Newshot

Undercover on Xiaohongshu for 48 Hours: What Are Top Tech Talents Talking About in the AI Era?

AI summary

At a Xiaohongshu campus recruitment salon, a group of top technical talents took young people's ambitions seriously.

Translation(model translation, not the source's own text)

Undercover on Xiaohongshu for 48h: What Are Top Tech Talents Talking About in the AI Era?

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

Tian Yanlin, reporting from Guafo Temple, QbitAI | WeChat official account QbitAI

QbitAI | WeChat official account QbitAI

"If a campus recruit has their own ideas, can they apply for manpower and compute resources?" "AI is getting better and better at writing code—what will we mainly be doing once we join?" "Will Leaders really listen to ideas raised by newcomers?" "In the AI era, how should I develop myself?"

Well now, today's fresh graduates ask practical questions that strike right at the core!

On November 14 and 15, QbitAI went deep into Xiaohongshu's Beijing office for two afternoons, experiencing a tech campus recruitment event with a rather different style.

This time, the "spud factory" moved the campus recruitment info session directly into the office area and turned it into the REDpath campus recruitment tech salon.

With little dry presentation from the stage, the salon left most of its time to the AMA (Ask Me Anything) session below the stage.

Campus recruits sat down with the tech Leaders they might one day work with, chatting about everything from technical directions and team atmosphere to performance and personal development.

For this offline exchange, some students traveled specially from Shanghai, Nanjing, Hong Kong, and other places, and there were even fresh graduates studying in Singapore who flew to Beijing overnight.

By QbitAI's rough count, nearly 200 applicants showed up to talk over those two days.

The students' questions were more down-to-earth one after another, and the tech Leaders held nothing back. One applicant who came from Singapore told QbitAI:

Compared with online interviews, this kind of face-to-face exchange lets you feel the real working atmosphere of a future team in advance. Pretty great!

At the event, it wasn't only students asking questions.

After hearing that one fresh graduate had earned over 10,000 yuan in profit from opening a shop on Xiaohongshu, a tech Leader's eyes lit up and he hurriedly asked whether Xiaohongshu's merchant workbench "Qianfan" was easy to use and what features still needed improvement.

Well now—originally the students came to learn about a future team, but as the conversation went on, it somehow turned into a user interview.

With AI, what does a tech newcomer's first job look like?

Over the two days of exchange, I found that whether in the engineering session or the algorithm session, AI was a topic applicants couldn't avoid.

In the exchange with the Trading and Core Data Warehouse group, one student voiced his anxiety.

During an internship at a major tech company, he had worked on AI efficiency-related projects, but as he went along, he began to worry: was he personally phasing out his future self?

In the Community Recommendation Algorithm group, the discussion was even more specific.

Students were concerned about what role AI could play in recommendation algorithms. Could large models combine a user's past behavior to help them discover a new interest they hadn't even realized themselves?

Even in an earlier campus recruitment exchange, a student asked Ke Xiong, head of multimodal algorithms on Xiaohongshu's self-developed large model Dots team: there are already so many large models on the market, why does Xiaohongshu still need to build its own?

Ke Xiong's answer at the time was: building a large model that understands life and can also help users solve complex problems is not easy—it may involve the work of several roles such as a tour guide, a designer, and a purchasing agent.

This is also one of the reasons Xiaohongshu insists on developing its own model.

The platform has massive multimodal content—image-text notes, videos, and voice in comment sections—and users often describe their life problems in a long stretch of natural language.

These unique content and user needs place higher demands on the model's understanding and generation capabilities.

But to truly build such a large model, compute, data, infrastructure, and upper-layer business systems such as search and recommendation all have to keep up.

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

In the past, frontend, client, and backend each had their own division of labor. But now, with AI, engineers have the chance to step out of their familiar technical modules and take part in everything from understanding product requirements and designing solutions to development and delivery.

QbitAI also learned on site that last year, the Trading and Core Data Warehouse group had a position called "Data Development," which has already become "AI Application and Development" in this year's hiring.

Moreover, to get employees to use AI and use it well, Xiaohongshu issues Tokens every month—apply for more when you run out, and the quota is ample.

Even for written tests for non-technical positions, Xiaohongshu focuses on assessing AI application ability.

Word is that in this year's campus recruitment, some non-technical positions require completing an AI Demo written test before there's a chance to get an offer. (Good, good, good—even Vibe Coding is getting competitive!)

This strong embrace of AI also gives engineers more room to stretch their wings.

Engineers have begun proactively building Agents and developing platform tools to help the business solve problems in actual production.

With AI, everyone is becoming a stronger individual, delivering systematic solutions to the business one after another.

At present, several of Xiaohongshu's tech teams are trying to let Agents take part in daily work.

For example, the recommendation algorithm team has already launched Agent-based interactive recommendations, letting users adjust the recommendation system according to their own preferences.

When users see content they're interested in, they can write directly in the comments: recommend more content like this. The Agent can recognize this and make the recommendation system understand the user better.

