2024 Year in Review: AI Going Global — Dreams, Guts, and Making $10M in a Year | A Conversation with Can Zhang of Linear Capital and Ning Gao of Linkloud

十字路口Crossing·January 24, 2025·1·0

How do you seize the new opportunities in 2025?

This week's Crossing — our final episode before the Lunar New Year — we're tackling the question on every AI investor and founder's mind: "going global."

In 2024, a growing wave of AI entrepreneurs set their sights on overseas markets, giving rise to a cohort of companies and products generating over $10 million in annual revenue. We'll walk through them on this week's show: Plaud, Monica, Opus Clip, HeyGen, Notta, Hix, and more.

The global push is unstoppable. Among the AI founders I've met recently, I'd estimate five out of ten told me outright they're building for global markets. The other five are still focused domestically, but they're restless — actively thinking, stockpiling resources, and working up the courage to make the leap.

So on this week's Crossing, we're asking: Why are AI founders going global? What opportunities and possibilities are they seeing that the rest of us aren't? For those AI companies that have already gone from zero to one and crossed the $10 million annual revenue mark, what did they get right? How can we study, emulate, and replicate their success?

👬🏻 This week, we're joined by two guests:

Can Zhang, Managing Director at Linear Capital: Linear invested in 11 AI projects in 2024, staying highly active. Over the course of the year, Can estimates he spoke with at least a hundred AI founding teams, keeping his finger on the pulse from the front lines.
Gao Ning, Co-founder of Linkloud: Over the past two years, he's accompanied more than a hundred entrepreneurs on trips to the United States and Japan, developing unique insights into how founders think and where markets are heading. About six months ago, Gao Ning joined us on Crossing for an episode called "The AI Products You Don't Know About Are Quietly Making Serious Money Overseas," introducing nearly twenty standout global AI products. Today, we'll continue that conversation with him — what new products have emerged in the past six months that are quietly raking in cash abroad?

Listen on WeChat:

Listen on Xiaoyuzhou:

2024 in One Word: Unlock and Discover

🚥 Koji

Welcome, both of you! If you had to describe your 2024 in a single keyword, what would it be?

👦🏻 Gao Ning

Thanks for having me, Koji. Really excited to record this second episode with Can and you, six months after our last one.

If I had to pick one word for 2024, it would be "unlock." This was year two of building my own company and community. I chose "unlock" because we started running our own salons in new places — the Bay Area, Tokyo — and met a lot of local founders there. The second reason is that we finally got to see some of the companies that went global with us in 2023 make real, substantive progress on the ground in 2024, especially in Japan. We were fortunate to be there for that zero-to-one journey.

🚥 Koji

Can, what's your one word for 2024?

👦🏼 Can Zhang

Thanks, Koji. For me, 2024 was about "discovery." Since at Linear we mainly invest in AI applications, and the entire AI application space is still in a very early stage.

There's often this feeling that we're standing at the beginning of a new era, like standing on a vast plain that gets bigger every day, searching for new species emerging.

To be honest, this feeling has been rare for a long time. There's a saying: "The wind rises from the tip of the green duckweed" — I think that's roughly what this feels like.

🚥 Koji

That's a vivid metaphor — standing on an expansive plain that grows larger by the day, searching for where new life is taking shape.

Plaud and Notta: How Two AI Products Reached Eight-Figure Revenue

🚥 Koji

We mentioned several cases that have crossed $10 million in annual revenue. What other global AI products are worth highlighting? Gao Ning, why don't you start.

👦🏻 Gao Ning

I'll start with one people may have already heard of: Gao Xu's PLAUD.AI[1]. It's an AI hardware device that attaches to the iPhone and records calls, then sends a summary afterward.

At this point it's not just annual revenue over $10 million — as we mentioned on social media, it's reached $10 million in monthly revenue. From a growth and product-definition standpoint, this is a very typical and representative AI product. It leveraged emerging channels like TikTok effectively, finding what looks like an unremarkable niche. It reminds me of Yusen's point from our last episode: find a demographic or need that seems unglamorous but is actually undervalued, then open up the market fast.

🚥 Koji

When people look at Plaud, many think there's nothing special about it — call recording with AI transcription. It sounds like a simple idea. And he probably faces fierce competition, even ultra-low-price knockoffs from Huaqiangbei.

Since you've spoken with him quite a bit, I have two questions: First, how did he become number one in that competitive environment? Second, how is he thinking about the tidal wave of copycat competition he's about to face?

👦🏻 Gao Ning

Let me answer the second question first: hardware is just the entry point; the real extension happens in the digital world.

Once you convert that voice and audio from the physical world into text, you have to think about what value you can provide to different user profiles.

He mentioned that because the AI software component is delivered through a subscription model, starting in the second half of last year, certain professionals — lawyers, dentists, on-site construction workers, insurance salespeople — have actually found it to be a highly effective tool for work, and they're very willing to subscribe and renew.

