A Look at 2024's AI Innovations in China: From Models to Hardware, From Applications to the Rise of Super-Individuals
Let's preserve the AI innovations that brought us excitement and encouragement in 2024.
As the year draws to a close, we're putting together a series of year-end retrospectives to revisit the major events of 2024 with all of you. This year, we've frequently discussed innovation connected to Silicon Valley, and we also ran our "AI in China" series. As we thought about how to wrap things up, we kept asking ourselves: what angle can we offer that's truly distinctive?
Recently, after seeing some data from domestic players and recalling what a guest said during our last year-end review — "we're not watching from the wings, we're center stage" — the idea for this episode took shape: a survey of AI innovation in China during 2024.
For this episode, we've invited Mars Ren, partner at Sky9 Capital and host of the podcast "AI Alchemy," who previously appeared on Crossing. Together, we'll trace the arc of 2024's Chinese AI landscape — from models to hardware, from applications to the young people at hackathons and the rise of "super individuals." Like our last year-end review, we hope this retrospective offers some momentum to carry you forward.
A quick note: in preparing this episode, we've undoubtedly missed things. If there's a project we overlooked, please tell us in the comments. Let's preserve together the AI innovations that excited and encouraged us in 2024.
Listen on WeChat:
Listen on Xiaoyuzhou:
Foundation Models: China's Cost Advantage
🚥 Ronghui
Hello everyone, and welcome to today's episode of Crossing. We're continuing our year-end retrospective series.
Looking back at 2024's major AI events, many people first think of keywords tied to Silicon Valley. While Silicon Valley remains the center of tech innovation, plenty of compelling developments and personalities also emerged in China's AI scene this past year. Inspired by something a guest said on our last episode: "We're not watching from the edge of the stage — we are the center of the stage."
Today we'll review 2024's Chinese AI innovations from multiple angles: models, applications, talent, and more.
We're delighted to welcome Mars Ren, partner at Sky9 Capital and host of the "AI Alchemy" podcast. We previously recorded an episode with Mars, whose observations on AI are always sharp and original. Welcome, Mars.
👦🏻 Mars
Hey everyone.
🚥 Ronghui
Today we'll survey 2024's Chinese AI innovations across models, applications, talent, and other dimensions.
Let's start with the closely watched foundation model space — specifically, developments that may have been overlooked but could prove consequential.
🚥 Koji
Sure, I'll jump in.
I recently came across an interesting analogy on Xiaohongshu. A user accustomed to Claude Sonnet — where 1 million tokens cost $15 — suddenly discovered that domestic DeepSeek charges just 2 RMB. He described the feeling as "like walking from Louis Vuitton straight into MINISO."
This reminded me of something Kai-Fu Lee has been emphasizing lately: Chinese foundation model companies' biggest advantage over Silicon Valley giants isn't being at the bleeding edge — we're roughly five months behind — but rather our formidable ability to build low-cost models and inference engines.
We may not be able to develop or spend our way to the absolute frontier of AGI right now, but we can keep pace with the first tier globally while offering far more economical inference. That may actually prove more valuable for nurturing a thriving application ecosystem. And China happens to excel at product-market fit. If inference costs drop and we combine that with our rapid PMF capabilities, we could very well leapfrog from behind.
Looking at historical precedent, we didn't quite pull this off from the mobile internet era through AI 1.0, but similar弯道超车 had happened twice before. So the current consensus is: drive down model prices, nail the applications, and trust the process.
👦🏻 Mars
On price cuts, I've been advising people to "abuse" tokens and compute.
The core of this AI revolution is injecting intelligence into products, and intelligence is measured in tokens. If tokens get ten times cheaper and three times better every year, then using more tokens in your product means that even if you change nothing, it'll perform better and cost less next year.
The more tokens you use, the more compounding benefits you capture. For application developers, this means getting better every day without trying.
But product design faces a crucial question: are you building something that rises with the tide, or something that gets buried like the Leifeng Pagoda? "Leifeng Pagoda" products get swallowed as foundation models improve — that's the pattern to avoid. By contrast, "rising tide" products integrate domain expertise, specialized workflows, and business systems in ways that grow stronger with each model upgrade. A simple PDF chat tool might get replaced by a model update; a deeply industry-integrated application keeps benefiting.
So I agree — China is genuinely good at applications. But from day one of building one, you need clarity: are you competing with large language models, or complementing them?
AutoGLM and Computer Use
🚥 Koji
Right — your gut reaction to each ChatGPT upgrade is revealing. Do you feel anxious or excited? That emotion is a direct diagnostic.
Also, there's one area where we genuinely "lead by a wide margin": globally, smartphones divide into iPhone and Android camps, and among Android manufacturers, Chinese brands are unquestionably at the forefront. This lets us innovate faster and more imaginatively than Apple Intelligence.
Take Zhipu's AutoGLM, released two weeks ago. After the initial excitement, opinions cooled on its practical utility, but it showed a first glimpse of what auto-agents could become — using natural language to book flights or train tickets on your phone, or finding recent unread WeChat messages and replying to them. These were previously imagined-only capabilities, and Zhipu made them real through AutoGLM, albeit in limited testing.
👦🏻 Mars
Right, AutoGLM and Computer Use really stuck with me. During that period, my phone kept buzzing with test messages from colleagues, all trying to get computers and phones to complete various tasks. Seeing everyone so actively experimenting, I felt this was genuinely a significant technical breakthrough.
I can't recall Computer Use's exact benchmark — whether it was 14% or 7% — but it's still completing only a small fraction of 100 common human tasks. Yet I think this may be the most important breakthrough of the year, because that number will only grow. If it handles 14 tasks this month, maybe 15 next month. The trajectory is both exciting and slightly unnerving.
It's like watching a clumsy little kid who's in third grade now. Next year fourth grade, then fifth — eventually they'll be in college. This isn't fundamentally a hard science problem; it's mostly engineering, and collective effort will solve it.
So I believe that whether it's Computer Use or AutoGLM, technologies that can manipulate our productivity tools represent the most significant development of 2024.
Meanwhile, foundational models like Claude Artifacts and o1 are also fascinating. China remains in catch-up mode at the model layer, but from an applications perspective, phone makers like Xiaomi are exploring more boldly than Apple.
🚥 Koji
Right, much bolder.
👦🏻 Mars
And probably more satisfying to use. I tried Apple Intelligence a few days ago and haven't found anything useful yet.
