Harry Wang: Where Are the Opportunities for Young People in the Coming Decade of AI?

On June 14–15, the 36Kr WAVES New Wave 2024 conference was held at Cangku, Langyuan Station in Beijing. WAVES means exactly what it suggests — a surge of momentum, and also a reflection of young people's courage to dream and create. This year's conference centered on four themes: Ideal and Reality, Technology and Humanity, Exploration and Depth, and Tradition and Transformation — each addressing topics that resonate deeply with young people. They say life is a wilderness, but in practice it feels more like the sea: sometimes calm and shimmering, other times raging with towering waves. The "Wavers" didn't simply drift with the current — they met the waves head-on with their own strength.

On the first afternoon of the event, Harry Wang, founder and CEO of Linear Capital, was invited to speak about the opportunities he sees for young people over the next decade. Below is a transcript of his talk.

---

Harry Wang: Thank you to 36Kr for the invitation. Today's theme is "Technology and Humanity," and I see many young people in the audience. I'll share Linear Capital's thinking on the opportunities emerging from the current AI wave, based on our experience investing in artificial intelligence, and particularly what these opportunities mean for young people. I'll focus on my understanding of this new generation of AI without getting into technical details — this is purely personal perspective. We'll explore how Linear Capital understands AI, what it means for us, what it means for the future, and the investment and utilization opportunities within.

My talk will have three parts: first, my understanding of AI; second, how to leverage AI; and third, some investment and entrepreneurship insights from an investor's perspective.

First, my understanding of AI:

The concept of AI is not new. As early as 2014, China saw a surge of AI investment enthusiasm. However, by 2018 and 2019, AI had gradually become a dirty word, because the AI of that era was primarily based on fitting given data — making judgments according to existing data cases. If the test content resembled existing cases, AI could produce decent answers; but confronted with unfamiliar cases, it was helpless. This was the infamous small-AI overfitting problem.

In November 2022, ChatGPT changed this. The new AI model first uses massive data and powerful computing to compress the world's knowledge into tens or even hundreds of billions of model parameters, forming a super-large model representing world knowledge. Then it answers questions posed by users. In reality, ChatGPT doesn't truly understand the meaning of user questions — it simply converts questions into tokens (a kind of text marker), then based on existing knowledge, guesses which tokens can "harmoniously" match with it, thereby providing corresponding text answers. Simply put, it's "guessing" — guessing based on world knowledge compressed in model parameters. There's nothing new under the sun; your question, or similar questions, have more or less appeared in world knowledge before, so AI can always guess convincingly.

In the past, our way of acquiring knowledge was embodied in traditional search: finding existing results and returning them to users; if none existed, the results were poor — including the experience of using Google or Baidu search. Today's AI answers through generation, through this "guessing" approach. This Q&A interaction style, in my view, is the new big-AI, giving everyone the opportunity to have a personalized, customized magic crystal ball — for any question you're curious about, you can ask it. Based on its historical conversation memory with you, plus world knowledge in the model, it gives the best possible answer. For AI, whether the answer is true or false doesn't actually matter, which is also why AI frequently hallucinates — because essentially it's just "guessing." However, after November 2022, ChatGPT and especially the subsequent GPT-3.5 were remarkably impressive. From 2017 to 2022, the vast majority of people didn't believe this approach would yield good results, until OpenAI achieved results through five years of painstaking effort. I don't think this represents massive methodological innovation, but rather the good result of persistent belief.

The future trend will definitely be increasingly powerful models at increasingly cheaper prices.

Over the past 12 months or so, the cost of calling large model APIs has dropped by roughly 100x, making them increasingly easy to use. This trend will continue over the next one to two years. We believe all of this is just beginning. It won't be as rapid as many optimists expect, but it won't be as slow as pessimists imagine either. Many things will eventually be AI-ified. However, there remains a huge practical gap between foundational large model capabilities and application capabilities built on top of them. Recently there's been heated discussion about whether today's AI development poses an existential threat to humanity's future, but I believe embracing big-AI doesn't necessarily mean human extinction. The emergent intelligence in large models currently cannot be explained scientifically, and with the current development approach, the emergence of truly new intelligent entities that could surpass and replace humans is unlikely. But if we don't embrace big-AI, we will certainly become unemployed. To avoid unemployment, how do we leverage the big-AI capabilities that have already emerged?

Today we're in a "hundred models battle" — China has over 130 large model companies, with roughly 10 well-known ones. From Linear Capital's perspective, we believe large models will become infrastructure commodities, unable to be monopolized, at least not by private enterprises — this is characteristic of all infrastructure. It will drive many upper-layer applications, and this driving has only just begun. Many people ask whether applications will face problems of low barriers and low ceilings. To address this, I believe AI applications must value three types of data, which is extremely important.

First, personal data. Today's AI applications differ greatly from those of ten years ago — a decade ago, apps were homogeneous standardized products with little personalization, except perhaps for personalized advertising; the product itself remained standardized. However, today's AI must collect user data as extensively as possible, because every question AI answers is based on the user's current input plus historical memory of past interactions with that user, giving what it believes is the best answer. Data accumulation is extremely important — if data accumulation reaches a certain barrier, even with similar large model capabilities, the experience will be completely different due to varying degrees of understanding of personal data.

