The Trump administration has used tariffs as a core diplomatic and economic weapon to drive manufacturing back to the United States. For companies, this will compress profit margins and erode cash flow step by step, an operational problem that is reflected in financial statements every day.

Under the impact of this geopolitical situation, many companies have begun to actively prepare for the battle and carefully consider what can be done when tariffs become a normal cost.

An overlooked fact: tariffs are not “irreversible” costs.

It is worth noting that through conversations with many companies, I have found an interesting phenomenon. Many people discuss tariffs with the assumption that they are a “pay it and forget it” cost. But the fact is that tariffs have never been irreversible under the US system.

The United States actually has a system that has existed for more than two centuries, called the Duty Drawback, which was originally designed to encourage exports and processing trade. In short, if a company paid tariffs at the time of import, and the goods are ultimately not sold in the US market but instead exported or transshipped to other countries after further processing, the company may be eligible to apply for a refund of the tariffs paid under the law.

What is often overlooked is that the scope of this system extends far beyond “processing for re-export.” In practice, merchandise that is returned and not resold, or legally destroyed in the United States due to inventory adjustments or quality issues, may be eligible for a tariff refund as long as it meets the relevant requirements.

The system has been revised and expanded over its 200-year history, and as the supply chain has become more complex and the regulations and circumstances have become more granular, tariff refunds have become an area that “only the experts understand.”

Why is it so difficult to carry out “tariff refunds”?

In fact, if you look deeper, you will find that the information required to apply for tariff refunds is often scattered in a variety of different systems and formats, from ERP and Excel to PDF invoices, customs declarations, and logistics records. Such information was not designed for tariff refund purposes, and organizing and comparing it is extremely time consuming.

As a result, it often takes more than half a year or even a year for a case to go from delivery to actual receipt of the tariff refund, and the relevant services are usually provided only to companies with sufficiently large tax refunds. In recent years, I have begun to observe a new shift. AI is beginning to be applied to highly regularized, data-intensive institutional problems. Tariff refunds are a clear example of this. Once dispersed, unstructured data can be converted into computable formats, and algorithms are then used to match rules with real-world contexts; processes that once required extensive human labor can be automated.

From Pax to Flexport, a trend toward AI-enabled tariff refunds is taking shape.

In the United States, several companies have begun to emerge that are trying to use technology to solve a problem that was previously considered “too complex.” Take Pax, a company specializing in customs refunds for businesses, for example. It designs its own algorithms and works with experienced domain experts to use AI to automatically process raw documents provided by companies. These include customs declarations, invoices, logistics records, and internal system outputs, all of which are converted into computable, structured data. Based on this foundation, algorithms are used to evaluate different tariff refund scenarios and identify the most advantageous combinations, after which professionals complete the compliance review and filing.

Under this model, processes that once took half a year or even a year can now be completed in just a few weeks. More importantly, because the calculations no longer rely solely on manual experience, many businesses are discovering for the first time that they are actually eligible for more tariff refunds than they had previously thought, and are often able to get back more than they had originally expected.

If Pax represents an innovative entry point for “specialized problem solving,” then Flexport proves the feasibility of this direction from another angle.

Flexport is a global digital freight forwarding giant that has grown over the past decade since its establishment. Flexport integrates sea, air, and land transportation through cloud-based software to make the complicated cross-border logistics transparent and digital, helping customers realize real-time cargo tracking, customs clearance, warehousing, and supply chain optimization on a single platform.

Over the past two years, Flexport has rolled out a series of AI-focused tools and systematically introduced AI into the customs refund process, aiming to overcome the fundamental computing limitations of traditional approaches.

For example, in the past, most tariff refund software relied on simple amount-based matching. In real-world import and export data, however, the number of possible matching combinations grows exponentially. In a public statement, Flexport noted that the volume of global import and export data is immense. Simply matching imports and exports can generate more possible combinations than grains of sand on Earth, a scale far beyond what human labor or traditional systems can handle.

To achieve this, Flexport has developed an algorithm that evaluates different import/export linkages and refund scenarios simultaneously in all possible combinations and automatically finds the combination that yields the highest refund, instead of being limited to a single, intuitive manual judgment.

Practice results show that with this type of AI-driven analysis, companies can actually obtain tariff refunds that are about 400% higher than by using the traditional method, and for the first time, processes that were previously considered too complex have the potential to operate at scale.

I believe that tariffs have brought the world back to a new era, and in this environment, there may be an opportunity to rethink cost management and cash flow strategies; companies should make better use of the power of AI to find resources that are rightfully theirs in a complex system.

For entrepreneurs, this is a pivotal realization. As access to powerful models becomes seamless, the question remains: what is your lasting advantage? A great idea used to provide a six-month head start. Now, the market can be filled with nearly identical products within weeks.

This is the new reality of the startup world. When everyone builds on the same foundational models, ideas are no longer the moat—because a competitor can replicate your work in a fraction of the time. The playing field has been completely leveled.

At this point, competition returns to the most fundamental element: people.