🚥 Koji
Xiaomi's recent launch distilled their vision into nine characters: "remember for you, find for you, fill for you." For example, you photograph a family member's ID, and later when booking flights, simply ask Xiao Ai to fill in your father's information — the phone handles it automatically. All sensitive data is encrypted and stored locally, protecting privacy. Chinese manufacturers may not be the most advanced here, but they move faster. Apple Intelligence, constrained by various considerations, has underwhelmed so far.
That said, Apple Intelligence did surprise me in one way. It automatically identifies people in your photo library — noticing I'd taken a dozen photos of Mars, for instance — and cartoonifies them.
This creates new social dynamics — I can share these cartoonified portraits with the person. All of this runs on local on-device models. After discovering this, I spent ages with friends checking how we appeared in each other's photos and what our cartoon versions looked like.
👦🏻 Mars
We recently came across an interesting AI video tool company. Initially, their product let users drop their own face into girl-group dance videos.
But that trick's become commonplace. Now they've switched to a pixel-art cartoon style — turning your avatar pixelated before it dances — and audience response improved dramatically. Because plain face-swapping has lost its novelty.
AI Breakthroughs on PC
🚥 Ronghui
We've covered foundation models and domestic manufacturers' innovation characteristics. Now let's talk about interesting changes in some familiar domains.
For example, Tencent recently launched a new product called ima, an AI-powered workspace. It's a PC-native application that uses search as its entry point, helping users write notes and edit content. I counted — Tencent isn't the first to build this kind of product. ByteDance's Doubao and Alibaba's Quark have already released similar PC-based AI search products.
While researching, I noticed two interesting keywords. The first is "PC client." When people talk about AI applications, they usually think of mobile apps or browser extensions first. Although products like ChatGPT support both mobile and web (and recently launched a desktop version too), overall, purely PC-native AI products still seem relatively rare.
🚥 Koji
I think they might just be rare in our field of view — in reality, the user base is probably quite large.
While preparing for this episode, I came across some surprising data. First, on the PC side — I found a report from Pacific Technology showing that Quark's PC download volume has already reached number one, surpassing both Doubao and Moonshot AI. Wasn't Quark primarily mobile-focused? How did its PC version do so well?
Digging deeper, I discovered Quark's mobile performance is equally impressive. QbitAI reported that as of October this year, Quark's cumulative mobile downloads reached 270 million, compared to Doubao's 140 million. Honestly, this data surprised me too.
After carefully studying their current feature set, I think there are two possible reasons:
- It adopts an "all-in-one" strategy, integrating various functions together;
- It provides substantial practical value for students, whether for searching problem solutions or making PPTs, offering one-stop solutions for various academic and daily needs.
Although Quark started as a search engine, it appears to have transformed into a comprehensive client integrating various AI functions.
🚥 Ronghui
This evolution makes perfect sense — AI is first and foremost a productivity revolution, and computers are our primary scene for using productivity tools. Now we're seeing both hardware and software manufacturers pushing in this direction: on the hardware side, there's AIPC and Apple's new generation of Mac computers; on the software side, companies like Quark are developing desktop applications.
Building a dedicated PC client is both challenging and ambitious. Because it means breaking down barriers between PC applications, enabling seamless connectivity across functions. This is no simple task.
After downloading and trying Quark's PC version, I found its design genuinely clever. You don't need to deliberately open the PC application at all — once installed, you can directly summon the AI assistant, feeling very much like Apple's Spotlight search. You can input text, use voice input, take screenshots, and draw lines — the entire interaction is remarkably smooth. When users want to use AI, they barely need any extra operations. This design makes the AI assistant incredibly accessible, feeling completely natural.
Then, besides the PC-native characteristic we just discussed, there's another keyword: "integration." This is truly a distinctive feature of domestic products. Look at Doubao and Quark — they've both integrated AI writing, AI PPT generation, and Quark has even added mind mapping, AI drawing, and other functions.
It feels like domestic products are all pursuing one-stop service, hoping users can handle most productivity needs on a single platform without switching between different applications. This integrated approach fits user habits quite well.
🚥 Koji
Right, I think Chinese products, starting from QQ, then WeChat, Alipay, have always loved making "mega-bundles" and "family buckets," packaging various functions together. This "N-in-one" model is indeed convenient, but it also reflects some interesting phenomena.
Previously, when I needed to recommend AI products to friends less familiar with the technology, I could only suggest mobile apps — like installing Moonshot AI or Doubao. On computers, there simply weren't good options. But the recent emergence of these PC clients has changed that situation. I recommended Quark to several friends, and their feedback was quite positive — they said it really meets their various imaginings of AI, whether for writing, making PPTs, or other work needs, basically handling it all.
This "N-in-one" product strategy has always been a particularly interesting topic to me. Why do overseas products take the specialized route, doing one function and doing it to extreme perfection; while in China, products are particularly enthusiastic about making "mega-bundles" and "family buckets"? Every product wants to become the user's one-stop solution.
Mars, what are your thoughts and insights after observing this phenomenon?
👦🏻 Mars
Overall, I think this is still because there are quite a lot of users who aren't very familiar with using tools, so the "family bucket"反而 becomes the better solution.
Speaking of how unfamiliar they are — didn't two memes go viral recently? One was "how do I open a URL," and the other was "are you trying to scam me into paying for decompression." This reflects how many users are still strangers to even the most basic software usage, some to quite extreme degrees.
Many people are generally unfamiliar with productivity tools. For example, needing to use ten different tools — just PPT alone is headache-inducing enough. Not many software installed on their computers, not many services known. In this situation, if there's an "N-in-one" solution, it can satisfy various user needs.
Let me give a few examples. There's a very popular AI PPT tool in China whose main function is actually helping users generate templates, rather than content generation.
Because perhaps the larger user group is stuck at the stage of "can't make a good-looking PPT," not yet at the point of needing AI to help refine content.
We might say, "wasn't this solved five years ago?" Indeed, China has many excellent software, like iSlide and such, but most users simply don't know about them.
Another very interesting example — I previously chatted with a founder of an African browser. Their browser has very high market share in Africa, with every function imaginable — PDF, Excel, any PC software you can think of, they've built it in, even Chrome's extension functionality is integrated.
I asked him: "Can you still call this a browser? What's your positioning? How do you compare to Chrome?"
