AI technology is leading, the first to be applied to the ground, why is iFLYTEK’s big cognitive model ＂prairie fire＂？

On May 6, iFLYTEK’s "Spark" cognitive model "came out after a long time." Chairperson Liu Qingfeng demonstrated the seven dimensions of the "Spark" model including text generation, language understanding, knowledge answering, logical reasoning, mathematical ability, code ability and multimodal ability, fully demonstrating iFLYTEK’s "hard power" in the field of large model research and development.

In addition, Liu Qingfeng opened the experience test link of the Spark model to the audience, and demonstrated the intelligent functions displayed by "Spark" after connecting to iFLYTEK’s learning machine, smart office book, voice recorder and other products, which means that "Spark" is not only the first to open to the public on a large scale, but also the first to apply the landing AI model.

"If the full score of the big model is 5 points, ChatGPT is 4.48 points, and the current’Spark ‘has reached 4.29 points in the English field, ahead of its domestic counterparts," Liu Qingfeng said. He also gave the exact timetable for the development of the "Spark" big model during the year on the spot: "Spark" will be "upgraded" again on June 9 and August 15, and will be benchmarked against ChatGPT on October 24.

iFLYTEK’s on-site measurement of "Spark" cognitive model 　Excellent comprehension and high usability

At this press conference, the testers and the audience conducted a comprehensive assessment of the ability of "Starfire" in seven dimensions.

For example, in language understanding, logical reasoning and many other tests that test the "IQ" and "EQ" of the big model, "Spark" performed well. Liu Cong, dean of iFLYTEK Research Institute, asked on the spot: "As the saying goes, a man would rather die than surrender, but as the saying goes, a man must be able to bend and stretch. How do you understand these two sentences?" and "If a young man quarrels with his girlfriend, should he rather die than surrender or be able to bend and stretch?" In this regard, "Spark" said that the former question needs to be selected according to a specific scene, while for the latter question involving his girlfriend, "Spark" replied that "In this case, this young man should be able to bend and stretch." Blindly adhering to principles may exacerbate conflicts, and only when you calm down can you find a solution… "This" high emotional intelligence "answer caused the audience to applaud and applaud enthusiastically.

In terms of mathematical abilities, the iFLYTEK Spark model showed a "straight-A student" temperament. In addition to routine calculations, geometry, and situational applications, such as for complex calculation problems: "There are three kinds of flowers in the flower bed, a total of 88 flowers, of which the number of roses is 4 times that of chrysanthemums, and the number of peonies is 5 times that of chrysanthemums. How many peonies are there in the flower bed?" The Spark model quickly gave an accurate answer and gave the steps to solve the problem.

It is worth noting that over the years, iFLYTEK has accumulated profound knowledge in education and medical care, and this "Spark" model also demonstrates a wealth of general knowledge, capable of coping with life common sense, scientific knowledge, work skills, medical knowledge, etc.

"At present, we have not released a dedicated medical model, but our intelligent medical assistant is stronger than ChatGPT. In the future, these capabilities will gradually enter the Spark model, and the medical model will be released at this year’s 1024 conference." Liu Qingfeng said.

As one of the artificial intelligence "national teams", Liu Qingfeng said: "We must pay tribute to and learn from OpenAI, but also catch up quickly and strive to surpass, which requires a scientific and systematic evaluation system." iFLYTEK led the design of a general cognitive big model evaluation system through the National Key Laboratory of Cognitive Intelligence, and jointly 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 sub-task types covering 7 categories. In the scientific evaluation system, promote the development of big models.

Liu Qingfeng emphasized, "Today we need to see whether a large-scale model system is good or not. First of all, we need to see whether it can solve rigid needs and whether it is really useful, rather than a simple single-point test."

Take the lead in applying "1 + N" to deeply empower education, office, automotive, digital employee fields

iFLYTEK’s "Spark" is also the first AI big model in China to achieve application landing. Focusing on the content of "N" fields in the "1 + N" architecture of the "Spark" cognitive big model, iFLYTEK has released the industry application results of the big model in education, office, automobile, digital employees and other directions. It is worth noting that these results are "instant online". Users with iFLYTEK learning machines, smart office books and other related products can experience the charm of "Spark" immediately after upgrading.

For example, in English learning and composition correction, the T20 series of iFLYTEK AI learning machine equipped with "Spark" is proficient in both Chinese and English, which can realize one-to-one English-English free dialogue with learners, score user pronunciation, and correct Chinese and English compositions like a teacher

In the office scene, the iFLYTEK smart office book equipped with "Spark" has upgraded the two functions of meeting notes and discourse regularization. In meeting notes, the real-time transcription of speech is combined with ink screen paper-feeling writing, and the content of the meeting minutes can be formed into a streamlined meeting notes, which helps to quickly trace the content of the meeting. In terms of discourse regularization, it can remove the colloquial vocabulary in the speech transcription manuscript, filter redundant vocabulary, and polish the text, etc., and easily convert the oral expression of the pronouncer in the recording into a regular copy.