On the first point, I think Gao Xu is one of the fastest movers I've seen after GPT came out — he grabbed the capability, found the opportunity, and made the best product integration. From a timing perspective, he was definitely among the first to launch. That user base and mindshare, if the first mover executes well, creates a massive advantage. Looking at subsequent keyword searches, video views, and other metrics, he's maintained significant long-tail dominance.

🚥 Koji

Let me also ask Can — among the global AI products you've seen cross $10 million in annual revenue, which ones struck you as especially worth mentioning?

👦🏼 Can Zhang

We invested in a company called Notta[2] — you may have heard of them. They mainly do AI meeting assistants.

Largely because this wave of AI made what they wanted to do actually possible. Unlike other companies, they started from the Japanese market. We watched them go from making a few thousand dollars a month in Japan to now generating over $10 million in annual revenue. They now hold over 75% market share among Japanese C-end users. After establishing themselves in Japan, they shifted from B2C to B2B, and also began globalizing — expanding from Japan to overseas markets. Now roughly 25% of revenue comes from outside Japan, with about 20% from the United States. These details probably aren't widely known.

🚥 Koji

Meeting transcription is an incredibly crowded space too. How did Notta get to where they are today?

👦🏼 Can Zhang

I think there are two critical factors. First and most importantly, Japan is a very special market. It has a natural barrier that blocks many competitors from the start — and they voluntarily stay on the other side of that barrier. Second, local competition in Japan is relatively weak; the entire environment is supply-constrained. So by doing SEO with a disciplined, professional approach in Japan, they were able to grow fairly smoothly.

🚥 Koji

When they did SEO in Japan, was it on Google? Did they use a local Japanese team? Or Chinese people who speak Japanese?

👦🏼 Can Zhang

In the very beginning, Chinese people who speak Japanese were sufficient — you could even find Japanese students studying in China. For B2C, there was no real need to be physically in Japan. But once you start moving down the B2B path, you need a Japan office, you need a Japan country lead — that's a step up in the fight.

🚥 Koji

We also mentioned another AI product that crossed $10 million in annual revenue: Hix[3]. This is a relatively low-profile company, so people don't typically discuss it much. Hix now has a product portfolio of several dozen products. I was surprised to hear you mention earlier, Can, that Hix's current state is something only the founder, Biao, could pull off.

Could you introduce Biao and explain what makes him so exceptional?

👦🏼 Can Zhang

What's most impressive about Biao is that through his past experience, he's developed a very complete methodology for traffic operations — from product selection to traffic execution, he has distinctive expertise in every step.

The same task can produce results that differ by orders of magnitude in traffic performance depending on whether it's handled by someone who understands operations or someone who doesn't. I think Biao is the soul of Hix.

🚥 Koji

I'm curious — looking back on this past year, what have you two done that you'd consider worth sharing?

Earlier on Jike, I teased Gao Ning by calling him China's Zheng He of 2024. Because he's taking entrepreneurs overseas, like a modern-day Zheng He.

👦🏻 Gao Ning

First, I think we've done something quite complete across the first and second halves of the year — making trips to Silicon Valley and Japan every six months. We've seen that whether it's founders of early-stage startups or founders of listed and soon-to-be-listed companies, they're all personally leading teams of senior executives on field visits. I saw what's called "founder mode" here in Silicon Valley — this spirit of founder-driven initiative — and it's become noticeably more pronounced compared to two or three years ago. That's a significant change.

Second, as we've continued to execute, we held our first SEO workshop in 2023. This year, building on topics that continued to develop and attract attention, we added two new tracks: one on PMF and one on social marketing.

Overseas promotion methods differ greatly from domestic ones. It's not just about different channels — you need to understand what users in local markets are thinking about and care about. This is a shift in mindset.

The last thing that made me quite happy was holding our first salon for Chinese people in Japan.

🚥 Koji

Can, I'm very curious — Linear Capital has been one of the most active dollar funds this year, with 11 investments. Could you share your main investment directions? If you're able to disclose, you could also introduce the specific companies and products you've invested in.

👦🏼 Can Zhang

This year we've made a total of 11 investments in the AI direction. Early in the year, we also launched a program called "Linear Bolt," mainly to help a new generation of globalized AI entrepreneurs. In the early stages of AI development, there are opportunities for both us and entrepreneurs. We hoped to create some new changes and provide help in a more agile way.

These 11 investments primarily target global markets, mostly in the To Consumer and To Professional directions. The vast majority of companies are headquartered in China. Our main investment directions include: education, companionship, productivity tools, and tech consumer products. These are also areas we'll continue to focus on. Fundamentally, we don't have too many restrictions on investment directions, but we do have this inclination, which is more a result of the sequence in which AI applications are landing.

🚥 Koji

Right, Linear made 11 AI investments. How many total investments did Linear make?