As an investor, I’ve noticed a consistent pattern: competitors can buy the same tools, the same models, and the same infrastructure. But the one thing they cannot buy is a founder’s judgment, their network, and the trust they have built over years.

AI has dramatically amplified productivity, but it has also sharpened the differences between individuals. That difference manifests first in judgment. I’m not referring to raw ideas—I’m talking about insight. AI can execute tasks with high efficiency, but it cannot yet identify where the true commercial opportunity lies. The instinct to pinpoint a genuine pain point has become the ultimate differentiator.

Equally critical are trust and distribution. In a market of increasing product parity, clients rarely choose you based on a superior model alone. More often, they choose you because of a long-term relationship that cannot be digitized. And as AI makes output easier to generate, the ability to get that output in front of the right stakeholders matters more than ever.

Third is speed of execution. In the pre-AI era, a slow-moving team might still have had the luxury of time. But when replication is nearly instant, winning is determined by whether you can reach the next frontier before others catch up. Adaptability is now the baseline, not the edge.

AI sets a higher bar for the ordinary, but it also amplifies excellence. If your only advantage is a “great idea,” this era will be challenging. But if your foundation is built on trust, judgment, and execution, AI will be the most powerful catalyst you’ve ever had.

In his dorm room at Michigan State University, David and two fellow students, Achyut Byanjankar and Soami Kapadia, decided to take matters into their own hands. Every weekend, they locked themselves in their room, brainstorming solutions in front of a whiteboard and their laptops. They even gave those sleepless, all-night sessions an ambitious name: “Dormathons.”

No one could have imagined that this dorm-room experiment—with virtually no budget—would amass over 2 million users within a year, including students as far away as Egypt, Syria, and India. By 2025, the trio had officially earned a spot in Silicon Valley’s top accelerator, Y Combinator. These three college students had truly turned their “impatience” into a business.

“We don’t want to waste any more time watching long lecture videos!”

When asked about the origin of the business, David admits that when they met in their freshman year, they simply wanted to solve their own problem. “At first, we just didn’t want to keep watching those outdated lecture videos, so we created a tool to help understand it quickly. But then we thought: if this feels frustrating to us, could other people be feeling the same way?”

Without any polished business plan, they decided to post their prototype on X, taking a “let’s see what happens” approach. Within a few days, tens of thousands of users were pouring in. For the first time, they realized that students all over the world were waiting for a tool that would enhance their learning efficiency.

Turning Learning Materials into “Interactive” Tutoring

So what, exactly, does YouLearn do? Simply put, it brings dense materials into something you can understand quickly.

In the past, when reading handouts or watching recordings, students had to painstakingly extract key points on their own. Now, students can simply upload a PDF, class recording, or YouTube link into YouLearn, and the platform transforms that material into an AI tutor that teaches you the topics. Users can ask questions anytime or generate quizzes and flashcards based on the content they just studied to check whether they truly understand it.

“We realized that students crave understanding—not just answers,” David explains. “So the core of our product design is to help people actually learn. Instead of just giving them the answers right away, we encourage active participation.”

How YouLearn Retained 2 Million Users through “De-Friction”

Looking back at the original version of the product, David admits that a lot of things simply didn’t work. Even the most basic functions like file uploads had issues. After multiple attempts, the team realized that it was time to return to a fundamental question: what do users really want?

The answer was simple: they want to start learning right away.

“We found that students didn’t want to spend time figuring out how to use a tool, so we focused on reducing friction,” David says. For example, “uploading materials should be simplified so it can be completed in just two steps, and uploaded content should be processed in real time so users can interact with the AI tutor as soon as they enter the platform.”

Additionally, the YouLearn team discovered that most learning tools are fragmented, requiring students to constantly switch between pages, drastically reducing efficiency. So they consolidated flashcards, notes, conversations, and quizzes into a single space. “Our goal was simple,” David says. “Minimize friction and make learning extremely easy.”

The second key lesson was to simplify their vision.

In the early days, the team experimented with bold ideas and explored multiple directions, but growth was limited. The real turning point came when they decided to stop chasing multiple paths and focus on one thing: observing how students actually used the product.

David admits that when the platform began its explosive growth, he and the team “could see the traffic, but we didn’t know what the users were doing.” So the team began to establish tracking metrics and analytics mechanisms to review user behavior, ensuring that every product iteration was backed by data.

The data revealed several key patterns: students repeatedly asked about the same concepts, spent hours completing quizzes, and even regularly listened to YouLearn-generated podcasts. These signals reinforced the team’s belief that students weren’t just looking for answers but seeking genuine learning opportunities.

However, insights alone weren’t enough. The sample would need to be larger if their hypothesis was to be validated. So the team began to consciously expand their visibility. Understanding that college students were more likely to trust creators than generic ads, they skipped traditional advertising and collaborated directly with micro-influencers on Instagram and TikTok.

(Image source: YouLearn Instagram )

“As students ourselves, we knew where people spent their time and what content kept them watching.” David explains that students were looking for “immediate value”—they wanted to know how they could learn a lesson in an hour or memorize content faster. So instead of talking about the brand’s vision, David and his team decided to demonstrate how the platform could enhance learning efficiency. “This deep understanding of the habits of our peers is what made YouLearn’s growth strategy truly effective.”

