Domestic Top 5 AI Startups, Who is China's Open AI?

GPT is soaring, AI investment is surging, and the entire venture capital market is heating up due to artificial intelligence.

Scarce companies can become unicorns in minutes, and new challengers are entering the game one after another. The AI competition is becoming increasingly fierce, and the "battle of the gods" of China's "big models" has already begun.

In February of this year, Meituan-W co-founder Wang Huiwen's "AI Hero List" made a high-profile entry into the "big model" field. Shortly thereafter, a group of big shots, including former Sogou CEO Wang Xiaochuan, former JD AI leader Zhou Bowen, and Alibaba's vice president of technology Jia Yangqing, who is known as the "number one Chinese in the AI framework field," rushed to join the AI entrepreneurship team.

There are also many start-up teams with gorgeous backgrounds emerging. For example, Tang Jie, a professor at Tsinghua University, founded Zhipu AI, Huang Minlie, a professor, founded Lingxin Intelligence, and Yang Zhilin, a professor at Tsinghua University, founded Cyclic Intelligence... They have all been given the hope of overturning the industry by top VCs, and the entrepreneurial train is speeding up again.

The technology media The Information has listed the top 5 Chinese AI startups. Who will be China's OpenAI?

1. MiniMax

Financing information: Angel round

Establishment date: 2021-11-03

Location: Shanghai

Affiliated enterprise: Dreaming (Shanghai) Technology Co., Ltd.

An AI company founded by Yan Junjie, former vice president of SenseTime and head of general intelligent technology, has already built the basic model architecture of the three major modes of text-to-vision, text-to-speech, and text-to-text. It is also the most valuable startup company in the current AI big model entrepreneurship wave.

On June 1, according to Reuters, MiniMax has completed a new round of financing of more than 250 million US dollars, and the company's valuation has exceeded 1.2 billion US dollars. In this round of financing, entities associated with Tencent participated, and the investment amount may be 40 million US dollars.

Previously, MiniMax had completed two rounds of financing, with investors including miHoYo, IDG Capital, Hillhouse Capital, Yunqi Capital, and Mingshi Capital. Yunqi Capital confirmed in a post in April that it had invested in MiniMax in 2021 and was the only early investment institution in the angel round.

On February 16 of this year, MiniMax revealed at a small media communication meeting held in Beijing that the team has more than 100 members, and the core technology research and development members of the company come from world-renowned universities and top technology companies around the world, with world-class industrial and academic experience in natural language processing, speech, computer vision, computer graphics, and other fields. One-third of the team members have doctoral degrees from the world's top technology labs.

It is reported that MiniMax starts directly from the bottom-level basic model and independently develops three foundation models-text to vision, text to audio, and text to text.

2. Langboat LanZhou Technology

Financing Information: Pre-A+ Round

Establishment Date: 2021-06-10

Location: Beijing

Parent Company: Beijing LanZhou Technology Co., Ltd.

LanZhou Technology's founder, Zhou Ming, decided to start his own business when the domestic AI market was at its lowest point.

At the end of 2020, Zhou Ming considered resigning from his position as Vice President of Microsoft Asia Research Institute. Many friends advised him not to leave, but he was determined to start a large-scale model business, believing that "large-scale models will become a kind of infrastructure in the future."

In 2021, Zhou Ming officially established LanZhou Technology and became an AI company incubated from 0 by Kai-Fu Lee's Innovation Works. Zhou Ming once pointed out that LanZhou Technology is committed to solving the problem of human language understanding and generation, providing open-source large-scale models based on NLP (natural language understanding) technology, as well as functional engines and applications focused on marketing, finance, cultural creativity and other scenarios.

The main products are a series of capability platforms and vertical scenario applications based on the core technology of "Mencius Large Model". Multiple technologies and products have been implemented, including Mencius Large Model, AIGC (Intelligent Creation) Platform, Machine Translation Platform, Financial NLP Platform, etc., which have landed in enterprises such as Tonghuashun and Huaxia Fund. Combining ChatGPT technology, LanZhou Technology has launched a dialogue robot MChat, which can help users complete various work tasks in specific scenarios through intelligent dialogue.

