The rise of AI agents over the past year is beginning to challenge that model. Some tasks can now be handed over entirely to AI, and this will only become more common. For businesses, AI is no longer just a tool. It is increasingly becoming a new form of labor.

Customer service is one of the clearest examples. In the past, companies bought customer service software to make their employees more productive. But once AI can take over part of the work itself, companies naturally begin to evaluate its value against what they would otherwise have spent on human labor.

We are beginning to see the same shift in professional services such as legal and accounting. This is why I believe the market opportunity for AI is so large: it is no longer limited to the software market; it is beginning to expand into the labor market as well.

This is also changing how some investors think about the value of AI companies. Instead of asking only how much businesses spend on software each year, they are increasingly asking how much work currently performed by people—and paid for through salaries—could eventually be handled by AI. That dramatically expands the potential economic value these companies can capture.

But once AI becomes a form of labor, companies will also need to think differently about how they allocate it.

When a new task comes up, should a company hire another person, assign it to AI, or have humans and AI work together? Companies already allocate work among employees based on skills and experience: senior employees handle tasks that require judgment, while more straightforward work is often assigned to junior staff.

I believe AI will evolve toward a similar division of labor.

A simple analogy is transportation. If you are only driving to a nearby store, you do not need to take a Lamborghini; a Toyota will get you there just fine. In the same way, companies do not need to use the most powerful—and most expensive—AI for every task. Some jobs require sophisticated reasoning, while others only require basic execution.

One of the most important capabilities companies will need to develop, therefore, is the ability to determine exactly how much intelligence a given task requires—and how much it makes sense to pay for it.

This shift will also affect how software companies charge for their products.

SaaS has traditionally been priced by the number of users or seats. But if a job that once required 50 people can now be done by 10 people working alongside AI, seat count becomes a poor measure of the value a product creates.

As companies increasingly pay not for a “tool,” but for the completion of a task, the unit of pricing will naturally change as well. We are likely to see more usage-based pricing tied to the amount of work completed, while services with clearly measurable outcomes may increasingly charge based on results.

At the current pace of development, AI may soon become embedded across the workplace as a new kind of “machine labor.” And that will raise questions that would have sounded unusual only a few years ago.

When a company hires an employee, for example, it pays not only a salary but also taxes and other employment-related costs. If the same work is instead performed by AI, the cost structure looks very different.

Some have even begun to suggest that AI usage could eventually be taxed, with the revenue used to reduce the tax burden on human workers. Whether such an idea would ever be practical remains an open question. But the fact that we are already debating whether AI “labor” should be taxed says something important about how its role is changing.

Another question companies will have to confront very soon is what kinds of people they should hire in the first place.

As AI takes over more foundational tasks, both headcount and job design will change—and so will the skills companies look for when they recruit.

The next challenge for companies, then, will not simply be learning how to use AI. It will be learning how to manage an entirely new kind of workforce.

My generation grew up hearing plenty of advice about which degrees would lead to a bright future and which careers held the most promise. Everyone around us seemed to have an answer. Most of that advice was well intentioned. But as we followed those paths, how many of our choices reflected what we actually wanted?

I recently watched a video about self-awareness. It made the point that some things come so naturally to us that we don’t even think of them as abilities. We may even assume everyone can do them. Only when different people keep telling us, “You’re really good at this!” do we begin to recognize our own strengths. I thought this was a valuable reminder, because we are often so focused on what we lack that we overlook what we already do well.

What excites me most is that advances in technology are giving people more opportunities to develop those strengths. In the past, an idea might never have gone beyond a passing thought because you couldn’t code or couldn’t find someone to work with. Today, with so many tools available, there is much more you can try on your own. You no longer have to wait until everything is in place to get started.

And creating even something small can help you figure out whether you actually enjoy doing it. Some abilities only reveal themselves once you try: “Oh, I can do this!” Often, that is how we gradually find our direction.

So I told these young people that the future would certainly bring plenty of change, along with things to worry about. But it would be a shame if all we talked about was which jobs would be replaced. We have so many tools at our disposal and more opportunities to turn ideas that once existed only in our heads into something real. I’m excited to see what new possibilities emerge when they discover what they truly want to do.

I can’t tell them which degree will guarantee a promising future. But to be able to start thinking about who they want to become at such a young age, and to have the tools to explore those possibilities for themselves, seems incredibly fortunate to me. That is why I feel even more confident about the future of the next generation.