It’s a hit! Now what?

The first wave of growth came quickly, and by the end of the school year, YouLearn had about 200,000 active monthly users. Even more remarkably, these users were spread across the globe, with students from Egypt, Syria, India, and beyond proactively sharing about their experiences with the tool on social media.

But the real challenge for the team wasn’t traffic—it was retention.

David mentioned that while college students today try out various AI tools, they often don’t stick with any single one for long. The team started observing user behavior more closely, tracking whether students completed the full learning process or came back for a second session, and where they tended to get stuck. What encouraged them to keep going? “Retention rate” became the core metric guiding product iteration.

The data also revealed that learning needs varied significantly by discipline. For example, David notes that for medical students who need to memorize a lot of information, the system emphasizes flashcards and spaced repetition. Law students deal with dense texts and case analysis, so the platform focuses on breaking down arguments and clarifying relationships between premises and conclusions. In mathematical fields, the platform includes step-by-step examples for calculation problems, helping students verify formulas and models. YouLearn’s intention is simple: reduce the cost of organizing information so students can spend their time actually thinking.

As learning methods were refined for different majors, YouLearn’s retention curve began to steadily rise. Gradually, more users upgraded from the free version to the paid version for a more comprehensive learning experience.

With a solid product foundation in place, external validation followed. In 2025, YouLearn was accepted into Y Combinator and completed its seed funding round, officially stepping onto a bigger stage.

Building An AI Tutor For Everyone

YouLearn also grew out of David’s own learning experiences. Reflecting on his university years, he believes that the most valuable skill he gained was “learning how to learn.”

“I experimented with more efficient ways of learning, such as deliberately spacing out review sessions and forcing myself to actively recall instead of rereading notes,” he explains. When he changed his study approach, the speed at which he could absorb knowledge increased dramatically. “The key is to understand, not just memorize. And if you can apply what you learn, you’ll remember it much longer.” In high school, he skipped hundreds of hours of class to build the e-commerce brands he was starting—while sharpening how he learns so he could move through coursework faster instead of passively sitting through lectures.

This experience has profoundly influenced the direction of YouLearn’s products, and David believes that the future of education will place less emphasis on rote learning and more on critical thinking and comprehension.

As AI reshapes education, David’s vision for YouLearn is bold yet simple: make learning feel as engaging as the shortcuts people wish existed, without skipping the work that makes it stick. ‘The goal isn’t to make learning effortless,’ he says. ‘It’s to make it personalized, just outside your comfort zone, and genuinely fun, so you learn smarter and faster and actually retain it.’

Ultimately, David wants to build an AI tutor for everyone.

In the past, choosing a “stable” path usually guaranteed a secure future. But in today’s world, technology moves at breakneck speed. The “safe bets” of yesterday can be overturned in just a few short years.

I believe that instead of obsessing over the “most promising” path, we should ask ourselves: Is there something so captivating that you lose all track of time doing it—and love every second of it?

A friend recently shared a story about a middle schooler obsessed with Roblox and Discord. At first, he was just like any other teenager who loved gaming. But gradually, he began researching, watching tutorials, joining international communities, and connecting with players worldwide.

Soon, being just a player wasn’t enough. He taught himself to code and developed tools to solve real-world problems within his community, making communication more efficient. Recently, he even fell in love with Japanese, asking his teacher to ramp up the difficulty just so he could converse more fluently with people from different cultures.

Watching him evolve from a player to a self-learner, then a creator, and finally a cross-cultural communicator, it’s clear that no one forced him to learn. (And thankfully, no one labeled his passion as “pointless gaming.”) He was simply driven by genuine fascination. When he talks about his work, his eyes light up. That level of focus and excitement is something no exam score could ever replicate.

The true engine of personal growth is precisely this state of “having a spark in your eyes.” When you find something that truly captivates you, skills like adaptability, communication, technical prowess, and cultural empathy develop naturally. These competencies might not have a formal job title at first, but they are quietly accumulating, becoming the bedrock of your future.

No matter what stage of life you are in, ask yourself: What are you willing to immerse yourself in, day after day? What makes your eyes light up again? Give yourself the grace to explore; allow yourself to take the “less linear” path. May we all find that version of ourselves—the one with the spark in our eyes.

His words challenge a deeply ingrained belief: that following a time-tested formula will inevitably lead to a stable and prosperous life. Yet, as technological change accelerates, even the most prestigious skills are seeing their “shelf life” contract at an alarming rate.

While Musk highlighted how specialized fields like medicine could be upended, he left the most critical question unanswered: How should we respond?

This question resonates with the pivotal transitions in my own career. My journey has taken me from the life of a national-level tennis player to elite academic institutions, and eventually into the worlds of entrepreneurship and venture capital. From the outside, this path may look like a seamless collection of high-status labels. But behind those accolades was a constant need for profound identity shifts and mental recalibration.

In my view, regardless of your field, the most vital asset is not your title, but the courage to constantly adjust.