👦🏼 Can Zhang

Linear made 17 total.

🚥 Koji

So the AI proportion is still quite high.

Then I also noticed another interesting detail — you said most of them are AI companies targeting global markets. How much is "most"?

👦🏼 Can Zhang

Probably all except one are global from day one.

🚥 Koji

I imagine this represents a sea change from your investment map of a few years ago — you wouldn't have had such a high proportion of companies going overseas. Was this deliberate, or did this distribution emerge unconsciously?

👦🏼 Can Zhang

We've actually thought about this question for quite a long time. The main reason, I believe, is China's engineer dividend. China's engineer dividend should be directed toward a larger market. Based on this idea, we started down the globalization route.

🚥 Koji

This year, both of your work appears to be overwhelmingly focused on AI going overseas. Are there any stories that left a particularly deep impression on you?

👦🏻 Gao Ning

In Tokyo, I attended the second global summit organized by WaytoAGI[4] and ComfyUI[5] Chinese organizers, along with some large model AI entrepreneurs.

What surprised me was that they not only invited the founder of ComfyUI but also many video AI large model trainers, as well as very talented creators, designers, and enthusiasts who all gathered in Tokyo. Everyone completed nearly 20 sessions of exchange over a full day. This was a global event initiated by Chinese people, and it made a huge impact on me.

Let me share a second thing. In the Bay Area, I met a Chinese entrepreneur working in AI healthcare, doing AI hospital informatization and automation services. Although this is particularly heavy work, it genuinely helps nurses and logistics staff in specific outpatient departments unlock significant capacity, allowing them to devote themselves to things that truly matter to patients and hospitals.

As a Chinese person doing SaaS or traditional domain entrepreneurship, I asked him whether he worried about encountering many challenges and barriers to integration. He gave me a very moving answer. He said he came from a PhD background and never thought he'd do AI + SaaS entrepreneurship. But what impressed him most deeply was that successful products like Notion[6], Airtable[7], and ComfyUI[8] all had Chinese founders behind them. These founders all spent a lot of time early on polishing products and understanding users, adjusting many times before ultimately succeeding. Their success not only inspired him but also gave VCs more confidence to invest in Chinese entrepreneurs doing SaaS or entering fields with high barriers. So he also very much hoped to have opportunities in the future to share his own experience and contribute to the community.

🚥 Koji

So you saw two things: one is the courage of entrepreneurs, and the other is the example and power that entrepreneurial predecessors have established for a new generation of entrepreneurs.

Actually, in one of our earlier "Crossing" episodes reviewing dollar funds, we spent a long time discussing dollar funds, talking about how U.S. dollar funds are now very actively investing in Chinese teams. We did some statistics at the time — for example, a16z's Speedrun incubator[9] invested in seven or eight Chinese teams. They even recruited investors with Chinese backgrounds to join their team, not ABC backgrounds, but people who did their undergraduate studies in China. This is a very rare type of hire in their history. I think the signal this sends is that they have high hopes for this new wave of Chinese entrepreneurs as they expand into overseas markets.

Can, I'd like to ask you to share any stories or people related to going overseas in 2024 that left a deep impression on you?

👦🏼 Can Zhang

I think I'll also share two things. The first was that I participated in a Japan trip organized by Linkloud[10]. This was probably the most inspiring trip of the year for me, and also my first time going to Japan for something like a business考察.

This experience made me realize that Japan is a very special market — it's quite different from both China and the U.S. I started thinking:

Often, when the first step of something is very high, people lose the interest and courage to explore because of the difficulty of crossing that step. But behind that step may be opportunity or paradise.

This discovery left a particularly deep impression on me and has influenced some of my current work to a certain extent.

Another thing is that when we launched the "Bolt" program early in the year, we also did some other work. We started taking our WeChat official account more seriously than in the past, using content as a starting point to write articles. The reason for this is that we feel at this point in time, everyone is thinking about future directions and their own positioning, but there are no standard answers because the world is changing too fast. So we hope to help entrepreneurs from these angles, sharing our thinking and the information we see. I think this is very important in this era.

A colleague told me before that they heard people praising our official account as very well-written on different occasions. Yesterday I happened to run into an old friend who said: "You started doing content this year, and it's pretty good." This was the first time I heard such an evaluation, and I felt particularly proud.

But what I really want to say is that I also especially encourage all friends who are currently thinking to share their own ideas at this time. Because in a world where everyone is eager to understand what others are thinking, the leverage effect of sharing will be particularly large.

🚥 Koji

I recommend everyone follow Linear Capital's official account. Actually, Can just mentioned that Linear has been very active this year — not just in investing, but also in content creation. So what have you felt on the front lines? How many projects have you looked at?