Second, context-based data. Autonomous vehicles constantly collect data from cameras, LiDAR, and millimeter-wave radar, making decisions to turn, accelerate, or brake within extremely short timeframes. They're constantly collecting this context data, while today's large language models remain largely in manual mode for this type of data collection. A good application should collect context data in the most convenient, least thought-requiring way, better serving users — such an application is a good application.

Third, collaboration data. Suppose an AI product hasn't yet built a barrier in personal data accumulation, and a superior competitor emerges in the market — users can easily switch. But if the AI product requires joint use with colleagues, switching becomes much more difficult because it requires coordinated switching, greatly increasing switching costs. We believe future AI applications must consider the curation of collaboration data.

Data curation plus the improvement of large model capabilities themselves creates opportunity for AI to achieve commercial value and monetization through Q&A approaches. If automation is involved, it inevitably involves interaction between new AI and the physical world — this is why the embodied intelligence field is so hot today.

So where do application opportunities lie? Let me give an example. When traveling, we use Ctrip and Fliggy, or book directly on airline websites/apps, make our own itineraries, or reference suggestions on Xiaohongshu. Imagine if there were an excellent AI analysis tool that deeply understood your preferences — you say you want to spend three days in Vienna, with three work meetings already scheduled, and these are your only free windows. A calendar AI agent knows your schedule, another AI agent understands your preferences for music or art, and based on your tastes suggests museums with artistic sensibility or concerts to attend. Once you've made your plan, the AI interacts with you to confirm specific accommodation and transportation options, then hands off to an execution AI agent that can automatically compare prices and book flights and hotels across different websites in the most cost-effective way. However, today's AI using Q&A approaches is still far from achieving what I've just described.

Therefore, we say 2024 is the true first year of AI applications. Over the past 12 months, costs have dropped nearly 100-fold, and capabilities have improved tremendously. Moreover, not all capabilities exist in today's big-AI — many capabilities have already been implemented in existing apps. For example, Baidu Maps, Google Maps, and booking functions on Ctrip and Fliggy — you can't hand these things over to AI. In the process of using AI, how to invoke mature external capabilities is an ability that only opened up at the end of last year.

Third, from an investor's perspective, how should we view opportunities in the AI era? Large models have made AI-native possible — you no longer need to fuss over large model problems, just use GPT well. The internet, smartphones, and payment systems have made global-native possible. A team headquartered in Shenzhen making an app for people in Boston, USA — this is completely normal. Whatever the app, users don't need to know who made it; what matters is that the product is good. For products, they can reach people anywhere well through the internet, smartphones, and the Apple App Store. Plus, global payment systems have made tremendous progress over the past decade. As I said, as long as the product is good enough, it can gain user recognition and payment.

Today, for friends engaged in application development, if they don't fully utilize these already-mature infrastructure elements to achieve global-native applications, that would truly be a waste. In the past, customer acquisition costs were high, but service costs were very low. Today, even service costs have become high, requiring massive computing power consumption. Therefore, we also believe that AI today must consider business models from the very beginning. The old "wool comes from the pig" scenario — we don't think it will happen again in the future.

What's the overall market outlook for AI applications? I believe large models are infrastructure, like highways built first, while applications are the cars and trucks driving on them. Without applications, the highways' value cannot be realized. The combined market size of these applications driving on top will far exceed the large model market size, and in my view this is a sufficient blue ocean. Because in the past 18 months, almost everyone's attention has been focused on the large model market.

The hundred-models battle will eventually, perhaps over the next five years or so, converge to ultimately leave 2–3 large model suppliers, including major enterprises like Alibaba and Tencent. Grasping application opportunities is what young people should focus on today. We need to re-solve all new and old problems in AI-native and global-native ways, using the most innovative products. Linear Capital has also established a dedicated investment project team focused on AI applications this year, Linear Bolt, hoping to support young people's innovation and entrepreneurship in the application field based on large model capabilities in this new era, in a lighter, faster, and more flexible way.

If interested, you can follow Linear Capital's WeChat official account and type "Bolt" for more information. I wish young people success in entrepreneurship in the big-AI era.

---

About Linear Capital

Linear Capital is an early-stage investment institution focused on "frontier technology + industry" — that is, frontier technologies represented by data intelligence, digital new infrastructure, next-generation robotics technology, and new technological transformations in traditional fields (such as biomedicine, materials, energy, etc.), applied across vertical industries to substantially improve industrial efficiency, empower solutions to pain points, and complete industrial upgrading, achieving excess returns through substantial increases in industrial value. It currently manages ten funds with total assets under management of approximately $2 billion.

Our investment stage focuses primarily on leading angel to Series A rounds, with individual investments ranging from $1 million to $10 million (or RMB equivalent).

To date, we have invested in over 120 entrepreneurial teams at early stages, including Horizon Robotics, Kujiale, Sensors Data, Tezign, Rokid, Guandata, and Agile Robots. The combined valuation of Linear Capital's portfolio companies is approximately $20 billion.

In the near term, Linear Capital is working to become the best "Data Intelligence Technology Fund," and in the long term, gradually build itself into the most influential "Frontier Technology Application Fund."