In March of this year, LanZhou Technology completed the Pre-A+ round of financing. This round of financing was led by Beijing Zhongguancun Science City Company, with follow-up investments from Sidao Capital and Innovation Works. Within less than a year, LanZhou Technology's total financing amount has reached hundreds of millions of yuan.

At the "New Opportunities from AI1.0 to AI2.0" trend sharing event held by Innovation Works on March 14, LanZhou Technology officially released the "ChatGPT-like" language generation model-Mencius MChat controllable large model.

Mencius MChat controllable large model emphasizes its own "controllable" characteristics-the model's capabilities are more flexible than other similar technologies, and landing in vertical fields and professional tracks will be more focused, and can make rapid adjustments according to industry, region and other needs. According to Zhou Ming, Mencius MChat controllable large model has the following characteristics:

Large models with 10B and 100B parameter levels will be launched successively;

It has multiple capabilities such as chat, question answering, translation, text generation, and information extraction;

It can integrate search results, domain data, and knowledge graphs; Controllability exists in terms of functionality, style, and human cognition.

When discussing the future direction of the industry, Zhou Ming frankly stated that the current ChatGPT technology still lacks reasoning, logic, mathematics and arithmetic, factual errors, and other aspects. In the future, the nine major problems related to large models are particularly worthy of attention, involving reasoning ability, factual correctness, Chinese processing ability, and other aspects.

3. Zhipu AI

Financing information: Series B

Establishment date: 2019-06-11

Location: Beijing

Affiliated enterprise: Beijing Zhipu Huazhang Technology Co., Ltd.

Zhipu AI was founded by Professor Tang Jie of the Department of Computer Science at Tsinghua University. The core members of the team have participated in the development of the "Wudao" project, a cooperative project between Tsinghua University and Zhuyuan Research Institute. In August 2022, the ultra-large-scale pre-training language model GLM-130B developed by the Knowledge Engineering Laboratory of Tsinghua University in cooperation with Zhipu AI was officially launched. It is the only global mainstream large model selected by Stanford evaluation in Asia that year.

GLM still performs very well in key indicators such as accuracy compared to large models of companies such as OpenAI, Alphabet-C Brain, and Meta, and surpasses GPT-3, Alphabet-C's PaLM, and Meta's OPT large models in MMLU, LAMBADA, and BIG-bench-lite index tests.

On May 16th of this year, 360 announced a strategic cooperation with Zhipu AI. The two parties jointly developed a trillion-level large model "360GLM". The two parties will refer to the cooperation model of "Microsoft + OpenAI" to combine large models with application scenarios.

Zhou Hongyi, CEO of 360, believes that China should establish a collaborative innovation model of production, research, and development between large technology companies and key research institutions to create a "Microsoft + OpenAI" combination that leads the development of large model technology in China. He said that this cooperation with Zhipu AI is precisely based on this collaborative relationship between production, research, and development. Regarding this cooperation, Zhang Peng, CEO of Zhipu AI, stated that Zhipu AI has always adhered to its vision of making machines think like humans and realizing the concept of Model as a Service (MaaS).

Currently, the training data volume of the model is 400 billion, with half in Chinese and half in English, and has 130 billion parameters. The training cost ranges from one to ten million RMB. As of May 1st this year, the model has received download and usage requests from more than 1,000 research institutions in 69 countries.

Based on GLM-130B, Zhipu AI has conducted supervised fine-tuning to obtain the ChatGLM model. ChatGLM model is currently the most advanced open source large model in China and has been opened for internal testing.

Can Zhou Hongyi's Microsoft dream come true? Can Zhipu AI become China's next OpenAI?

4. Light Years Away

Financing information: Angel +

Establishment date: 2023

Location: Beijing

Wang Huiwen, co-founder of MEITUAN-W, has founded a new AI start-up company called Light Years Away, which has attracted much attention since its inception.

In February, Wang Huiwen, Li Zhifei, founder of Chumenwenwen, and two partners from ZhenFund, Dai Yusen and Liu Yuan, had dinner together. The four of them witnessed the changes brought by ChatGPT, and Wang Huiwen's attitude was "must participate".

