【Dahe Caifang Reporter Zhang Keyao Wen Tu】iFLYTEK (002230. SZ) set off a new round of "100-model war", with a number of newly released applications and products, but also quietly joined the "unlimited game" in the field of large models.
On August 15, iFLYTEK Cognitive Model V2.0 Upgrade Conference arrived as scheduled, iFLYTEK Chairman Liu Qingfeng and Research Institute President Liu Cong released upgraded versions of code capabilities and multi-modal capabilities, and released a number of applications and products equipped with upgraded Xunfei Spark Cognitive Model V2.0 capabilities. iFLYTEK and Huawei have cooperated again to jointly release the Spark all-in-one machine.
Coincidentally, on the same day of iFLYTEK's press conference, the first normative policy for the generative artificial intelligence (AIGC) industry in mainland China, the Interim Measures for the Management of Generative Artificial Intelligence Services (hereinafter referred to as the "Measures"), jointly issued by the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Education, and the Ministry of Science and Technology, was officially implemented.
iFLYTEK is not a "lone brave" in the field of domestic large models. With the three six zero (601360. SH), Xiaomi Group (01810. HK), China Literature Group (00772. HK) and other enterprises to increase the size of large models, the mainland generative artificial intelligence industry is expected to accelerate development.
iFLYTEK Spark code capability upgrade,
In the first half of next year, GPT-4 was benchmarked
Code is the ability of large models to hit hard, and this upgrade is also the flag set by iFLYTEK when it first released the iFLYTEK Spark cognitive large model on May 6 this year.
"Code data can improve the 'wisdom' of cognitive large models, and code capability is an important indicator of the intelligence of cognitive large models." Liu Qingfeng said that code capabilities are also an effective means to build and link the digital world, which can greatly reduce the threshold and cost of entrepreneurship in the digital economy.
The iFLYTEK Spark Cognitive Model V2.0 upgrades code capabilities in five dimensions, including code generation, code completion, code correction, code interpretation, and unit test generation.
In the live demonstration, Liu Cong used Python to draw red heart lines, draw three-dimensional diagrams of saddle equations, and generate mini games with code. "Using Python to process video spark .m4v, extract the 2nd to 10th second, save it as a gif image, and even this need to use small videos to make memes can be quickly solved." Liu Cong said.
Liu Qingfeng introduced that according to the code ability public test set HumanEval built by OpenAI, the effect of the Spark V1.5 Python language is only 41 points, but V2.0 has reached 61 points, close to ChatGPT. According to the test set of code real-world scenarios constructed by the National Key Laboratory of Cognitive Intelligence, iFLYTEK Xinghuo Cognitive Big Model V2.0 has surpassed ChatGPT in the dimension of code generation and completion. According to the plan, the capabilities of each dimension of iFLYTEK Spark code will surpass ChatGPT on October 24 this year, and benchmark GPT-4 in the first half of next year.
Multimodal capability upgrade,
Spark large model live "reading picture"
Took a Meitu and want to send a circle of friends with pictures and texts "in seconds"? Open your mouth and want to draw a picture that comes to mind? Just enter a piece of text and want to generate a small video with a single click? These functions can be realized by iFLYTEK Spark Cognitive Model V2.0.
At the press conference, Liu Qingfeng released the multi-modal capabilities of iFLYTEK Xinghuo, and iFLYTEK Xinghuo upgraded its image description, image Q&A, image recognition, text and image generation, and virtual human synthesis. Liu Cong continued to measure the real machine on the spot, let the Spark model "read the picture" on the spot, generate various photos such as people and landscapes, generate the class hand-copied newspaper, and "paint" according to ancient poems.
"Multimodal capability is a rigid need to empower the industry, and it is also the only way to achieve general artificial intelligence." Liu Qingfeng said 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 learn, train and improve at the product terminal, including more flexible and autonomous industrial robots, better autonomous driving, and companion robots that enter the home.
It is understood that in the past three years, iFLYTEK has won 17 international authoritative evaluation champions in the field of multi-modality, and at the beginning of 2022, iFLYTEK released the "iFLYTEK Super Brain 2030 Plan" including core capabilities such as multi-mode perception, deep understanding, multi-dimensional expression, and motor intelligence, allowing robots that understand knowledge, are good at learning, and can evolve into every family.
The upgrade of multimodal capabilities will also bring industrial opportunities to AIGC. At the press conference, the product after the multimodal capability upgrade - iFLYTEK Smart Work 2.0 was officially released, whether it is video post-processing or creative video production, iFLYTEK Smart Work 2.0 equipped with the core capabilities of iFLYTEK Spark can be easily done. iFLYTEK Smart 2.0 can further lower the threshold of short video production, meet more diversified video production needs, and promote the development of AIGC industry.
iFLYTEK joins hands with Huawei,
Build a base of localized computing power
On July 6 this year, iFLYTEK announced that iFLYTEK Spark will join forces with Ascend AI to create a new foundation of universal intelligence based on China's independent innovation. At this press conference, iFLYTEK and Huawei once again cooperated to jointly release the Spark all-in-one machine, which allows enterprises to privately deploy large models more conveniently, autonomously, safer, and controllably on the platform of domestic independent innovation.