Today I’m happy to announce that Cherubic Ventures has closed its sixth fund at $68.88 million.  The fund size reflects the auspicious meaning of the number eight in East Asian cultures, where it is traditionally associated with prosperity and good fortune. With Fund VI in place, total assets under management across our six funds have passed $500 million.

Fund VI means more to me than the amount raised. Ten years in, I’m even more certain about why I love early-stage investing. My greatest satisfaction comes from working alongside exceptional founders as they find their way through uncertainty and ultimately change an industry.

Before the world believes

Companies that genuinely change the world rarely start with good numbers, a finished product, or market consensus. Sometimes it’s a few people and an idea that sounds too early, too hard, or a little crazy.

The best founders, I think, are often the ones who notice a shift before most people do, and see an opportunity others haven’t recognized yet. They use new technology and new approaches to turn that view of the future into something real.

Industries and technologies have kept changing since Cherubic started in 2015, but I’ve always been looking for the same kind of founder: someone who sees new possibilities in change and has the conviction to make them real.

That’s why we remain committed to investing early: to be there with founders while the product is still unproven and the direction is unclear, working through trial and error together until we find the way forward.

Ten years alongside founders, finding the way through uncertainty

Ten years ago, when I started investing in early-stage startups globally from Taipei, no one knew whether a venture capital firm from Taiwan could reach the center of Silicon Valley, or whether the solo GP model could hold up across different market cycles.

I built Cherubic as a solo GP so I could judge independently, decide quickly, and be there early when I truly believe in a founder. Over ten years, that has shaped how I invest:

Judge independently. Don’t wait for signals.
As a solo GP, I don’t wait for an investment committee or for the market to move. When I believe in a founder, I can decide fast.

Start with the founder.
What I spend the most time understanding is how a founder sees the change underway, what they want to redefine, and why they’re the one with a shot at doing it.

From the first check.
I want to be there when a founder most needs someone to believe in them, and to stay with them from the earliest stage onward.

Stay through the uncertainty.
Early-stage investing happens before anything is settled. From the first check, I want to stay close to the founder and keep adjusting as we learn.

Ten years on, Cherubic has backed more than 200 companies worldwide. I was lucky to meet Hims & Hers, Flexport, Calm, Paidy, 91APP and Astranis while they were still early. Many of the startups I’ve backed have since gone public, become unicorns, or been acquired, and have built global influence.. Back then, all of them were just getting started, with their futures still full of uncertainty.

AI is reshaping industries, and the next generation of great companies is already forming

We’re at another major technological turning point. AI will be like the internet, smartphones, or cloud computing, becoming an underlying platform that reshapes industries. It will make existing companies more efficient and make possible a new set of companies that were previously hard to imagine.

Software is starting to do work on its own. Machines are starting to understand the physical world. Healthcare is moving toward earlier predictions. A company that once took hundreds of people to build may one day be built by a team of just over ten, and the problems founders take on will change with it.

Through this technological shift, I continue to look for founders who are redefining markets. Since 2024, the AI-native companies I’ve backed span infrastructure, developer tools, enterprise software, healthcare, physical AI and robotics, exploring what AI makes possible.

Fund VI is still early, but we’re already seeing some promising signs. The portfolio has attracted more than $500 million in follow-on funding. The companies I’ve backed include Entire, the developer platform founded by former GitHub CEO Thomas Dohmke; the robotics company Sudo AI; the AI patent technology platform Patlytics; and Max AI, Generation Lab and therapiAI, which are using AI to rework healthcare workflows and drug development.

Entire raised one of the largest seed rounds ever in developer tools. Sudo AI, founded about two years ago, is now valued at close to $2 billion. Its sudo R1 robotic system is trained through virtual simulation and can reliably handle objects it has never encountered without relying on real-world manipulation data.

Conviction before consensus

The past ten years have made me even more certain that believing in a founder before the answers are clear, and staying with them, is an important part of early-stage investing. To me, that’s conviction before consensus. It’s what Cherubic has done for ten years, and what we’ll continue with Fund VI.

Over that time, we’ve built a wider global network of founders, and some of my most important investments have grown out of relationships built over years.

Fund VI is the start of Cherubic’s second decade. Two things won’t change in the decade ahead: stay early, and keep building the future we want to see alongside exceptional founders.

The next companies to truly change the world may not have taken shape yet. If you’re working to make the future you believe in real, I hope I get the chance to build it with you.

Investors across all six funds include leading global institutional investors and foundations, as well as publicly listed companies, family offices, successful entrepreneurs and high-net-worth individuals.