Tennis, entrepreneurship, and investing are vastly different arenas. Each transition forced me to dismantle and rebuild my mental frameworks from the ground up. My appetite for risk and innate curiosity repeatedly pulled me into unfamiliar territory, expanding both my resilience and my worldview.

Had I stayed on the tennis court, I might be a professional coach today—an honorable and stable career. But had I clung to that singular identity or refused to adapt to the unknown, I never would have survived the volatility of business, let alone entered the high-stakes world of investing, which demands constant macro-level judgment.

This journey led me to a realization: the most successful individuals possess an exceptional degree of cognitive agility. When the landscape shifts, they don’t linger in frustration. They acknowledge when their original assumptions are no longer valid and move swiftly to identify the next opportunity. They experiment boldly, fail fast, and course-correct even faster.

Elon Musk’s commentary is certainly thought-provoking. However, more than any specific career choice, it is this ability to shatter our own assumptions and lean into change that will serve as our compass through the technological storms ahead.

On April 2, U.S. president Donald Trump announced the launch of large-scale “reciprocal tariffs.” Once the news came out, corporate pricing, inventories, procurement rhythms, and market expansion plans were thrown into disarray. The U.S. and China quickly entered a cycle of retaliation and counter-retaliation, straining cross-border supply chains and plunging the entire world into unprecedented uncertainty.

Half a year later, countries were stuck at the negotiating table, while businesses were struggling to adjust their survival strategies under mounting cash flow pressure. In such an environment, any system that can reduce costs and improve cash flow flexibility has become more important than ever.

Precisely at this time, a mechanism that had been overlooked for more than 200 years was once again discussed: “duty drawback.”

“Many companies don’t even know that the tariffs they pay are refundable,” said Penny Chen, founder of Pax, a startup that uses AI to help companies reclaim taxes. In her interviews with clients, she sees the same thing over and over again: companies pay tens of billions of dollars in tariffs to the government every year, and only about 20 percent of that money is actually refunded. The remaining 80%—amounting to nearly $15 billion—just sits there unclaimed. “It’s free money left on the table!” she added with a sense of helplessness.

This huge disconnect unexpectedly became an entry point for Pax. With AI at its core, Pax uses algorithms to help companies identify 15% more refundable tariffs than traditional service providers, streamlining processes that would otherwise take more than six months to just over ten working days. For the first time, organizations are realizing that AI tools can transform tariff refunds from a burdensome process into an immediate source of cash flow.

Companies Don’t Have to Overpay Tariffs! But Who Is Eligible for a Refund?

A “duty drawback” is a refund administered by U.S. Customs and Border Protection (CBP) that allows businesses to recover import duties they have already paid; it is distinct from the more familiar income tax refund. Under U.S. law, any business that has paid duties on goods imported into the United States can recover part or all of the duties if the goods are subsequently re-exported, re-exported after processing, or destroyed within the United States, among other qualifying conditions.

The most common case comes from the manufacturing industry: if a company imports raw materials from overseas, processes them in the United States, and then exports the finished products, the tariffs previously paid can then be refunded in accordance with U.S. law. Another typical example comes from

Another typical example involves retailers and distributors: if a company imports merchandise from overseas and re-exports it without selling or using it in the United States, the duties previously paid can also be refunded under U.S. law.

Cross-border e-commerce companies and large retailers, which have grown rapidly in recent years, are also frequently eligible for tax refunds. Shipments moved from within the U.S. to overseas warehouses are considered exports. If a consumer returns shipped goods that are then destroyed in the U.S., duties on those goods can also be recovered.

In other words, many routine logistics operations, such as moving warehouses, returning goods, or destroying products—which on the surface may not seem directly related to revenue—can in fact be a source of substantial tax rebates. As long as companies can clearly track the flow of their goods, they can reclaim funds that already belong to them.

Complicated Processes and Outdated Tools: Why Traditional Tariff Refunds Are So Difficult

Although the system itself is not difficult to understand, the high complexity of implementation is the most challenging aspect for companies. Penny Chen explained that, in addition to different commodities and various import and export scenarios, a bigger headache is that the information required for tariff refunds is often scattered across PDFs, Excel files, and ERP systems. Companies are tasked with organizing data from piles of invoices, customs declarations, and logistic records—which use different formats and layouts—into the specific format required by government agencies, an extremely time-consuming process.

Due to this complexity, many companies choose to outsource to specialized service providers, but this does not necessarily make the problem any simpler. Traditional providers continue to rely on software that has been in use for more than 20 years, and the process still depends heavily on manual labor. Each case has to be manually entered and compared line by line, and a company applying for a refund for the first time may need to wait an entire year from the time it submits the relevant documents to when it actually receives the refund.

Because the process is so cumbersome and slow, many service providers are only willing to take on large clients with potential refunds of $100,000 or more per year. As a result, even though SMEs are eligible for tariff reimbursement, they are often unable to find anyone who is willing to handle their cases.

