The wave of large models has swept faster than any technological revolution in history.
At the beginning of the year, OpenAI released Sora, and Musk and Zhou Hongyi did not hesitate to praise it, and the frenzy spread all over the world. Just as everyone was still marveling that "the future is here", Anthropic announced the official release of Claude-3 and announced that it surpassed ChatGPT-4 in the AI logic benchmark. At the same time, Google and Musk jumped into the open source war at the same time, launching Gemma and Grok respectively.
The competition boom of the global model has spread outward from the niche circle and spread to the world like ripples. An infinite competition for large models has begun. The global tech giants are carrying large sums of money and technical talents with beautiful resumes to burn the battlefield even hotter.
But if you look at it calmly, most of the world's recognized leaders in large models are young AI companies: OpenAI, with a valuation of $29 billion, was founded in 2015; Anthropic, a large model company known as OpenAI's "strongest competitor", was founded in 2021; The "European Rookie" Mistra has just been established for a year. The combined number of employees of the three shining AI star companies may be less than 1,000, which is only the number of people in a department of a major technology company.
Why is this chase for the crown of the big model not within the range of the giants?
01 Why does Glory belong to an AI company?
First of all, to dispel a misconception: the big model is not an arms race that relies solely on resources.
After ChatGPT came out, a widely circulated claim was that the key to OpenAI's success was to rely on tens of thousands of A100 cards on Microsoft's Azure cloud, costing hundreds of millions of dollars. What's more, they believe that the winning rate of a large model depends on the abundance of resources.
However, in March this year, the American startup Databricks suddenly announced its open-source large language model DBRX, which is claimed to be the world's most powerful open-source large model, with a parameter scale of 132 billion, and its performance surpasses Meta's Llama2, Mistral AI's Mixtral, and Musk's xAI company's just-open-sourced Grok-1.
What's more, they only spent 2 months and $10 million, surpassing GPT-3.5 in performance across the board, while training in a fraction of the time and cost of GPT-3.5.
In fact, resources are important for competition in any field, but resources are not the only factor in the field of large models. Compared with technology giants, AI companies have a unique advantage, which lies in the flexibility of technological exploration.
Google used to be the undisputed king of AI deep learning. In 2016, Alpha Go, which defeated human Go champion Lee Sedol, came from Google Deepmind, and Google was also far ahead in the field of natural language models.
However, in 2022, ChatGPT was born. In fact, it is the disagreement on the technical route that widens the gap between the positions. The natural language model pursued by Google should be a series of vertical classes, with relatively small parameters and a relatively narrow application scenario, while OpenAI believes that it should be a general super model trained with massive parameters and massive data.
OpenAI当年的梦想看来是天方夜谭。 但即便在与谷歌的较量中长期落于下风，OpenAI也没有放弃将GPT作为唯一路线。 阿尔特曼的一句话给出了答案——“创业公司做什么都很难，那不如抓住大机会。 ”（Startups are very hard no matter what you do , you may as well go after a big opportunity.）
Compared with the flexibility of AI companies, it is difficult for tech giants to bet on the technology route so desperately, which also causes the slow action to a certain extent. This is also destined that the latest direction of AI is likely to be explored by AI companies.
Overseas AI star companies are lined up and star-studded. So the question is, where are the AI star companies in the Chinese field?
02 "Chinese Large Model Five Tigers" surfaced
A SuperCLUE review list has revealed the secrets of China's large model Five Tigers.
Excluding BAT, the list shows the strongest AI companies in China, namely Baichuan Intelligence, Zhipu AI, Dark Side of the Moon, MiniMax, and Cloudwalk Technology.
Baichuan Intelligence was founded in April 2023 by Wang Xiaochuan, the former CEO of Sogou. In just a few months after its establishment, Baichuan Intelligent has become one of the unicorns of science and technology, and has continuously released 11 basic large models since its establishment.
Founded in 2019, Zhipu AI has completed a new round of financing earlier this year. The founder, Zhang Peng, graduated from the Department of Computer Science and Technology of Tsinghua University. In the past few years, the company has continuously released GLM series large models, ChatGLM, CodeGeeX code large models, etc., and has become one of the earliest and most experienced enterprises in the research and development of large models in China.
Founded in 2021, Yan Junjie, who graduated from the Institute of Automation of the Chinese Academy of Sciences, was responsible for building the toolchain and underlying algorithms of deep learning, as well as the technical development of general intelligence, during his tenure as vice president of SenseTime. The co-founder, Yang Bin, is an alumnus of Yan Junjie at the Chinese Academy of Sciences.
Also menacing is Moonshot AI - in 2023, the Dark Side of the Moon will complete two rounds of financing totaling nearly 2 billion yuan in succession, with a valuation of more than 2.5 billion US dollars. The Dark Side of the Moon was founded in March 2023 by three Tsinghua alumni, led by Yang Zhilin, a post-90s student.
At this point, with the addition of Cloudwalk Technology, China's AI model Five Tigers is about to emerge.