👦🏼 Can Zhang

I did a count before coming here — we have roughly over 100 projects in our database. By "in our database," I mean projects that we've formally communicated with and need to log in. Actually, because people know we're quite open, there are even more projects at the first-contact stage. I'd like to share a few fairly clear observations:

This year the overall pace has indeed slowed down — there's no need to rush for projects. The reason for this phenomenon may be different from what people think of as "investment cooling." I think it's because in a time when everything is changing rapidly, founders are thinking carefully, and investors are also thinking carefully. Only when both sides' thinking reaches a point of alignment does investment happen. This is actually a more normal state. Of course, in this process of change, thinking takes more time — unlike before when everyone was in a fixed state for a long time, and this thinking was already done, so you could invest directly when you saw something. Now I think many times everyone needs to consider carefully.

On another note, I think the cost of trial and error for entrepreneurship has significantly decreased now. Because of AI, the cost of trial and error is lower than before, and the speed of trial and error is faster than before. To some extent, I think everyone is expecting founding teams to be able to demonstrate something convincing. Because it doesn't take very long — at least from our perspective, many things can be quickly validated — this may also be one reason.

Coming back to it, I think this is the normal speed of investment. The entire market, at least from what I feel, absolutely isn't lacking money. As long as founders can deliver good results at whatever stage they're in — whether Pre-product or at the PMF stage, whether you have special ideas and prototypes at the Pre-product stage or beautiful data at the PMF stage — I don't think you need to worry about fundraising. That's my feeling.

🚥 Koji

Gao Ning is actually a former investor himself, and now probably like being an investor, still spends every day with entrepreneurs. What have you felt over this past year?

👦🏻 Gao Ning

Actually everyone has noticed that some of the products we discussed earlier, including Plaud and Hix, have developed quite well. Whether from a user or revenue perspective, they've had good growth, but these companies haven't raised funding (bootstrap). We've indeed seen more and more companies like this this year — it's not that they'll never raise funding from start to finish, but that they don't need to raise funding from the start.

Because the market cost of rapid trial and error has really become very low, some entrepreneurs may be just one or two people, or a small team, and can quickly validate certain products.

They become good indie developers or small teams and achieve a certain revenue scale. In this situation, they can choose not to raise funding.

New Species: Innovative Products Quietly Growing Overseas

🚥 Koji

Six months ago, Gao Ning joined us for an episode about "overseas companies making serious money under the radar." That episode got huge numbers — everyone wanted to find the wealth code in it.

Now another six months have passed. Gao Ning, have you spotted any new overseas products quietly raking it in?

👦🏻 Gao Ning

First, something interesting in the image space. When we talked in the first half of last year, there were lots of text-to-image products — Cutout.pro, Fotor, and so on. These projects grew rapidly after open-source image models like Stable Diffusion emerged, and they were pretty representative of the overseas expansion playbook. Looking back from 2020 to now, I see these companies have evolved from pure image processing into multimodal plus video formats. The wave from text-to-image to image-to-image to video — I think this dividend is far from over. Products like OpenArt[11] and Polyverse[12] are doing very well.

These products have made especially big breakthroughs in revenue growth, audience segmentation, even geographic expansion. For example, targeting fitness enthusiasts, photo retouching users in different countries, different age groups, plus poster-making needs for SMEs (small and medium-sized enterprises), even specialized scenarios like restaurants — all of this has driven revenue gains beyond what anyone imagined.

Second, an interesting area: religious apps. There's a product developed by Chinese founders called Blessed, focused on Christianity, offering scripture reading and guided prayer services. Its impact isn't just functional — it's more about spiritual uplift. I understand this product has seen major growth. Also, Xmind[13], that very successful mind-mapping product we previously featured at one of our salons, launched a Buddhism-integrated product about a year ago. They recently developed an AI tarot app too. They shared that it's mainly through exquisite color palettes and character design that they've attracted paying users who resonate with their aesthetic.

There's also a relatively mature company, Ceche[14] — I'm sure many of us have used their mini-program. They started with psychological tests, personality assessments, MBTI, and now combine AI for better interactivity and interpretation, achieving significant growth over the past year or two.

Third, I think AI-native hardware has especially big opportunities in the next year or two. At the recent CES we already saw many imaginative products, with founders or teams mainly from China and Japan. I'm more excited to see products emerging from psychological or emotional companionship angles, targeting specific demographics and countries.

🚥 Koji

Actually, Xing Wang rarely posts on Moments, but yesterday he shared a CES product — a pretty hilarious one. It's called Nékojita FuFu.

Why FuFu? Because it perches on your cup and goes "FuFu," blowing your boiling water or milk or coffee cool.

👦🏻 Gao Ning

Right, I can add — I read an article about CES hardware, and this $25 FuFu left the deepest impression on me.

🚥 Koji

I think they're still selling joy, selling a story. Regarding the new overseas products with solid revenue that Gao Ning just mentioned, let me try to summarize two themes:

One is the improvement in model capabilities unlocking many new possibilities — whether text-to-image or text-to-video, open-source or closed-source, these advances create lots of opportunities.

