In the past few days, I have seen news reports that Hu Guoping, senior vice president of iFLYTEK, released the annual plan of iFLYTEK Spark in the Zhongguancun Forum. It is planned that on June 9, open-ended questions and answers will be broken, multi-round dialogue capabilities will be upgraded, and mathematical skills will be upgraded again. It is planned that on October 24, the generic model will benchmark against ChatGPT (Chinese surpassing, English equivalent).
Why does iFLYTEK Spark have the confidence to surpass ChatGPT?
Up to now, the iFLYTEK Spark model is still limited to the ecological partners of iFLYTEK Open Platform for experience, and has not been officially announced.
As early as the 5.6th of this year, iFLYTEK announced the Spark model, which showed that in order to scientifically PK ChatGPT capabilities, iFLYTEK took the lead in designing a general cognitive large model evaluation system through the National Key Laboratory of Cognitive Intelligence, and discussed with the Chinese Academy of Sciences Artificial Intelligence Industry-University-Research Innovation Alliance and the Yangtze River Delta Artificial Intelligence Industry Chain Alliance to form 481 subdivided task types covering 7 categories.
The scene of the press conference
First of all, I understand the iFLYTEK Spark cognitive model
iFLYTEK's next-generation cognitive intelligence model has cross-domain knowledge and language comprehension capabilities, and can understand and perform tasks based on natural dialogue. Continuous evolution from massive data and large-scale knowledge to achieve a closed loop from proposing, planning to solving problems. The following are some of the application scenarios currently supported by the Spark cognitive large model:
1. Language understanding
l Machine translation: translate text in multiple languages, including English, Chinese, French, German, Spanish and other common languages
l Text summary: Extract a concise and accurate summary based on text to quickly understand the core ideas of the article
l Grammar check: Check grammar errors and provide correct grammar suggestions to make writing more standardized and professional
l Sentiment analysis: analyze the emotional color in the text, such as positive, negative or neutral, to better understand the content perspective and attitude
2. Knowledge Q&A
l Life knowledge: provide advice on daily life, such as diet, exercise, travel, etc
Job skills: Provide knowledge of the job, such as communication skills, time management skills, teamwork and other advice
Medical knowledge: provide basic health care knowledge and advice on disease prevention, diagnosis and treatment
l History and humanities: provide copywriting on historical events, cultural inheritance, celebrity stories, famous sayings and aphorisms, etc
3. Logical reasoning
l Thinking reasoning: Reasoning out answers or solutions by analyzing the preconditions and assumptions of a problem, and giving new ideas and insights
Scientific reasoning: the use of existing data and information for inference, prediction and verification and other basic tasks in scientific research
● Common sense reasoning: Use existing common sense knowledge to analyze, explain and respond to users' questions or needs when conducting dialogue and communication
4. Solve mathematical problems
l Equation solving: including univariate quadratic equations, binary primary equations, ternary primary equations, etc
l Geometry problems: planar geometry (e.g. properties of lines, circles, triangles, etc.) and solid geometry (e.g. volume, surface area, projection, etc.)
Calculus: Deals with calculus-related problems such as derivatives and integrals, involving basic concepts such as limits, continuity, derivatives, etc
l Probability and statistics: involving random variables, probability distributions, hypothesis testing, etc
5. Code understanding and writing
Code understanding: help users understand most programming languages, algorithms and data structures, and quickly give the required answers
Code modification: modify or optimize existing code, provide suggestions and guidance, find out potential problems and provide solutions
Code writing: Help users quickly write some simple code snippets, such as functions, classes or loops
l Step compilation: Provide documentation and tools about programming languages, such as syntax rules, function libraries, autocomplete code tools, etc
From the web
From the web
From the perspective of release, the current Xinghuo general capability is obviously ahead in the industry. In terms of Chinese, the current iFLYTEK Spark cognitive large model has surpassed ChatGPT in the three major capabilities of text generation, knowledge question and answer, and mathematical ability (the actual situation has to be officially announced, and the majority of users will make a conclusion after the actual use), and will catch up with ChatGPT as a whole at the end of October. However, Liu Qingfeng said at the press conference that iFLYTEK's current language understanding ability is still slightly behind ChatGPT, but it has surpassed similar domestic products.
One presentation is more interesting, enter a illustrated English menu, Spark not only immediately gives a Chinese version of the menu, but also can introduce the basic information of the dishes that have not been eaten on request, and users can even designate virtual people to introduce themselves to these menu contents with a specified synthetic voice.
In addition, Xinghuo's mathematical logic ability is still good, and a complex calculation problem was thrown at the press conference: there are three kinds of flowers in the flower bed, a total of 88, of which the number of monthly flowers is 4 times that of chrysanthemums, and the number of peony flowers is 5 times that of chrysanthemums 2 less, so how many peony flowers are there in the flower bed? The large model quickly and accurately gives the answer and gives the steps to solve the problem.
