aifollow.news
返回 量子位 AI 报道
量子位 AI 报道· · 原发布时间 AI 评分47

小红书上线基于 Agent 的交互式推荐

AI 辅助摘要

用户看到感兴趣的内容时,可在评论区写“多推类似内容”,让 Agent 识别偏好并调整推荐。报道还提到,小红书今年首次面向校招生开放产品工程师专项,招聘岗位也开始体现 AI 应用需求。

正文 · AI 翻译

At Xiaohongshu’s campus recruitment salon, a group of tech experts takes young people’s ambitions seriously

Tian Yanlin reporting from Aofeisi

QbitAI | WeChat public account QbitAI

Tian Yanlin reporting from Aofeisi

QbitAI | WeChat public account QbitAI

“If campus recruits have ideas of their own, can they apply for staff and computing resources?” “As AI gets better at writing code, what will we mainly do after joining?” “Will leaders really listen to ideas from newcomers?” “How should we develop ourselves in the AI era?”

Well, today’s new graduates ask practical questions that go straight to the heart of the matter!

On November 14 and 15, QbitAI spent two afternoons immersed in Xiaohongshu’s Beijing office, experiencing a tech campus recruitment event with a somewhat different feel.

This time, the “Potato Factory” brought its campus recruitment presentation directly into the office area and turned it into the REDpath campus recruitment tech salon.

There were few dry presentations from the stage. The salon reserved more time for the AMA (Ask Me Anything) session with the audience.

Campus recruits sat down with technical leaders they might one day work with and talked about everything from technical directions and team atmosphere to performance reviews and personal development.

For this in-person exchange, students traveled specially from Shanghai, Nanjing, Hong Kong and elsewhere. One graduating student studying in Singapore even flew to Beijing overnight.

By QbitAI’s rough count, nearly 200 applicants came to talk over those 2 days.

The students’ questions were unfailingly practical, and the technical leaders spoke frankly. One applicant who had come from Singapore told QbitAI:

“Compared with an online interview, this kind of face-to-face conversation lets you get a feel in advance for what the future team’s working atmosphere is really like. It’s great!”

“Compared with an online interview, this kind of face-to-face conversation lets you get a feel in advance for what the future team’s working atmosphere is really like. It’s great!”

At the event, the students were not the only ones asking questions.

When a technical leader heard that a new graduate had opened a store on Xiaohongshu and made more than 10,000 yuan in profit, their eyes lit up. They eagerly asked whether Xiaohongshu’s merchant workspace, “Qianfan,” was easy to use and what features needed improvement.

The students had come to learn about their prospective teams, but as the conversation went on, it turned into a user interview.

During the 2 days of conversations, I found that AI was an unavoidable topic for applicants at both the engineering and algorithm sessions.

In a conversation with the transactions and foundational data warehouse team, one student voiced his anxiety.

During an internship at a major tech company, he had worked on using AI to improve efficiency. But as the work went on, he began to worry: was he helping make his future self obsolete with his own hands?

In the community recommendation algorithm group, the discussion became more specific.

Students wanted to know: What role can AI play in recommendation algorithms? Can a large model use a person’s past behavior to help them discover a new interest they did not even know they had?

At an earlier campus recruitment event, a student had even asked Ke Xiong, head of multimodal algorithms on Xiaohongshu’s in-house Dots large-model team: With so many large models already on the market, why does Xiaohongshu still need to build its own?

Ke Xiong’s answer at the time was that building a large model that understands everyday life and can help users solve complex problems is difficult. It may involve work spanning several roles, including tour guide, designer and purchaser.

That is one reason Xiaohongshu has persisted in developing its own model.

The platform has vast amounts of multimodal content, including posts with text and images, videos, and audio in comment sections. Users often describe problems in their lives in long passages of natural language.

This distinctive content and these user needs place greater demands on a model’s ability to understand and generate content.

But to actually build such a large model, computing power, data and infrastructure, as well as business systems built on top of them, such as search and recommendation, all have to keep pace.

This year, Xiaohongshu opened a dedicated PE (Product Engineer) track to campus recruits for the first time.

In the past, front-end, client-side and back-end engineers each had their own responsibilities. Now, with AI, engineers have an opportunity to move beyond their familiar technical modules and take part in everything from understanding product needs and designing solutions to development and delivery.

QbitAI also learned at the event that a position in the transactions and foundational data warehouse group called “data development” last year had become “AI applications and development” in this year’s recruitment.

And to encourage employees to use AI well, Xiaohongshu issues tokens every month. Employees can request more when they run out, with ample allowance available.

Even in written tests for nontechnical positions, Xiaohongshu places particular emphasis on the ability to apply AI.

Apparently, in this year’s campus recruitment, applicants for some nontechnical roles must complete a written test involving an AI demo before they have a chance at an offer. (All right, even Vibe Coding is getting competitive!)