He said: "We're fundamentally different from Chrome. In Africa, browser equals the tool for 'browsing everything.' Don't expect users to have other software installed on their phones — if you don't do it all for them, they simply can't use these functions."
So I think China may currently be in a transitional period between the earliest stage and developed countries. There are still massive numbers of users unfamiliar with various software, so giving them a complete solution all at once is反而 more effective. On the positive side, this might be because we've developed too fast. In developed countries like the United States, they progressed step by step: this software was popular for a couple years, that software for another couple years, users gradually learned how to connect APIs between different software, how to use them in combination.
But in China, many people only started encountering Office or WPS during the mobile internet era. Suddenly having twenty apps on their phones, it felt too troublesome to learn. It's like giving students all the textbooks for middle school and high school at once — they might just want one "Complete Middle & High School Three-Year Exercise Collection" and throw everything else away. But if there's sufficient time, if the pace of development is slower, it's like learning grade by grade, naturally enabling more detailed, subject-specific learning.
From a business perspective, China's traffic competition is indeed fierce, far more cutthroat than overseas. The upside is that this compels product teams to really deliver on service, to satisfy all user needs. From the product itself's perspective, the goal is "once users arrive, don't let them leave." This saves users effort while maintaining platform stickiness — a win-win situation.
🚥 Koji
Actually, Mars's Phoenix Browser story is quite amazing — I had no idea such a thing existed. What I'd add is that although Chinese people do "N-in-one," whether it's Alipay, WeChat, or Quark we just discussed, none of them are simply mashing N functions together.
The amazing thing is that each of the N functions in their N-in-one approach is actually done quite meticulously and well.
China's Force in Video Generation
🚥 Ronghui
Now that we've covered foundation models and innovations in familiar scenarios, let's look at innovations in video generation in 2024.
Recently, many people were sharing Venom special effects videos. The company behind this effect is a Chinese company.
🚥 Koji
PixVerse from AISphere. Early this year, during the Lunar New Year period, following Sora's thunderclap demo release, Chinese companies quickly caught up. Whether it's Kuaishou's Keling AI, MiniMax's Hailuo AI, or Shengshu Technology's Vidu, the final results have all received extremely high praise.
Speaking of special effects, the recent viral effects on Xiaohongshu — like inflating yourself to fly away, or disintegrating into ash — all came from Pika. Pika's founder is Chinese-American Demi Guo. Another star company in the video space, Heygen, is also founded by Chinese-American Joshua.
Notably, many leading companies in this space have Chinese founders and CEOs. For example, Viggle AI was built by Chu Hang, and Akool by Lü Jiajun. These companies are not just domestically but globally in the first tier, their products receiving pixel-level attention. Although these companies target global markets, their core engines and founders are Chinese.
Speaking of Chinese entrepreneurs, perhaps most noteworthy is Alexandr Wang's Scale. While not a video company but rather doing data annotation, its achievements are remarkable — born in 1997, its latest funding round reached $1 billion, at a $13.8 billion valuation. That's really impressive.
👦🏻 Mars
I was amazed when I saw Keling AI last year, because Sora wasn't available yet. Using Keling AI, I found the effects surprisingly good, so for a while I kept using Midjourney to generate images, then processing them through Keling AI.
What's worth mentioning is that now beyond meme entertainment, we're seeing many domestic tools becoming increasingly controllable. Compared to before when models were quite unruly and it was entirely blind luck, prompt adherence has been improving noticeably — the progress is clear.
🚥 Koji
There's an activity called Video Battle, where a theme is set each week for people to generate videos and compete, with the top three selected.
At first, we invited Li Dan to set the theme. Everyone had to make a 5-second video based on "egg," and the imagination on display was extraordinary. For instance, someone transformed the egg into the boulder Sisyphus pushes, or shaped it into a hole in barbed wire on a Middle Eastern battlefield, through which you see children playing — telling a story of the egg's destruction and rebirth. This event has now run for dozens of sessions. As a judge, I can clearly see the progress from the first session to now.
One recent week was particularly shocking: the event received 1,800 submissions. The theme was a person standing before a wardrobe, opening the door to see a special world. Someone saw Doraemon fly out for a hug. Someone saw their younger self abandoned in the rain without an umbrella, and walked over to embrace them. Another saw cities floating within cities. The sheer creativity of these 1,800-plus entries was breathtaking, reflecting how rapidly video models have developed to make such imagination reality.
The contestants weren't all experts. Take the Li Dan session: we surveyed several dozen selected participants and found they had been using AI to generate video for as little as three days, and the oldest was around 50.
This made me think about creative democratization — tools are becoming easier to use. Just as writing was once only for the literate, it later became a skill everyone possesses.
The same is happening with video creation. It's no longer only directors and screenwriters who can use these tools.
Among these 1,800 works, many used domestic tools. Keling AI excels at facial expressions, HeyGen solved hand and mouth issues, and Hailuo AI stands out for large movement amplitude.
👦🏻 Mars
The Keling AI and Hailuo AI you mentioned do work very well. But for many people, creativity itself remains a barrier. So more and more companies are moving toward "meme-ification." Pika, for example, directly packages effects into specific functions like "squeeze a person until they pop."
I saw one company turn "hug your past self" into a dedicated tool: upload childhood and current photos, and your present self embraces your younger self. Others made functions letting you hug celebrity idols. These are all about concretizing application scenarios — not providing universal tools, but specifically doing one thing.
Similarly, Viggle lets users upload a photo and video to dance along. One product we incubated turns uploaded photos into pixel-style dancing figures.
🚥 Koji
I think there's so much to do with video, especially on the application side.
Some time ago I met Huang Shuo, founder of Remini. The company, which went viral for its clay filter and was later acquired by Bending Spoons, started from a single use case — helping users restore old photos. They designed special immersive interactions that let users revisit memories repeatedly.
Although Remini has been acquired, their domestic mini-program "You Wo Dang Nian" was not sold, and still generates 500,000 RMB monthly revenue. Why do users keep subscribing after restoring their photos? Because they provide an immersive viewing experience that puts users into a special reminiscence mode, fully engaging their emotions. By contrast, if you simply save photos to an album without a dedicated viewing environment, it's hard to enter that emotional state.
This shows there's enormous potential in visual content. It's not just about using AI to generate video, but further productizing the viewing and reminiscing process, designing special interactive experiences.