Liu Qingfeng said that the industrial field is also a very important landing scenario, lowering the programming threshold for ordinary engineers and workers without software background, and we can expect a breakthrough in capabilities in August. At the same time, it said that the iFLYTEK Spark model is not only far ahead in the domestic system, but also surpasses ChatGPT. It also said that at the developer conference on October 24, Spark will benchmark against ChatGPT, surpass it in Chinese, and reach a level comparable to it in English.
What level of ChatGPT?
Let's take a look at the development of ChatGPT
Developed by the American company OpenAI, this chat tool has caused a global sensation since its release in November 2022. Within five days of launch, the number of users exceeded 1 million, ruthlessly crushing Facebook's record of breaking one million users in 10 months. In just two months of launch, the number of ChatGPT users exceeded 100 million, and it took TikTok 9 months to reach 100 million users, becoming the fastest growing consumer app in history.
November 30, 2022
ChatGPT goes public.
December 5, 2022
OpenAI founder Sam Altman announced that ChatGPT has surpassed 1 million users in just 5 days.
End of January 2023
ChatGPT exceeded 100 million users, making it the fastest-growing consumer app in history.
End of January 2023
Microsoft announced an additional $10 billion investment in OpenAI, which launched ChatGPT.
February 2, 2023
OpenAI released ChatGPT Pilot Subscription Plan, ChatGPT Plus, available for $20 per month, where subscribers get more stable and faster service than the free version, as well as priority to try out new features and optimizations.
February 2, 2023
Microsoft announced the integration of ChatGPT across all of its products.
February 6, 2023
Google parent company Alphabet has announced that it will launch the chatbot Bard, unlike ChatGPT's database, which is only updated until 2021, and Bard will collect the latest content.
February 7, 2023
ChatGPT's official website said that a large number of users flocked to the website, causing the website to be paralyzed, and users asked questions on the ChatGPT page showing "too many times in an hour, please try again later".
February 8, 2023
Microsoft announced the latest versions of the artificial intelligence search engine Bing and Edge browsers powered by ChatGPT.
March 15, 2023
In the early morning of March 15, OpenAI released the large-scale multi-mode model GPT-4, ChatGPT Plus, according to the official introduction of OpenAI, GPT-4 is a large-scale multi-mode model that can accept image and text input, output text, and show human-level performance on various professional and academic benchmarks. After trying the new version, many people said that it was much stronger than the GPT-3.5 of ChatGPT used earlier, once again refreshing the understanding of AI.
From the web
Compared to the previous generation, GPT-4 has a wider range of knowledge and stronger problem-solving skills, and performs better in creativity, visual input, and long content. Used in creative projects, it helps users create songs, write screenplays, or learn their writing styles together. It is worth noting that GPT-4 can directly read 32,000 tokens, which is equivalent to giving it 25,000 English text background information, and can quickly give conclusions - that is, professions like paralegals are really about to be replaced. The previous version could only read 4096 tokens, equivalent to 3000 English text information, and the progress of this version is leapfrogging. In addition to text, you can also use images as input to GPT-4, which not only recognizes objects in the image, but also further processes the content based on this information.
Make a website in 10 seconds
A video of the GPT4 launch event circulated online, and the two-minute video presentation reads:
1. Make a very rough sketch with pen and paper on a draft book;
2. Take a picture and tell GPT that we want to make a website, the effect is as shown in the figure, let it generate the website code;
3. After the website is completed, it takes about ten seconds in total.
In the official demonstration, GPT-4 can generate a complete front-end HTML code of a website in real time and produce a website based on a sketch, almost in about ten seconds.
Make a game in 60 seconds
The relevant demonstration is not from the official, but netizens have experimented. According to reports, netizens talked to GPT-4 and asked it to do a pinball game, which took about 60 seconds. In the end, without repeated communication, GPT-4 completed the game in one go.
Another netizen asked GPT-4 to make a snake game, which took about 20 minutes to successfully write and debug the entire snake game. Although GPT-4 cannot complete the operation in the reply at once, it is still completed after many conversations, and netizens only need to reply "continue" during the whole process.
Accurate identification of memes
This is also a qualitative leap for GPT-4, that is, the beginning of processing images. Previously, people could use it to process text, that is, to generate text based on a given context, such as generating articles, poems, dialogues, etc. The content supported by GPT-4 is no longer limited to text, but begins to accept images as input media. According to the official demonstration, in the face of a "meme map", GPT-4 accurately described the content of the picture and thoughtfully explained why this picture is funny.
According to OpenAI, GPT-4 shows "performance at the human level under a variety of professional and academic indicators": similar to mobile phone scoring software, when ChatGPT first came out, many people used it to "brush questions and run scores", under the GPT-3.5 version model, its SAT score can only rank at the bottom 10% level, but the GPT-4 model can exceed the level of 90% of candidates.
