(Image source: edited and photographed by Titanium Media App)
On May 20, it was reported that Titanium Media App exclusively learned that Tencent is planning to invest in the dark side of the moon (Moonshot AI), a large model startup, hoping to establish a deeper cooperation between WeChat and the dark side of the moon.
Titanium Media App sought a response from Tencent and the dark side of the moon on this matter. However, as of press time, neither Tencent nor the dark side of the moon have commented.
It is worth mentioning that before this, the dark side of the moon has received $1 billion in financing from Alibaba Group, Ant, etc. Once the above financing is completed, the dark side of the moon will become the largest single round of financing in the domestic AI field, and the fourth large-scale model company (after Zhipu, Baichuan, and MiniMax) to receive investment from two giants, Alibaba and Tencent.
At present, the new round of pre-investment valuation of the dark side of the moon has reached 3 billion US dollars.
It is reported that the dark side of the moon was established in March 2023 and is a key player in the field of large models in China. Its core team has participated in the research and development of many large models such as Google, Pangu, and Zhiyuan Wudao.
Yang Zhilin, founder and CEO of the Dark Side of the Moon
Yang Zhilin, founder and CEO of the Dark Side of the Moon, graduated from Tsinghua University with a bachelor's degree, graduated from Carnegie Mellon University with a Ph.D. in computer science, worked in Google Brain (Google Brain) and the American start-up FAIR, and studied under Ruslan Salakhutdinov, the head of artificial intelligence at Apple.
He is also the most cited researcher in the field of NLP (Natural Language Processing) under the age of 35 in China, and the first author of two important papers, Transformer-XL and XLNet, both of which are core technologies in the field of large language models. The other two co-founders, Zhou Xinyu and Wu Yuxin, have more than 10,000 Google Scholar citations.
In terms of personnel, as of March this year, the Dark Side of the Moon team has more than 80 people.
At the product level, the dark side of the moon has completed the layout from the general large model to the upper-level application. Among them, the basic layer of the large model, the dark side of the moon, has trained a self-developed general large model at the level of 100 billion, and has obtained the filing and approval of the domestic large model; Application layer, in October 2023, the dark side of the moon launched the world's first intelligent assistant product Kimi that supports the input of 200,000 Chinese characters, supporting long text input of 200,000 Chinese characters, focusing on lossless memory and "long text (Long Context)", known as ChatGPT Chinese replacement by netizens, good at reading long texts, searching web pages, and can be used in scenarios such as meeting minutes, assisted programming, and copywriting.
Recently, the Kimi intelligent assistant on the dark side of the moon began to test the paid function in a small range of grayscale through the method of "tipping". At present, there are 6 Kimi reward schemes, the prices are 5.2 yuan, 9.99 yuan, 28.8 yuan, 49.9 yuan, 99 yuan and 399 yuan, and you can get 4 days, 8 days, 23 days, 40 days, 93 days and 365 days of peak priority after tipping, and click "Go to pay" to pop up the WeChat payment page.
A few days ago, Yang Zhilin revealed that multimodality is the key direction of the company's continuous investment, and the company hopes to strike a balance between commercialization and technological development.
"We hope to strike a balance between 'climbing stairs' and 'seeing the scenery', and there is a lot of room for improvement in the ability of large models, and we hope to focus our energy and priority on 'climbing stairs'." Yang Zhilin said that as a commercial exploration of "seeing the scenery", the company is also trying and releasing some new commercial functions.
At the closed-door event of MiraclePlus Space on May 16, Yang Zhilin said that the most important product capability of AI is the model, especially in the next two or three years, the most important thing is the model, and the model determines your boundary. Trust between humans and AI is particularly important, but today, humans don't trust AI enough because AI is not robust enough.
"The model has two core competencies, one is how many things you can do, and the other is how many points you can do in each thing. Applying the thinking framework of traditional products, it is divided into several parts, including function, interaction, growth and other directions. The functional direction is basically determined by the model capability, which refers not only to the algorithm, but also to the product. Of course, as a new product, the function is very important, and the interaction growth must also be done well. Yang Zhilin said.
When it comes to commercialization, Yang Zhilin said that commercialization will determine how many users your product can grow, and sometimes the product is not necessarily a bottleneck, the focus is to find a scalable growth method. There could be some great business innovations here.
For example, he said that there are currently two business models for large models: subscription model and commission model.
"Subscription model: It is a business model that has been proven and established now, and it may be of different sizes in different markets in each country, but it has a limited ceiling: according to how many people use it to charge, there is no way to create more and more value with the product, and there is no way to create more and more commercialization. So I don't think this model is the final model.
Draw mode: 1. From the user's time attention to the traditional advertisement. Advertising has been verified by the Internet now, and its certainty is higher. But because people's attention and time are limited, the opportunities for this business model may not be that great; 2. Take a cut from GDP. Essentially, your product is creating a new GDP, and a large part of GDP is created by intelligence, and intelligence used to be generated in only one place, and now suddenly there are new places that can be generated, so it is essentially taking a cut from the incremental GDP. Yang Zhilin said.
Talking about business decisions, Yang Zhilin pointed out that the scale of control cannot be controlled for the sake of controlling the number of people, and the essence is to cut business. For entrepreneurs, the most important thing is to focus and not do too much. If we all do the direction of OpenAl, the team may feel safe, but it is not a good thing. A lot of times we have to make some good decisions, and although they do it, we don't have to do it one by one so that we can play to our advantage.
Regarding the financing mentality, Yang Zhilin said frankly, don't be afraid of rejection, the mentality should be strong, entrepreneurship is a new thing, there must be few people who can understand, and the mentality must be good. The second is the specific strategy, and it is particularly important to arrange the order of how to meet investors in each round of financing, and the mentality is the most important when chatting.
Yang Zhilin emphasized that he believes in the first principles brought by the Scaling law, and for a long time in the future, as the model becomes stronger, the conversational Al boundary will be expanded, not only language interaction, but also multimodal interaction, and graphical interfaces can be generated at any time, so it is possible to directly generate an interactive interface you want to solve this problem.
"The next most important thing is the go-to-person logic. It is quite possible to start with the productive forces and then slowly extend to life. The user will give some new context every time, and each time it doesn't need too many things, it can give the user a preference, and make a real go-to-person entrance. Yang Zhilin said.
(This article was first published on Titanium Media App, author | Lin Zhijia, editor | Hu Runfeng)