This strong embrace of AI has also given engineers more room to put their skills to work.

What do they do with the time and energy it frees up?

Engineers have begun proactively building agents and developing platform tools to help business teams solve real production problems.

With AI’s help, people are becoming more capable individuals, delivering comprehensive solutions to business teams one after another.

At present, several Xiaohongshu technical teams are trying to bring agents into their daily work.

For example, the recommendation algorithm team has launched agent-based interactive recommendations, allowing users to adjust the recommendation system according to their preferences.

When users see content that interests them, they can write directly in the comments, “Show me more content like this.” The agent can recognize the request and help the recommendation system understand them better.

The transactions and foundational data warehouse team has also built a Data Agent. Each person develops different Skills and builds a knowledge base for the business area they are responsible for. They then compare the accuracy of their agents and gradually combine the capabilities that work well.

Seen this way, the student who had worried about “making his future self obsolete with his own hands” can probably rest easy.

Of course, as AI becomes more capable, engineers cannot “go looking for nails because they have a hammer.”

Ji Hao, who leads the video side of the recommendation algorithm team, does not advocate using large models for every problem.

“To solve the same recommendation problem, one person might use a large model and expend a great deal of computing power, only to get business results roughly the same as someone else using a simple strategy. Is using the large model worth it then?”

In his view, identifying a problem is important, and so is deciding which technology should be used to solve it.

That is something he pays particular attention to when interviewing young algorithm engineers now.

Ji Hao believes students just out of school have an advantage: they have not yet been constrained by the industry’s established ways of working, so they are more likely to come up with different ideas.

That creative spark should not be diminished by AI.

Still, having ideas is 1 thing. Once young people join Xiaohongshu, will they really get to make decisions about those ideas themselves?

This is also an important consideration for many technical professionals choosing a company.

At the salon, a doctoral student in computer science told QbitAI that, more than simply comparing salaries, he cared about what kind of dependable growth he could achieve over the next 2 years.

Xiaohongshu has also prepared a tiered development system, with corresponding training programs at every stage from new campus recruit to team leader.

But for applicants holding more than 3 offers, training programs are 1 consideration; more important is how much they can develop their ability to handle real business work.

At the event, Xia Qi, an algorithm engineer in Xiaohongshu’s multimedia technology department, gave her answer.

Shortly after Xia Qi graduated from Tsinghua University and joined Xiaohongshu, she faced a tough challenge: the 2026 World Cup.

For Xiaohongshu, livestreaming a major event like the World Cup is entirely new territory.

Xia Qi’s Media Lab regularly works on video clarity and smoothness so that users can still see clearly and scroll smoothly even with a weak network connection.

For the World Cup, the difficulty rose sharply.

Xia Qi was responsible for coordinating image quality enhancement algorithms and designing quality tiers. Soccer broadcasts have many wide shots, in which a player may occupy only a few pixels. Applying existing algorithms directly instead erased details such as players’ heads and limbs.

To find the problem, she and her team displayed the footage on a large TV and checked it frame by frame. They then retrained the model with a large amount of soccer match footage, finally preserving the details of the players.

Xia Qi said frankly that taking part in the technical effort for such a major event was a considerable challenge for her and an unforgettable experience.

But the World Cup happens only once every 4 years. Do other young engineers get opportunities like that?

Haolin, who works in the same department as Xia Qi, joined through the 2024 campus recruitment intake after graduating in computer application technology from Peking University.

Less than a year after joining, Haolin led Xiaohongshu’s technical exploration of 3D spatial photos, taking part from the initial technical validation all the way to the product launch.

Previously, 3D spatial photos mostly stayed in users’ phone albums. Getting photos with spatial information to circulate on a social platform meant solving quite a few problems.

For example, 3D photos contain far more data than ordinary images, and spatial photo formats and capabilities are not necessarily compatible across different phone models.

For an engineer who had graduated only recently, this was no easy project.

But Haolin thought that images, videos and Live Photos had been on traditional social platforms for many years. Could 3D spatial photos become a new form of content? He wanted to try.

The team chose to support his exploration.

To make that possible, the group rearranged its work assignments to free up more of Haolin’s time for research.

But early technical research can fail. Haolin said the team did not require every attempt he made to produce an immediate business result. Failure could also build technical knowledge.

In the end, the project succeeded. From proposing an idea and solving technical problems to bringing the product to launch, a young engineer had created a signature piece of work of his own.

Xiaohongshu really is willing to champion young people’s good ideas, even putting them on outdoor digital billboards.

Of course, getting support does not mean results come easily.

Technical challenges have to be overcome, resources coordinated across teams, and whether a product can actually launch has to be tested repeatedly.

It is hard and challenging, but that is exactly what some people are looking for.