👦🏻 Mars
What you described leans toward the humanistic, emotional direction. Another technical route explores 3D Gaussian splatting, 4D Gaussian splatting, and similar directions. These researchers emphasize precise control and accurate modeling, aiming to build a world that obeys physical laws — water bottles maintain relatively fixed sizes, object collisions must be accurate. Some propose starting from modeling, but diffusion fundamentalists consider this regressing to old ways. Yet judged by actual results, many stunning demos have already emerged: it's like speaking reality into being, where whatever you say the world responds to precisely — want the box to move left and it moves left, finish filming a short drama and remove the props, swap faces at will.
Both schools share a common view: what they're doing isn't simple video generation, but world creation. Take the sentence "I'm holding a red apple in my hand" — linguistically fine, but a machine might interpret it as impaling the apple on a finger. Only through visual presentation can the model truly understand what "holding in hand" means.
So many video model researchers believe they're building genuine world models. They argue language models remain in conceptual space, unable to derive real world models. Whether the diffusion or 3D modeling route, both emphasize they're creating world models, not merely generating video.
AI Hardware Innovation: From Plaud Note to Smart Glasses
🚥 Ronghui
Alright, let's move on to AI plus hardware. This is one area where Chinese entrepreneurs excel — not only can they reduce product costs, but they can fully leverage China's supply chain advantages.
So what distinctive Chinese hardware products have you seen this year?
🚥 Koji
The first that comes to mind is the Plaud Note that Gao Xu built — a card-sized AI recorder that sticks to the back of your phone. It automatically records during calls or video conferences, and provides AI transcription and summarization. Sales have already reached 300,000 units, with strong user reviews. One user bought five units — for home, office, bag, and to give to friends — showing the product truly becomes indispensable.
This is a textbook case: Chinese entrepreneurs skillfully using supply chain advantages to control costs, while seizing first-mover advantage in branding and marketing, with well-executed product design and tone.
🚥 Ronghui
Mars, what have you seen on your end?
👦🏻 Mars
I've seen many, but the biggest example was already taken by Koji (laughs).
Before Rewind launched Limitless, there were already quite a few necklace-style devices overseas. But searching on JD.com at the time, you could find similar products with richer functionality for just 98 yuan — simply add a communication module to connect to your phone, and use cloud-based large models for processing. It makes you sigh: shouldn't those overseas teams visit Huaqiangbei? Then they wouldn't have made such bulky devices. This truly demonstrates China's supply chain advantage.
But the current challenge is how to break through. There are already countless smart glasses and earbuds on the market. Yet a product like Plaud is very focused — not flashy enough, but before reaching its current sales scale, it was genuinely hard to predict such success.
🚥 Koji
This is indeed a赛道 that requires courage. Entering means inevitably facing fierce competition, with unclear product differentiation. The outcome could be a hundred schools contending for small slices of market, or one product ultimately breaking through.
👦🏻 Mars
China is genuinely strong in hardware. But not every product needs to pursue efficiency. What left a deep impression on me was the Jibo robot from a few years back — a cute big round face with a bouncy big eye. According to the founder, they invested most heavily in the neck and hip joints for multi-axis rotation. These designs seem to have no practical value, but let the robot tilt its head and wiggle — incredibly adorable.
Watching that video, the robot was purely selling cuteness. My wife immediately said, "How much is this, let's buy one." I said it wasn't on the market yet, I'd figure something out. Later seeing the actual product didn't match the video, I gave up. But if it could really achieve that level of cuteness, it wouldn't need other functions — cuteness alone would be enough.
🚥 Koji
Selling cuteness is indeed good business. I often can't resist buying blind boxes when I pass Pop Mart. Speaking of robots, Japan's LOVOT has a physical store at Shanghai's Grand Gateway Plaza, priced at $10,000, with most costs going into eye design. I've visited several times. Knowing it's a robot, I was completely drawn in by its blinking big eyes, subconsciously wanting to say "come here for a hug." Only afterward did it feel strange — why would I react to it like a baby or puppy? Its eye design creates that fleeting illusion, as if facing real life.
Back to Chinese innovation — I recently saw the Beijing team behind Ropet, preparing to launch on Kickstarter. Compared to LOVOT's $10,000, Ropet achieves similar functionality at just $300-400, only 3-4% of LOVOT's price. This is a promising direction.
Recently some heavyweight entrepreneurs have entered this space too. For instance, one person (identity temporarily undisclosed) received investment from three top-tier funds and is also building AI-driven companion robots. This shows the potential of Chinese-style innovation: combining emotional value, Huaqiangbei manufacturing capability, and AI technology could very well open new territory.
🚥 Ronghui
So within AI hardware, there are already two directions with increasingly more products. Should we talk about glasses and earbuds?
👦🏻 Mars
Most glasses products are largely similar, all oriented toward recording the external world. But recently I saw a very unique concept — unverified yet, but quite novel.
This concept inverts the traditional approach — not seeing the world through glasses, but observing the wearer through glasses.
🚥 Ronghui
I find this fascinating, because people can rarely see themselves. Beyond mirrors, including myself — if friends hadn't told me, I wouldn't know that my expression is very serious when speaking normally.
Wearing glasses isn't about seeing the world better, but about seeing yourself from the world's perspective. Then combined with your calendar — "I see you're pretty stressed in meetings with your boss" — and combined with your heart rate, your pupils, your expressions, it can actually do much deeper analysis.
For example, we often use iPhone apps, something called Mood Tracker or similar mood software. No need to record manually anymore — it can see, it can record. So I think this is a truly magical concept.
I've been thinking about what tools could help collect more data. Previous ideas were limited to carrying hardware to record external sounds and video, rarely considering recording one's own state in reverse — including heart rate, pupils, breathing, and facial expressions.
Combining and analyzing this data, we can better understand our relationship with the world. For instance, if you want to increase happiness, the system might discover "you're in a good mood at xx place, but get irritable at Starbucks," or "when chatting with someone, you feel happy when they wear white clothes and low when they wear black." Though these correlations seem mystical, they genuinely help us make better choices. This user-centric perspective truly refreshed my thinking.
👦🏻 Mars
Right, this is precisely using glasses as a mirror.
🚥 Koji
The current AI glasses market faces intense competition. According to NuMiao's Li Nan's analysis, Huaqiangbei production costs run 300-700 yuan, with final retail prices likely anchored at $99, and inventory pressure potentially pushing them down to $49. In this environment, making homogeneous products offers no advantage.
