Collaborating with Huawei to release the Spark all-in-one machine, iFlytek releases the Spark Cognitive Model 2.0 with the ability | Model | Spark
"Many insightful individuals and industry leaders in China are very anxious. In this era of general artificial intelligence, does China still have the opportunity and hope? Since December 15th last year, iFlytek has launched the research and development of the" 1+N "Spark Cognitive Model." 1 "is a benchmark for the ChatGPT General Cognitive Model, and" N "is an application in various industry tracks. Today, we not only have full confidence in algorithms, but also have sufficient guarantees in computing power. I think China has stood firm in the key strategic link of general artificial intelligence!" On the afternoon of August 15th, the iFlytek Spark Cognitive Model 2.0 release conference with the theme of "liberating production capacity and unleashing imagination" was held. In Hefei, Liu Qingfeng, Chairman of iFlytek, gave a passionate and passionate introduction.
Compared to version 1.0 on May 6th and version 1.5 on June 9th, what are the features of the Starfire 2.0 that have jumped?
[Code capability will significantly lower the threshold for entrepreneurship in the digital economy]
It is reported that the various abilities of Starfire 2.0 continue to improve, with a benchmark of 100 points, 72 points for text generation, 78 points for language understanding, 70 points for knowledge Q&A, 60 points for logical reasoning, and 72 points for mathematical ability. "As long as it passes 60 points, it can empower many fields and improve work efficiency," said Liu Qingfeng.
Multiple on-site tests were conducted, a welcome speech for the press conference was written, and a mathematical problem involving trigonometric functions, equations, and arithmetic sequences was answered. After being set as a high emotional intelligence, high intelligence, and experienced foodie personality, a chat was conducted, and the Spark 2.0 version performed well.
After serving these "appetizers", let's take a look at the "coding ability" of one of the "main dishes".
Code ability is a key dimension that supports the intelligence of cognitive models, and it is a "hard hit" ability that is considered an important indicator of the intelligence level of large models. "Code proficiency will significantly lower the threshold for entrepreneurship in the digital economy. It is not necessary for everyone to be a programming expert. As long as you have imagination and a deep understanding of application scenarios, you can start a business." Liu Qingfeng introduced.
In on-site testing, the Starfire 2.0 version successfully completed coding tasks such as "handwriting in the air", "snake eating mini game", "drawing a three-dimensional diagram of the saddle surface equation using Python language, and setting gradient colors". In the comparison of actual application scenarios before this, the "code generation" and "code completion" capabilities of Starfire 2.0 have surpassed ChatGPT, and are very close to ChatGPT in terms of "code error correction", "code interpretation", and "unit test generation". "By October 24th this year, the code capabilities in these five dimensions will surpass ChatGPT," said Liu Qingfeng.
IFlytek has released a newly developed "Intelligent Programming Assistant" based on this code capability on site, which can seamlessly integrate its code capabilities into integrated development environments. A few weeks ago, over 2000 iFlytek employees tried the "Intelligent Programming Assistant", with a code adoption rate of about 38%, coding efficiency improved by more than 30%, and overall efficiency improved by 15%.
Liu Qingfeng gave an example of a school's academic affairs office wanting to generate an application to address the issue of non-standard leave requests from students. Although they do not have any programming experience, the teachers of the Academic Affairs Office only need to have a conversation with the "Intelligent Programming Assistant", and the assistant can automatically perform requirement analysis, generate data models, application pages and processes, and quickly complete application construction and launch. "The traditional development cycle for such an application is about 17 days, but now it can be completed in just 1 day through conversational development."
[New Upgrade of Multimodal Capability]
Multimodal ability is another main dish. This is iFlytek's established long-term strategy for artificial intelligence technology, and in the past three years, it has won 17 international authoritative evaluation championships in the multimodal field.
Through multiple tasks on site, the abilities of Starfire 2.0 in image description, image understanding, image reasoning, image recognition creation, text and image generation, and virtual human synthesis were demonstrated.
In actual testing, for a picture with a polar bear and a penguin, Spark 2.0 vividly described it as follows: "The polar bear stands on an ice block, while the penguin stands next to it, seemingly observing each other, or perhaps just enjoying this cold moment." When asked about any strange places in the picture, Spark 2.0 found that "polar bears mainly live in the Arctic region, while penguins mainly live in the Antarctic region. Therefore, it can be inferred that this penguin may have migrated from the Antarctic region to the Arctic region, or they may have appeared in zoos or aquariums." Migrating from Antarctica is unreliable, but appearing in zoos or aquariums at the same time is possible. Possible.
Write creative copy for Mao Feng of Mount Huangshan Mountain and create video of Mid Autumn Festival blessing. "The" iFlytek Smart Work "with over 3.75 million users has also been upgraded." Traditional dubbing and editing require at least one or two hours, but now with the help of AI, it can be solved in 5 minutes, "said Liu Qingfeng.
Dedicated to creating a "Spark Language Companion" for oral practice, the number of topics has increased from 73 to 393. The oral mock exam has added intelligent evaluations, and situational communication also supports custom scenarios for images/documents.
Developers increased by 282% year-on-year
In Liu Qingfeng's view, the key elements for the deep application of cognitive models in the industry are: safety and controllability, scenario driven, and exclusive models. The current content security issues of the cognitive big model mainly include dirty language materials and "serious nonsense". iFlytek, which is responsible for building the "National Engineering and Technology Center for Speech and Language," has formed a three-dimensional guarantee mechanism for this. In terms of computing power security, iFlytek and Huawei are jointly building a computing power cluster for training super large models. At the press conference, iFlytek and Huawei jointly released the Spark all-in-one machine, allowing enterprises to deploy large models on independent innovation platforms.
It is reported that the iFlytek Spark cognitive model has driven the flourishing development of the developer ecosystem, with a year-on-year growth of 282% among developers, with the highest number serving the enterprise. "Whether it's a team or an individual, you may not have professional development experience. As long as you have creativity and imagination, understand user needs, and insight into customer needs, you can use our platform to generate the required applications with just one click, greatly reducing the threshold for innovation and entrepreneurship."
"Regarding the future, whether you think it's possible or not, you will eventually be right." Liu Qingfeng quoted a famous quote to express his and iFlytek's determination. "Although the Spark Universal Model has many abilities that are already stronger than ChatGPT, overall it is still benchmarking against it. October 24th will be a milestone, and Spark's Chinese ability will surpass ChatGPT, while its English ability will be comparable. Next year, we will benchmark against GPT4."