The other, the relationship between religion and AI, is fascinating. These products don't brand themselves as AI products — you won't see "AI" anywhere on their sites. So why does AI help religious products take off?

I think there are two reasons: First, religious products need massive amounts of content, which was extremely expensive to produce in the past, and AI solves the content supply problem. Second, these products need to acquire customers and get traffic, and AI helps enormously here too, especially with SEO.

The similar products I've observed share two characteristics: First, AI can generate content with high quality and efficiency. Second, AI can help acquire traffic — solving both the supply side and the traffic side. There are actually quite a few entrepreneurial opportunities like this.

There's another case from the past six months that impressed me, though it has nothing to do with AI — an app called StressWatch[15]. Because app revenue is relatively transparent, you can see on Sensor Tower[16] that StressWatch is now doing $600,000 to $700,000 per month. This reminds us: while we constantly discuss AI, Chinese teams are also continuously building new products in other fields that global users recognize and are willing to pay for.

👦🏻 Gao Ning

When we observed app-side products last year, we saw some small teams doing quite well. They borrowed scenarios already validated on the web, found similar user groups and opportunities on mobile, and moved fast to develop mobile applications. I believe this will be a pretty significant opportunity this year.

🚥 Koji

Actually, many of the profitable companies we mentioned are web-based, and there's a traffic dividend at play here. Google has existed for decades, but the past decade-plus of mobile internet development has actually made PC traffic competition weaker than mobile. Companies can very efficiently acquire free traffic through SEO — something that's impossible in the mobile era, even right now. Getting free traffic on mobile is extremely difficult.

That's why we're seeing many profitable companies in this wave building for the web. And because AI naturally has productivity-enhancing properties, new opportunities are emerging on the web. Can Zhang, anything to add?

👦🏼 Can Zhang

Since we mainly focus on China-based startups, we've found that the profitable tracks are relatively crowded right now. The successful products we see are basically concentrated in a few major categories: images, video, and some AI companionship or chat products. These are all large markets, but indeed very crowded. We've done well largely because we already had deep research in these areas.

Video was naturally China's strength — because of Douyin there's CapCut, then various AI video tools evolved from that. We're also eager to see similar opportunities in other fields and markets. This echoes the driving forces we discussed earlier, like new models bringing new opportunities or new model capabilities.

I can give an example: many previously quietly profitable non-AI domains can actually be transformed by AI while maintaining their stable profitability characteristics. This year is probably the first year that AI speaks like a human. If we apply this technology to phone services, we find there's a particular domain overseas: services that literally make phone calls to handle billing disputes, returns, and such for you. This was something you didn't want to do yourself, and previously companies would outsource it. Now AI can handle it. We've seen teams trying this, and they're ranking well on Product Hunt too.

This is an industry where profitability is clearly established. The project is called Pine, built by a former Agora team.

🚥 Koji

That's very interesting.

👦🏻 Gao Ning

Similar to DoNotPay[17].

👦🏼 Can Zhang

Right, similar to DoNotPay and Bill Shark[18]. We're looking forward to seeing more projects in these already-mature commercial domains get innovated and transformed with AI.

Devin vs. Cursor: Differentiated Positioning

🚥 Koji

What other products have you seen recently that gave you lots of inspiration?

👦🏼 Can Zhang

Two areas made the deepest impression on me: coding and multimodal. On the coding side, I've tried most AI coding products on the market, and my best experience has been with Cursor. What's interesting is that these products actually target different audiences and customers.

This gave me a deep feeling about product strength. Discussing product strength is really about getting the product to the degree of "just right." Today, because AI handles most of the foundational work, what teams really need to do is spend lots of time on research, giving users precisely the right amount of control, so the product outputs exactly what's needed with minimal interaction. Sounds magical, but AI can actually do this today.

For example, as an engineer using Cursor, I need a certain degree of control. I need it to work according to my ideas. Now it basically outputs the right result 80-90% of the time, while still giving me appropriate room for control.

By contrast, another product called bolt.new targets non-programmers. I find it very awkward to use because it always feels like I'm "missing the mark" — slightly off to the left, slightly off to the right. But for its target users, this might be a perfectly calibrated product. Because for them, it's entirely natural language interaction, getting a complete result without needing to do any technical work.

🚥 Koji

Right, it's actually extremely subtle. On the surface, when people discuss Cursor they group it with bolt.new, Windsurf, Lovable — but after actually using them, you discover their target audiences and design philosophies are fundamentally different. Devin is even more distinct from the others.

👦🏼 Can Zhang

Right, I spent $500 on Devin, but it felt like it stripped away all my joy and left only unhappiness. It completely removed my creative capacity. I gave it a simple requirement, but it made me wait for a very long time. Finally it produced a result, and I still needed to very carefully check whether it met my requirements, then continue interacting with it. This process stripped away all the pleasure and left me only with the unpleasant parts.