Li Nan chose a different path: going straight for the premium segment, pricing several times higher, targeting affluent consumers. These users won't settle for ordinary mainland products — they demand superior design and distinctive taste. Of course, this requires delivering on both hardware and software quality.
I'm quite looking forward to Li Nan's new product launch. Based on his track record with hardware, I believe these AR glasses or earbuds will maintain his signature premium feel and delightful surprises — expensive, but worth every yuan.
🚥 Ronghui
Koji, did you also try the Ola Friend?
👦🏻 Mars
How was it?
🚥 Koji
I wore it for a full day and night — literally 24 hours except when sleeping, just to properly test it. I thought it was great because the design made me constantly forget I was wearing it. It takes zero effort, so you can comfortably keep it on as a long-term companion.
The wake-up experience is remarkably smooth. At a café, even if you whisper "Doubao Doubao" very quietly, it wakes up successfully every single time. Then there's the natural language tone of its responses, plus efficiency and speed improvements. So it genuinely feels like having a friend, an assistant by your ear that's always on call and responds to you almost instantly.
Beyond the obvious use cases everyone thinks of — like seeing a tree and asking what species it is, or encountering an unfamiliar historic building and asking "what's this thing in this tourist spot, tell me the background" — there was one moment at a café when it got really crowded. Two women sat down next to me chatting extremely loudly, and I couldn't focus on work anymore. On a whim, I said "Doubao, play me some white noise" — no unnecessary chitchat, rain immediately started in my ears, and the world suddenly went quiet. I found that pretty delightful.
And I see they're hiring various influencers on Xiaohongshu to promote it, all using the "friend by your ear" angle. After using it myself, I genuinely get that feeling.
2024 Viral Hits: From Dating Keyboards to AI Mice
🚥 Ronghui
Moving on from hardware, I really want to discuss a phenomenon in the AI space: products that create FOMO. When you see something blowing up across social feeds, you think "what's this?" and rush to find out, afraid of missing the next big thing. This happens both in the United States and domestically.
So in your view, what products suddenly went viral in 2024? Let's focus mainly on domestic ones.
👦🏻 Mars
Keling AI, which we mentioned earlier, was all over my feeds for a while.
🚥 Koji
Speaking of recent viral products, while there haven't been many that truly reached saturation-level buzz, some innovative ones did leave an impression. For instance, Mars told me about "Lianxiaoyu" dating keyboard, a chat input method.
🚥 Ronghui
Dating keyboard (laughs).
🚥 Koji
Here's how it works: when you're chatting with someone you're romantically interested in, if you might reply with something low-EQ, it suggests alternatives to help you respond with more emotional intelligence — making the other person feel cared for, delivering emotional value, keeping the conversation flowing naturally.
This is classic Chinese-style innovation — the idea itself isn't new, but Chinese teams executed it with formidable operational strength into a practical input method product. They've promoted heavily on Xiaohongshu and Douyin, and have already gained substantial user adoption.
👦🏻 Mars
Speaking of hot products, AI PowerPoint tools definitely had their moment. For office workers, writing documents and making PPTs are two major pain points. While writing feels somewhat manageable, PPT creation remains perpetually headache-inducing. Numerous AI PPT tools emerged — dedicated products like "AI PPT," plus WPS integrating AI PPT functionality.
🚥 Ronghui
Why does the AI PPT赛道 feel so crowded? As Mars said, this is genuinely high-frequency demand for working professionals. And PPT production plays perfectly to two major strengths of large models: summarization capability and structured information output — it's a very typical application scenario.
In this space, PC-side products like Quark have also integrated AI PPT features. For many office workers, an aggregator product that consolidates commonly used work functions is genuinely more convenient. This lets users complete all work tasks on a single platform, improving efficiency.
🚥 Koji
For AI PPT tools, users do need a multi-function integrated scenario. Right now, while we've heard about many AI tools that can make PPTs or draft documents, actually finding these tools is always a hassle. If you go search on Baidu, everyone knows how many of the first-page results are ads — it's especially troublesome to find what you need. So an all-in-one client like Quark is genuinely valuable, particularly recommended for friends wanting to use more AI at work. At least you won't be scrambling around searching for tools when you need them.
Speaking of AI PPT companies, there's another fascinating marketing case — Zhao Chong and team made an AI mouse, have you heard of it? This was truly an inspired promotional move. They discovered that in many street-level 3C digital accessory shops, the best-selling product was mice. So they launched an affordably priced AI mouse, distributed through these shops. Buy a mouse, get AI PPT service included — brilliantly executed marketing.
👦🏻 Mars
A guest on my podcast mentioned something interesting. He asked if I knew about AI mice, I said yes — iFlytek and AI PPT companies are making them. But he meant something different: an ordinary mouse with one added button — click it and a hao123-like webpage pops up, aggregating various AI products including Moonshot AI.
Their pitch was clever: "iFlytek mice only work with iFlytek's AI, ours gives you one-click access to 100 large models." This mouse is priced at 120 yuan, costs only 20 yuan to make, and sells extremely well.
Hearing this, I was shocked — I realized I need to maintain more humility and keep learning about the market. Never imagined this kind of market existed.
🚥 Koji
That feels like Chinese-style innovation in AI hardware too, worth promoting globally.
👦🏻 Mars
It genuinely solves real needs for many people. Just as people early on needed navigation sites like hao123, or deal aggregators like Tuan800 during the group-buying wars, now people need an entry point aggregating various AI tools — yet no such platform exists in the market.
While you could build a web-based navigation aggregator, it requires heavy marketing spend and is difficult to promote. But through the hardware route, people feel 120 yuan for a mouse isn't expensive, making it more acceptable and easier to spread. And the hardware-as-traffic strategy, with costs around 20 yuan, leaves attractive margins.
This marketing approach really opens your eyes and forces us to maintain more market humility. This kind of innovative thinking is truly worth learning from.
AI Education: Chinese Founders' Domain
🚥 Ronghui
We previously noticed that in one AI entrepreneurship sector, roughly half the players are Chinese founders.
Rather than tell you directly, let's pose a question to listeners: do you know which AI field has attracted so many outstanding Chinese entrepreneurs?
Koji and Mars, why don't you share what particularly impressive Chinese startups you've seen in this direction?