When talking about the rise, Wang Huiwen picked up his mobile phone and announced his attitude of joining the game in a high-profile manner:

With a self-funded 50 million US dollars, he hopes to join a suitable company in the ChatGPT craze.

Two days later, when more entrepreneurial details were disclosed, Wang Huiwen's idea had obviously changed from "seeking teammates" to "I am organizing the team".

This commander-in-chief not only announced the recruitment of "top technical talents", but also boasted that everyone can rest assured to display their talents and leave the trivial matters to him to handle, and there is no need to worry about funding-

In addition to his personal investment of 50 million US dollars based on a valuation of 200 million US dollars, "the next round of financing has already received VC subscriptions of 230 million US dollars"!

Since then, Wang Huiwen has successively posted recruitment notices for product managers, algorithm engineers, interns, etc.

Entering March, the second "big shot" who invested beyond Lightyear appeared - Wang Xing, the founder of MEITUAN-W and an old friend of Wang Huiwen.

This MEITUAN-W big shot joined the team by "investing" and participating in the A round of investment, and serving as a director beyond Lightyear. Wang Xing and Wang Huiwen are classmates and roommates at Tsinghua University. They have cooperated in entrepreneurship several times and created many well-known brands such as the campus network and MEITUAN-W.

Once again fighting side by side, Wang Xing explained the reason behind it simply in the ticket circle:

The AI big model makes me excited about the huge productivity that is about to be created, and worried about its impact on the whole world in the future. Since Lao Wang is determined to embrace this big wave, I must support him.

On April 6th, the operation beyond Lightyear officially began, and Wang Huiwen updated his dynamics in his circle of friends.

The "Double Wang" sign of MEITUAN-W also makes it easier for "Beyond Lightyear" to attract funds in the venture capital market. After all, venture capital institutions should also want to seize the opportunity to "recreate MEITUAN-W".

Wang Xing and Wang Huiwen, with MEITUAN-W, survived the battle of a thousand groups and a hundred groups, but it took nine years to achieve full-year profitability for the first time. This time, can they survive until the birth of "MEITUAN-W" in ChatGPT version?

5. Yang Zhilin

In this intense big model "arms race", there are many outstanding post-90s entrepreneurs. Information mentioned Yang Zhilin, the co-founder of Recurrent Intelligence.

Yang Zhilin studied computer science at Tsinghua University and was taught by Tang Jie, the founder of Zhiju AI. He graduated with excellent grades in 2015. Afterwards, he went to the Language Technology Institute (LTI) of Carnegie Mellon University, which ranks first in the world in natural language processing (NLP) research, to pursue a PhD. He was taught by Ruslan Salakhutdinov, the head of Apple AI research, and William Cohen, the chief scientist of Alphabet-C. He obtained his PhD degree in four years. During his PhD, Yang Zhilin collaborated with Turing Award winner Yoshua Bengio to release the "HotpotQA" dataset, and published XLNet and Transformer-XL as the first author, which had a significant impact in the field of NLP, becoming one of the highest cited papers at NeurIPS 2019 and ACL 2019, with Alphabet-C's academic citations directly exceeding ten thousand...

In 2016, Yang Zhilin founded Recurrent Intelligence, a company whose main business is to use AI technologies such as NLP, speech, multimodality, and large models to create "sales technology" solutions. He led multiple AI projects for Zhipei AI and the Tsinghua research team, and Huawei's "Pangu" large model was also jointly launched by Yang Zhilin's team and Huawei Cloud.

Currently, Recurrent Intelligence has completed its Series B financing and achieved over 200% revenue growth for three consecutive years.

Epilogue

However, the burning speed of AI large model research and development is also a "gap" that cannot be ignored by all participants.

The cost of a single training of the ultra-large-scale language generation model GPT-3 launched by OpenAI is as high as 4.6 million US dollars.

Senior AI research expert Tian Taoyuan further explained the cost in a media interview: "Training GPT3.5 once requires a cost of 3-4.6 million US dollars, which is only the cost of computing power, not the cost of talent. OpenAI has about 375 employees, and the annual salary expenditure alone is 200 million US dollars. The AI computing power expenditure is 500 million US dollars, which requires strong capital support."

No matter who ascends to the throne of China's AI startups, a new round of "blood and tears" is inevitable.