The Trading and Core Data Warehouse team has also built a Data Agent. Each person develops different Skills and builds knowledge bases according to the business direction they're responsible for, then compares the accuracy of their respective Agents and gradually integrates the capabilities that work well.

Looking at it this way, the student who worried earlier about "personally phasing out his future self" can rest easy.

Of course, AI is getting more and more capable, but engineers can't be "hitting nails with a hammer."

Ji Hao, head of the recommendation algorithm video direction, doesn't advocate throwing large models at every problem.

"To solve the same recommendation problem, someone uses a large model, invests a huge amount of compute, and in the end the business benefit may be about the same as someone else using a simple strategy. So is that large model worth using?"

In his view, discovering problems is important, and judging what technology should be used to solve them is equally important.

This is also something he pays special attention to now when interviewing young algorithm engineers.

Ji Hao believes that students who have just come out of school have an advantage: they haven't yet been constrained by the existing working frameworks of industry, so they tend to come up with different ideas.

This kind of "spark" shouldn't be affected by AI.

It's just that young people having ideas is one thing. Once they join Xiaohongshu, can these ideas really be theirs to decide?

At Xiaohongshu, newcomers have a "piece of work" in their first year

This is also an important consideration for many tech talents when choosing a company.

At the salon, a PhD student in computer science told QbitAI that rather than simply comparing salaries, he cares more about what certain growth he can achieve over the next two years.

Xiaohongshu has also prepared a tiered training system, from campus recruits to team Leaders, with a corresponding training program at each stage.

But for applicants holding more than 3 offers, training programs are only one aspect; more important is how their business capabilities can be honed.

At the exchange event, Xia Qi, an algorithm engineer in Xiaohongshu's Multimedia Technology Department, gave her answer.

Xia Qi, a graduate of Tsinghua University, encountered a tough battle not long after joining Xiaohongshu—the 2026 World Cup.

For Xiaohongshu, live broadcasting of a major event like the World Cup is a brand-new battlefield.

Xia Qi's Media Lab usually grapples with video clarity and smoothness so that users can see clearly and scroll smoothly even on weak networks.

When it came to the World Cup, the difficulty escalated directly.

Xia Qi was responsible for coordinating image quality enhancement algorithms and designing quality tiers. Football broadcasts have a large number of long shots, and a player may occupy only a few pixels in the frame. Directly applying existing algorithms instead erased details such as players' heads and limbs.

To find the problem, she and her team put the footage on a large TV and checked it frame by frame, then retrained the model with a large amount of football match footage, and only then were the players' details preserved.

Xia Qi admitted frankly that being able to take part in the technical battle for such a major event was a huge challenge for her, and also an unforgettable experience.

However, the World Cup only happens once every four years. Do other young engineers have opportunities like this?

Haolin, in the same department as Xia Qi, is a 2024 campus recruit who graduated from Peking University's Computer Application Technology program.

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

Previously, 3D spatial photos mainly stayed in users' phone albums. To let this kind of photo with spatial information circulate on a social platform, quite a few problems needed to be solved.

For example, the data volume of 3D photos is far larger than that of ordinary images, and the formats and capabilities of spatial photos may not be compatible across different phone models.

For an engineer who has only recently graduated, this is no easy subject.

A Spider-Man 3D spatial photo shared by a user

But Haolin felt that images, videos, and Live Photos on traditional social platforms have existed for many years—could 3D spatial photos have a chance to become a new content format? He wanted to try.

To that end, the team rearranged the division of work to free up more time for Haolin to do research.

But the technical pre-research stage carries the possibility of failure. Haolin said the team doesn't require every attempt to immediately produce business results. Failure can also bring technical accumulation.

In the end, it worked out. From proposing the idea, to solving technical problems, to driving the product launch, a young engineer now has a signature piece of work of his own.

Article title: 48 Hours Undercover on Xiaohongshu: In the AI Era, What Are Top Technical Talents Talking About?

Body: When young people have good ideas, Xiaohongshu is genuinely willing to support them—even arranging outdoor big screens for them.

Of course, getting support doesn't mean results can be achieved easily.

Technical difficulties have to be tackled on your own, cross-team resources have to be coordinated on your own, and whether a product can truly launch also needs continuous validation.

This is hard and very challenging, but some people come precisely for this.

At the salon, Xia Qi met a candidate who proposed an idea during the conversation: users only need to say "help me make this image clearer" to invoke the underlying algorithm capability.

This left a deep impression on Xia Qi. Before even joining the company, the person had already started thinking about how to turn technology into a product.

Xiaohongshu is still growing rapidly, with one major battleground after another, and newcomers can also step directly onto the field.

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

From initially working on recall and fine-ranking for recommendation algorithms to now being responsible for search advertising relevance and prediction model technology and leading his own team, he describes this experience as "wild growth."

Now, when he leads newcomers, what he most hopes to see is young people daring to propose ideas—even daring to say "no" directly to him.

