Your Position Home AI Technology

The dark side of the moon Kimi suspends “burning money” launch, and insiders say that the number of App users is growing naturally now

(图片来源:GuShiio.comAGI编辑林志佳)

(Photo source: GuShiio.comAGI Editor Lin Zhijia)

GuShiio.comAGI reported on February 18 thatIt is reported that under the influence of DeepSeek, Kimi on the Dark Side of the Moon decided to postpone burning money.

In this regard, GuShiio.comAGI learned from people familiar with the matter that the core reason for this action is the DeepSeek craze, which has led to the recent surge in natural new users of Kimi smart assistants. This means that Dark Moon has adjusted its promotion actions accordingly.

In addition, GuShiio.comAGI also learned that among the recent large models of the Six Tigers (Dark Side of the Moon, Baichuan Intelligent, Intelligent AI, Zero One Everything, MiniMax, Step Star), in addition to the recent launch of Step Star, the remaining five companies have reduced or suspended advertising.

It should be noted that Kimi on the Dark Side of the Moon has no new products recently, and it is normal not to place advertisements. For example, MiniMax was only briefly launched and promoted before and after the new product was released in December last year. Before that, MiniMax would not be released in channels such as Xiaohongshu and Douyin.

From the perspective of the Xiaohongshu platform, currently, ByteDance Bean Bag and Baidu WenXiaoyan (Baidu Library) are both important head-advertising parties. AI application and hardware companies such as Keep have also begun to gradually invest in Xiaohongshu, Douyin and other channels.

It is reported thatDark Side of the Moon was established in 2023 and mainly develops domestic AI models and applications. The founder and CEO of the company, Yang Zhilin, graduated from the Department of Computer Science of Tsinghua University and has a doctorate in computer science from Carnegie Mellon University in the United States. He has worked on Meta AI and Google AI research and development teams. He is also the first author of two important papers, Transformer-XL and XLNet. Both of them involve core technologies in the field of large language models and are the most cited researchers in the field of NLP (Natural Language Processing) in China under the age of 35. After returning to China, he led his team to participate in the research and development of the core technology of Pangu Model, and also led the product research and development of Circular Intelligence of the NLP Company he co-founded.

At the product level,The dark side of the moon has completed the layout from a general model to an upper-level application.

Among them, at the basic level of the large model, the Dark Side of the Moon has trained hundreds of billions of self-developed universal large models, and has obtained the filing and approval of domestic large models; at the application level, in October 2023, the Dark Side of the Moon launched the world’s first intelligent assistant product Kimi that supports input of 200,000 Chinese characters, supports long text input of 200,000 Chinese characters, and focuses on lossless memory and long text. In October last year, the Dark Side of the Moon also released the latest mathematical model k0-math, benchmarking two publicly usable models in the OpenAI o1 series: o1-mini and o1-preview.

Yang Zhilin told GuShiio.comAGI in October last year that AI assistant Kimi earned more than 36 million a month in October and continued to grow faster.He also emphasized that improving retention is Kimi’s current core goal.

On February 18, Dark Side of the Moon announced the launch of a new model, Kimi Latest.It will benchmark the model currently used by Kimi’s smart assistant and upgrade it simultaneously with the smart assistant product updates (the model name remains unchanged). According to reports, Kimi Latest always uses Kimi smart assistant products to use the latest Kimi model, which may contain features that have not yet been stabilized; the context length is 128k, and the 8k/32k/128k model will be automatically selected as the billing model based on the context length; Support automatic context caching. Currently, Kimi Latest only supports calling the Kimi normal version model, and the Kimi k1.5 long thinking model does not support API calls for the time being.

FinancingAli, Tencent, and Ant have successively invested in the large-scale model startup Moonshot AI. As of now, the new round of pre-investment valuation of Moonshot AI has exceeded US$3.3 billion (approximately RMB 20 billion), making it the highest-valued company among the six big models, but lower than DeepSeek’s US$150 billion valuation.

It is worth noting that.comAGI previously conducted statistics based on AppGrowing platform data,Currently, in the entire domestic App market, in October last year, Pinduo ranked first in advertising spending, reaching 1.159 billion yuan; Douyin placed 971 million yuan in advertising, ranking second, but ranked first in the number of advertising ideas, far higher than Pinduo, while Kimi ranks 16th, which is about one-tenth of Pinduo’s monthly advertising spending, reaching 199 million yuan. However, Kimi Smart Assistant is a free App application and currently has no payment plan.

As of October 29 last year, AI applications such as kimi smart assistant, ByteDance Bean Bag, and Tencent Yuanbao spent more than RMB 300 million in advertising (streaming) on the entire network in October.

ButAs the DeepSeek craze hit, the promotion of tap water flow has become mainstream, which allows Dark Moon to reflect on the water flow. On February 17, Flood Sung, a researcher on the Dark Side of the Moon, revealed in a sharing that the effectiveness of long thought chains was actually known more than a year ago, but in order to save money, they preferred to tackle long texts rather than long thought chains (Long-CoT); after the launch of OpenAI o1 in September last year, Dark Side of the Moon decided to extend the thought chain.

Flood Sung said that they realize the importance of long contexts, but prioritize making the text long and do not pay enough attention to long thought chains. ldquo; The main consideration is the cost issue. He said that long context is mainly for long text input, with pre-filling and Mooncake (Kimi’s underlying reasoning platform that has been gradually open source). The cost and speed are controllable, while the long thought chain is long text output, which costs much higher and the speed is much slower. In this case, making the output longer is not a high option.& ldquo; The release of OpenAI o1 was shocking and the effects exploded. The effectiveness of Long-CoT made me think.” rdquo;

However, GuShiio.comAGI learned that for platforms such as Zhihu, Xiaohongshu, and Station B, the discussion popularity of Kimi, Doubao and DeepSeek is still higher than that of other AI products.

Flood Sung said that AGI (Artificial General Intelligence) is indeed a matter at hand, and now he is thinking about ASL (Artificial Super Intelligence) again. For people who do reinforcement learning (RL), they have never regarded achieving AGI as a goal. Now give AI a measurable goal, and then let AI explore by itself, and improve the model through reinforcement learning. In the future, this process is just a process that is constantly replicated into more complex scenarios.

 

Popular Articles