iFLYTEK Liu Qingfeng: The cause of improving people's livelihood and social development needs to be rooted in its own computing power security platform

Reporter Chen Jialan reported from Hefei

For domestic large model manufacturers, their goal is not only to catch up with ChatGPT in algorithms, but also to strive to achieve sufficient guarantee of computing power, and even benchmark NVIDIA chip performance.

On May 6, iFLYTEK released the Spark Cognitive Big Model, which demonstrated in real time to the outside world, including text generation, language understanding, knowledge question and answer, logical reasoning, mathematical ability, programming ability, multi-modal and other capabilities. On June 9, iFLYTEK Xinghuo Cognitive Model V1.5 was launched, and its capabilities such as open-ended knowledge quizzes, logical reasoning mathematics, and multi-round dialogue were upgraded.

On August 15, the reporter of China Business News learned from iFLYTEK that the iFLYTEK cognitive large model was upgraded to V2.0 again, the code upgrade and multi-modal capabilities were improved, and Huawei launched the iFLYTEK Spark all-in-one machine, began to layout the computing power of the large model, and paid attention to the security of computing power.

"On October 7 last year, iFLYTEK was once again included in the US Entity List for maximum pressure. Later, the United States gave restrictions on the computing power and bandwidth related to Chinese intelligence. In this case, we need to continue to open the industrial curtain of general artificial intelligence and improve people's livelihood and social development, which needs to be rooted in our own computing power security platform. Liu Qingfeng, chairman of iFLYTEK, said that iFLYTEK is working with Huawei to build a domestic computing power cluster for ultra-large-scale parameter large model training, and the platform benchmarks the performance of NVIDIA A100 in terms of single-card computing power.

The ability upgrade of the two aspects of the large model accelerates the benchmarking of ChatGPT

The reporter learned that iFLYTEK Xinghuo cognitive large model V2.0 has been upgraded in two aspects: code upgrade and multimodal capability.

The iFLYTEK Spark V2.0 upgrades code capabilities in five dimensions, including code generation, code completion, code error correction, code explanation, and unit test generation.

Liu Cong, Dean of iFLYTEK Research Institute, used Python to draw red heart lines, draw three-dimensional diagrams of saddle equations, set gradient colors, and generate mini games with code.

In addition, the Spark model can also use Python to process video. During the demonstration, Spark V2.0 used Python to process a 2-minute video, extract the second to tenth seconds, reduce the picture by half, speed it up by 5 times, and then save it as a GIF image. During the live demonstration, the Spark model also completed the code generation for the "Snake" game and made the game run smoothly.

Liu Qingfeng said that code capabilities are not only applied to the digital base of iFLYTEK Education, but also widely used in hospitals, universities, enterprises, governments and other different institutions, through iFlyCode to achieve rapid construction and low-cost rapid presentation, "This is why we say that general artificial intelligence will profoundly change today's business model with manpower and time as the main logic, and achieve complete subversion and upgrading of the entire industry."

According to Liu Qingfeng, at present, iFLYTEK Spark Model V2.0 based on Python and C++ code writing ability has been highly approaching ChatGPT, the gap is only 1% and 2%, and GPT4 will be officially benchmarked in the first half of next year.

In terms of multimodal capability upgrade, it is embodied in multiple dimensions such as image description, image Q&A, image recognition, text and image generation, and virtual human synthesis.

"Create a Liqiu lyrical essay and generate a video with the image of a girl with short hair and a Republican style." Liu Cong asked on the spot, and then the Spark model generated a short video of a virtual person, in which a girl dressed in the style of the Republic of China stood in front of the constantly switching autumn scenery and told the story.

"Multimodal capability is a rigid need to empower the industry, and it is also the only way to achieve general artificial intelligence." Liu Qingfeng emphasized that the current multimodal capability has been widely used in education, medical, industrial, automotive, robotics and other fields, it can obtain more and more data from the real world, and there is learning, training and improvement in the product terminal, including more flexible and autonomous industrial robots, better automatic driving, and companion robots that enter the home.

Jointly launched iFLYTEK Spark all-in-one machine layout large model computing power

On December 15 last year, iFLYTEK officially launched the special research plan of iFLYTEK Spark Cognitive Big Model 1+N, which benchmarks ChatGPT general cognitive large model and N empowers the application of various industry tracks. Liu Qingfeng also said many times that the general model of Xinghuo cognitive large model will be directly benchmarked against ChatGPT on October 24.