Fund VI maintains the firm’s early-stage focus, investing in AI-native companies across infrastructure, developer tools, enterprise software, healthcare, physical AI and robotics. Sudo AI, a robotics startup in the portfolio, has reached a valuation of nearly $2 billion two years after its founding, joining the ranks of unicorns.

“After ten years, I am more certain than ever about why I chose to invest at the earliest stages,” said Matt Cheng, Founder & Solo GP of Cherubic Ventures. “Working alongside exceptional founders, finding a path through uncertainty, and ultimately changing an industry is what keeps driving me.”

Investing Across AI, From Infrastructure to Industry Applications

As AI reshapes industries, Cherubic Ventures continues to look for founders using the technology to build new products and redefine markets. Since 2024, the firm’s AI-native investments have spanned infrastructure, developer tools, enterprise software, healthcare, physical AI and robotics.

In robotics, Sudo AI was co-founded by Hao Su, a leading researcher in embodied AI and 3D vision and co-author of PointNet, and serial entrepreneur Robin Han. Its sudo R1 robotic system is trained through virtual simulation and can reliably handle objects it has never encountered without relying on real-world manipulation data. This addresses a key bottleneck to deploying robotics at scale. Cherubic Ventures was its earliest institutional investor.

Cherubic Ventures is also an early investor in Entire, the developer platform founded by former GitHub CEO Thomas Dohmke. The company raised US$60 million earlier this year, the largest seed round ever for a developer tools startup.

While Fund VI is still at an early stage, its portfolio companies have already raised more than $500 million in subsequent funding. Other notable investments include AI-powered patent technology platform Patlytics, along with healthcare and drug development companies Max AI, Generation Lab and therapiAI.

A Decade Alongside Founders, Supporting the Next Generation

Founded in 2015, Cherubic Ventures was among the first venture firms in the world to adopt the solo GP model. It has invested in more than 200 companies globally, with early investments including Hims & Hers, Flexport, Calm, Paidy, 91APP and Astranis. 

Across its portfolio, Cherubic Ventures has been the earliest institutional investors in dozens of companies that went on to become unicorns. Hims & Hers is listed on the New York Stock Exchange and 91APP on the Taipei Exchange, while Paidy was acquired by PayPal for US$2.7 billion.

Fund VI marks the beginning of Cherubic Ventures’ second decade. “The past ten years have made me more certain that believing in founders before the answers are clear, and backing them through uncertainty, is at the heart of early-stage investing,” Cheng said. “In the next decade, we will continue to ‘Stay Early’ and work with the most exceptional founders to build the future we want to see.”

The song was “come again”, released in 2001 by Japanese hip-hop group m-flo. It brought the 2-step sound that was sweeping the UK at the time into Japan’s pop scene and became one of the defining J-pop anthems of the early 2000s. Then, like so many hits before it, the song gradually faded into karaoke catalogs and nostalgia playlists as the years went by.

That changed in 2024, when m-flo performed the song with Sho Sakurai on the annual music special Best Artist 2024. The following year, a dance practice video by girl group IS:SUE reignited interest in the track, pushing the original “come again” music video past 50 million views. A 24-year-old “old” song had found a new generation of listeners who had not even been born when it was released.

The release of those 127 remixes marked the culmination of the song’s remarkable comeback. The competition was jointly organized by m-flo and a Japanese startup founded less than five years earlier: SURF Music.

150 Unreleased Songs Gave Rise to SURF Music150 Unheard Songs Gave Rise to SURF Music

SURF Music is a digital marketplace founded by veteran Japanese musician Ken Kobori  that focuses primarily on unreleased music. Creators upload tracks that have yet to find a home, while record labels, A&R teams (the departments responsible for discovering artists and selecting songs), and film and television production companies use SURF Music to scout promising material.

Kobori has spent more than 20 years in Japan’s music industry. He is best known for “Story,” a 2005 song performed by singer AI. The song remained in the top 10 of the Oricon chart for 73 consecutive weeks. Nine years later, Disney chose it as the ending theme for the Japanese version of Big Hero 6, cementing its status as a classic familiar to audiences across Japan.

But being able to write a song like that does not mean you know how to market yourself. After his long-term management contract with Dentsu ended, Kobori went independent. “For a while, I didn’t know how to meet new artists or people at record labels, and I’m really bad at promoting myself,” Kobori admitted. “I felt a little lost.”

To solve the problem, he built a simple website, uploaded around 150 of his unreleased songs, and gave several record labels and management companies access to them. Three months later, three of those songs were picked up—and all three had been written 15 to 18 years earlier.