🚥 Koji

Yes, the joy you're talking about is mainly the joy of creation. With Cursor, you're still creating — basically interacting with it every 30 seconds to a minute. With Devin, you hand it a task, then come back hours later to check the result.

👦🏼 Can Zhang

It takes forever, half an hour minimum, and honestly it really can exceed half an hour, an hour.

🚥 Koji

Right now I use both Cursor and Devin. They serve very different purposes, but I've found ways to use each of them effectively in different scenarios.

For Cursor, I use it to solve very specific problems. For example, I just wrote a simple program the day before yesterday: every morning at 10 AM, it checks my Gmail for any new invoice emails from the previous day. If there are any, it downloads the attachments or links from those invoice emails and saves them to a designated location in my Google Drive. That way I can batch download them for finance at the end of each month. For this kind of specific technical task, even though Devin could probably handle it, I tend not to delegate it to him.

So what tasks are suitable for Devin? Research tasks. For instance, I've been thinking about a project recently: if we were to develop a second language version, which language should we choose? Should we enter the Spanish market, the Japanese market, or the German or French markets? I gave this task to Devin. He's excellent at breaking a task down into different components — it's like hiring a McKinsey consultant who sets up a rigorous framework based on my objectives. The process itself is enlightening. What I care about isn't the result Devin gives me, but rather how he deconstructs the task, which shows me how I can break down my own thinking process.

👦🏼 Can Zhang

Actually, Devin really isn't just a tool — he's more like an agent, a junior employee. In certain areas, his performance isn't actually junior at all; it's quite good. You just hand him the task and come back to check on it later.

From Notebook LM to Multimodal Interaction

🚥 Koji

The Devin I use is on a team account that a friend purchased, with six or seven people in it. Since team members can see each other's usage, I noticed that almost none of the programming tasks people assigned to him succeeded — the final results were mostly unsatisfactory. So now almost everyone has adopted the same approach as me — mainly assigning research tasks to him. This phenomenon is quite interesting.

Devin is one product you mentioned that's impressive and inspiring. Are there others?

👦🏼 Can Zhang

Speaking of multimodal, another product I really love is Google's Notebook LM[19]. For those unfamiliar, let me briefly introduce it: it started as a research assistance tool. What I particularly want to talk about is its Audio Overview feature, which can convert materials (whether PDF or other formats) into a several-minute podcast summary.

So where's the innovation in this product? Although TTS technology is already quite mature, and people might think this sounds like simple work. But in reality, I've tried many approaches and researched many methods to replicate it, and very few can reach Notebook LM's level. It handles certain details exceptionally well. Like when we're recording this podcast — when I hear Gaoning say something particularly exciting, I'll jump in; during the conversation, because Koji is actively listening, he'll make brief response sounds like "mm-hmm" or "yes"; when getting excited, the speaking pace speeds up or slows down. These details are actually very difficult to achieve with existing other technical approaches.

🚥 Koji

Yes, it's been about four months now, with many replication attempts, but none have reached even 60% of its level. This is indeed interesting. Ronghui and I previously recorded a podcast episode specifically about ten interesting use cases we discovered for Notebook LM.

Recently Notebook LM released an even more interesting new feature: you can listen to two people conversing, and at any time insert a question — they'll stop to answer. And before answering, the two people will even exchange a few words with each other, picking up on what the other said — the entire process is extremely natural. I find this unimaginable. There must be a great deal of innovation here beyond AI models alone — likely deep user insights from product managers and engineers, combined with engineering capabilities, that achieved this effect.

To prepare for the previous episode, we researched their team behind the product. One interesting discovery: among the core members is someone who is neither an engineer nor a product manager, but rather a very influential bestselling author in the United States. He participated in this project from beginning to end, and it's likely his thinking provided some wisdom that's difficult to capture. These subtle details make Notebook LM's product a level above other similar products, and to this day no one has truly been able to replicate it.

👦🏼 Can Zhang

Yes, that author you mentioned was on an episode of Hard Fork. Additionally, their product team also recorded an episode on another podcast, Training Data. Both resources are great for understanding their thinking and productization approach. In comparison, other replication attempts we've seen mostly stay at the technical level and code level.

🚥 Koji

Right, the product manager and core engineer of this product have now left to start their own company. Seeing this news, I also had a thought: Silicon Valley is indeed a free place, with no non-compete agreements. Once you've made an excellent product, you can very quickly and freely go do the next thing you want. In comparison, many friends around us have to bear two-year non-compete restrictions — that's really deadly, I think it stifles innovation tremendously.

👦🏼 Can Zhang

There's indeed no precedent for the results of innovation. This is a legal-level difference — California law prohibits non-compete restrictions.

🚥 Koji

Actually, law is also a reflection of everyone's culture and values.

👦🏻 Gaoning

Of course.