🚥 Koji
Because many Chinese founders come from "small-town exam grinder" backgrounds, their first instinct when starting companies is helping others grind problems more effectively. This has made AI education a sector where Chinese founders dominate.
Take AI problem-solving — the previous generation had products like Yuanfudao. Now with more powerful AI technology, answer completeness and accuracy have improved dramatically. So products like Answer.AI, Question.AI, GauthAI, and Solvely rapidly emerged and swept global markets — all built by Chinese entrepreneurs.
👦🏻 Mars
We're relatively good at exam grinding, and this advantage aligns well with large language model characteristics. Because LLMs currently aren't great at creative tasks, but in education — transmitting accumulated human knowledge to new learners — they're particularly well-suited.
I used to think that with so many major companies already in this space, startups would struggle to break through.
But I'm increasingly feeling that the AI era may differ from the mobile internet era. Because everyone's "weapons" are too similar, major companies can't leverage their advantages effectively.
For example, the education company we incubated, AskXia, though smaller than companies like Answer.AI, we've found their accuracy on math problems is sometimes actually higher. This phenomenon is quite puzzling — technically hard to explain. But thinking more carefully, it's probably because everyone uses the same large language model foundation. What matters is who's better at tuning the model — Chain of Thought optimization, various techniques. At this level, team size may genuinely not be the decisive factor.
Of course, major companies definitely have advantages in question bank resources and such. But if you focus on R&D directions like improving effectiveness, there are actually still opportunities in this seemingly crowded赛道. Small teams can build better features — this was genuinely surprising to me. So I think education is indeed a massive赛道, and even now, when it looks so packed, there still seem to be openings.
I've been thinking lately — if you could combine Notebook LM, audio, video, beyond just problem-solving education products, there might be even more opportunities. These are precisely what we Chinese are best at.
🚥 Koji
While major companies may not outperform startups by much in final results, they do have entry point advantages. In preparing for this podcast episode, we researched many Chinese-style innovation cases, such as Quark's PC client. A substantial portion of its features center on education, problem-solving, and academic scenarios — including academic search, screenshot problem-solving, and more.
Because Quark has already embedded itself deep into many students' desktops, they'll likely just use what's closest to them. If the experience is solid, users form a habit. Add in cloud storage services, and it probably closes the loop on the entire usage scenario. So the competitive trajectory in this space is genuinely hard to predict.
That's also why many of the Chinese entrepreneurs we mentioned earlier who command "half the market" have chosen to go overseas. The traffic landscape abroad is more open — it's truly a hundred flowers blooming.
👦🏻 Mars
What Koji just said about traffic entry point advantages — that advantage is just massive. Traffic entry points are absolutely critical.
Entrepreneurs need to be extra careful about their positioning now: small features might get swallowed by aggregator products like Monica, aggregators might get swallowed by browsers, and browsers might get swallowed by operating systems.
Pure technical problems can still be partially solved through smarts, but when it comes to occupying entry points, that's genuinely a brutal fight.
The Silent Soak: When AI Melts Into Life
🚥 Ronghui
Alright, speaking of applications where AI functionality has become "the silent soak" — so seamless you don't even notice it — we discussed some great examples earlier. Some products have integrated AI to the point where users don't sense its presence at all; it's as natural as air.
So which ones did we talk about that are particularly strong in this regard?
👦🏻 Mars
Before naming apps, let me share something interesting. I've often heard that once AI is done right, you stop thinking of it as AI.
When you pull into a parking lot, you take it for granted that the gate recognizes your license plate, lifts automatically, and starts billing. But that's all AI technology — high tech just twenty years ago. There are countless examples like this: when we use Tencent Meeting, we're all getting AI beauty filters, eye contact correction that makes us seem to stare at the screen, even audio quality processed by AI. These AI features have melted into daily life, yet we barely notice them.
Recently I took my family to Jingdezhen. My wife shot a video of our baby using a blowtorch to make a glass cup. The editing was remarkably polished, but she said she just used "Douyin's one-tap video generation." Looking closely, it wasn't simple cutting — it could extract the baby from the frame and place them against different backgrounds. Behind this lies something like Meta's open-source Segment Anything model for entity recognition, but users don't need to know any of this.
When evaluating AI products now, we need to look at multiple layers:
- The underlying technology itself
- What capabilities the technology enables
- What tasks those functions help users accomplish
- And finally, understanding and fulfilling user intent
Mature products like CapCut have done exceptionally well, connecting all these layers. It's genuinely hard to compete as a startup in this space because existing products have already reached such a high bar.
🚥 Koji
In image and video processing, products like Meitu and CapCut are already world-leading. Looking at Meitu's public company data, you can see how adding AI capabilities on top of their existing strengths in image and video processing has made them even more formidable.
This is indeed typical Chinese-style innovation, but for entrepreneurs, it also means this track is now extremely difficult to enter. These giants have been cultivating the space for years, and now with AI integrated in, they've built very high barriers.
👦🏻 Mars
AI technology is now amplifying the accumulated technical advantages of established companies like Meitu and Douyin. For instance, before, if I made a product with just 5% of Meitu's functionality, most users might not feel the difference because those complex features were rarely used anyway.
But now it's different. With AI features like "one-tap video generation," users don't even need to articulate what they want — the system understands and automatically achieves their goals. These big companies can orchestrate all the complex features in the backend, and users enjoy 100% of the product's capabilities without needing to learn anything. So the barrier suddenly becomes very high, and small companies struggle to compete.
Because of this, entrepreneurs should now be doing newer, more "nonsensical" things.
For example, Viggle lets you upload a video and AI makes the person in it pole dance — a function that never existed before, so users have nothing to compare it against. Or having Suno sing songs — these things that might seem useless actually have more room to grow.
There are also applications that were previously technically impossible and for which no market existed — like helping blind people "see," or teaching deaf people to speak. These are new opportunities.
So entrepreneurship is best aimed at opening new markets; doing something nice-to-have in old markets probably isn't worth it.
Because while the flowers you add may be beautiful, cultivating that underlying "brocade" really takes years of accumulated effort.
Interesting People and Encounters in 2024
🚥 Ronghui
Now let's move to the last topic we're especially eager to discuss: who and what stood out as particularly interesting in 2024?
One thing that left a deep impression was the inaugural youth hackathon, Adventure X. We've had two of its organizers on our show to share their story, plus ten winning projects to introduce their products.