“Those songs sounded terrible to me. There’s no way I would let anyone hear them now,” Kobori said with a laugh. “But then I thought, if I have more than a hundred songs sitting around, producers all over the world must have tons of tracks gathering dust on their hard drives. If there were a platform where everyone could upload their music and record labels could choose from that catalog, it could help creators all over the world.”

The website he built for himself eventually grew into a company. SURF Music officially launched in 2023 and now has more than 40,000 creators from 195 countries and over 450 B2B clients. Major Japanese record labels have joined the platform, while music discovered through SURF has been featured in projects by Pokémon and Disney. SURF Music has also helped nine songs reach the top five on charts in Japan and South Korea.

In 2023, SURF Music was named a finalist for the SXSW Innovation Awards, and in 2024, it was recognized by Fast Company as one of the world’s most innovative companies. In the first half of this year, the company also completed a $7.9 million pre-Series A funding round. 

Apart from Cherubic Ventures, nearly all of its new investors are major players in the music industry. They include YG Plus, a subsidiary of South Korean entertainment giant YG Entertainment; leading Japanese music production company agehasprings; and companies such as Enigmo, Rainbow Entertainment, and Hip Land Music.

Before SURF, Who Heard Your Music Depended on Who You Knew

To understand the value of SURF Music, you first have to look at what it is trying to replace.

In Japan’s music industry, Kobori explained, finding a great song often starts with an email list. Every A&R executive at a record label has their own private “email list,” a network of contacts built up over the years. When an artist needs new material, the A&R executive sends a brief to everyone on that list, then listens to the submissions one by one and selects the most suitable tracks. A call for songs for a major artist can attract hundreds of submissions at a time.

“Even if you listen to only 15 seconds of each song, getting through them all is still grueling,” Kobori said he spent about a year working in A&R at Universal Music, and that entire year was spent listening to demo after demo.

And this system creates two problems at once: veteran A&R executives have inboxes constantly flooded with demos, many of which do not match what they are looking for, while junior A&R executives without industry connections have no list at all and must rely on senior colleagues for introductions, slowly building their networks one contact at a time.

That email list is exactly what SURF Music aims to replace. New creators from around the world contribute music to SURF’s catalog every day. This means junior A&R executives no longer have to wait for senior colleagues to make introductions, and record labels in Tokyo no longer need intermediaries to connect with producers in Los Angeles or Seoul.

Using AI to Solve the Search Problem

But as the catalog grows, another problem emerges: with tens of thousands of tracks to choose from, how do you find the one that fits best? Kobori explained, “Once users log in, they can use SURF Music’s AI search to find a specific kind of track. They can simply type in a sentence like ‘a lonely girl has just broken up with her boyfriend and is walking home alone,’ and the system will find songs that sound exactly like that.”

Three songs for Disney+ productions have been selected this way: a director simply copied and pasted the script for a scene that needed music and had SURF Music recommend suitable tracks.

Building a “Google Workspace” for Cross-Border Music Collaboration

Kobori himself is one of the most active creators on the platform.SURF Music analyzes the work uploaded by each creator and connects people around the world who make similar kinds of music. That is how Kobori found the collaborators he works with today.

“Actually, I really enjoy working with Korean music creators, but before this, it simply wasn’t possible,” Kobori said. “We didn’t speak the same language, so we needed an interpreter.”

That challenge led to the creation of SURF Sessions. Creators can start a session and bring together collaborators for arranging, songwriting, and other roles. Everyone in the chat chooses their preferred language, so Kobori can type in English while a producer in Seoul instantly sees the messages in Korean. All the files are also kept together in a cloud drive dedicated to that song. “We used to send project updates back and forth by email,” Kobori said. “So we were constantly asking, ‘Where did that folder go?’ Now everything is centralized, and everyone knows exactly where the project stands.”

“They Said I Was Living in La-La Land”

These features may seem like obvious ideas today, but when Kobori first took the concept to friends in the industry, they all gave him the same answer: it would never work. Everyone who said so worked in the music industry. It was precisely because they understood how the business operated that they were so certain.

“But I’ve been like this since I was a kid,” Kobori said. “When someone tells me something can’t be done, it only makes me want to do it more.”

Their skepticism came down to two questions. First, why would top creators trust an unfamiliar platform with their unreleased music? Creators are fiercely protective of their work. As Kobori put it, “A song you create is like your child.” Second, why would record labels pay to use the platform? They could already get songs from the best producers.