“This Is a Fun Math Problem!” How Pax Uses Algorithms to Help Companies Recover 20% More in Tariff Refunds

Chen’s earliest exposure to tariff refunds came when she worked as a researcher at Flexport. She quickly realized that the process involved extensive data cleansing and rule comparison. Through exchanges with industry experts, she sought to understand the market’s real pain points. “I found that everyone’s dilemma was almost exactly the same,” she said. “They were eligible to recover the tariffs they had already paid, but because they didn’t understand the system, they didn’t have the tools, or they couldn’t find providers who were willing to support them, they ended up with nothing.”

Penny Chen holds a Ph.D. from the Massachusetts Institute of Technology (MIT), where she specialized in algorithmic design. “From my point of view, tariff refunds are actually a very interesting mathematical problem—it’s just that no one has ever approached them algorithmically!”

Simply put, Pax is like TurboTax for corporate tariff refunds. TurboTax, the most widely used tax filing software in the United States, breaks down complicated rules into a standardized process that allows taxpayers to file returns with a single click. Pax aims to re-create the same experience for corporate tariff refunds, helping businesses reclaim tariffs without needing to understand the rules, organize the data, or waste time and money on a long, drawn-out process.

However, the first and most difficult step in achieving this is addressing a fundamental pain point: “fragmented data.”

Currently, service providers handling tariff refunds spend most of their time manually preparing data. To proceed to the next steps, companies often have to submit all necessary documents up front, requiring them to be organized in a uniform format. This stage alone demands a significant amount of time and manpower, and it’s the reason why most companies find their first experience with tariff refunds so frustrating.

Pax’s approach eliminates this “manual review” step. Companies no longer need to prepare any documents in advance; they simply provide the raw data to Pax, and the system automatically reads and extracts the relevant information, transforming unstructured data into structured, calculable formats, thus saving countless hours.

The next step is the algorithm. Chen and her team design their own algorithms that enable the system to identify which permutations will maximize tariff refunds. Many businesses, after submitting their cases to Pax for calculation, have been able to recover an additional 15% to 20% compared with the amounts previously determined through manual review.

The final step is actually submitting the information. After calculating the refundable amount using its algorithm, Pax has its in-house tax experts verify the results and then submits the documents directly to the government. Because Pax is authorized for U.S. submissions, the entire process eliminates the need for the traditional iterative submission process.
Under this model, processes that used to take six months to a year for companies can now often be completed in just 10–15 working days, resulting in a significant efficiency boost.

A Niche but Overlooked Market, Reimagined by AI

Tariff refunds have long been regarded as a fringe aspect of the trade system, but in reality, they are part of a system that has existed for over 200 years, involving the entire cross-border supply chain of import, processing, and export. Any company engaged in import and export activities may be eligible for refunds, meaning that the range of industries covered is far broader than generally imagined.

That’s why it’s a mature but fragmented market: even though more than a dozen service providers in the U.S. have been operating for years, the process is still highly dependent on manual labor, leaving a large volume of eligible refunds untouched for long periods—like forgotten “sleeping cash” waiting to be reawakened.

Pax tackled this problem by combining algorithms with experienced domain experts. Within just over a year of its founding, it was selected for Y Combinator, a leading Silicon Valley accelerator, received $4.5 million in early stage investment, and processed tariff refunds totaling around $10 million. Following Trump’s tariff announcement this year, demand from businesses surged, and Pax’s revenue tripled as a result.

With a team of fewer than 10 people at its founding, Pax achieved these results not only through the strength of its product but also thanks to favorable policy conditions. The evidence is clear: there is a huge and vastly undervalued market for “duty drawback,” and Pax is leading the way in addressing this gap.

Looking ahead, Penny Chen points out that the United States has adjusted the relevant laws and regulations numerous times. This system, which has existed since the founding of the nation, has been continuously revised over the past two centuries, becoming increasingly complicated as a result. “I think tariffs will only continue to increase,” Chen said. Amid supply chain restructuring and geopolitical tensions, it is difficult for companies to go back to the past, and the demand for tariff refunds is only likely to grow stronger.

In a rapidly changing world, “duty drawback” deserves to be better understood and more effectively utilized than ever before. Chen expects that through AI and automation, Pax can help businesses of all sizes transform what was once a cost burden into cash flow resilience, giving them greater control in the new economic landscape.

Looking back through history—from the steam engine to the computer, from the carriage to the automobile—every technological revolution has made humanity more efficient. But AI is different. For the first time, technology is directly replacing human thinking. As AI agents become widespread, even the act of “execution” will be automated. When technology shifts from being a “tool” to becoming a “competitor,” the speed and depth of this wave of impact will far exceed any previous industrial revolution.

Even more worrying is AI’s impact on the education system. For over a century, there has been a stable pathway from education to the workplace: graduating from school, taking an entry-level position, learning on the job, and gradually advancing through the ranks.

But the rise of AI is disrupting this pathway. Companies now prefer to buy a few more AI tools rather than invest time in training newcomers. The Federal Reserve Bank of New York warns that the U.S. unemployment rate for recent graduates climbed to 5.8%. Meanwhile, research by the Stanford Digital Economy Lab reports a sharp decline in employment among 22- to 25-year-olds, especially in software development, customer service, and clerical roles.

