**: Hunting Cloud Selection, Wen Wang Fei
At a time when OpenAI's Wensheng ** model SORA has attracted widespread attention, the large model company founded by Tsinghua Xueba, a post-90s generation in China, continues to be sought after by well-known institutions.
Today, Moonshot AI was revealed to have completed a new round of financing of more than $1 billion, with investors including Sequoia China, Xiaohongshu, Meituan, and Alibaba, followed by old shareholders, with a post-investment valuation of about $2.5 billion.
On February 3, just before the Spring Festival, the dark side of the moon was revealed to be in the process of raising $200 million, jointly invested by Ant and Alibaba Group, with a pre-investment valuation of $1.5 billion.
In response to the "latest round of financing", Moon Zhidian said to Lieyun.com, "Thank you for your attention! It is not convenient for the company to comment on the specific financing information for the time being. The dark side of the moon has been insisting on promoting the underlying key technology progress and product innovation in China's AGI field in the past and future, and will continue to match the capital strategy corresponding to the company's development stage, and look forward to more good news to share with you in the future. ”
In fact, the dark side of the moon, which was established in April 2023, has "few clear words" when it comes to financing disclosures. In response to the first round of financing that was completed only two months after its establishment, founder Yang Zhilin "corrected" in October 2023 that the company has received nearly 2 billion yuan of investment from well-known institutions such as Sequoia Capital, Today Capital, and Lisi Capital.
And this is also the only accurate financing disclosure since the establishment of the dark side of the moon nearly a year ago.
The reason why the dark side of the moon was able to get the head VC bet at the beginning of its establishment has a lot to do with the post-90s Yang Zhilin's "scholar" identity and rich experience.
During his studies at Tsinghua University, Yang studied under Professor Tang Jie, the head of the Knowledge Engineering Laboratory (KEG) of the Department of Computer Science at Tsinghua University, and the academic associate dean of KLCII and the leader of the Enlightenment Project. In the end, he passed all programming courses with perfect marks and graduated first in his grade.
Then, in 2015, Yang joined the Language Technology Institute (LTI) at Carnegie Mellon University (CMU), where he studied with Ruslan Salakhutdinov, head of AI at Apple, and William W., chief scientist of AI intelligence at GoogleCohen is pursuing a PhD.
After graduation, Yang worked at Google Brain Research Institute and Meta (Facebook) Artificial Intelligence Research Institute, and was the first author of Transformer-XL and XLNet. Among them, the XLNet model has achieved better results than Google BERT in 18 natural language tasks, and is one of the popular international cutting-edge models in the NLP field at that time.
According to incomplete statistics, Yang Zhilin has published more than 20 articles in computer summits such as ICLR, NEURIPS, ICML, ACL, EMNLP, etc., and his research results have accumulated more than 17,000 Google Shcolar citations.
Currently, Yang is also an assistant professor at the Institute for Interdisciplinary Information Sciences at Tsinghua University, with research interests in large-scale pre-training, natural language processing, natural language understanding and generation, few-shot learning, zero-shot learning, and multimodal learning.
Yang Zhilin, born in the 90s, is well-known in the field of large models: Circular Intelligence, Zhipu AI, and Zhiyuan Research Institute ......His name and figure are everywhere.
At the same time, Yang Zhilin and his team have also participated in the research and development of large models such as Google Bard, Gemini, Einstein, Pangu, and Wudao as core R&D members, and invented milestone achievements in the AI field including Transformer XL, XLNet, Rope, Detectron2, and Group Normalization, which have been adopted by models such as Google Palm and LLAMA.
Tianyancha APP information shows that the dark side of the moon is 78 owned by Yang Zhilin97% with absolute control. The entrepreneurial partners around him are also extraordinary and should not be underestimated.
Zhou Xinyu, the co-founder of the dark side of the moon, owns 10% of the company's shares, and he, along with Yang Zhilin and Zhang Yutao, is a 2011 undergraduate student of the Department of Computer Science and Technology at Tsinghua University. In his senior year, Zhou Xinyu joined the Megvii internship that met his standards in all aspects, and officially joined after graduation, and the work content is the mass production of algorithms, which is to increase the production efficiency of algorithms many times.
As a stake 596% of the third largest shareholder, Wu Yuxin, co-founder of the Dark Side of the Moon, graduated from Tsinghua University and Carnegie Mellon University, and was nominated for the best of the best at the 2018 European Computer Vision Conference (ECCV). In October 2018, Iyswim was the only team out of six teams to crack the facial recognition algorithm at the GeekPWN International Security Geek Competition. Wu Yuxin participated in the competition as an iyswim team at the time, and according to him, "I (who signed up for the competition in my own name, and a teammate did not come to the scene) used Google's Facenet open source ** model to break the algorithm."
In addition, Yang Zhilin's fellow brother Zhang Yutao currently holds 5% of the company's shares. According to public information, Zhang Yutao studied in the Department of Computer Science of Tsinghua University. His research direction is heterogeneous data fusion and knowledge graph construction, and he has published many articles at top computer conferences such as KDD and CIKM. As the technical leader, he was involved in the development of the technology big data analysis platform aminer.
With a luxurious team lineup and deep accumulation, the dark side of the moon was established less than half a year ago, and it was announced in October 2023 that it had achieved a breakthrough in the field of "long text".
According to Yang Zhilin, in response to the "application difficulties caused by the limited input length of large models", the dark side of the moon officially launched the first large model Moonshot that supports the input of 200,000 Chinese characters, and Kimi Chat, a smart assistant product equipped with this model.
Subsequently, he gave a detailed introduction with some practical use cases of Kimi Chat. Taking the entire book "The Moon and Sixpence" as an example, Kimi Chat can read it with users to help them better understand and apply the knowledge in the book:
Compared with the current large model services on the market based on English-based training, Kimi Chat has strong multilingual capabilities. For example, Kimi Chat has a significant advantage in Chinese, and the actual use effect can support the context of about 200,000 Chinese characters,25 times that of Anthropic's Claude-100K (measured about 80,000 words), 8 times that of OpenAI's GPT-4-32K (measured about 2.).50,000 words).
At the same time, through innovative network structure and engineering optimization, Kimi Chat realizes a lossless long-range attention mechanism under 100 billion parameters, and does not rely on "shortcut" solutions that have great performance damage such as sliding windows, downsampling, and small models.
On January 26 of this year, Kimi Chat released the latest "V13 Spring Festival Edition": The basic model capabilities have been comprehensively upgraded, and UFIDA's online search capabilities, contextual learning capabilities, literary creation capabilities, and language translation capabilities have been ......The mini program version of the Kimi smart assistant already supports voice input in Chinese and English.
What is needed is that recently, OpenAI's Wensheng ** model SORA has continued to attract attention with its breakthrough one-minute duration, coupled with the high realism and high quality of the demo**.
According to multiple sources, the dark side of the moon is also secretly developing a general multimodal model, which is expected to be launched within this year.
In other words, the dark side of the moon, which has made phased progress in the field of text-based large language models, will also compete with domestic and foreign counterparts such as OpenAI in the field of image-based and **-based multimodal models in the future.
The "frequent actions" of the dark side of the moon in the capital market seem to be making more adequate preparations for higher training costs, larger amounts of capital and talent needs.
Taking the team size as an example, the dark side of the moon has about 50 people in October 2023, and the company's team has more than 80 people.
In the future, we will continue to pay attention to whether the dark side of the moon can occupy a place in the field of "multimodal models".