Straight after ChatGPT? Hundreds of billions of AI giants, quickly upgrade the big model!

China Foundation News reporter Feng Yao

Just over a month after the first launch of the multimodal large model, iFLYTEK has been upgrading its "Spark Cognitive Large Model" non-stop.

On June 9, iFLYTEK announced the new progress of its general large model, releasing the V1.5 version of the "Spark Cognitive Big Model", which has made breakthroughs in open-ended question and answer, upgraded multiple rounds of dialogue and mathematical capabilities, and improved text generation, language understanding, and logical reasoning capabilities. In addition, iFLYTEK also put the "Spark Cognitive Big Model" on the mobile terminal this time and released its Spark APP.

As early as a month ago, iFLYTEK had planned to achieve the benchmark of ChatGPT's capabilities by October 24. According to iFLYTEK, the next step will be to open the multi-modal interactive upgrade of the "Spark Cognitive Big Model" on August 15.

Targeting the "three major flaws" to be overcome

On May 6, iFLYTEK first announced the "Spark Cognitive Big Model", and Liu Qingfeng, chairman of iFLYTEK, set a goal for it: to surpass GPT in Chinese on October 24 this year and reach a comparable level in English.

34 days later, iFLYTEK upgraded it and announced the V1.5 version of the "Spark Cognitive Model". Liu Qingfeng introduced that this version has made breakthroughs in open-ended question and answer, upgraded in multiple rounds of dialogue and mathematical skills, and continuously improved text generation, language understanding, and logical reasoning capabilities.

Especially in terms of pan-domain open Q&A, version V1.5 of "Spark Cognitive Big Model" aims at the "three major defects" that need to be overcome in pure large model technology: new knowledge is difficult to update, factual questions and answers are easy to "Zhang Guan Li Dai", and historical facts and traditional classics are easy to "make up plots".

At the same time, the leap in multi-round dialogue capabilities makes the dialogue experience of iFLYTEK Xinghuo more suitable for real people. Multi-round dialogue is the traditional problem of large models, which is simply "no memory".

At the meeting, iFLYTEK also gave a live demonstration of the "Spark Cognitive Big Model". When talking about the question "What are the new trends in artificial intelligence in China". "Spark Cognitive Big Model" mentioned that on June 3 this year, the Yangtze River Delta Entrepreneur Alliance Industrial Digitalization Summit issued the "General Artificial Intelligence Yangtze River Delta (Hefei) Declaration" and "General Cognitive Intelligence Big Model Evaluation System".

In fact, the Spark big model has been finalized in May, and the answer given by the "Spark Cognitive Big Model" has included relevant policy trends in June, which also indicates that the big model has been updated and learned in real time. It is worth noting that the "Spark Cognitive Big Model" also further gives the gap faced by domestic artificial intelligence in the answer.

"It is meaningless for a large model to give the same answer as search, but to give a constructive solution through professional knowledge and reasoning ability," Liu Cong, president of iFLYTEK Research Institute, also said bluntly at the meeting. In addition, the "Spark Cognitive Big Model" also accurately answered the math and language questions in this year's college entrance examination at the meeting.

Next node: multimodal interaction and upgrade

According to the plan, iFLYTEK will carry out three rounds of iterative upgrades within this year, and the goal is to benchmark ChatGPT on October 24. In addition to June 9, the next upgrade node will be on August 15, mainly to break through the code capability and multimodal interaction and then upgrade. In terms of multi-modality, including virtual human synthesis, graphic understanding, etc., it will also be open to customers at that time.

Liu Qingfeng, chairman of iFLYTEK, previously said that at present, iFLYTEK's code capabilities focus on the industrial Internet and many applications within the enterprise, and the future goal is to allow large models to generate various codes without the need for programmers. However, Liu Qingfeng also admitted that there is still a big gap between this function of the Spark model and ChatGPT, and the key function of the next upgrade is also in this field.

Liu Qingfeng revealed at the meeting that in more cutting-edge fields, iFLYTEK will also explore more potential artificial intelligence technology routes, such as game intelligence, brain-like intelligence and neural network models.

In addition to further improving the capabilities of the big model, iFLYTEK also released the "Spark Cognitive Big Model" to further commercial implementation progress in the fields of learning, medical care, industry, and office, including the launch of the Spark APP and the Spark Language APP.

At the same time, iFLYTEK also further aimed at subdivisions and launched Xinghuo cognitive large model + medical post-diagnosis management platform, Xinghuo cognitive large model + industrial Internet platform, Xinghuo cognitive large model + iFLYTEK hearing smart screen products. In the eyes of industry insiders, this move is intended to promote its commercialization in subdivided fields, and the scenarios that are expected to break through are the above-mentioned medical, industrial manufacturing and office fields.

At the same time, in addition to developing demonstration application products for different application scenarios, iFLYTEK is also recruiting ecological partners for AI developers, upstream and downstream enterprises of large models, and entrepreneurial teams.

In fact, referring to the development history of OpenAI, the premise of large model research and development is that the development, training, and application of small models are mature enough, and the products at the beginning of OpenAI are only vertical small models in the game field, and after fully understanding the development and landing of small models, the number of parameters is continuously expanded and finally GPT3 is formed.

Domestic large models open "100-model war"

Since March this year, domestic general-purpose large models have sprung up one after another. Among them, Baidu took the lead in releasing Wen Xin's words, Ali followed closely and officially announced the Tongyi Qianqian, and even scientific research institutes such as Tsinghua University, Beijing KLCII Artificial Intelligence Research Institute, and Shanghai Artificial Intelligence Laboratory also released their own AI large model results.

According to the statistics of relevant research reports of Minsheng Securities, at least 30 large models have appeared in China, covering Internet giants, AI concept listed companies, server leading enterprises, scientific research institutes and primary market startups, and the parameter scale of some large models has approached, or even exceeded the scale of ChatGPT (hundreds of billions).

IDC forecasts that the spending scale of the Chinese intelligence market will increase to US$14.75 billion by 2023, accounting for about one-tenth of the global total. In the long run, the innovation and iteration of AI technology has driven the further landing of application scenarios, and hot spots represented by AIGC, digital human, multi-modality, AI large model, and intelligent decision-making have brought more imagination and possibilities to the market.

IDC predicts that China's AI market will reach a market size of US$26.44 billion in 2026, with a five-year compound growth rate (CAGR) of more than 20% from 2021 to 2026. CSC believes that the domestic boom in the R&D and application of large models continues to rise, and the development of large models is accelerating in an all-round way, but the current global large model industry is still in the early stage of exploration, and it is necessary to cooperate with downstream scene enterprises to establish a business model of large models.

Editor: Captain

Review: Xu Wen