In other exams, it also shows a similar contrasting effect, if it is just a "running score", ChatGPT-4 belongs to the level that can be admitted to Harvard and Stanford.
OpenAI's latest generation AI language model, ChatGPT 4, completed and passed several accounting exams in the United States, including Certified Public Accountant (CPA), Certified Management Accountant (CMA), Certified Internal Auditor (CIA), and Certified Tax Agent (EA), with an average score of 85.1.
ChatGPT not only "passed" the U.S. Medical Licensing Exam, but also published an oncology paper as a worker.
Not only has ChatGPT passed the most challenging professional exams in the United States, the MBA, the legal qualification exam, and the U.S. Medical Licensure Exam, ChatGPT has also passed the Chinese Database System Engineer exam. Previously, ChatGPT passed a Google coding interview for a Level 3 engineer with a salary of $183,000.
Elon Musk, one of the original co-founders of OpenAI, previously tweeted: "This is a new world. Goodbye homework! So far, it seems that Musk's claims have been confirmed. When it comes to writing papers, ChatGPT is really good. Antony Aumann, a philosophy professor at Northern Michigan University, recently named the best paper in his class for his World Religions class, only to learn that it was written by students using ChatGPT.
One of the most prominent features of ChatGPT is the use of ethically focused training to say "no" to unsolicited questions and requests, following a pre-designed code of ethics. Once it is found that the text prompt given by the user contains malicious intent, including but not limited to violence, discrimination, crime and other intentions, it will refuse to provide a valid answer. This allows users to clearly feel the "smartness" of ChatGPT during the interaction.
It is worth noting that ChatGPT's database has only been updated to 2021 and is currently non-networked, which is quite different from Google's upcoming Bard. Google emphasized that Bard will use the latest data, which is a big advantage for Google as a latecomer to compete with ChatGPT.
Today, a number of technology giants have launched related product planning, and a super "involution" artificial intelligence competition is being staged.
Microsoft is a major supporter of OpenAI, the producer of ChatGPT, and has used ChatGPT-related technology for its Bing search engine. In January this year, Microsoft further announced the expansion of cooperation, will invest 10 billion US dollars (about 67.902 billion yuan) in OpenAI.
According to media reports, Google regarded ChatGPT as a "red alert" level threat, and specially invited back co-founders Sergey Brin and Larry Page, who left a few years ago, to discuss countermeasures. On February 6, Google's parent company Alphabet announced that it will launch the chatbot Bard, which is said to be initially only used by some testers and then widely promoted.
In addition to self-developed products, Google is also making great efforts in the external brain. According to media reports, Google invested nearly $400 million (about 2.716 billion yuan) in artificial intelligence startup Anthropic in early February. Anthropic's AI assistant Claude is also expected to be a competitor to ChatGPT. When releasing his latest earnings report in early February, Google CEO Sundar Pichai said that Google is in a good position in the field of artificial intelligence because it has ushered in an inflection point.
The data shows that the number of robots deployed by Amazon is also growing rapidly, reaching an increment of about 1,000 per day.
In addition, Facebook's parent company Meta also plans to invest an additional $4 billion to $5 billion in data centers in 2023, all of which is expected to be spent on artificial intelligence.
In addition, the domestic Baidu company Wenxin Yiyan cloud service was launched on March 27. On May 28, Baidu Chief Technology Officer Wang Haifeng demonstrated the "Wen Xin Yiyan" function of generating video through text, intelligent summary chat history, and intelligent programming function that has not yet been publicly launched in the Zhongguancun Forum. "Flying Oar" and "Wenxin" are jointly optimized, and the inference performance has been improved by 10 times in the past 1 month.
ChatGPT's rapid advance has sparked a new round of artificial intelligence competition.
The confidence of iFLYTEK Spark model
In 2010, iFLYTEK Open Platform was officially launched, integrating multiple capabilities such as speech synthesis, speech search, natural language processing, and speech dictation, becoming the earliest open intelligent voice platform in the industry at that time. It is understood that after more than ten years of development, iFLYTEK Open Platform has opened 318 AI capabilities and solutions to the public, linking more than 2 million ecological partners and supporting more than 2.8 billion terminals.
In the answer to investors' questions at the 2022 annual results briefing, iFLYTEK mentioned that iFLYTEK already has rich experience in Transformer deep neural network algorithms, which are also widely used in iFLYTEK's speech recognition, graphic recognition, machine translation and other tasks and have reached the international leading level.
Among them, the core technology has maintained the international leading level. In the stage of artificial intelligence technology from perceptual intelligence to cognitive intelligence, common sense reasoning is an important part. In 2022 alone, iFLYTEK has won 13 world championships in the field of cognitive intelligence technology.