As for why he stayed for five and a half years, Huanling's answer is also very down-to-earth: Xiaohongshu keeps moving forward, and I always have new things to learn and new things to do.

Five and a half years later, Huanling still feels that there are many things he has not yet tried.

On Xiaohongshu, users set new problems every day

Young people have the opportunity to create representative work, but the premise is that the company still has enough technical problems worth exploring.

Xiaohongshu now has 400 million monthly active users.

Every day, some people search for life experience here, some share their own experiences, and some go all the way from discovering products to placing orders.

The real lives of 400 million people set new problems for the technical team every day.

How can recommendation algorithms know what users truly like, rather than making recommendations more and more similar? How can search understand people's true intentions and connect discovery with transactions? When AI can also write, draw, and edit, what kind of content on the platform is more worth encouraging?

How can recommendation algorithms know what users truly like, rather than making recommendations more and more similar?

How can search understand people's true intentions and connect discovery with transactions?

When AI can also write, draw, and edit, what kind of content on the platform is more worth encouraging?

These questions have no standard answers.

What tests technical people even more is that sometimes, even when an algorithm has clearly improved the data, the related actions cannot be rolled out fully.

Ji Hao told Quantum位 that some algorithm optimization actions at Xiaohongshu may even reduce watch time.

If a recommendation strategy makes users stay longer but pushes more content that does not align with the community's principles and may harm the community content ecosystem or user experience, then even if the data looks beautiful, it cannot be launched directly.

Sometimes, the team also actively reduces exposure to some content that does not align with the community's content direction, even if this will cost watch time in the short term.

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

In Ji Hao's view, the search, advertising, and recommendation technology levels of China's leading internet companies are not as far apart as people imagine. Once a company validates an effective technology, other companies can quickly catch up.

What is truly difficult to replicate is the business judgment and adherence to values behind the algorithms, and the long-term accumulation that has built today's thriving community ecosystem.

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

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

At that moment, he even got goosebumps.

After many years of work, it was the first time Ji Hao strongly felt that technology could create this kind of value.

"Technology itself has no right or wrong, but the people who do technology need judgment and values."

In a community of 400 million users, this statement carries real weight.

What content algorithms ultimately present to users, which experiences are worth insisting on, and what kinds of gains can be given up all require technical people to participate in judgment.

Of course, having enough problems to explore also requires the opportunity to get hands-on.

Xiaohongshu is still growing rapidly, with new businesses and new directions constantly emerging, and many projects are still in the 0-to-1 stage.

This means that after newcomers join, they have the opportunity to get involved in core business earlier, and even take charge of part of a project themselves.

The aforementioned Xia Qi, Hao Lin, and Huan Ling are ready examples.

A graduating student who had interned at several major tech companies told Quantum位 that he had originally expected that after joining a company and choosing a technical research direction, he would need to deliver results through business metrics as quickly as possible.

But through this exchange, he found that Xiaohongshu does not want young people to only follow arrangements and bury themselves in work; instead, it values whether they have their own ideas and whether they dare to actively explore new directions.

If you have ideas, the company is willing to give you space, and it also hopes that young people can truly carry their interests and technical exploration forward.

Looking back at those two days of campus recruitment technical salons, students dared to ask, and department Leaders dared to answer.

This sense of real people is also quite Xiaohongshu.

"Xiaohongshu is very free, and it also has things it insists on." A graduating student sighed after the exchange ended.

Thirteen years after its founding, from starting out with 7 PDFs to where it is today, that entrepreneurial drive at Xiaohongshu is still there.

  • Large models can control robots—how much further can Human Data go? 2026-11-13
  • OpenAI's new funding reaches $11 billion, less than half of the original plan 2026-11-07
  • To remove the AI flavor, humans have started actively writing bad articles... 2026-11-07
  • Beating DeepSeek to the punch! Silicon Valley Chinese researchers have built the next generation of Agent Harness!? 2026-11-07

Related Reading

Xiaohongshu discloses its AI technology system for the first time, going all out for its largest campus recruitment

Xiaohongshu experts also answer questions live online

Multimodal experts all open Xiaohongshu

At 19:00 on April 20, a livestream produced by the Xiaohongshu technical team, where many technical experts explain multimodal technology applications in detail. Come watch~

Xiaohongshu's large model wins gold with a perfect score at IMO, and its solution to the third problem makes champion contestants call it elegant

A Chinese large model receives IMO official gold medal-level certification for the first time

Conversation with Han Xu: After the dual listing, elite campus recruitment starts at 3 million

"Autonomous driving is still the sexiest direction"

I stumbled into a technical dream team event where everyone had 10 top offers, and I finally understood the choice logic of top AI talent

Actually, companies are waiting for them to reveal their decisions

CVPR2023 | Xiaohongshu proposes the OvarNet model: a new SOTA for open-set prediction, and "recognizing everything" gets a new way to play

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 量子位
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