👦🏼 Can Zhang

Especially since their laws are all built through case litigation.

👦🏻 Gaoning

I want to mention two products. The first has become my daily tool — YouMind[20], developed by Teacher Yubo.

As a creator, I have high sensitivity toward every article I want to read, knowing which content is useful to me. There are two elements here: First, I hope to quickly save article snippets, including images, videos, and even highlights from videos. This is the so-called Save function, an important starting point for building my future creative materials.

The second element is: during the thinking process while reading an article, I might record some personal thoughts, or I might use AI to explore technical areas I'm not familiar with. It's like having someone beside me to exchange ideas with — I really like using the plugin form to capture a passage of text and then have a dialogue with AI.

Another relatively mature product is my favorite AI Native product — Gamma[21]. The founder is a Chinese-American born and raised in the United States. This year, one feature particularly impressed me: beyond AI-generated PowerPoint and slides, it's their innovation in the Prompt domain. I think it excellently solves the problem of how to let users not have to write prompts.

For example, when making slides you need a lot of supporting images. Gamma's current feature is: after you input content, it has an Optimizer that automatically parses your simple description into two parts — entity description and style environment. These are all automatically generated based on the user's initial rough prompt description.

I've used many similar image generation products, but what surprised me was that although Gamma isn't a specialized image generation tool, its results are excellent — usually no more than two tries to select a suitable supporting image. This demonstrates true product strength: they deeply understand that users want to generate good content but don't know how to write good prompts, and they've invested heavily in this area.

Land of Opportunity: The Uniqueness of the Japanese Market

🚥 Koji

This sounds like a relatively simple thing, but they really built it into the product and delivered a great user experience. We've also been working on an AI blog writing project recently.

To give blogs good SEO performance, you need to insert supporting images, tables, table of contents, and other elements into articles. The image part is quite nuanced — I require three images per blog post, and I want the AI to choose the most visually evocative positions to insert them based on context. Then it writes prompts according to context and calls the Recraft API. We've completed several hundred now, and every one looks quite pleasing to the eye.

Earlier when we were discussing going global, both of you mentioned Japan more than once — this seems to be a keyword that left a very deep impression on you throughout this year. What opportunities do you see in the Japanese market, or what possibilities excite you so much?

👦🏼 Can Zhang

Japan's biggest characteristic is that it's a market where demand exceeds supply. The main reason is that Japan's own IT capabilities are relatively weak, and its AI strength isn't strong either.

But overall digitalization demand, especially in the past two to three years, has become extremely strong. There are many opportunities embedded here, and the market lacks supply.

I think there's another important reason: the entry barrier to the Japanese market — its first step is particularly high. This causes many entrepreneurs and companies to automatically filter it out when considering options, thereby creating this unique market situation where demand exceeds supply.

👦🏻 Gaoning

Japan's market barrier is relatively high, mainly because they're quite conservative and not very willing to change easily.

But there's an interesting phenomenon: once you enter the market, if you can bring users revolutionary products or entirely new quality experiences, and demonstrate your commitment to localization and steadfast development in the Japanese market, making people recognize that you're there to serve them, then their acceptance of you rises dramatically.

From another perspective, Japan's capital environment is also relatively friendly. If a company seeks to go public or exit, on one hand there are larger conglomerates or CVCs that can serve as acquisition targets; on the other hand, the revenue threshold is quite flexible. If your revenue reaches tens of millions of dollars, there are corresponding platforms for listing. When revenue reaches the range of nearly a hundred million dollars or several hundred million RMB, there are different listing channels available. Notably, among this year's larger software or tech-related IPO projects, quite a few came from Chinese or Greater China teams.

👦🏼 Can Zhang

Speaking of the capital environment, beyond being exit-friendly and having low listing barriers, Japan has another interesting characteristic:

Japan's early-stage investment is relatively weak.

Whether in terms of overall capital volume or the level of assistance provided, it's somewhat smaller than China and the United States. From this perspective, Chinese entrepreneurs entering the Japanese market may actually have a certain competitive advantage.

🚥 Koji

So Chinese startup teams going to Japan don't just have the engineer dividend — they also have the capital dividend. Yes, so it's this double buff.

👦🏻 Gaoning

To give a concrete example of startup opportunities, one particularly interesting point is this: even with today's foundational language models and translation capabilities, it's still surprisingly hard to find a product that can do high-quality Japanese-to-Chinese or Japanese-to-English translation for the Japanese market. That genuinely is unexpected.

While there are plenty of translation products out there, real-time translation depends on rapid contextual understanding. Because once you understand what comes later, you often need to go back and revise what you translated earlier. And Japanese itself has so many unique expressions and inverted structures — these linguistic characteristics create massive localization opportunities in real-time translation.

2025 Outlook: The Certain and the Uncertain

🚥 Koji

And after everything we've discussed today, both of you have shared so much information and so many insights.