Cloudwalk was once the first A-share stock in the "AI Four Tigers", and it was also the only company with a fully domestic background in the domestic AI troika. In the industry, Yuncong Technology, SenseTime Technology, Megvii Technology, and YITU Technology are known as the "AI Four Little Dragons", but in terms of halo and popularity, the youngest Yuncong Technology is the most "top-notch" existence.
"AI National Team Player", "Incubation Enterprise of the Chinese Academy of Sciences", "China's First Echelon AI Enterprise", "The First Artificial Intelligence Enterprise to Undertake the Construction of Major Projects such as the National Development and Reform Commission's Artificial Intelligence Basic Platform, Application Platform, and the Open Platform of the Ministry of Science and Technology, and Participate in the Formulation of National and Industry Standards" and other titles, so that Cloudwalk Technology has always been in the industry spotlight since its birth in 2015.
Today, Cloudwalk Technology is striding forward in the era of large models.
03 An honor student who is favored by the times
In May last year, Cloudwalk Technology, which ranks in the first echelon of artificial intelligence in China, officially released a strategic product in the field of AI agent (AI-agent) - "calm multi-modal large model".
Sun Jin, product director of Cloudwalk Technology Research Institute, said in an interview with the media: The large model has undergone many rounds of iterations internally. In version 1.5, balancing context length, model performance, and inference cost is the focus of iteration. Version 2.0 of the Graceful Model has been completed, and version 3.0 focuses on multimodal capabilities – skipping text and directly processing data of different modalities.
Not only is it a conversational experience, but it can also be used for programming, writing, problem solving, etc. In the process of answering the same real question, the answering speed of the large model is faster, but the reasoning ability and semantic understanding ability have exceeded GPT 3.5 and slightly lower than GPT 4.0.
After comprehensive evaluation by third-party organizations such as SuperClue and C-Eval, the comprehensive performance of the large model ranked among the top five in the world. At the same time, the large model has multi-modal capabilities, and has refreshed the world record 10 times in the field of vision and cross-modality.
According to Cloudwalk, the company has laid out dozens of industry models, and has developed a number of pan-AI intelligent applications such as DataGPT, intelligent customer service, and AI mouse, which have become an important starting point for Cloudwalk Technology's layout of AI agents.
At the same time, Cloudwalk and Huawei Ascend jointly proposed solutions to the challenges of intelligent computing infrastructure, opening a new layout of "localized computing power + intelligent computing". At present, the two parties have cooperated to launch a large-scale model application base - Calm Large Model Training and Push All-in-One Machine, and together with partners such as Tianjin Port Group, Shouchain Technology, Jinshiyuan, State Grid Shandong, China Telecom and other partners, they have successfully helped customers in ports, medicine, manufacturing, electric power, banking and other industries to implement generative AI application scenarios.
Cloudwalk has stepped into two waves of AI in a row, which is inseparable from its profound scientific research background.
Zhou Xi, the founder of Cloudwalk Technology, has been introduced back to China as an expert of the "Hundred Talents Program" of the Chinese Academy of Sciences, and served as the deputy director of the Information Institute of the Chongqing Institute of Green and Intelligent Technology of the Chinese Academy of Sciences. In half a year, he set up a team with more than 20 professionals across the country, and was selected as a Class A strategic leading science and technology project of the Chinese Academy of Sciences, becoming the only face recognition team among them.
In the future, this team became a national team in the field of computer vision, and its technical achievements have been used in many provinces, and before alpha go and Lee Sedol made AI explode overnight, it quietly brought this technology to the lives of ordinary people.
Until April 2015, 33-year-old Zhou Xi gave up the "iron rice bowl" that others envied, chose internal entrepreneurship, and took the lead in establishing Yuncong Technology, specializing in the field of face recognition.
"It's surprising because it's rare to see scientists in this field who are willing to give up their current positions and go all in on entrepreneurship." A person from Haitong Securities said in an interview with the media, "Haitong, as the first project of Yuncong, has many difficulties in application implementation, and Zhou Xi brought the entire R&D to Shanghai for a week, and there has been no problem since the system was launched." ”
Since 2015, face recognition, as the easiest track in the field of computer vision, has gradually formed an outlet. Seven years after the blue wisp of the road, Yuncong handed in his answer sheet: as the only fully domestic-funded AI company, it successfully landed on the Science and Technology Innovation Board, and was known as the "AI national team".
Looking at the prospectus of Cloudwalk Technology, we can see that most of the founding teams of Cloudwalk Technology are from the Chinese Academy of Sciences and the University of Science and Technology of China. The company has nearly 600 scientific research personnel, with R&D personnel accounting for more than 50%, and the core team has won the championship in the field of artificial intelligence at home and abroad for 10 times.
Nowadays, the global large-scale model competition is surging, and a new wave of science and technology has swept all walks of life, and the field of Chinese artificial intelligence has once again reached a crossroads in anxiety and catching up.
How to write this new chapter is not only a new topic for Cloudwalk Technology, but also an urgent mission for all Chinese AI companies.