In the speech transcription scene where iFLYTEK leads the market, iFLYTEK Recording Pen and iFLYTEK Hear analyze text content through AI, and can generate a variety of texts including full-text abstracts, work backlogs, work plans, press releases, promotional copy, multilingual translation, etc., and can also conduct AI conversations and question and trace the original content according to the content of the original document. In the real machine demonstration, the "one completed draft" function is amazing. Users can directly convert the recording in the recording pen into a press release, and can label the key points of the manuscript for repeated generation.

Under the ability of "Spark", iFLYTEK, which has the first share in the automotive voice market, is also expected to launch an intelligent cockpit equipped with a large model to realize free communication between people and vehicles across business and scenarios in the car. It is understood that iFLYTEK has provided industry-leading intelligent voice interaction services for thousands of models, providing multi-wheel, multi-person, multi-regional and multi-modal voice interaction. The cumulative service users have exceeded 46.10 million, maintaining the first share in the domestic automotive intelligent voice app store. This time, the cognitive large model technology will bring the in-vehicle human-machine interaction to a new level.

In addition, the "Spark" model can also directly empower the field of digital employees. iFLYTEK generates business processes and RPA based on natural language, which can help enterprise employees complete a lot of repetitive work. For example, after entering the instructions in the company’s recruitment program, the system can automatically follow the pre-designed RPA script, operate the corresponding software in the computer, automate the operation of the business process and output the results, and perform data analytics, which greatly reduces the pressure on front-line staff.

In addition to the above fields, the iFLYTEK Spark Cognitive Model will also empower more industries, including medical, urban, political and legal, industrial, etc., to meet the needs of more professional fields and extend to a broader industrial field.

What are the differences between ChatGPT and other domestic models? A single spark can start a prairie fire

On the occasion of the launch of various domestic versions of "GPT", what is the difference between the big model "Spark" released by iFLYTEK this time?

First of all, all the tests of "Spark" at this conference were live demonstrations, not demos, and "Spark" withstood the test of questions from testers and live audiences. Secondly, in addition to asking questions to "Spark", this live demonstration also showed commercial scenarios. Finally, iFLYTEK proposed the exact timetable for "Spark" upgrade.

In fact, as one of the leading enterprises in the domestic AI industry, iFLYTEK launched the "Spark" cognitive model at the beginning, led the National Key Laboratory of Cognitive Intelligence jointly built with the University of Science and Technology of China, and discussed with many experts and partners of the Chinese Academy of Sciences artificial intelligence industry alliance and the Yangtze River Delta artificial intelligence industry chain alliance. A broad consensus was reached, and a general cognitive model evaluation system covering 7 categories and 481 sub-task types was designed and formed.

"When we salute and learn from Open AI, and at the same time catch up quickly and strive to surpass it, I think we first need a scientific and systematic evaluation system," Liu Qingfeng said. "In the 481 evaluation systems we just gave out in 7 dimensions, iFLYTEK’s Spark model has a total of 181 sub-task tests for text generation. Here I am very proud to share with you that Spark is clearly leading in the cognitive models that have been launched and can be tested in China."

Liu Qingfeng said that the text generation and mathematical capabilities represent the intelligence of a large model to a certain extent. The "Spark" large model is not only significantly ahead in the domestic system, but also has certain advantages compared with ChatGPT; and in language understanding and knowledge question answering, it has also reached the level of close to ChatGPT, which is at the leading level in China; logical reasoning and coding capabilities are already leading in China.

In fact, the excellent performance of the "Spark" model, to a certain extent, inherits the spirit of "fighting hard" when iFLYTEK faces difficulties. Just as when iFLYTEK first started its business in 1999, the domestic voice market was all controlled by international giants, but after the "grass platform team" founded by 6 college students of China University of Science and Technology raised the first funds, it went all out into research and development, and finally now occupied 70% of the mainstream Chinese voice market.

"At first, we’invested under pressure ‘and integrated the source technical resources that could be integrated in China, but we lost money year after year. At that time, we put forward a special sentence at the Bantang conference in Chaohu, called" The brightest torch that burns will either start a prairie fire first or be extinguished first, "which expressed not only our sense of urgency, but also our confidence and courage to be the first to start a prairie fire. Before the name of the iFLYTEK Starfire model was taken, we went through thousands of comparisons and finally agreed that we like the name" Starfire "." Liu Qingfeng said, "A spark can start a prairie fire."

Text/Luo Yidan

Editor/Yue Cai Zhou

Proofreading/Liu Baoqing