Yet what concerns me even more is this: where will those who are replaced go?

While most companies are busy using AI to cut costs and increase efficiency, another market with huge potential is emerging. McKinsey & Company predicts that by 2030, more than 400 million people worldwide will need retraining or career transitions. This means that for every person displaced by AI, another will need to return to the workforce. I believe this is not just a crisis—it could become the next trillion-dollar industry.

Amid this wave, countries around the world are taking action. Our government plans to train 200,000 AI professionals by 2028, building a workforce ready to meet industry demands. Japan has gone a step further: starting in 2024, it launched a five-year reskilling support program, investing a total of one trillion yen to help companies and workers relearn in the areas of AI application and digital transformation.

In the United States, startups are also entering this field. Inference.ai is developing an AI-driven, human-centered “employment infrastructure” designed to help displaced white-collar workers reenter the job market. The team began by focusing on high-demand positions such as machine learning—roles that have long faced talent shortages but have high entry barriers. The Inference.ai system functions like a “driving school for the AI era,” using AI to scan global job postings, break down required competencies, and build skill trees and personalized training maps.

Leveraging its proprietary GPU partitioning technology, Inference.ai enables thousands of participants to gain hands-on experience in real computing environments at low cost, guided by mentors from leading U.S. tech companies and AI-based coaching systems. Participants then use simulated question banks and AI interviewers to validate their skills and prepare for job applications.

Without any publicity, Inference.ai has already attracted more than 1,000 engineers and professionals to join its community, which continues to grow rapidly each week. This shows that “helping people become needed again” is emerging as a central theme of the new workplace.

The AI revolution is advancing quickly, and new forms of employment, education, and social order are already taking shape. To me, this is not merely a labor market crisis—it is a global experiment in how humanity can coexist with AI, a question that we must all take part in answering together.

In recent years, I’ve had many conversations with people from different generations. What’s striking is how much earlier—and how much more intensely—this sense of uncertainty is showing up. Many people have taken countless courses and earned every certification they could, yet still find themselves asking the same question: What does it actually mean to be ready?

These conversations have pushed me to rethink the learning paths we’ve long taken for granted. Traditionally, the sequence was clear: choose a major, spend years accumulating knowledge and skills, then enter the workforce and draw on what you’ve learned when real problems arise.

That model worked because industries evolved slowly and access to knowledge was expensive. If you didn’t prepare in advance, many doors simply remained closed.

But AI is fundamentally changing that assumption. Today, learning a new skill no longer requires years of upfront investment. As long as you have a sense of what you want to do, the relevant knowledge and tools can often be filled in later—with the help of AI. In that sense, knowledge itself is becoming inflated. Simply accumulating skills is no longer enough to create a lasting advantage.

Against this backdrop, the idea of “being fully prepared before you begin” feels increasingly outdated—and in some cases, inefficient. As we move from learn first, then apply to apply first, then learn, the real differentiator may no longer be how many skills you’ve mastered, but whether you’re clear about the problem you want to solve.

This reversal in learning order may feel counterintuitive, but it often leads to greater clarity. That doesn’t mean foundational knowledge is no longer important. Rather, it should function as a map—helping you identify good problems—rather than as the sole weapon you rely on.

Those who can identify meaningful problems early tend to see their learning efficiency grow exponentially. On the other hand, even vast amounts of knowledge can become scattered and unfocused if there’s no clear problem guiding it.

As we look ahead to 2026 and begin setting new learning goals, perhaps the better question to ask is this: What problem is worth solving in the coming year? When direction comes first, learning tends to follow naturally. And perhaps, this way of preparing for the future can make the year ahead feel more purposeful—and more exciting.

The reason is simple. AI now performs many of the skills that used to belong exclusively to trained professionals. If you spend your entire youth mastering a single, rigid skillset, by the time you finally gain expertise, there’s a good chance AI will already be doing it faster, cheaper, and at scale.

This isn’t to say expertise doesn’t matter — it does. But expertise alone is no longer your greatest competitive advantage. What will set you apart in the years ahead is flexibility: the ability to learn across domains and adapt when the world shifts under your feet. Over the next decade or two, industries we once thought were stable will be reshuffled. And that won’t stop just because you happen to be good at one thing.

I often tell students: don’t just learn knowledge — learn how to learn. It sounds abstract, but in the age of AI, this may be the most practical skill of all. AI can generate endless answers, but it cannot define the right questions. It can show you many possible paths, but it cannot decide which one you should walk.

That’s why the value of a generalist becomes even clearer. When your perspective is broader and your interests span multiple fields, you’re more capable of cross-disciplinary thinking — of combining ideas that don’t usually sit together to create something new. AI can give you all the pieces, but you need the ability to see patterns, make connections, and even challenge assumptions.

Looking back, my own career has unfolded the same way. I didn’t follow a straight professional track. I’ve been an athlete, a founder, an investor, and now I work deeply in education. These roles may seem unrelated, but precisely because I never confined myself to a single identity, I’ve been able to navigate major transitions and keep finding new directions.