"We not only have full confidence in the algorithm, but also taken an important step in computing power." Liu Qingfeng said. It is worth noting that in addition to continuously upgrading large models to achieve ChatGPT benchmarking, iFLYTEK also cooperated with Huawei to launch iFLYTEK Spark all-in-one machine, and began to lay out large model computing power, benchmark NVIDIA chip performance, and ensure sufficient computing power.

"The era of deep empowerment of cognitive large models has arrived, and healthy development urgently needs to build safe and controllable guarantees." Liu Qingfeng said: "Safe and controllable", "scenario-driven" and "exclusive model" are the key three elements of the in-depth application of cognitive big models in the industry. He believes that to realize the "security and controllability" of large models, it is necessary to solve the two major problems of "computing power security" and "content security".

In terms of computing power security, iFLYTEK Xinghuo and Huawei Ascend jointly released the Xinghuo all-in-one machine, which provides industry-leading large model training and inference capabilities based on Ascend AI hardware and Ascend AI open source framework, providing a solid independent innovation computing power foundation for the whole process innovation of large models.

"iFLYTEK has a self-developed large model training platform, which has the advantages of full-process design of training and data closed-loop, integrated design of large model training and inference, compatibility of large-scale heterogeneous computing power, and easy expansion of hybrid cloud architecture. Liu Qingfeng said.

It is reported that the Xinghuo all-in-one machine can provide 5 customized optimization modes such as dialogue development, task orchestration, plug-in execution, knowledge access, prompt engineering, etc., as well as more than 10 kinds of out-of-the-box rich scene packages such as office, code, customer service, operation and maintenance, marketing, procurement, etc., and support 3 model sizes for users to choose.

Liu Qingfeng said that with these capabilities, every enterprise, every industry, every school, and every hospital can have the opportunity to build its own exclusive big model. At present, the data of 9 exclusive large models in the fields of finance, government affairs, and automobiles completed by iFLYTEK Xinghuo shows that the average effect in scene task optimization and private domain knowledge enhancement can be increased by 20%.

It is worth noting that on August 15, the Interim Measures for the Management of Generative Artificial Intelligence Services (hereinafter referred to as the Measures) promulgated by the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration of Radio, Film and Television (hereinafter referred to as the "Measures") officially came into force. The Measures put forward that the state adheres to the principle of attaching equal importance to development and security, promoting innovation and governing according to law, and takes effective measures to encourage the innovative development of generative artificial intelligence. From a policy perspective, fastening a "seat belt" for the development of artificial intelligence is the general direction.

"The April 28 Politburo meeting, and the seven-department document on generative AI, which came into effect on August 15, require us to balance both development and security." Liu Qingfeng told reporters that the "Measures" gave very clear standards for the generative artificial intelligence industry, such as the main scope and content of the seven departments concerned about security, marked the direction of enterprises' independent efforts, and what needs to be certified by the Cyberspace Administration of China, forming a boundary for everyone.

Drive the growth of the C-end market The growth space of the B-end market needs to be opened

After two months, iFLYTEK Spark has successively promoted the implementation of real and visible applications in education, office, automotive, finance, industry, medical and other industries.

With the empowerment of the Spark model, iFLYTEK learning machines, office and other consumer hardware products have achieved rapid sales growth in a short period of time.

According to the iFLYTEK report, from May to June 2023, iFLYTEK's C-end hardware GMV reached a record high, doubling year-on-year. Among them, the GMV of iFLYTEK AI learning machine equipped with the iFLYTEK Spark cognitive large model increased by 136% and 217% in May and June, respectively. iFLYTEK AI hardware (AI learning, AI office, AI health) sales increased by 125% year-on-year during the "June 18" period. At the same time, the proportion of the C-end of smart education in the first half of the year has increased to 40%.

For the performance of the C-end market, Liu Qingfeng told reporters that the C-end market can reach users more easily than the B-end market, such as Spark Companion, users can use and solve problems as soon as they get it, and they can see immediate results.

"The B-side market needs to be applied, involving how to import a more complex internal knowledge base into learning and training, the investment cost is relatively high, the decision-making chain is relatively longer, and the B-side scene sense is temporarily not as direct as the C-side, but the B-side scenario has become more and more clear." Liu Qingfeng introduced that for the B-end market, once everyone understands it, with a set of scientific testing methods, find out the application scenarios can be evaluated, input and output results, the market will rise quickly, the current iFLYTEK is building some B-side typical cases, allowing users to see its value.

Liu Qingfeng told reporters that the fields of digital pedestal and medical post-diagnosis management will have clear growth in the next three months to half a year, especially after the commercial application is fully liberalized, "it is a mule or a horse, take it out and try it."

(Editor: Wu Qing, Proofreader: Zhai Jun)