Koji went to the event and saw young people gathering together, spending three days and nights with full energy using AI to create whatever they wanted. It's a wonderful platform for young people to fully unleash their creativity.
What's especially rare is enabling so many people to have genuine offline connection and exchange during this period. Every time I mention this event — whether discussing it with tech people or sharing with others — everyone agrees it's deeply meaningful. Activities that let young people showcase talent, exchange ideas, and create together are truly worth encouraging.
So I wanted to ask Mars: who else and what else did you find particularly interesting in 2024?
👦🏻 Mars
I did meet many fascinating people. For example, Xiuhann Hu of "Nie TA" introduced me to the magical world of ACG subculture, making me realize that beyond handsome guys and pretty girls, there are bizarrely popular character concepts like "cockroach girls." These young people building niche, eccentric products gave me lots of inspiration.
Another person who left a deep impression was Li Zhen. He originally worked in the metaverse, then used AI to cut costs and boost efficiency at his own company, and later exported that methodology to other companies. Sounds like standard knowledge monetization, but what's interesting is he doesn't just lecture for money. After class, he leads workshops, organizes the workflows, and uses Dify to build and deliver a working SaaS system on the spot.
This made me realize a potential transformation: if Dify develops for another six months or a year and becomes more powerful, why buy off-the-shelf SaaS at all? They can build you a functional system right there on site — unthinkable in the past.
We've always assumed customization is expensive, but if R&D and customization costs drop to the point of same-day completion, it could massively disrupt the entire SaaS market. That trend is quite sobering.
🚥 Koji
I'm curious — did he have one or two specific examples of these customized SaaS builds? What scenarios were they for?
👦🏻 Mars
Yes, Li Zhen showed me a very practical resume screening workflow: from job boards to a dedicated email for receiving resumes, using Zapier to import emails into the system. I haven't deployed it myself yet, but seeing his flow was striking — using Google Sheets paired with AI to score resumes across multiple dimensions of competency models. The system automatically filters out resumes below 60 points, generates evaluations, and can automatically send emails to schedule interviews.
Sounds uncomplicated, but think about it — many HR management or recruitment workflow SaaS products charge three or four hundred yuan monthly for exactly these functions. But now it can be built for you on the spot. Based on your specific needs — your talent requirements, approval processes — he can drag out a couple of boxes and implement it right there.
They first use Coze to drag-and-drop a prototype, then refine it with Dify's development tools. That workflow is pretty remarkable.
🚥 Koji
AI has indeed created many "super individuals," enabling one person to accomplish what previously required a team. Among this year's interview guests, Chunxiang Zhao and idoubi are great examples. When recording their podcast episodes, they not only introduced their products but also shared how one person can work efficiently. They've both built widely acclaimed products with substantial traffic and revenue — all done solo. This phenomenon demonstrates AI's empowerment of individuals, letting one person achieve what previously required multi-person collaboration.
Speaking of super individuals, many are doing overseas business. Even without particularly strong language skills, AI assistance has made Japanese markets, English markets, and others viable. I've seen many impressive super individuals running overseas projects.
This leads us to wonder: are Chinese entrepreneurs truly at the global center? Teams like Heygen and Scale that we mentioned earlier have indeed become globally watched startup stars, while numerous super individuals are also making their mark on the world stage.
Looking at global AI To C product rankings today is quite interesting. In AI New Rank's global download rankings, two of the top four are Chinese products — Quark and Doubao. That's genuinely shocking.
This shows Chinese products are catching up in the new era. Though many don't see it, massive numbers of global users are already using Chinese AI products. Compared to the last era, this is substantial progress — Chinese products have quietly earned global user recognition.
🚥 Ronghui
Seeing these phenomena reminds me of a saying that may sound exaggerated: "Geniuses always appear in clusters." As Koji said, with AI empowerment, many people can accomplish incredible things.
In our last episode's 2024 wrap-up with Big Smart, he quoted Hyacinth's words, which gave us great inspiration and encouragement:
"The world isn't a stage far away from us — we are at the center of the stage."
Speaking of 2024, many people said AI applications would explode, but now feel it hasn't happened. I think behind this sentiment is everyone's persistent search for an "iPhone moment" or "iPhone 4 moment" benchmark. Because the iPhone 4 was indeed the iconic product that truly popularized mobile devices.
Still, after reviewing so much of China's innovation in 2024, I personally feel deeply inspired and encouraged. I'm curious how you all feel?
🚥 Koji
It's true that some people say large language models are progressing slowly and that product-market fit remains elusive. But preparing for this episode, we discovered so much exciting innovation. Whether it's established companies, startup teams, or super-individuals — people are getting things done.
Seeing this much vibrant innovation and breakthrough makes you more hopeful about next year, and gives us a lot of confidence and motivation.
👦🏻 Mars
First off, I don't think large models are slow at all. Especially since o1 came out — you can literally see progress happening every day. Take some of the products we just mentioned: at the start of the year, you'd say something and wait a beat for a response. Now you put on those headphones and it's instant. At the beginning of the year that would've felt like black magic. Now it's just the default — "Shouldn't you respond immediately?" People already take single-turn conversational voice at the 4o level for granted. That's actually pretty fast progress.
Then there's everything in video and audio we just touched on. We've been looking at a lot of opportunities in this space lately, and things are moving quickly there too. On the graphics and image side, people aren't necessarily only doing large models — Stable Diffusion, DIT, 3D Gaussian splatting, all of these keep advancing. Countless open-source models and papers come out every single day. It's incredibly vibrant.
And we just talked about applications. We're very good at building applications — connecting scenarios with technology. The technology itself, not everyone can master. Speaking as an incubator at this very early stage, we can't master it either. But we can piece things together. Once the technology exists, we can adapt it, repackage it — that's not hard. I think we still have great opportunities in applications. It's still fun.
The New Generation of Young People: Opportunities AI Brings to Entrepreneurship
🚥 Ronghui
I came across a line recently that said "nowadays AI practitioners are everywhere, all making 5,000 RMB a month." Many people also compare this period to the early days of mobile internet. That analogy has probably been made too often, so I won't dwell on it.
But one fact I think we've observed is that many young people are doing genuinely interesting, creative things. Just the other day I saw Liang Zhu in my Moments feed — they've formed something like a community, almost a communal space, and they're hosting their own Demo Day there.