The first issue was the classic chicken-and-egg problem faced by every two-sided marketplace: without a strong catalog, record labels would not come; without record labels, creators would not upload their work. SURF Music chose to build up the supply side first, and Kobori took the most hands-on approach possible: reaching out to creators one by one. Drawing on the network he had built over the years, he contacted producers individually. “I probably had close to a thousand Zoom meetings with different producers,” he said. Eventually, he persuaded them to upload their unreleased work.

After the platform attracted around 1,000 creators—including Grammy winners and producers behind Billboard No. 1 hits—Kobori no longer had to persuade every new creator personally. As the catalog expanded and record labels began signing on, creators’ attitudes shifted: if those companies were using SURF, the platform had to be legitimate.

Giving Up Commissions Earned Industry Trust

The second issue nearly backed the company into a corner and taught Kobori an important lesson: in a mature industry, how others categorize you can sometimes matter more than what you actually do.

SURF Music’s original business model seemed perfectly reasonable: take a 10% cut of royalties when a song was selected, a standard approach for many marketplaces. But in Japan’s music industry, that commission sent a clear signal. “The record labels said, ‘If you take a cut of royalties and match songs with opportunities, then you’re a publisher—you just don’t call yourselves one.’”

In Japan, a music publisher signs creators, places their songs with artists or labels, and takes a share of the royalties in return. Once SURF Music was categorized as a publisher, it was no longer seen as a neutral tool, but as another party taking a cut from the royalty chain. Record labels would have to pay an additional share, while top creators who already had publishers would see yet another slice taken from their royalties.

Kobori therefore scrapped the original business model and switched to a pure subscription model. Once the positioning changed, so did the record labels’ attitudes—and doors began opening one after another. “In the end, they saw the value we were bringing and got behind our mission.”

One of the people who had initially told him it was impossible was working at Sony at the time. Four years later, Kobori showed him what he had built. His response was: “Oh my God, you actually did it.” Today, that person works at SURF Music too.

Bringing Everyone Who Makes Music Together on SURF

Looking further ahead, Kobori’s vision for the next five years goes far beyond a “music marketplace.” He wants everyone in the industry who works with music to have a SURF account, with the platform connecting them to one another. “That’s our core goal: to connect everyone so people can create great music together.”

What makes it all worthwhile for him is watching creators on the platform go from relative obscurity to landing opportunities to work with major artists, with many of them now even launching companies of their own. “Watching them grow like that is the most rewarding part for me.”

In the end, the platform Kobori originally built for himself really did solve his own problem. Now 46, he sets aside just one day a week for music. On that day, he comes up with five song ideas, then hands them off to more than a dozen producers he regularly collaborates with through the platform to develop into finished songs. At his most productive, he has completed as many as eight songs in a single week.

“I’m probably making more music now than I ever did before,” he said.

“Sovereignty” may be the most valuable thing in the AI era — and the most easily surrendered. Figma at least owns its technology and its customers; it took a punch, but it can still punch back. The ones really losing sleep are the companies built entirely on top of someone else’s model. For them the model is the lifeline: one price hike, or one lookalike product, and they may have to shut the doors.

From these conversations, I’ve come to see that this kind of deep entanglement happens one step at a time. First you adopt the platform’s model. Then you wire in your data and your workflows, layer by layer. Eventually the product experience itself has grown roots in it, and pulling free means digging up the soil along with them. In the beginning, AI feels like a pair of wings — everything moves faster. But once you’re bound tightly enough, the day you want to switch, you discover it’s no longer a matter of changing a line of code. In the worst case, you have to tear down half the product. Those wings, without anyone noticing, have become handcuffs.

Lately, people have started taking this seriously. Some have moved to open-source models and built stacks they fully control. Others are pulling their most important data back in-house, because they understand that whatever you hand over easily today may well be raising your competitor tomorrow.

Perhaps this is a collective awakening now underway in Silicon Valley: founders rethinking the balance between borrowed power and independence. Foundation models remain extraordinarily efficient leverage — but making sure that lever stays in your own hands is becoming something every founder has to settle before typing the first line of code.

The reason is that what a robot needs is completely different from what generative AI needs. Large language models can understand the world through text, images, and video from the internet, but what a robot has to learn is manipulation. Take the same clip of a person setting a cup down on a table: generative AI can recognize that “someone placed a cup on a tray,” whereas a robot needs to know from which direction the hand should approach, how much force to apply so the cup doesn’t slip, and how to tell when it has a secure grip. This kind of information doesn’t appear in the data on the internet, and it can’t be harvested in bulk by web crawlers.

