Why are AI developers so crazy? (II)

all-night online battles, hundreds of models competing, and thousands of companies challenging each other|Blogs|Tech Web|

2023-09-06 16:47:06 883

Only by moving forward

In fact, whether it's ChatLaw, the large-scale legal model developed by Peking University, or WPS AI by Kingsoft, or even Baidu Wenxin Yiyu, an app with over a million downloads in 24 hours after its full release, these may all be successful phenomenon-level applications. However, relying solely on these apps is not enough to drive the domestic AI industry onto the fast track.

As the saying goes, "One tree does not make a forest, and a garden is filled with flowers."

In this regard, I have to mention the "Wenxin Cup" entrepreneurial competition initiated by Baidu on May 31st this year, which is the first domestic AI application-based startup competition.

As the saying goes, "When the large-scale model rings, there is a fortune of gold." The cost of training a large-scale model can easily start at millions of dollars, which has become an industry consensus.

Originally, developers who wanted to use applications based on ChatGPT or Wenxin's large-scale model had to pay for it, which was also an industry consensus.

However, in the Baidu "Wenxin Cup" entrepreneurial competition, invitations were extended to all AI entrepreneurs to promote the implementation of native AI applications. Baidu made its Wenxin large-scale model and Baidu Intelligent Cloud Qianfan platform open for free access by all qualified participants. The Baidu Intelligent Cloud Qianfan platform is even "ready to use."

Moreover, according to the competition rules, the winner of the highest award will also receive an early-stage investment worth 10 million yuan.

As a benchmark in the AI field, Baidu has created a stage with both technology and funding. Entrepreneurial teams from China and abroad have responded enthusiastically. In less than a month, nearly 1,000 projects have fiercely competed with each other.

Yesterday (September 5th), after more than three months, the "Wenxin Cup" entrepreneurial competition is coming to an end. After authoritative evaluations and layers of selection by well-known investors and AI experts, 15 promising projects have ultimately stood out.

Among these 15 entrepreneurial projects, application-layer startups account for over 80% of the total. Both To C (consumer-oriented) and To B (business-oriented) applications have a nearly equal share. There are AI-native applications such as intelligent assistants, community social platforms, and productivity tools, as well as the reconstruction and transformation of AI applications in industries like healthcare, education, finance, and law.

For example, Buysmart is an intelligent shopping assistant specifically designed for overseas e-commerce. It functions like an AI shopping guide, continuously narrowing down options based on user preferences and ultimately helping buyers make purchasing decisions.

Another example is FantomTech, which focuses on the research and development of new materials. Leveraging materials informatics and integrating AI, computational physics, and high-throughput experiments, they combine data from various disciplines to enable forward prediction and reverse design, accelerating the industrial implementation of new material research and development.

Another startup called Genie AI focuses specifically on the second-dimensional (2D) community. The founders themselves are niche 2D enthusiasts who enjoy creating derivative works of anime characters and engaging in discussions with like-minded individuals within the community. However, creating such derivative works requires a certain level of artistic skills, posing a high barrier for those who are less skilled in hands-on creation.

For instance, in the game "Genshin Impact," without AI assistance, 2D enthusiasts can only admire characters created by others. However, with the help of Wenxin's large language model and the Stable Diffusion (SD) image-based generative model, Genie AI only needs simple natural language inputs to make it possible for characters to transcend time and space. For example, they can make Hu Tao become a classmate sitting next to them, or transform Ganyu into a member of the school's cheerleading team by putting her in a school uniform. Xinyan can also be dressed in jeans and dance street dance by the roadside, and so on.

According to the research report "Mapping Research on China's Large-scale Artificial Intelligence Models" released by the Ministry of Science and Technology at the end of May, there have been 79 large-scale models with over 1 billion parameters in China, almost entering a "hundred-model war."

With so many large-scale models available, one of the reasons why entrepreneurs in the application layer choose Wenxin's large model is not only due to the generous offerings from Baidu in terms of providing free access to the models and startup funds, but also because Wenxin's GPT-3.5 excels in four core artificial intelligence abilities: understanding, generation, logic, and memory. It ranks among the top tier internationally.

The authoritative institution IDC released the "AI Large Model Technical Capability Assessment Report, 2023," which shows that Wenxin's large model achieved a perfect score in seven aspects. It received a unique full score in the two core indicators of "algorithm models" and "industry coverage."

According to an article in "China Science Newspaper," comprehensive evaluations were conducted on three authoritative benchmark tests: AGIEval (benchmark test for human-like abilities of basic models), C-Eval (comprehensive evaluation for Chinese language models), and MMLU (a multiple-choice test comprising 14,000 questions across 57 topics covering natural sciences, social sciences, arts, humanities, and supporting multiple languages). In these evaluations, Wenxin's GPT-3.5 model surpassed ChatGPT and also outperformed GPT-4 in Chinese testing projects.

Moreover, the iteration speed of Wenxin's GPT-3.5 is astonishing. Compared to the 3.0 version in March, the training speed of GPT-3.5 has tripled, and the inference speed has increased by over 30 times. Furthermore, the 4.0 version is expected to be launched by the end of this year.

Whether it's the explosive technological iterations, the "hundred-model war," or the competitive development in various application fields, everyone is dedicated and giving their all in this pursuit.

It is difficult for us to imagine any other industry that would disregard the late hour and compete relentlessly to release their products. It is also hard to fathom the excitement in the hearts of practitioners when they press the final decision button in the early morning hours.

To be continued...

Prev:: Why are AI developers so crazy? (III)

Next:: Why are AI developers so crazy? (I)

Why are AI developers so crazy? (II)

Advisory Service