Liu Qingfeng said that the era of deep empowerment of cognitive big models has arrived, and healthy development urgently needs to build safe and controllable guarantees. In his view, the deep application of cognitive big models in the industry has three key elements: first, safety and controllability, general artificial intelligence to improve people's livelihood, empower society, and development should be rooted in independent controllable, computing power security platform; The second is scenario-driven, which can produce real application value in visible and tangible scenes, and can use statistical data to prove application results; The third is special training, building a privatization exclusive model, protecting users' proprietary data and intellectual property rights, and improving the application effect of the industry.
In terms of security and controllability, Liu Qingfeng introduced that iFLYTEK and Huawei are jointly tackling the problem of "stuck neck" computing power. iFLYTEK has a self-developed large model training platform, which has advantages such as 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.
At the meeting, iFLYTEK Xinghuo and Huawei Ascend jointly released the Spark all-in-one machine, which can provide five customized optimization modes such as dialogue development, task orchestration, plug-in execution, knowledge access, and prompt engineering, as well as more than 10 kinds of out-of-the-box rich scenario packages such as office, code, customer service, operation and maintenance, marketing, and procurement, and support three 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. According to the data of 9 exclusive large models in the fields of finance, government affairs, and automobiles, iFLYTEK Xinghuo shows that the average effect in scene task optimization and private domain knowledge enhancement can be improved by 20%.
AI policy implementation,
"100 Model Wars" upgraded to "Unlimited Game"
On the day that iFLYTEK held its new product launch conference, the aforementioned Measures were officially implemented, which is also the first normative policy for the generative artificial intelligence industry in the mainland.
The Measures propose: "Encourage the innovative application of generative AI technology in various industries and fields, generate positive, healthy and upward high-quality content, explore and optimize application scenarios, and build an application ecosystem." ”
James Cass, a professor in the Department of History of Religion at New York University, argues in his book The Game of the Limited and the Infinite that finite games are played within boundaries, and that infinite games are played within boundaries. From the current competitive situation in the field of global large models, the applications and products launched by various companies continue to break through the "boundary" of people's imagination of artificial intelligence, and the "hundred model wars" of the artificial intelligence industry have been upgraded to "unlimited games" without suspense.
At the 2023 All-China Federation of Industry and Commerce Chairman's High-end Summit and National Excellent and Strong Private Enterprises to Promote Henan's High-quality Development Conference, Zhou Hongyi, founder of 360, said in an interview with Dahe Finance Cube that the big model represents an opportunity for an industrial revolution and will greatly improve the management and production efficiency of the government and enterprises.
"The enterprise-level market represented by cities and government departments is an important direction for the development of the big model, and each city will have its own proprietary GPT in the future." Zhou Hongyi said in an interview with Dahe Finance Cube that now 360 has signed a strategic cooperation with Zhengzhou City, which will rely on the data advantages accumulated by 360 search, terminal and other businesses, integrate Zhengzhou's proprietary data, and create a "safe and good, credible and controllable" urban GPT brain for Zhengzhou.
The day before the new product launch of iFLYTEK, Xiaomi Group officially announced its embrace of large models. On the evening of August 14, Lei Jun, founder, chairman and CEO of Xiaomi Group, officially announced at the Xiaomi new product launch conference that the large model on the mobile phone side initially ran through, and Xiaoai upgraded the AI large model and opened the invitation test.
Lei Jun introduced that Xiaomi's self-developed large model with 6 billion parameters ranked first in the same parameter magnitude on the C-EVAL authoritative list, and ranked first in CMMU Chinese large model; Xiaomi's self-developed end-side large model has been running through the Snapdragon platform, and the effect of the self-developed 1.3 billion parameter end-side large model can be comparable to the industry's 6 billion parameter cloud large model in some scenarios.
In June this year, China Literature Group released its organizational upgrade plan and established four business divisions: Content Ecological Platform Division, Film and Television Business Division, Intelligence and Platform R&D Division, and Enterprise Development Division, the core purpose of which is to open up the "content + platform" and use AIGC to improve IP incubation and ecological efficiency.
One month after the organizational business upgrade, China Literature Group released China's first large-scale model in the web literature industry, "Reading Essay Miaobi", and the application product "Writer's Assistant Miaobi Edition" based on this large model, which played a positive role in improving creative efficiency and upgrading operational tools, which is also the first step for China Literature to embrace AIGC and promote the IP industry to a new look.
Responsible editor: Wang Shidan | Reviewer: Li Jinyu | Reviewed: Li Zhen | Executive producer: Wan Junwei