For example, after OpenBookQA won the championship, iFLYTEK upgraded X-Reasoner, a unified understanding framework for the integration of winning systems, knowledge and large models, and launched X-Reasoner++ and won the top QASC list in 2022, surpassing the human average for the first time in the world.
In 2022, they also open-sourced a series of Chinese pre-trained language models in 6 categories and more than 40 general domains, with an average monthly call volume of more than 10 million related model libraries, ranking first and far more than second in the number of stars of similar Chinese on the Github platform. In addition to the accumulation of core algorithms, iFLYTEK has accumulated more than 50TB of industry corpus and active applications with more than 1 billion user interactions per day in years of R&D and promotion of cognitive intelligence systems.
Third-party data seems to have a low threshold for access, but it is not easy to obtain massive amounts of high-quality data on a large scale, which requires long-term standardized accumulation and must also have certain guarantees in data compliance, which is why iFLYTEK's innovation in the field of large models is worth paying attention to.
In the field of education, iFLYTEK's related educational products have been applied in 32 provincial-level administrative regions across the country, covering more than 50,000 schools, 130 million teachers and students, and possessing massive data such as phonics, transcripts, and question banks.
In the medical field, iFLYTEK is the only artificial intelligence system in the industry that has passed the national medical practitioner qualification examination, exceeding 96.3% of medical candidates, and has provided more than 580 million artificial intelligence auxiliary consultations for grassroots doctors, with an average of more than 700,000 times per day.
Similarly, as the first batch of national new generation artificial intelligence open innovation platform, it has been used by more than 5 billion people per day, and has also provided massive text corpus and user feedback data for large models for many years.
In terms of computing power related to cognitive big models, iFLYTEK has built its own industry-class data center at its headquarters, and has built a deep learning computing platform in four cities and seven centers, laying a good hardware foundation for the construction of large model training platform.
In December 2022, iFLYTEK began the special research of Xinghuo's cognitive intelligence big model, which can achieve a rapid breakthrough in cognitive large model in five months, which is inseparable from the company's long-term solid accumulation. It is understood that based on the steady development of iFLYTEK's business base for many years, iFLYTEK Spark Cognitive Big Model has been implemented in many industries and products such as education, office, automotive, and digital employees. For example, iFLYTEK AI learning machine can not only help students practice speaking, but also correct essays like teachers and accurately point out mistakes; iFLYTEK Intelligent Office can directly generate meeting minutes based on handwritten keywords; iFLYTEK can realize the function of "one-click writing of a recording".
At the same time, iFLYTEK Xinghuo cognitive big model also empowers 4.45 million developers to build a "spark" ecology of artificial intelligence through 560 AI capabilities such as open graphic recognition, face recognition, and voiceprint recognition, and solve the rigid needs of the industry in education, medical care, justice, automotive and other livelihood fields.
Liu Qingfeng also demonstrated iFLYTEK Xinghuo's language translation, logical reasoning and other capabilities at the scene, and shared the upgrade and iteration milestone plan of iFLYTEK Xinghuo's cognitive large model: on June 9, it will break through open-ended question and answer, and upgrade its multi-round dialogue ability and mathematical ability; On August 15, the code capability will be broken, and the multimodal interaction will be upgraded again; On October 24, the generic model will benchmark ChatGPT (Chinese surpassed, equivalent in English).
Something
When ChatGPT came out, in fact, I was more optimistic about Google, the domestic Baidu, in fact, the model has open source, in addition to technology research and development of large models, the most important thing is learning, that is, large-scale training, in addition to the algorithm itself, also need higher requirements of hardware and huge data, these are Google, Baidu's most advantage, especially the data. But Google just came out and smashed it, Baidu's Wen Xin has a lot of criticism, and it did not reach the height of expectations. On the contrary, ChatGPT is growing rapidly in terms of update speed and data volume, which is indispensable to Microsoft's strong support, including technology and funds, Microsoft has invested a lot of money, and has begun to implant ChatGPT into Microsoft's own products, and also closed the meta-universe project, which can be seen. Even this cannot be said that there are no opportunities in China, there are definitely but whether they can be grasped, such as:
Geographical advantages: After all, the domestic user base is large, iFLYTEK products are many (you can go to understand iFLYTEK's products, there have been AI products many years ago), and there are a lot of data, if you can collect BAT, the success rate is greatly improved, but it is not realistic, and the reason goes without saying.
Policy advantages: the state vigorously promotes artificial intelligence, local enterprises have huge advantages, in addition, foreign platforms into the country has many policies and security restrictions, for example, you use ChatGPT to do business in China, data has a transit risk, which is a great hidden danger. Innate advantages of domestic companies.
In short, there are opportunities, how fast, how high are the walls?
Who do you think will win in the end?
Welcome to give your answer in the comment area.