Many people instinctively assume that if data is what’s lacking, the answer is simply to collect more. But a recent study in robot learning found that what truly limits a robot’s capability may not be the volume of data, but whether that data comes from a sufficiently wide range of different environments. The research found that within a single scene, model performance quickly hits a ceiling even as the number of training repetitions keeps rising. Conversely, simply increasing the diversity of objects, environments, and manipulation scenarios improves generalization far faster—even without a large increase in the amount of data.

This result isn’t hard to understand. Even for the same act of picking up a cup, today it might be a glass, tomorrow a stainless-steel one; today the tabletop is level, tomorrow someone might have just bumped the table. Humans live amid these subtle variations every day, so we never realize just how difficult they are for a robot.

For this reason, many startups are working out how to expose robots to a greater variety of situations. The most direct method is to record human operation in full. Quite a few companies make heavy use of teleoperation, having humans control the robot while every joint angle, force profile, and step of the process is saved—the hope being that the more different ways of operating a robot has seen, the better it can adapt to the real world.

Another path is to let the robot accumulate experience on its own. NVIDIA recently unveiled a learning system called ASPIRE, which analyzes the cause after each failed task, revises its control program, and then saves the successful method into a “skill library.” The next time it encounters a similar situation, the robot can correct itself.

A third approach is more imaginative still: training to the extreme directly inside a high-fidelity virtual world. Sudo AI, a U.S. startup backed by Cherubic Ventures, takes a pure-simulation route, having robots repeatedly face countless combinations of materials, lighting, backgrounds, and placements in a virtual environment before being deployed straight into reality.

At this June’s CVPR conference in the United States, this robot took on random challenges from the audience on the spot. Faced with transparent objects, metals, soft materials, and even tiny pills, it could pick up nearly all of them with composure. The seemingly improvised challenges from the audience had, in fact, already been rehearsed countless times in the simulated world. This year SUDO AI closed a new funding round at a valuation of over $2 billion—a bet on exactly this path.

Over the past few years, the success of large language models led everyone to believe that as long as there’s enough data, AI will keep improving. But robotics has developed in the opposite direction: the thorniest feature of the real world is its unpredictable variability. For a robot, repeating a perfect motion ten thousand times in a perfect laboratory brings no real-world application; only by experiencing ten thousand different challenges in an environment full of randomness—finding order within chaos—will the real solution be found.

This is what a prediction market is: it turns future events that have not yet happened into tradable contracts. Over the past two years, prediction markets have grown from a niche experiment in the crypto world into one of the few areas of finance that is still accelerating. Yoso, built by James Shen and his team, is one of the new players in this space.

What is a “Prediction Market”?

Will the Federal Reserve raise interest rates at its next meeting? Will evidence of extraterrestrial life emerge this year? Will Taylor Swift get engaged before the end of the year?

These seemingly unrelated questions are all turned into propositions that people can trade with money on “prediction market” platforms such as Kalshi and Polymarket, where you can bet on whether events will or will not happen. If you believe that an event will happen, you buy in; if it comes true, the contract is redeemed at full value. If you are wrong, it’s worthless. As long as there is someone willing to buy your position, you can sell and exit at any time. Thus, for the first time, the “probability of an event” has a price that changes in real time. The more people bet on a certain outcome, the higher the price becomes. This is why, during elections, prediction market prices are regarded by financial media as a more sensitive indicator than opinion polls.

At this point, most people’s first reaction is probably: Isn’t this just sports betting? In fact, the key difference lies in “who you are betting against.” When you buy lottery tickets, you are betting against the bookmaker, who is on the opposite side of the trade; if you win, they lose. But the prediction market is more like an exchange, in which your contract is traded with “someone who holds the opposite view.” The platform itself does not participate; it earns revenue from trading volume.

Because it resembles an exchange, the price can fluctuate at any time, and positions can be sold at any time, without having to wait for the end of the session or for the final vote count. This status is also reflected in regulation. In the United States, these contracts have an official name, “event contracts.” They are treated as a type of financial derivative and regulated by the Commodity Futures Trading Commission (CFTC) rather than by state gambling laws.

The 2024 US Presidential Election Lit the Fuse, and Other Categories Followed

Once the regulatory green light appeared, traffic poured in. During the 2024 US presidential election, the single market asking “Who will be elected president?” generated over US$3.6 billion in trade volume for Polymarket. The market was initially worried that interest would cool down after the election, but the enthusiasm quickly found new outlets. Sports supported Kalshi, while politics and crypto were Polymarket’s main focus. In addition, there was a constant stream of topics such as the Fed’s interest rate hikes, corporate earnings, the weather, Netflix rankings, and celebrity activities, as there was always another event to predict.