At the salon, Xia Qi met a candidate who proposed an idea during their conversation: a user could simply say, “Make this picture clearer for me,” and the system would call on the underlying algorithm.

That left a deep impression on Xia Qi. Before even joining, the candidate was already thinking about how to turn technology into a product.

Xiaohongshu is still growing rapidly, with one major challenge after another, and newcomers can step straight into the action.

Huanling, who grew from a campus recruit into a team leader, has been at Xiaohongshu for 5 and a half years.

He started out working on candidate retrieval and fine ranking for recommendation algorithms. Now he is responsible for search ad relevance and prediction model technology and leads his own team. He describes the journey as “wild growth.”

Now, when he mentors newcomers, what he most hopes to see is young people who dare to propose ideas and even dare to tell him “no” directly.

As for why he has stayed for 5 and a half years, Huanling’s answer is practical: Xiaohongshu keeps moving forward, and there are always new things for him to learn and new things to do.

After 5 and a half years, Huanling still feels there is much he has yet to try.

For young people to have the chance to produce signature work, the company first needs to have plenty of technical problems worth exploring.

Xiaohongshu now has 400 million monthly active users.

Every day, some people search there for advice on everyday life, others share their experiences, and still others go from discovering a recommended product all the way to placing an order.

The real lives of 400 million people present the technical teams with new questions every day.

How can a recommendation algorithm tell what users really like, instead of making its recommendations more and more alike?

How can search understand people’s real intent and connect the path from product discovery to purchase?

When AI can also write, draw and edit, what kinds of content on the platform deserve greater encouragement?

How can a recommendation algorithm tell what users really like, instead of making its recommendations more and more alike?

How can search understand people’s real intent and connect the path from product discovery to purchase?

When AI can also write, draw and edit, what kinds of content on the platform deserve greater encouragement?

There are no standard answers to these questions.

AI-generated

What tests technical professionals even more is that sometimes an algorithm clearly improves the metrics, yet the associated changes cannot be rolled out to everyone.

Ji Hao told QbitAI that some algorithm optimizations at Xiaohongshu might even cause viewing time to fall.

What does that mean?

Suppose a recommendation strategy keeps users on the platform longer but pushes more content that does not align with the community’s values and may harm its content ecosystem or user experience. No matter how good the numbers look, it cannot simply be launched.

Sometimes the team will also deliberately reduce exposure for content that does not align with the community’s content direction, even if it costs viewing time in the short term.

For algorithm engineers accustomed to pursuing click-through rates and viewing time, this is somewhat counterintuitive.

But Xiaohongshu has its own principles.

In Ji Hao's view, the leading domestic internet companies are not as far apart in search, advertising and recommendation technology as one might imagine. Once a technology proves effective at one company, others can quickly follow suit.

What is truly hard to replicate is the business judgment and commitment to values behind the algorithms, which have built today's thriving community ecosystem through years of accumulation.

What Xiaohongshu wants to preserve is precisely those real people and their genuine sharing.

Ji Hao still remembers how, after Xiaohongshu launched its voice comment feature, some users with disabilities gained a new way to express themselves, giving more people a chance to hear their voices.

At that moment, he even got goosebumps.

After years of work, Ji Hao felt strongly for the first time that technology could create value like this.

“Technology itself is neither right nor wrong, but the people who build it need judgment and values.”

In a community of 400 million users, those words carry real weight.

Technical staff need to help decide what content the algorithms ultimately show users, which experiences are worth preserving, and what kinds of revenue can be forgone.

Of course, besides having plenty of questions to explore, they also need opportunities to work on them firsthand.

Xiaohongshu is still growing rapidly, with new businesses and directions constantly emerging, and many projects still at the stage of going from 0 to 1.

This means newcomers have a chance to get involved in core business areas earlier, and even take charge of parts of projects themselves.

Xia Qi, Haolin and Huanling, mentioned earlier, are ready examples.

A recent graduate who had interned at several major tech companies told QbitAI that he had originally expected to choose a technical research direction upon joining a company and then quickly deliver results measured by business metrics.

But through this conversation, he found that Xiaohongshu does not want young people merely to follow instructions and keep their heads down. Instead, it places greater value on whether they have ideas of their own and dare to explore new directions proactively.

If they have ideas, the company is willing to give them room and hopes young people can truly pursue their own interests and technical exploration.

“That really struck a chord with me.”

Looking back at the campus recruitment tech salons over those two days, students dared to ask questions, and department leaders dared to answer them.

That sense of real people being present is also very Xiaohongshu.

“Xiaohongshu offers a lot of freedom, and it also stands by its principles,” a recent graduate reflected after the conversation.

It has been 13 years since Xiaohongshu was founded. From its beginnings with 7 PDFs in hand to where it is today, it still has that entrepreneurial drive.

发现内容有误?提交纠错

本文目录