Text丨Sina Technology Zhou Wenmeng

It will only be one year since ChatGPT was launched. What is the current state of development of the big model industry in China? Recently, Sina Technology's “100 Science Innovators” interviewed Liu Cong, vice president of iFLYTEK and director of the research institute, to discuss “the progress of large-scale domestic model research and production applications” and discuss the current state of development and ecological development of iFLYTEK's large-scale models.

During the conversation, Liu Cong pointed out, “Currently, there is still a gap between the top domestic and international cognitive intelligence models, such as comprehensive strength and platform capabilities. We must face up to the gap and aim for the most advanced level to catch up.” In this process, we will encounter difficulties and obstacles such as algorithms, chips, high-quality data, and system iteration; however, there has been long-term exploration and accumulation in various technical fields of the big model, and systematic innovation in core technology has also “blossomed”. In addition, the country is paying great attention to this field, “we need to have courage and confidence in the long run of the big model of cognitive intelligence.”

According to Liu Cong, the emergence of large cognitive models of intelligence will bring new opportunities to solve the immediate needs of humanity. Specifically, it is reflected in changing information distribution and acquisition models, innovating content production models, fully interacting with nature to complete tasks, realizing expert-level virtual assistants, disrupting traditional manual programming methods, and becoming an accelerator for scientific research work. The historical significance of big models in artificial intelligence technology and applications is no less significant than the birth of the Internet and personal computers.

The big model “straight charge” must have courage and confidence

Last October, iFLYTEK announced that the company's Spark cognitive model v3.0 fully benchmarks ChatGPT (GPT3.5), and has surpassed ChatGPT in areas such as Chinese language ability, design plan generation, and code completion. English proficiency is comparable to that. By the first half of 2024, InfoFly Spark will officially benchmark GPT 4.0.

In the process of promoting the iFLY Spark model to catch up with leading international big model products such as ChatGPT, the iFLY Research Institute at HKUST, led by Liu Cong, is the main force responsible for the development of the iFLY Spark model. In communication with Sina Technology's “100 Science Innovators”, Liu Cong pointed out, “The Starfire Model is benchmarking GPT 3.5. At the beginning, it was a comprehensive overall benchmark, and it did not completely pursue transcendence from a single technical point. Of course, in the specific process of advancing technology, iFLY will prioritize fields with technological and scenario advantages to make breakthroughs, seeking technology implementation while promoting technological catch-up.”

Liu Cong pointed out that in the process of comprehensively benchmarking GPT 3.5 and even GPT 4.0, the algorithm challenge will be a huge gap. Taking GPT 4.0 as an example, everyone is talking about GPT 4.0 because in the process of solving some complex problems, it will actively call some programs. This kind of call is not simply called manually, but it can know how to call it, and the instructions can support it.

“For most people, the capabilities of GPT 3.5 are sufficient; but for some more demanding people, they may have used GPT 4.0 to complete some complex tasks, and there are many plug-ins, resources, etc. behind it, so they are used to using GPT 4.0. Under such circumstances, if we want to fully surpass ChatGPT, the gap brought about by the implementation of core algorithms and the industrialization of core algorithms will become a huge gap.” Liu Cong said.

Furthermore, all aspects of problems, such as computing power limitations, high-quality data, system iterations, and changes in the skills of technicians raised by large models from a single mode to multiple modes, will all become obstacles on the way for large domestic models to catch up with leading levels abroad.

Of course, Liu Cong also pointed out, “Domestic exploration and accumulation has been carried out over a long period of time in various technical fields of the big model, and systematic innovation in core technology has also 'blossomed fruits'.” Coupled with the country's great attention to this field, “we must have courage and confidence in the long run of the big model of cognitive intelligence.”

According to Liu Cong, the current “direct charge” of large domestic enterprises aiming for advanced levels abroad is bound to be a “protracted war.” On this path, industry needs to come first, and innovative applications then drive the overall development of the ecosystem; “momentum” is also essential if we want to benchmark our goals and keep the gap narrowing. After achieving phased results, industrial development and scientific research and innovation must go hand in hand, continue to iterate on each other's paradigms, and promote each other. Only then can long-distance driving impetus be continuously injected into the development of China's large model of autonomy and achieve innovation and transformation.

In the future, through close collaboration between industry and academia, the field of scientific research will also create large-scale open source models from various fields, so that scientists in the research community can carry out more cutting-edge and innovative work on the basis of these model platforms. Research tasks and goals in the fields of speech, image, and natural language processing may also undergo new changes.

“The emergence of big model intelligence will bring new opportunities to solve the immediate needs of humanity”

According to Liu Cong, “The emergence of large cognitive models of intelligence will bring new opportunities to solve the immediate needs of humanity. Specifically, it is reflected in changing information distribution and acquisition models, innovating content production models, natural interaction to complete tasks, realizing expert-level virtual assistants, disrupting traditional manual programming methods, and becoming an accelerator for scientific research work.”

“The historical significance of big models in artificial intelligence technology and applications is no less significant than the birth of the Internet and personal computers.” Liu Cong said. He pointed out, “Currently, large domestic models are in full bloom. Various artificial intelligence companies have increased their investment one after another, and released and iterated on big models and related products one after another. With the continuous introduction and guidance of relevant national policies, it will also further regulate competition in the development of large model markets.”

In the increasingly fierce competition of big models, the competitive advantage depends on what you do and what you want to do.

According to Liu Cong, “For the basic general model, the investment is very large. Based on whether the enterprise can support long-term investment and the different business model choices, it may end up being several major basic models. This type of enterprise itself is an infrastructure, and some new ecosystems will spawn on this basis. In the field of industry applications, innovative and entrepreneurial companies can also find entry points for applications and products, and have achieved very good results.”

In line with iFLYTEK's practice of promoting the integration of big models with the industry, Liu Cong described iFLYTEK's current achievements in the education industry.

In the education industry, there are currently two major problems in language learning — writing and speaking.

In the essay correction scenario, assuming that there are 40 students in a class, the teacher simply checks each student's essay and performs in-depth analysis and comments at the same time. It takes at least 30 minutes per person, and a class of 40 people takes 20 hours. If equipped with the “iFLY Spark” iFLY AI learning machine, it is possible to complete annotations and essay reviews “automatically”, and give students timely, in-depth, and detailed feedback “one-on-one”, which saves teachers more time and allows students to learn more accurately and faster.

In the English learning scene, the fastest way to learn spoken English is to have a tutor to accompany you anytime, anywhere. The InfoFly AI learning machine is one such teacher. You ask and answer questions, continuously strengthen standard pronunciation and grammar, and learn authentic expression in “Sea Chat.” In addition, the InfoFly AI learning machine is also fluent in Chinese and English. If students encounter words they can't say, they can directly speak Chinese. It is immediately translated into English and answers in fluent English to guide students to continue their practice.

According to Liu Cong, in addition to the education scene, iFLYTEK is also exploring a combination of big models and industry scenarios in 11 scenarios, including automobiles and medical care.