By 2025, Polymarket and Kalshi had become the two giants of this space. The former started with blockchain and uses stablecoins for settlement, while the latter is a regulated exchange licensed by the US CFTC. Market capital has also made heavy bets. Intercontinental Exchange (ICE), the parent company of the New York Stock Exchange, has invested a total of US$2 billion in Polymarket, while Kalshi completed a new funding round at a valuation of approximately US$22 billion.

Not Building the Next Polymarket! Yoso Enters a Market the Giants Have Overlooked

But however large the two giants may be, they cannot cover every topic. When everyone was asking, “Besides these two, who else has a chance?” James Shen’s answer was a completely different map.

The focus of this map is not US elections, nor is it the NFL or the NBA. These worldwide trending topics are also available on Yoso’s platform, but only as side dishes. The real home field for James and the Yoso team is an area the two giants have barely taken seriously: cricket.

The move into cricket was not a spur-of-the-moment idea, but the result of a series of deductions. “We believe that beyond Polymarket and Kalshi, unmet demand must still exist,” James said. “And that demand is for things that local people care about, not for global events.”

Following this line of thought, several things came together. First, cricket is the third most popular sport in the world, behind only football and basketball in terms of audience size and commercial value. The Indian Professional League (IPL) is the world’s largest cricket tournament, with hundreds of millions of people in one country obsessed with it. Second, Indians are as enthusiastic about prediction as anyone else. Of course, the two giants also list major events such as the IPL, but their focus has always been on American sports and politics. This overlooked demand is where the opportunity lies.

So, the Yoso team first released a batch of questions to test the waters. What truly surprised James was that demand began to grow on its own. Cricket is a common language across the Commonwealth, and once the questions went live, users from Australia, New Zealand, South Africa, Bangladesh, and Pakistan emerged one after another.

Cricket is just the entry point. From there, Yoso has developed an entire localized question bank, including kabaddi, a local Indian sport, followed by India’s version of real world assets. Tradable topics include local blue-chip stocks, the NSE Nifty (known as the “Indian S&P 500”), and oil prices. On top of that, Yoso adds another layer of culture and traffic-driven topics, including weather forecasts and celebrities’ posts and activity on social media.

What truly makes this strategy work are the two engines hidden beneath the question bank.

The first engine is an AI deep learning system. The most labor-intensive part of prediction markets is not just matching trades, but also determining “what questions to ask.” When a major event occurs, someone needs to break it down into a clearly defined question that can be objectively settled. Yoso entrusted this task to a group of AIs. One group of bots monitors competitors’ posts, sports events, and local media around the clock, suggesting which topics to create at the first sign of movement and capturing traffic before the two giants do. The other group is responsible for translating users’ imaginative “wishes” into clearly defined questions that can be settled. Finally, additional AI agents help automate market making and trade monitoring in order to meet traders’ needs. For a platform that seeks to win through being “local and real time,” this effectively turns reaction speed into a moat.

The second engine is a financial mechanism designed by James. When new users enter the platform, they receive an amount they can trade with directly, and they can only withdraw the portion they have earned after meeting certain conditions. The result is straightforward: the cost of acquiring a new user has been reduced to almost negligible levels, which is the foundation that allows Yoso to grow rapidly despite competition from the giants.

The results speak for themselves. The product, which was officially launched in February this year, has achieved an annualized trading volume of over US$100 million in just a few months with almost no marketing costs. It records about US$2 million in weekly trading volume and has accumulated around 50,000 registered addresses.

Lessons Learned from Multiple Startup Journeys

James Shen studied mechanical engineering in college, but what truly fascinated him was mathematics, psychology, and philosophy. “I am a person who is curious about many things, so I always want to search for the underlying logic of things,” James said. This curiosity led him into the world of blockchain. He is a serial entrepreneur, an active angel investor, and one of Binance’s earliest investors. Along the way, he has accumulated several judgments that have been tested repeatedly.

The first is about timing. When he was developing B2B products in the past, he believed that features had to be complete and fully polished before launch. When he shifted to Yoso, which targets the mass market, he completely overturned that approach. For a product that is racing against time in a competitive market, “timing” is the key to success.

The second is about users. He discovered that what developers think is important is often different from what users actually care about. The team once spent great effort making every number in the database precise and accurate, but users did not notice at all. What truly keeps users engaged are things that seem minor but directly affect the experience, such as whether deposits are smooth and whether order execution deviates from the price users see.

