Ali Baidu 360 shot overnight, and the commercial battle of large models began
Fengse Mingmin from the Au Fei Temple
量子位 | 公众号 QbitAI
Overnight, domestic large models started a long text war.
From last night to the early hours of this morning, major factories lined up to officially announce the ability to follow up on long texts:
Baidu Wenxin said
The version will be upgraded next month, and the length of 2 million to 5 million will be opened
360 Smart Brain
5 million words are being tested internally, and as soon as it is completed, it will be settled in the 360AI browser
Ali Tongyi asked
As soon as you make a shot, it will be 10 million words, and you don't have to wait, you can start on the spot.
The involution is launched, and the large volume is specially rolled.
You know, the follow-up of long text ability means that you have to burn money harder...... And manufacturers seem to have no hesitation in making this decision.
Why?
All of this has to start with Kimi, who has recently been out of business.
Explosive Kimi, five expansions are still down
Kimi,全称kimi Chat,来自月之暗面(Moonshot Eye).
It is the first report card of AI genius Yang Zhilin after starting a large model. It debuted in early October last year, with hundreds of billions of parameters, and the main selling point is long text, supporting 200,000 Chinese character input processing, and a copy of "The Three-Body Problem" and "Musk's Biography" can be read casually. Coupled with the fact that it has always been free to use, Kimi has quickly amassed a wave of fans.
Its explosion happened this week, and at one point it burst into the fifth place in the App Store's free version of the app, surpassing WeChat.
According to statistics, the number of daily active users of the Kimi model has reached 1 million (600,000 in Mini Programs, 340,000 on the web, and 50,000 on the App), and the retention rate continues to rise.
The number of downloads on its App (free list of efficiency apps) has also squeezed into the top 10 from the top 100 in February, second only to Doubao (Douyin AI intelligent assistant) and Wenxin Yiyan.
Why did it suddenly explode?
Just this Monday, Kimi brought a major iteration, expanding the long context window by x10 times to 2 million words. Claude3 and GPT-4 Turbo are beyond their reach.
What used to take 10,000 hours to become an expert now takes only 10 minutes, and Kimi is close to the junior expert level in any new field.
Now netizens are having fun.
Nearly 800,000 words of "Dream of Red Mansions" and "Romance of the Three Kingdoms", and 1.5 million words of "Zhu Xian...... Throw them all in, play dialogue with your favorite characters, and it's free.
△ Source: Digital Life Kazik
With the continuous influx of onlookers after wave of onlookers, Kimi soon "couldn't hold on":
On Thursday afternoon, less than 72 hours after the release of the new feature, its mini program, App, and web version suddenly showed that they could not be used normally.
Ask it questions, and without exception tell you "I'm tired, come back later".
Official Explanation:
Since yesterday morning, we have detected a continuous and unusually high flow rate on the Kimi system, far exceeding our expected resource planning.
We've done 5 expansions.
Why is everyone having such a good time?
On the one hand, Kimi's underlying model capabilities are really strong.
Last month, Kimi ranked first in the "AI ChatBots" of large model startups:
The number of visits is up 107.6% from January (to 3.05 million, which is expected to increase to 5 million this month).
The official is also quite confident, and at the previous media communication meeting, they directly put a stack of thicker than a stack of "Traditional Chinese Medicine Internal Medicine", "Traditional Chinese Medicine Diagnostics", NVIDIA's 2023 annual financial report, and the script version of "The Legend of Zhen Huan" on the table for everyone to experience.
On the other hand, Kimi has been free since its launch, and at the same time, the product iteration is fast, and the long text that is "far ahead" has become a competitive differentiator.
So this big update directly made Kimi out of the circle.
Recently, Kimi concept stocks have appeared. At the close of trading on March 22, a number of A-share companies such as Huace Film and Television, Zhongguang Tianxue, and Huayang Lianzhong had daily limits. Among them, Huace Film and Television and Zhongguang Tianxue have achieved three boards.
(A number of companies on the current list have denied it)
The investment institution China Securities Construction Investment suggests that for the AI application sector, it is necessary to focus on scenarios that involve long text processing, such as legal texts, contract texts, knowledge base learning, reading, customer service, and other scenarios that require text knowledge output.
Therefore, it is not difficult to understand why friends followed up overnight.