So if you’re still in school, don’t rush to label yourself as “a finance person,” “an engineer,” or “a legal professional.” Instead, train yourself to pick up new domains quickly and apply your knowledge in more flexible, creative ways.

If I could leave you with one message as you step into the future, it would be this: don’t lock yourself into one specialty — build the ability to cross boundaries. Be a generalist, not a specialist.

Taiwan-based TherapiAI, founded by a group of AI experts and medical specialists, has built an AI platform that helps biotech companies, pharmaceutical manufacturers, and CDMOs (Contract Development and Manufacturing Organizations) save raw materials and speed up production lines—allowing pharmaceutical production to run at ten times the speed, no longer dragged down by cumbersome traditional processes.

The structural bottlenecks that have accumulated in the biopharma industry over the years are clear: fragmented experimental data, slow R&D processes, shortages of specialized manpower, and complicated, time-consuming regulatory documentation cycles. TherapiAI aims to solve these four major pain points. They have built AI infrastructure focused on CDMO and CMC (Chemistry, Manufacturing, and Controls) needs, enabling faster R&D, optimized process parameters, and accelerated documentation and compliance workflows across the board.

Just as smartphones require an operating system to allow apps to work together, TherapiAI functions as an “AI operating platform” for pharmaceutical companies: connecting fragmented data, automating complex workflows, and transforming expert knowledge into reusable AI agents. Work that once relied on specific doctoral experiences to move forward can now be accelerated and scaled.

Crossing from the Courts into Pharmaceuticals: Building Foundational Cross-Domain Capability

Therapi AI, formerly known as Akousist, was founded by Michael Han in 2018. Interestingly, its early work had nothing to do with biopharma. “We previously handled AI-driven automation of electronic case files for 36 courts across Taiwan,” Han explained. Using AI to help court professionals automate large volumes of routine documents was the difficult challenge they were focused on at the time.

This seemingly unrelated experience turned out to be a key capability that allowed Therapi AI to enter the biopharmaceutical industry. The challenges faced by pharmaceutical companies are, in fact, very similar to those faced by the courts: data scattered across ERP, equipment, and laboratory systems; departments seeing different sets of information; and clinical and experimental documents containing sensitive information that cannot leave secure environments.

Through its work with the courts, the Akousist team developed a rigorous approach and deep expertise that enabled AI to “act as an agent” for professionals—performing routine tasks under strict data-security and interoperability constraints. These are precisely the foundational structures most lacking on pharmaceutical production lines, and thus became one of TherapiAI’s core strengths.

However, to understand why this AI infrastructure matters, we must first return to the question: what are the pain points of CDMOs?

The Four CDMO Pain Points: Data, Process, Manpower, and Regulations

First is the fundamental issue of data silos. On pharmaceutical manufacturing floors, critical experimental data is often scattered across instruments, paper records, and systems used by different departments. Formats are incompatible, and it is difficult to compare everything on a single platform.

Second, the high variability of cell and drug processes makes scale-up especially challenging. These processes are extremely sensitive to parameters and environments, where even tiny deviations can cause multi-million-dollar experiments to fail. While cells exist in relatively simple conditions in laboratory settings, once they enter large reactors, every parameter must be recalibrated. Pharmaceutical companies often rely on repeated trial and error and additional batches of raw materials to identify a stable process for production.

The third pain point is the manpower bottleneck created by heavy reliance on experts.

CDMO projects are complex and lengthy, taking an average of 18 months just to sign a contract. Much of the workload falls on a handful of senior VPs, BDs, and PMs. Take PMs as an example. CDMOs commonly face shortages of experienced PMs, high turnover rates, and overwhelming workloads. Even more challenging, key process know-how often sits only in the minds of PhDs, so when a core PhD leaves, the entire line has to be rebuilt from the ground up.

The final and most difficult pain point is data sensitivity and regulatory pressure.

In CDMO operations, much of the data is highly confidential and inherently unsuitable for the public cloud. A deeper issue is that “the whole industry shares a common misconception: that the first step to adopting AI is to centralize all the data,” Han explained. “But centralized data processing is slow, expensive, easily costing millions, and it may not even be effective.” As a result, although most pharmaceutical companies understand the potential benefits of AI, many prefer to maintain the status quo rather than take risks.

What Does TherapiAI Do? Building the AI Operating Platform for the Biopharmaceutical Industry

TherapiAI’s technical architecture consists of two core layers: the underlying “Knowledge Layer Model” and the front-end “AI agents.” At the knowledge layer, the team integrates publicly available global datasets with CDMO internal production data, enabling AI to genuinely understand highly specialized pharmaceutical knowledge. Such adjustments turn the models into a credible base for pharmaceutical companies and lay the groundwork for subsequent automation and application capabilities.

On top of this foundation, TherapiAI goes on to build AI agents that can “actually do the work.” These agents are not designed to serve a single function but rather to address the three interlocking stages of the pharmaceutical process: research, exploration, and exploitation.

