news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Text/Wang Shuoguo Editor/Yan Ziwei

The two large model companies in which Huawei Hubble has quietly invested have released new developments one after another.

In mid-August, Mianbi Intelligence announced that its MiniCPM series of large language models has been downloaded over one million times since its launch in February.

At the end of July, Shengshu Technology launched the Wensheng video model Vidu globally, with performance comparable to Sora.

Mianbi Intelligence and Shengshu Technology were founded in 2022 and 2023 respectively, and their core teams are both from Tsinghua University. The CEO of Mianbi Intelligence is Li Dahai, the former CTO of Zhihu; the CEO of Shengshu Technology is Tang Jiayu, who studied computer science at Tsinghua University for both his undergraduate and master's degrees.

The AI ​​track is booming, and Huawei Hubble is optimistic about the potential of these two young talents.

New team

Habour Investment is an investment institution wholly owned by Huawei.

According to Qichacha, it has two entities, namely Hubble Technology Investment Co., Ltd. and Shenzhen Hubble Technology Investment Partnership.

Previously, Huawei Haobo's investment focused on hard technology and it invested in many semiconductor chip companies. This year, it has successively supported two Tsinghua start-ups, demonstrating its emphasis on the AI ​​track.

According to ITjuzi data, Hubble has made a series of bets on the field of artificial intelligence this year, with a total investment of 132 million yuan.

In the direction of large models, it prefers elites from prestigious universities.

The two companies in which the company invested have similar founding team structures and are both backed by Tsinghua University.

The core members of Shengshu Technology come from the Institute of Artificial Intelligence of Tsinghua University, and the founding team of Mianbi Intelligence was born out of the school's Natural Language Processing Laboratory (THUNLP).

A number of Tsinghua graduates make up the top management of Shengshu Technology. In addition to the CEO, its chief scientist is Zhu Jun, vice president of the Tsinghua Institute of Artificial Intelligence, and its CTO Bao Fan is a fellow student of Tang Jiayu and a member of Zhu Jun's research team.

From June to August last year, Shengshu Technology completed two rounds of financing, raising more than 100 million yuan in total. In June this year, Hubble invested in the company and participated in its A+ round of financing.

The growth path of wall-facing intelligence is similar.

Its co-founder Liu Zhiyuan is a doctoral supervisor in the Department of Computer Science at Tsinghua University, and his research direction is computer natural language processing; the company's CTO Zeng Guoyang, 26 years old this year, was a competition recommended student and entered the Tsinghua Natural Language Processing Laboratory in his sophomore year.

According to Li Dahai, the Mianbi Intelligent research team has more than 100 people, 80% of whom are graduates from Peking University and Tsinghua University. The average age is only 28 years old, and they have published more than 100 papers in authoritative journals and conferences.

Before founding Mianbi Intelligence, Li was the CTO of Zhihu. He built the search and recommendation business for the platform from scratch, launched the AI ​​"smart community", and increased the number of monthly active users.

Last year, ChatGPT was very popular and everyone was talking about AGI (artificial general intelligence). He talked to the core members of the company and decided to join after confirming that AGI was everyone's belief.

His former employer supported his entrepreneurship. In April of the same year, Mianbi Intelligence received tens of millions of RMB in investment from Zhihu. A year later, Hubble became a shareholder.

New products

Both startups launched new products in a relatively short period of time, which was an important reason for attracting investment from institutions such as Huawei and Habo.

In late April, Shengshu Technology and Tsinghua University jointly released a large video model - Vidu. At the end of July, Vidu was launched globally, opening up two core functions: text-generated video and image-generated video, providing two length options of 4 seconds and 8 seconds, with a maximum resolution of 1080P.

Vidu generates a 4-second clip in just 30 seconds. Currently, users can register directly with their email address to experience Vidu.

It is reported that the videos generated by Vidu are smooth and coherent, without obvious frame insertion, and have rich lens language, and can switch between different lenses such as long shot, close-up, and close-up.

"Vidu performs very well in terms of 16-second long-term retention and semantic understanding," said Zhu Jun, chief scientist of Shengshu Technology.

Facing the Wall Intelligence is also constantly launching new actions.

In May, its large model Luca was released. Li Dahai said that Luca's multiple language model capabilities are comparable to ChatGPT.

In the same month, the MiniCPM-V2.0 was launched, which can accurately identify street scenes with intricate details and read ancient handwriting on the Tsinghua Bamboo Slips from more than 2,300 years ago.

As early as last year, Mianbi Intelligence launched the ChatDev intelligent software development platform. Users who need to make small games, website development, creative design, etc. only need to describe the project name and related ideas through ChatDev to quickly realize them.

In Li Dahai's words, an ordinary user can make a small software "in the time it takes to drink a cup of Coke and at a cost of less than a dollar."

Commercialization has been initially implemented. Mianbi Intelligence has joined hands with China Merchants Bank, Shuke Network, Zhihu, etc. to apply technology to finance, education, smart terminals and other scenarios.

For example, at the end of June, the artificial intelligence-assisted trial system developed by the company was put into operation in the Shenzhen Intermediate People's Court, covering processes such as case filing, file review, trial, and document preparation.

Since its trial operation in January this year, the system has assisted in the filing of 291,000 cases and the generation of 11,600 draft documents.

A unique approach

Li Dahai and Tang Jiayu have one thing in common: they do not blindly believe in the paths taken by their predecessors.

For example, GPT emphasizes "great effort makes miracles happen", while the Wall-Facing Team's approach is to predict the performance of large models through small models: first train on a model with a parameter size of 0.009B to 0.03B, then extrapolate to a 2.4B model to predict performance, and finally, train the 2.4B model.

This method is more efficient and can reduce training costs by first conducting experiments and adjusting parameters on a small model.

Before the company was founded, as a member of the "Wudao" project of Beijing Zhiyuan Artificial Intelligence Research Institute, the Mianbi team began training large language models in 2020.

Past experience lets them know what kind of data is needed for large models.

"It is easy for people to fall into a misunderstanding and focus too much on the absolute amount of data. In fact, the quality of data, how to use data, and the understanding of data are more important things," said Zeng Guoyang.

Due to limited resources, the team has long used distributed acceleration, parameter fine-tuning and other methods to reduce costs. In 2022, the parameter fine-tuning work of the wall-facing team was also published in a Nature journal.

Shengshu Technology’s approach is similar.

In terms of technical route, Shengshu adopts the same fusion architecture as Sora, but the two are different in product path.

The Sora team chose to go all in on long videos, with the technical strength of Open AI and the computing power support of Microsoft behind it. The conditions of the startup company Shengshu Technology cannot be compared with them.

Tang Jiayu's team chose to start with 2D images and then expand to 3D and video fields.

Video is essentially an amplification of images in time series and can be viewed as multiple consecutive frames of images. The engineering work on images, such as data collection, cleaning, labeling, and efficient model training experience, can be reused.

Throughout 2023, Shengshu's main resources were placed on images and 3D. It was not until January this year that 4-second short video generation was launched. After the release of Sora in February, the company's progress accelerated and it was able to generate 16-second short videos in April.

In the large model race, teams from home and abroad competed to show off their strength. These two Tsinghua teams have just started warming up and are looking forward to achieving good results.

By then, Huawei Hubble, which invested early, is expected to obtain excess returns.