And the bottom reason for this wave of collective actions is that this year has come to the moment of commercialization of large models to meet each other.
The battle for commercialization of large models begins
In the first three months of 2024, the latest actions of many large model manufacturers, including the dark side of the moon, Zhipu AI, Baidu, etc., are more focused on the field of commercialization.
First, let's look at the dark side of the moon.
In the past month, it has announced a lot of new news in terms of technology, commercialization, and financing.
On March 21, 3 days after the latest upgrade of the Kimi model, Hand Information revealed that it had paid attention to the fact that the Kimi model had started the AIGC platform docking test, and was currently looking for some scenarios to explore the possibility of landing.
△ Yang Zhilin, the founder of the dark side of the moon
Immediately after the person in charge of the dark side of the moon responded, there will be a preliminary plan for commercialization within the year:
The demand for commercialization is gradually becoming very strong, and some users will say that during the peak period, Kimi will prompt you to wait and ask if we can pay to solve these problems, so there will be a preliminary plan in this area this year, and you will see.
In terms of multimodal large models, the dark side of the moon has been revealed to be in the R&D layout and will be launched this year.
In terms of financing, in February this year, the dark side of the moon was revealed to have reached a valuation of $2.5 billion.
According to people familiar with the matter, the dark side of the moon won a new round of financing of 1 billion US dollars, of which 800 million US dollars came from new shareholders (Alibaba, Xiaohongshu, etc.), and 200 million US dollars were followed by old shareholders (Sequoia China, Lisi Capital, etc.). Meituan Dragon Ball led the company's last round of financing.
At the same time, the dark side of the moon is undergoing a large-scale launch, and Kimi is clearly perceived by the outside world.
On platforms such as Station B and Douyin, as well as airports and elevators, Kimi's advertisements can be seen (so Kimi's explosion in the past two months has something to do with this).
△ Screenshot from the up main dragon Tenghu Yuelong veterinarian video
Secondly, Zhipu AI, which is also a Tsinghua background, almost replicates the delivery route of the dark side of the moon. There was a fight in Station B and the elevator.
And this month, on the first anniversary of the ChatGLM 100 billion conversation model, Zhipu AI also disclosed a series of commercialization results.
This is the first time that Zhipu AI has comprehensively interpreted the commercialization progress and strategy in the trend of large models.
In addition to showcasing rich experience in landing cooperation and benchmark customer cases, Zhipu AI comprehensively interprets different landing models for different customers.
This kind of clear and unambiguous display is not only to show its own strength, but also to attract more companies to further promote the commercialization process.
In terms of financing, Beijing Artificial Intelligence Industry Investment Fund has just participated in a new round of financing completed by Zhipu AI at the beginning of this year. Previously, Zhipu AI said that it had received more than 2.5 billion yuan in financing in 2023 (as of October).
In addition to large-scale model start-ups, Baidu's latest large-scale model movement also focuses on the implementation of the industry.
The newly released five large models no longer simply volume parameters, but pay more attention to the cost performance and domain fit of the model, and further reduce the price.
In the latest news, Apple has also been revealed to be discussing the use of Baidu AI technology. Apple is looking for a local generative AI provider in China, and the conversation with Baidu is still in its early stages.
As soon as the news came out, Baidu's U.S. stocks rose more than 6% pre-market.
……
It is not difficult to see that in the first year of large-scale model application, manufacturers are no longer simply showing off their technical muscles, and every time they disclose public information, there will be commercialization considerations.
Especially in terms of product application experience, it has always been regarded as the strength of the domestic market. Some voices also believe that Kimi may be a turning point, and the next thinking of the domestic large-scale model market may have to shift from the evolution of basic capabilities to the understanding of the details of the product itself and customer needs.
△ Screenshot from the article of @小熊Run's fast public account
A number of manufacturers have officially announced overnight that they will enhance their long text processing capabilities, perhaps to promote a new trend.
All of the above trends indicate that the field of large models is about to enter a more crazy stage of burning money.
Volume length text processing capabilities, team expansion, large-scale advertising ...... These accounts, no matter how they are calculated, are astronomical.
So next, how to balance the income and expenditure of the main free domestic large-scale model tool -
Is it to start payment, rely on To B to drive To C, or burn financing?