First of all, in the research stage, AI must follow a “better none than wrong” principle. This working environment demands high precision and has nearly zero tolerance for error. If the model lacks sufficient supporting data, it will simply respond with “I don’t know,” avoiding incorrect inferences. This allows researchers to treat AI as a trustworthy information partner rather than a black box requiring constant verification.

Next is the exploration stage, during which researchers seek not just answers but AI-assisted reasoning. At this point, the AI agent uses its built-in knowledge and cross-system data to propose possible parameter ranges, hypothesis paths, or potential causes of anomalies, helping researchers shorten experimental iteration cycles. This marks the key transition from “looking up information” to “thinking together.”

Finally, in the exploitation stage, AI formally “enters the production line,” transforming research outcomes into operational workflows—such as performing cell-line screening; automatically generating contracts, process documents, and GMP reports; and querying FDA regulations—all directly addressing CDMO pain points.

For pharmaceutical companies, research directions converge faster, manufacturing can identify stable parameters with fewer raw materials, and regulatory teams can reduce document turnaround and revision cycles. Tasks that once required weeks of cross-checking by senior PMs or PhD-level personnel can now produce a preliminary version in minutes, ready for expert review.

Han notes that TherapiAI’s core is stripping away all non-domain noise from large language models, leaving only pharmaceutical-relevant knowledge. Combined with its three-stage agent workflow, the model not only “answers questions” but can “actually get work done” on R&D and production floors. That is why TherapiAI offers not just standalone tools but AI infrastructure that compresses the entire development cycle to a fraction of its previous length.

Practical Applications of AI Agents: Accelerating “Magic Bullet” ADC Drug Development and Tracking Regulatory Changes

TherapiAI’s AI agents are already being applied in a number of highly specialized domains. Among them, the most representative case involves the development of ADCs (Antibody-Drug Conjugates), a field that has drawn much attention in recent years. Because ADCs can precisely deliver payloads without damaging normal cells, they are regarded as a major breakthrough in cancer therapy and nicknamed “magic bullets,” with licensing deals often reaching billions of dollars.

In this trial-intensive, highly complex domain, TherapiAI has built an AI agent specifically for ADC development. After researchers pose questions in natural language, the system automatically integrates cross-system and cross-literature data—covering antibodies, linkers, payloads, and other core components—and organizes design factors that influence efficacy and toxicity. This enables teams to evaluate the feasibility of different strategies early on, without repeatedly cross-checking literature and databases.

More importantly, the ADC AI agent highlights parameters likely to require adjustment later, helping teams to quickly narrow down their direction. For pharmaceutical companies and CDMOs, this reduces the number of unnecessary experiment cycles, dramatically shortens the traditional 2–3 year early exploration phase, and enables earlier entry into development stages with commercial value.

In the highly sensitive regulatory domain, TherapiAI uses its GMP AI agent to challenge manual workflows and help companies stay current with global regulatory changes.

The system can instantly search, compare, and interpret international regulatory texts such as FDA and ICH guidelines, and automatically detect inconsistencies or gaps across department documents, ensuring alignment and preventing delays caused by version errors.

The GMP AI agent can also simulate questions likely to be raised by reviewers using its built-in risk-prediction model, marking high-risk sections on a “risk map” so documents undergo a round of pre-review before submission.

TherapiAI has already been deployed at multiple pharmaceutical companies and CDMOs across Taiwan, Japan, and other places, supporting use cases such as early ADC design, cell-line screening, automated process documentation, and GMP regulatory comparison. Some customers have integrated AI agents directly into their production workflows for cross-department collaboration and document generation; others, after adopting the system, have proactively requested new features, hoping to shift more critical process steps from manual work to AI automation. These collaborations have allowed TherapiAI to evolve into an essential operating layer in the pharmaceutical production chain.

Are You Selling a “Solution” or a “Tool”?

Many deeptech start-ups encounter the same early-stage blind spot: strong technology does not automatically translate into perceived value, and having many tools does not mean that customers want to assemble them themselves. TherapiAI faced the same issue.

“We started with the technology, expecting customers to operate everything on their own,” Han said. However, most pharmaceutical teams are overwhelmed with work. They simply don’t have the time, and therefore won’t pay for “tools.” Whether a tool is powerful or not is a separate issue from whether it solves real on-site problems.

The team later realized that “experts don’t need a powerful tool, yet they need their pain point solved directly.” Han explained, “Within two weeks, we pulled back all the separate tools, stopped selling technology, and transformed it into AI agents that could directly complete work.” For example, their previously standalone OCR module was integrated into a GMP document AI agent. Once the “tool” became an “agent” capable of delivering outcomes, its value became immediately clear: customers were willing to pay and more willing to adopt.

Customers do not want separate tools; they want foundational capabilities that truly integrate R&D, manufacturing, documentation, and regulatory workflows. TherapiAI is not merely creating tools to solve individual problems; it is building an AI operating platform capable of reshaping the entire biopharma sector.

Not every AI revolution begins in Silicon Valley. TherapiAI, a team from Taiwan, offers a glimpse of another future possibility: the next major breakthrough will not depend on geographical coordinates but on who can bring AI into the world’s most complex and critical operational sites.