These are facts we're seeing, things happening right now. In what is a relatively pessimistic macro environment, there are still many young people working hard and building things.
👦🏻 Mars
Like Ronghui just mentioned, he's been emphasizing young people, and I quite agree. Any disruptive technology benefits young people first. Because at my age, you pick up new things more slowly.
New technology doesn't typically replace young people — it replaces the mature experience of older people.
I have a vague recollection. In the earliest textile mills, you had to be very senior because it required so much experience; the machines were difficult to operate. But after the textile machinery and steam engines were improved, young girls and even children could operate them. On the dark side, child labor could work. But you'd find that people who previously had no qualifications to work there now could. The biggest losers were the senior technicians who'd been earning high wages — "only I know how to operate this machine, and now a child can do it." So this is the best opportunity for young people to rise.
Because AI replaces the experience of older people. Take podcasting: Ronghui and Koji are veteran content creators — they're 90 out of 100 on topic selection. I'm a 70. A young person might be a 60. But now, with AI, that 60 can instantly become an 80 — better than me.
What about recording and editing? Ronghui and Koji can hire excellent editors — 90 out of 100. But a young person who was a 60 can now use AI editing tools and jump to 80. The gap narrows. Spend a bit more effort on content, and you might surpass them. Previously, to learn editing and topic selection, you really had to study hard and catch up on all that accumulated experience. The veterans were already ten steps ahead, and you had to walk those ten steps while they kept moving.
But now with new technology, it's like flipping the table and starting a new game — "Everyone's using computers and new tools. You don't know how, I don't know how, but I'm young and I learn faster than you."
So I've genuinely been reflecting on what things are actually burdens. People who are slightly older definitely have a lot in their heads, and much of it becomes a burden that prevents us from learning. But young people don't have this burden, and then they nullify our experience. So this might be the greatest opportunity.
For example, there's something I haven't figured out recently — whether it's a burden in my own mind.
We have a Silicon Valley partnership that's clearly a business process collaboration. Say customer service handles a certain volume daily, certain calls, then needs to check certain systems, respond to users in certain ways. I immediately said, "This is so familiar — it's business process mapping," then looked at how to structure the model, how to do RAG. But my Silicon Valley friend just stared at me blankly and said: "What era is this? Why do it that way? Shouldn't you just feed all the data to AI and let it learn itself?"
I started doubting myself. Young people, Silicon Valley friends — they just assume your whole approach of decomposing workflows belongs to the previous generation, that you haven't caught up. I'm not sure if his radical view is correct, but you do find that as productivity changes, our mental frameworks for how we work and how we work effectively may all need to change. Previously, you had to learn how to break down workflows — I might be better at that than you. But that might immediately become a burden.
People who break down workflows become the "old guard" — you're the ones using rifles as clubs. But newcomers will learn the proper way to use rifles.
So speaking of young people, I genuinely envy them now. Our CTO often looks at me and says: "I envy you so much — you don't know anything. It's so convenient for you to use technology now. So much of what we learned has gone to waste." I say, are you complimenting me or insulting me? (laughs)
🚥 Ronghui
Right, when we previously talked about young people's creativity, I kept thinking of a song by Jiulian Zhenren called "Don't Bully the Young for Being Poor" — about a young person's ambitious spirit heading out to work.
👦🏻 Mars
Now look at it from another angle — for young people, this isn't about working for someone anymore. Because if you position yourself as just doing a job and delivering tasks, you're still a means of production. AI is also a means of production — you contribute intelligence, AI contributes intelligence.
In this scenario, you're napa cabbage, and AI is baby bok choy. Baby bok choy keeps getting cheaper and better quality, so napa cabbage definitely won't fetch a good price.
From a young person's perspective, you should be thinking "I should become a chef, open a restaurant." You're sourcing ingredients — baby bok choy, potatoes, sausages all getting cheaper and better value, and you cook them into dishes to sell. That's when you benefit most.
Only chefs are happy when vegetable prices drop. If you're the cabbage, you're sad when prices drop. So we must make ourselves into chefs.
Previously you'd think, "I want to open a restaurant, but I have no money, no people, can't afford two helpers, I only saved 3,000 RMB." But now you can gaslight AI, because AI is cheap — you can hire cheaper labor. Previously, being a boss, you'd think "I only have 3,000 RMB in savings." But now you find that 3,000 RMB can hire ten workers. You can organize them to spot opportunities. You can actually be more "boss-like."
Everyone wants to be a boss now. You might feel like the boss is exploiting you — well, now you can exploit AI. Finally there's an opportunity where everyone can define what they want to do, and through very limited resources organize some productivity to help accomplish what you want to accomplish. This is rare in history. Historically you needed to save at least a million before you could think "I'm going to do something significant." Now 3,000 RMB is enough. That's a great opportunity.
🚥 Ronghui
Jensen Huang said in a podcast recently too that the future is this AI super-factory model of working.
Alright, let's wrap up here for today. Thank you so much Mars for joining us again at Crossing and sharing so many insights and observations on the industry. We'll stop here for today.
👦🏻 Mars
Great, thanks. Bye.
🚥 Ronghui
Alright, bye.
Subscribe to the "Crossing" Podcast
🚦 We follow the industry transformations and entrepreneurial opportunities brought by the new wave of AI technology. Crossing is Steve Jobs' metaphor for Apple — standing at the crossroads of technology and liberal arts, where great products are born. AI is transforming every industry. We seek out, interview, and bring together "active agents" of the AI era. Together with them, we explore and embrace the new changes and new possibilities.
👦🏻 Host Koji: Co-founder of The Fair and Tangdao. I believe technology, especially AI, will fundamentally transform society and empower humanity. Welcome to chat with me, bounce ideas, and connect on the next possibility. Koji on Jike[1], Koji's website[2]
👧🏻 Host Ronghui: Works at a tech VC, former Silicon Valley correspondent for CBNweekly. Ronghui on Jike[3]
Join the "Crossing" Membership Community
☀️ First-hand AI news and insights
👫🏻 We encourage everyone to date/make friends/find future collaborators
🦀 Add our assistant on WeChat to join: Rwkfbcianvd, or scan the QR code below
🎄 This podcast is supported by the Sound Forest Podcast Plan.
References
[1] Koji on Jike: https://okjk.co/0JSUes
[2] Koji's website: https://koji.super.site/
[3] Ronghui on Jike: https://okjk.co/0cbnYV