo can retrieve the new information that happened in June this year on its own, and extract common knowledge from this content to supplement insights.
In the judicial field, the Spark model can also help users generate complaints. When asked, "I have a friend named Zhang San, who was bitten by Li Ming's dog on the night of January 5, 2023, and spent 2,000 yuan on the early treatment, and lost about 1,500 yuan after delaying work." Zhang San wanted Li Ming to compensate him for all his losses, but several times the communication was fruitless, if you want to sue him, please list the materials that need to be prepared", Xinghuo generated materials including communication records, witness testimony, etc., and asked questions again It can also directly generate a complaint.
The big-model open-ended Q&A shows a greater imagination, combined with its natural language processing capabilities and expertise to power industries.
2. Upgrade logical reasoning and mathematical ability, and comprehensively apply mathematical methods to solve
The scenario-based logical reasoning and mathematics of the large model based on the thinking chain are very close, so the improvement of logical reasoning ability is also the basis for the improvement of mathematical ability.
When asked, "The farmer needs to take the wolf, sheep, and vegetables across the river together, only the farmer can row, and the boat is relatively small, the farmer can only bring one thing across the river at a time, if the farmer is not there, the sheep will steal vegetables, the wolf will eat the sheep, please design a way so that the farmer can safely bring everything across the river", Spark can not only give a plan, but also explain the intention of each step.
The classic puzzle Spark can also be easily handled, such as "There are three people in a boat, but there are two fathers and two sons, what is going on?" ”
There are also junior high school math problems that examine permutations and spatial imagination, when asked, "How many intersections can six straight lines have?" "Spark will list the thought process and find the final answer.
Many mathematical problems, such as trigonometric functions, will contain a large number of pictures, etc., and cannot be entered by voice or text. iFLYTEK Spark APP realizes image input based on OCR capability.
For example, a mathematical problem that examines a polynomial, based on the mathematical problem solving assistant in the iFLYTEK APP, can give a complete solution step.
The Spark model can not only sort out some very winding logical phenomena, but also comprehensively apply mathematical methods such as equations and column combinations to solve problems.
It is worth mentioning that Sohu Technology previously selected Baidu Wenxin Yiyan, Ali Tongyi Qianqian, iFLYTEK Xinghuo Model, 360 Wisdom Brain, and ChatGPT to test the first 10 fill-in-the-blank questions in the Shanghai Mathematics Paper of the 2023 College Entrance Examination. The results show that the Spark Large Model is up to 50% correct.
Answers to five AI big model college entrance examination math questions (Source: Sohu Technology)
3. Upgrade your dialogue skills in multiple rounds and become an interviewer and a children's writer
In general, people and people need to cooperate to complete tasks that require multiple interactions, and complex tasks can rarely be completed through a single interaction, and the same is true in human-computer interaction.
The "iFLYTEK Smart Recruitment Interviewer" in the Xinghuo APP can simulate the interview scene and support full voice interaction. Recent college graduates who do not have interview experience can conduct simulation exercises. For example, if you say "I want to interview for the position of product manager", the assistant will ask the user about previous work experience, etc., and finally give an overall evaluation and recommendation based on the content of the answer.
A more interesting assistant is the story creation assistant, which generates a story when the user enters a title, such as "The story of the little rabbit traveling to Huangshan Mountain". If the child is not satisfied with the story, they can also supplement the demand, such as adding "Little Bunny Meets a Partner Pikachu".
In this way, the stories generated by the Spark model can be continuously enriched and updated based on the needs of users.
There is also an assistant that helps users generate stories called story dialogue co-creation, after the user enters the title, the Spark Assistant will not generate the entire article, but first give a paragraph to introduce, and then the user enters the next story direction, and so on to continue the story.
Second, the equipment in education and office scenes has been upgraded, and the number of industrial and medical services has increased significantly
In the education industry, iFLYTEK's previously released Xinghuo AI speaking assistant has been upgraded to the Spark Language APP, making it more convenient for users to use. The APP is aimed at all English learning enthusiasts such as primary and secondary schools, college students, business people, etc., in addition to general communication, it can also find users' pronunciation and grammar problems, help correct errors, and support voice and picture translation software.
If you encounter a problem that will not reply in English, the user can switch to the Chinese, and the Chinese and English mixed input Xinghuo language companion APP can also be accurately recognized, and at the same time, the user can also adjust the difficulty of the language partner in the setting interface.
In order to create a face-to-face dialogue scene with real people, Spark also supports virtual human dialogue.
In the office field, the Xinghuo cognitive large model is equipped on the iFLYTEK hearing smart screen, the pickup range of the device can reach 10 meters, and it can also access sound, video, etc. After the recording is completed, iFLYTEK hears that the smart screen supports fast transcription of recordings, and can also organize drafts based on the capabilities of large models and generate meeting minutes.
For B-side scenarios, the Spark model has been applied in the industrial and medical industries.
Previously, iFLYTEK released the Antelope Industrial Internet Platform, which runs through the research, production, supply, marketing and service management scenarios of enterprises in the industrial field. In the enterprise service of industrial scenarios, labor is required to meet the service supply required by the enterprise, but the amount of demand and supply docking completed by manual is very limited. At present, the accurate understanding, analysis, and recommendation of AI+big data in the integrated application of industrial scenarios have greatly improved the efficiency of demand and supply docking. At present, the total number of users of the Antelope platform has reached 322,000, and the number of platform service enterprises has exceeded 721,000.
Based on the large model of Xinghuo, iFLYTEK launched the antelope machine, which can recommend the business opportunities and policies of the industry to users.