The third is about “doing what others are not doing.” He does not compete with the giants over who has the lowest fees. “When you have something that others do not, you get to set the price.” That is why he started with cricket and localized topics. Rather than cutting prices in a competitive market, it is better to differentiate yourself from others.

All of these judgments ultimately point to the same greater conviction. In James’s view, prediction markets will not be just a passing trend. The human desire to place bets will not disappear, and regulations in different countries will eventually catch up and define this market clearly. As the market matures, this place that turns the “future” into an asset will grow from a niche activity into a legitimate part of the financial world.

Parents are willing to spend almost anything on their children’s future. So why are education companies valued so little? In the first half of this year, nearly 80% of global venture funding flowed into AI. Edtech, by comparison, attracted only a tiny fraction of that amount, falling to its lowest level in a decade.

The more I thought about it, the stranger it seemed. Knowledge that once required expensive private tutors is now only a prompt away for anyone with an internet connection. By all accounts, education should be becoming a better business than ever. Instead, the opposite has happened.

After thinking about it for a while, I came to a simple conclusion: AI made knowledge accessible to everyone, but it didn’t make people want to learn. Technology can improve efficiency, lower the cost of learning, and make information almost effortless to access. What it can’t do—no matter how powerful it becomes—is create the desire to know.

Yet much of edtech is still built on the opposite assumption. Teams pour their energy into making better products, believing that if the tool is good enough, people will naturally want to learn. But is that really true? After my conversation with the founder, I realized that before AI, the biggest barrier to learning was not knowing. Today, the bigger barrier is not wanting to know.

Think back to the period when you studied the hardest. Chances are it was during your school years, and not because you loved learning. Most of us were driven by exams, class rankings, our parents’ expectations, or the desire to earn other people’s approval. Once those external pressures disappeared, many people stopped learning altogether. The habit didn’t disappear because knowledge became less valuable; it disappeared because the motivation was never really our own.

So where does that motivation come from? Just watch a child. They ask hundreds of questions every day without anyone forcing them to. They learn faster than anyone else because they genuinely care about the answers. That instinct never really disappears. When learning begins with a question you truly want answered, motivation has a way of taking care of itself.

AI has placed an entire gold mine of knowledge in front of us. Never before has so much information been available to so many people. But a gold mine is only valuable if someone is willing to dig. You can stand on extraordinary wealth and still walk away empty-handed. In that sense, it’s entirely possible to starve on a gold mine.

My hope is that we don’t just inherit this extraordinary mine of knowledge, but hold on to the one thing that gives it any value at all: the curiosity to keep learning.

As I read it, one question kept nagging at me: if judgment is humanity’s last moat, where does it actually come from?

A seasoned engineer can spot a flaw in a second because he has personally written tens of thousands of lines of code, building intuition through countless rounds of correction. A renowned doctor can identify an ailment at a glance because she has treated tens of thousands of patients. Judgment doesn’t appear out of thin air; it is an intuition cultivated slowly, layer upon layer, over years of accumulated knowledge and experience.

In other words, the “critical thinking” so revered in the age of AI rests on the foundation you already have. If that foundation isn’t solid, then the judgment we humans take such pride in may be little more than another kind of “hallucination.”

Yet the current pace of technological progress is hollowing out that very foundation.

Scientists running experiments have found that people who habitually let AI do their writing show the weakest neural connections—some can’t even recall what they supposedly wrote. Put another way, if we grow too accustomed to outsourcing the thinking work of reading and analysis to AI, before long we’ll lose even the ability to verify the answers. In a sense, we are gradually becoming the “new illiterates” of this high-tech age.

AI’s power is beyond question. But I often remind myself: the basics that look the most inefficient are precisely the ones with no shortcuts. Reading a long report cover to cover yourself, running the numbers by hand, thinking a logical gap through from end to end—these seemingly primitive, even foolish acts are exactly the “gym” where judgment is built. (By all means, outsource the busywork to AI—but the things that train your thinking, please keep for yourself.)

For the past few years, no matter how busy I’ve been, I’ve insisted on walking ten thousand steps a day. Faster transport is always within reach, and walking really is the slow choice—but it is precisely this “inefficient” habit that has earned me something money can’t buy: my health. Thinking works the same way. When answers are available at our fingertips, perhaps we should ask ourselves whether we are still willing to stop, and leave a little room for the slow craft of thinking for ourselves.