news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Hengyu from Aofei Temple
Quantum Bit | Public Account QbitAI

In the AI ​​2.0 era, at least two of Wang Xing, Huang Zheng, and Zhang Yiming are definitely already working on it.

This is one of the most popular phrases about big model entrepreneurship recently.

Although it is an analogy of lamenting that “each era has its own opportunities”, some people are indeed already creating information flow products for the AI ​​era and creating the "Toutiao" of the AI ​​2.0 era.

A product calledAsk XiaobaiThe App was quietly launched, which has the shadow of an assistant for products in the AI ​​era, and also provides new-era thinking for information flow products.

The creator behind it is also a trend-setter in the era of mobile Internet recommendation algorithms.

And Kuaishou’s founder Su Hua has already placed his bet.

Ask Xiaobai, what kind of App is it?

Wen Xiaobai has just been launched online, but it has taken shape so far.

Its positioning isInformation AssistantUsers can actively ask questions or passively accept content pushed to them.

In addition to the user settings interface, there are two main interfaces. One interface combines the information flow recommendation of Chatbot and AI-generated content, and the other interface is a review of recent conversations.

First of allChatbot FeaturesIt has the basic functions of similar chat robots on the market and supports voice input.

According to the official introduction, one of the outstanding features of the Wen Xiaobai Chatbot part is its very low latency, with "the average first-word delay being within 2.3s."

The reasons behind this are also vaguely revealed. First, the team behind the scenes has developed its own data capture-quality screening and filtering process.Build a trillion-level high-quality database。

thenCreated a knowledge link data synthesis algorithm,This linking algorithm synthesizes and filters out billions of high-quality knowledge link data from trillions of high-quality data to help improve the efficiency and quality of knowledge extraction.

Quantum位 has learned more in-depth information that Yuanshi has been following the MoE route since it started developing its own large model in the first half of 2023.

The reason is that the team saw early on the technical trend that MoE is a more efficient architecture for large models, and has better performance and cost when implemented on the C-end.

How early? The founder said that he had practical experience in MoE since 2018...

Let's look at it againAI-generated content information flow recommendation function, currently a single-column article display format is used.

It is worth noting thatAll recommendations are generated by the team’s self-developed large model。

And each person has a different face.

This is not just about different recommendations for each person, but more importantly,Recommended content is closely related to the conversation between the user and Chatbot。

If users often chat with Chatbot about content related to throwing eggs, the recommendation function will be more inclined to recommend in-depth knowledge related to throwing eggs rather than simply hot news related to throwing eggs.

In other words, the more you ask Chatbot, the better the recommendation function will understand you.

At this point, perhaps you can feel a little more truth in what the leader of the Xiaobai team said:

What we want to do is to create a 2C large-model content recommendation product that focuses more on mid- to long-term content and covers a wider range of topics.

so,

Who created Ask Xiaobai?

Li Yan,Founder and CEO of Yuanshi Technology, former head of MMU (Multimedia understanding) at Kuaishou.

He is a veteran of Kuaishou, and his employee number is around 75.

In 2015, with the support of Su Hua, the then CEO of Kuaishou, Li Yan took the lead in establishing the first deep learning department within Kuaishou, with the goal of building an algorithm model to identify illegal and irregular video content.

This team later evolved into Kuaishou's MMU department today, which not only deals with security and compliance issues, but is also responsible for the research and development of multi-modal algorithm models such as voice, text, and images.

In other words, Li Yan personallyResponsible for building Kuaishou’s AI algorithm and search department with about 500 people from 0 to 1.

In 2022, Li Yan quietly chose to resign, and the news was not disclosed to the public. It was not until half a year later that a media report claimed that "Li Yan, the former core figure of Kuaishou AI, started a business and devoted himself to the big model track."

So Yuanshi Technology was bornThe office is located in Tsinghua Science Park, Wudaokou, one of the places with the highest concentration of large-scale model startups in the country.

However, Li Yan's entrepreneurial journey, like his departure from Kuaishou, has always been low-key. For more than a year before the product was unveiled, Li Yan and the company had been silent and no news could be found on the Internet.

Apart from being exposed as being bet by Su Hua and Jingwei, what exactly did he do? What track did he choose? What is the name of his product? Asking around the industry, not many people know.

An investor close to Li Yan told Quantum位 that Li Yan is a true i-person. Since starting his business, many media outlets have sought to report on him but were turned down. Being a man of few words, he decisively chose to focus on development and launch products as soon as possible, hoping to let the products speak for themselves, "otherwise there would be no point in speaking out."

Now, everything Yuanshi Technology is doing has finally surfaced:

DoLarge Model Products, the initial implementation form is to ask Xiaobai.

Today, the Yuanshi Technology team has about 50 people, and many core team members come from the speech, language, and visual model fields of the Chinese Academy of Sciences, as well as Carnegie Mellon University, the University of Maryland, College Park, and the National Engineering Laboratory for Brain-Inspired Intelligence of the University of Science and Technology of China.

We need to create large-scale products that “do not exchange for short-term user happiness”

According to Li Yan himself, on the first day of starting his business, he knew exactly what he wanted to do.That is a single large-scale toC application, that is, a large-scale model product。

When he entered the market, the domestic Hundred Model War was in full swing, and making "large model products" was not a particularly sexy story.

However, even though it is determined that in order to make large-model products, Yuanshi Technology must have its own large models, it has not changed its words and called itself a large-model company.

Eventually, the long-term vision became clearer and the product goals became more explicit.

Li Yan said that Chatbot and information flow recommendation have been widely implemented, but Yuanshi wants to make productsIt’s not just a combination of two tools.。

Yuanshi’s ultimate goal is to help users consume mid- to long-term content that makes them feel comfortable and happy (rather than in exchange for short-term happiness) through technological innovation, enter a flow state, and fight against the mental information entropy brought about by the information explosion in the AI ​​era.

It can be said,Yuanshi used a very strong subjective consciousness to define the question of Xiaobai。

So, will Ask Xiaobai be the super application created by Yuanshi Technology?

We did not seek a clear answer.

However, according to Yuanshi’s internal statement, the company’s positioning of it is “very important” and the so-called"Large model products" are different from the products launched by the current large model companies that follow the modeling integrated route.。

Its conception includes both new thinking about information flow in the era of recommendation algorithms and new practices of smart assistants in the era of AI 2.0.

In Li Yan's words, asking Xiaobai at least answered what kind of super application Yuanshi wants to make.

Again:

In the AI ​​era of information explosion, AI can answer all questions while providing mid- and long-term content that understands users in order to alleviate anxiety.

This is also why Quantum位 learned from an unnamed employee of Yuanshi that the difficulty in advancing Xiaobai is not in developing a large model on its own, but in "how to convert product language into technical language."

Let me say a few more words. During Quantum位's continuous exploration of Yuanshi Technology, I once heard a rumor that "Li Yan will promote multimodal products."

But since Yuanshi was founded, the only product it has launched is Ask Xiaobai.

At present, this product has only entered the field of AI content using text and images.

However, considering Li Yan's previous experience,In the future, Wen Xiaobai will adopt more cross-modalRich media format?

Yuan Shi's mouth was really tight-lipped, and Quantum Bit never tried to find out from him, either directly or indirectly.

However, according to my usual understanding, not saying "no" directly may mean that it may be possible in the future, which means please stay tuned (dog head).

Finally, let's play a joke——

Li Yan used to be in charge of multimodal AI at Kuaishou. Is it because he doesn’t want to promote more modal products now that he doesn’t want to?

Hey, it’s really because he doesn’t want to (slip away) for the time being.