For example, asked, "We are a home appliance manufacturer, the workshop has more than 20 large and small equipment, is looking for solutions that can further save energy and reduce consumption", Lingji will make suggestions for users according to the characteristics of the industry, and give some solution cases, resources and corresponding technical experts.
Liu Cong said that based on the capabilities of the Spark model, iFLYTEK has developed tools such as product introduction and news writing for enterprises, and in the next step, iFLYTEK will also open up the capabilities of this tool.
The Spark model can be combined with the knowledge of the industrial field to form an industrial brain, and then combined with the knowledge base of the enterprise to form the knowledge brain of the enterprise, and all aspects of production, research and development, service and marketing of the enterprise can interact with the enterprise brain, locate problems more accurately, and find targeted solutions.
In addition, in the workshop scene, if some new employees encounter equipment failure, they can also ask the antelope machine to move, "in the process of debugging the whole machine in the final assembly workshop, there is a problem that the touch screen is not working, what are the reasons." When the user enters the information of the surrounding working environment, the Antelope will analyze it again and give suggestions. After that, Lingji will also provide users with touch screen replacement suggestions, first retrieve from the company's internal inventory, and then go to the outside to find the right touch screen, realizing the whole process of fault inquiry, purchase advice, test verification
In the medical industry, iFLYTEK's intelligent medical assistant can be applied to pre-diagnosis and guidance to help doctors prevent misdiagnosis, missed diagnosis, and post-diagnosis rehabilitation. Wu Xiaoru revealed that the intelligent medical assistant system has been applied to 31 provinces across the country, completing 629 million auxiliary diagnoses and correcting 127 doctors' first diagnoses. And the combination of intelligent medical assistant and voice outbound calls has completed 1.1 billion telephone follow-ups.
Under normal circumstances, the continuity of services after patients are discharged will be missing, and the vast majority of patients face immediate care after discharge, and the mismatch in the number of doctors and patients makes doctors unable to timely guide patients on medication and dietary safety. Wu Xiaoru gave an example that 12% of stroke patients will be discharged again within 30 days after discharge, and 50% of these patients can actually avoid readmission through post-diagnosis rehabilitation management.
Based on the above phenomenon, the Spark model can analyze the whole process of the case and quickly generate a rehabilitation plan, and this plan will be synchronized to doctors and patients in real time.
The doctor-side interface of this post-diagnosis management platform has patient admission, discharge, consultation orders, test records, inpatient records, etc., and will generate a 90-day rehabilitation plan for users based on these data, including six dimensions, including doctor reminders, medication guidance, rehabilitation exercises, and dietary recommendations.
After being reviewed and approved by the doctor, this plan will be synchronized to the patient's mobile phone, and the patient will upload the checklist after the review, and the data on the doctor's side will also change synchronously, and the health management plan will be adjusted based on the changes in the patient's body data, forming a process of follow-up, follow-up, and tracking the recovery of closed-loop patients.
In addition to the development of rehabilitation plans, intelligent medical assistants can also answer more open-ended questions from patients, and also give answers based on patients' medical records, such as whether patients with fever can take fever-reducing drugs with long-term medication.
Wu Xiaoru said that rehabilitation management doctors have improved the efficiency of rehabilitation by more than 10 times after diagnosis, and real-time management has increased patients' dependence on doctors by 2.4 times, and their satisfaction with the hospital has increased from nearly 90% to more than 98% because patients receive immediate hospital follow-up, timely response to problems and guidance.
3. Open seven capabilities of large models and 200+ assistant development interfaces
In the early stage of the release of the Spark model, iFLYTEK will work with industry partners to build a large model "Spark" ecology. Liu Qingfeng revealed that there are currently more than 4 million development teams on iFLYTEK's artificial intelligence open platform.
He announced that he would open the development interface of the Spark big model, including seven dimensional capabilities and 200 Spark assistants, and support multi-terminal access for rapid integration, and enterprises with higher requirements for data security also support private deployment.
When the Spark model was released on May 6, iFLYTEK set a timetable for upgrading to V1.5 on June 9 and benchmarking ChatGPT with 1024 this year, achieving Chinese surpass and English equivalent. Liu Qingfeng said that this is because iFLYTEK adheres to independent research and development of each key module in the research of general artificial intelligence, and its "1+3+1" innovation system includes a self-developed safe and controllable large model training base, a high-performance large model inference platform that integrates software and hardware, and key modules such as data, modeling and reinforcement learning.
This is also the key for iFLYTEK to give a clear time iteration rhythm.
In the future, iFLYTEK will explore more potential paths in the fields of brain-like intelligence, new algorithms for neural network large models and game intelligence, which have previously been cross-researched.
In terms of talent training, iFLYTEK has joined hands with the first batch of 22 key universities in China to launch the Spark Model Scene Innovation Competition and the iFLYTEK University AI Spark Camp to allow more students to participate in the general artificial intelligence industry.
Conclusion: Focusing on technology, application and ecology, "Spark" is upgraded again
A little "spark" has become a fire. The combination of iFLYTEK Spark model and industry applications is getting closer and closer, and its capabilities are more fully displayed, including not only devices in education, office and other scenarios with higher user perception, but also application upgrades in the industrial and medical industries.
iFLYTEK, known as the AI national team, has set a timetable for the research and development of large models with its technology accumulation, and its blueprint for technology research and development, application landing and ecological construction with large models as the core is slowly unfolding. The development of technology is not difficult to achieve overnight, and it is necessary to comprehensively consider many factors, including the progress of technology research and development, application landing, etc., but iFLYTEK based on its self-developed large model base, as well as data security, modeling, reinforcement learning to achieve security and control, all make it more relaxed in the wave of large models.
With the continuous upgrading of the Spark model, the Spark model is moving forward on the road to catch up with ChatGPT.