AsianFin--Tencent is mulling investing in Moonshot AI, a large model startup, in an effort to establish deeper ties between its social media platform WeChat and Moonshot AI.

TMTPost reached out to both Tencent and Moonshot AI for comments on this matter. As of the time of publication, neither company has responded.

Moonshot AI had previously secured $1 billion in financing from Alibaba Group and Ant Group. Upon completion of this financing round, Moonshot AI will become the fourth domestic large-model company to receive investments from both Alibaba and Tencent, following Zhipu, Baichuan AI, and MiniMax, with the highest single-round financing in China’s AI field.

Currently, Moonshot AI's pre-money valuation has hit $3 billion.

Founded in March 2023, Moonshot AI is a key player in China’s large-model field. Its core team had engaged in the development of multiple large models at Google, Pangu, and Zhiyuan Wudao.

Yang Zhilin, Moonshot AI's founder and CEO, graduated with a bachelor's degree from Tsinghua University and a PhD in computer science from Carnegie Mellon University. He had worked at Google Brain and the U.S. startup FAIR, under the mentorship of Apple's AI chief Ruslan Salakhutdinov. Yang has extensive entrepreneurial experience and has co-authored papers with several Turing Award winners, contributing to the early development of large models like Pangu and Wudao in China.

Yang is also the most cited researcher in the NLP field under the age of 35 in China and the first author of two important papers, Transformer-XL and XLNet, which are core technologies in the large language model field. His co-founders, Zhou Xinyu and Wu Yuxin, each have over 10,000 citations on Google Scholar.

On the product side, Moonshot AI has completed the layout from general large models to upper-layer applications. At the foundational level, Moonshot AI has trained a self-developed general large model with hundreds of billions of parameters and obtained domestic large model registration approval. As to application, in October 2023, Moonshot AI launched the world's first intelligent assistant product, Kimi, capable of supporting the input of 200,000 Chinese characters.

Kimi, known as the Chinese alternative to ChatGPT, excels in reading long texts and web searches, and can be used in scenarios such as meeting minutes, coding assistance, and copywriting.

Recently, Kimi began testing paid features in a small-scale gray test through "tipping". Currently, six tipping options are priced at 5.2 yuan, 9.99 yuan, 28.8 yuan, 49.9 yuan, 99 yuan, and 399 yuan are available, which provide priority access during peak periods for four days, eight days, 23 days, 40 days, 93 days, and 365 days, respectively. After clicking "Proceed to Payment," the WeChat payment page pops up.

Yang revealed that multimodality is a key focus area for the company, aiming to balance commercialization and technological development.

"We hope to find a balance between 'climbing up the mountain' and 'enjoying the view.' The capability of large models has significant room for improvement, and we hope to focus our energy and priorities on 'climbing the mountain','" Yang said.

He noted that the company is also exploring and releasing new commercial functions as part of its commercialization efforts.

At the closed-door event of Miracle Mile Creation Space on May 16, Yang said that the most important product capability for AI is the model itself, especially in the next two to three years. The model determines your boundaries. Trust between humans and AI is crucial, but today, human finds AI not trustworthy due to insufficient robustness of AI.

"The model has two core capabilities: how many things you can do and how well you can do each thing. Applying traditional product thinking, it is divided into several aspects: functionality, interaction, and growth. Functionality is primarily determined by the model's capability, which includes not only the algorithm but also the product. As a new product, functionality is important, but interaction and growth must also be well managed," Yang said.

Regarding commercialization, Yang pointed out that commercialization will determine how many users a product can attract. Sometimes the product itself may not be the bottleneck; the key is to find a scalable growth method. This may lead to significant commercial innovations.

He cited two current business models for large models: subscription and commission models.

"Subscription model is a proven commercial model. It may vary in scale in different markets, but its ceiling is limited. It charges based on the number of users, which cannot grow proportionally with the increasing value created by the product. So, I believe this model is not the final one.

Commission model entails taking a cut from users' time and attention. Advertising has been tried and tested to be a feasible approach. But since human attention and time are limited, the opportunities for this business model may not be as large; or entails taking a cut from GDP. Essentially, launching a new product is creating GDP. Most of the GDP was previously generated by intelligence in one place, but now new places can suddenly generate it. So essentially, it is taking a cut from the incremental GDP," Yang said.

Regarding business decisions, Yang mentioned that controlling scale should not just be about controlling numbers but fundamentally about cutting business. For entrepreneurs, the most important thing is to focus and not do too many things. “If we did everything OpenAI did, the team might feel secure, but it wouldn’t be a good thing. Often, making good decisions means not doing what others do, to play to our strengths,” he said.

When it comes to financing strategy, Yang stressed the importance of resilience and not fearing rejection. Entrepreneurship is about something new, and few people will understand it initially. Having a right mindset is crucial. The specific strategy for arranging investor meetings in each financing round is also very important, and maintaining the right mindset during discussions is paramount.

Yang emphasized his belief in the first principles brought by scaling laws. In the long run, as the model's capabilities strengthen, the boundaries of conversational AI will expand to include not only language interaction but also multimodal interactions, where graphical interfaces can be generated on demand. This could directly generate the desired interaction interface to solve problems.

"The next most important thing is still the logic of 'go-to-person.' It's likely to start from productivity and gradually extend to everyday life. Users will provide new contexts each time, and it won’t take much to give users a preference, making it a true go-to-person entry point," Yang said.