I'm personally curious — as investors and founders, how do you maintain this kind of sharpness about the industry? Are there any particular information sources or methods you'd recommend?

👦🏻 Gaoning

Last year I mainly used two channels: Newsletters and Podcasts. Both platforms offer immersive reading experiences on the web. With Youmind, I've also gotten better at collecting and organizing all kinds of information.

Lately I've been paying close attention to Newsletters from well-respected investors, including the weekly from Ed Sim, founder of Boldstart Ventures, and the SaaS weekly from Jamin Ball at Altimeter.

As for Podcasts, the content is even richer. I follow shows about AI, software, and tech, and this year I'm especially focused on one-on-one deep interviews with industry leaders. Last year we saw a lot of founders start appearing on podcasts — teams from Cursor, Devin, Anthropic, Notebook LM and others all made appearances. I think you can extract a lot of valuable information from these one-on-one deep conversations.

👦🏼 Can Zhang

As I mentioned earlier, I also consume a lot of Newsletters and Podcasts, but what's somewhat unusual about me is that I'm passionate about hands-on experimentation. By actually building things, I get a more intuitive grasp of what's possible, and I understand what can be done and to what degree.

🚥 Koji

So what's the most interesting thing you've built recently? Can you share?

👦🏼 Can Zhang

I recently tried rebuilding Notebook LM. Before that, I had an interesting experience: the weekend before 4o launched, I built a local "speech-to-LLM" system. It could convert speech to text, process it through a model, then output voice via TTS.

At the time I wrote an internal post arguing that only by running locally with extremely low latency could you achieve natural interaction patterns like interrupting mid-conversation. We believed this direction had enormous potential, and a week later 4o launched.

🚥 Koji

Yes, and now the team provides APIs, and costs have come way down.

👦🏼 Can Zhang

Exactly, which is why this space attracts me. I believe next year and beyond, the entire multimodal space — especially voice modality — will see a lot of innovation, because the technology has opened that door.

🚥 Koji

Final question: looking ahead to 2025, what do you think will definitely happen in AI applications? And what remains uncertain but you're excited to see?

👦🏻 Gaoning

I'll go first, and I welcome being proven wrong. On the more certain side, I hope to see a Personal Agent that proactively pays attention to what I care about, letting me converse with it anytime. Its role would evolve from pure productivity assistant toward something that provides companionship and emotional value in daily life.

As Yusen once mentioned, can we move from being a mere "pen pal" — where you feed it content to revise — toward something like Devin, that can think and act autonomously, only coming to you when it encounters something it doesn't understand. That's the direction I'm more excited to see.

On model capabilities, whether search, reasoning, or influence, we're seeing improvements. What people notice more are breakthroughs in scientific domains like physics and math. But I'm somewhat concerned that in emotional intelligence, companionship, and personality traits, there hasn't been much academic progress. Frankly, any product today is still quite far from truly intelligent conversation. I'm unsure whether this is an engineering problem or requires more research and capability advancement. Though perhaps that's also a good startup opportunity.

🚥 Koji

Can Zhang, what are your hopes for 2025? What's certain, what's uncertain?

👦🏼 Can Zhang

On the certain side, as I said, I'm quite confident that multimodal — especially voice modality — will see more concrete落地 [commercialization].

By more concrete落地, I mean it can be commercialized, and there may be various carriers, both software and hardware — I think that's one direction. Of course it's not just voice; other modalities too, these end-to-end models will open new possibilities.

On the relatively uncertain side — actually extremely uncertain — is that up to now, virtually all AI opportunities we've seen are still product-type or tool-type. But will there be new platform-type opportunities? Because each of those would be enormous. Another uncertainty is what role Chinese entrepreneurs can play in all this?

🚥 Koji

Right, I think that final point is quite interesting. Everything we've seen, everything we've discussed today, has been tool-type. Whether platform-type opportunities emerge — that's worth thinking about and watching closely.

Alright, let's wrap here for today. Thank you both for your time.

We also welcome everyone to follow "Crossing." We periodically take stock of overseas AI products and companies that are quietly making serious money, hoping to provide more inspiration. Happy New Year, and enjoy the holiday 🧧

👦🏻 Gaoning & Can Zhang

Happy New Year 🧧🧧🧧

Subscribe to the "Crossing" Podcast

🚦 We track the industry transformations and new entrepreneurial opportunities brought by the new wave of AI technology. "Crossing" is Steve Jobs's metaphor for Apple — standing at the intersection of technology and liberal arts, where great products are born. AI is changing every industry, and we seek out, interview, and gather the "active doers" of the AI era. Together with them, we explore and embrace new changes, 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 reach out, exchange ideas, and connect on what's next. Koji on Jike[22], Koji's website[23]

👧🏻 Host Ronghui: Works at a tech VC, former Silicon Valley correspondent for CBNweekly. Ronghui on Jike[24]