AI Thinking Part Three: One-Year-Old vs. Eighteen-Year-Old AI | VOICE

5,700 words, 12-minute read

Author: Bingjian Wu, Partner at Heart Capital

On AI, I've previously written A Framework for Thinking About AI Application Entrepreneurship in the Age of GPT and Known and Unknown Cards in AI. Continuing that thread here.

Six months into the AI entrepreneurship wave, beyond the progress of large language models, the whole market is waiting for AI startups to deliver their report cards — breakout hit applications. But there don't seem to be enough. Why?

The essential reason is whether foundation models are at age 1 or age 18. Having gotten our hands dirty and invested in projects at both the model layer and application layer, we've gained new perspectives. We're more convinced than ever that AI holds massive entrepreneurial opportunity, and we have a better feel for the timing of AI's development.

Three Inflection Points

Right now there are four AI applications with over 10 million DAU: ChatGPT, Midjourney, Character.ai, and GitHub Copilot. But there still aren't enough fully validated applications. Why? Because of the industry's age — age determines the capability framework.

Mobile internet development had three critical inflection points:

2007: iPhone 1 launched. Mobile turned one. The starting gun fired, and the mobile internet supply chain began building from zero.

2010: iPhone 4 launched. Mobile had matured to 18. It had a clear capability framework. Today's iPhone differs from the iPhone 4 only in quantitative performance improvements, not qualitative ones. That same year, Meituan and Xiaomi were founded.

2012: This was the year mobile internet applications exploded. Smartphone penetration crossed 20%. That same year, ByteDance, DiDi, and Xiaohongshu were founded; Kuaishou pivoted to community; Ele.me began scaling; and in 2013, Meituan started its food delivery project.

During Spring Festival 2012, PC traffic dropped and never bounced back. The baton passed to mobile, kicking off seven or eight years of explosive growth. 2012 was the "1587" moment for PC internet — a hinge year, connecting what came before and after, with hidden threads stretching far into the future.

AI's development will follow a similar pattern of three inflection points.

2022: GPT-3.5 launched. Generative AI turned one. The supply chain based on large models began building from zero.

Today: We're at roughly mobile's 2009, waiting for AI's "iPhone 4 moment" — a more mature model with a clearer capability framework. AI differs fundamentally from the internet. The internet supported many application scenarios starting from dial-up, even at tortoise speeds. But AI needs to reach a certain IQ threshold before applications work. The difference between a model scoring 50 and 60 is, to users, the difference between 0 and 1.

What capabilities should an 18-year-old model have? It needs stronger reasoning, longer text memory, less hallucination, and multimodal understanding and generation — especially video generation. It also needs lower compute costs, faster inference speeds, gradually achieving sub-second response times. Large models will have their own "Moore's Law" — annual percentage improvements in capability and token cost reductions that will soon become clear.

China's leading model makers will basically reach GPT-3.5 level by year-end. GPT-3.5 is sufficient for ad copy and summary applications, but more complex applications require stronger reasoning. They'll reach GPT-4 level by the second half of next year, which will open up the application development landscape significantly. OpenAI will also release its next-generation model next year. Optimistically, in six months to a year, models may reach 18.

After iPhone 4, the iPhone's capability ceiling essentially plateaued — no more qualitative leaps, just quantitative improvements for years after. But large models may not plateau in the same way. From GPT-1 to GPT-4 Turbo, each iteration brought qualitative change. Scale-up hasn't hit a ceiling yet. Currently, large model capabilities roughly equal those of a typical intern. They'll evolve to scientist level, eventually surpassing all human intellectual limits. This is the vertical evolution of models.

Horizontally, large language models are just one part of the AI landscape. The first principles of foundation models have emerged — "predict next token." This principle could spawn other models:

If future transformers or alternative algorithms can accurately predict the next frame, video models will emerge, unlocking the next Douyin-level content platform. If they can accurately predict the next action sequence, embodied intelligence models will emerge, unlocking general-purpose robots. If they can accurately predict the next protein sequence, protein models will emerge, advancing drug discovery significantly. If they can accurately predict the next pixel, 3D models will emerge, unlocking metaverse construction.

Once the full landscape unlocks, we'll see multiple foundation models, with marginal costs in many directions approaching zero, continuously unlocking new application-layer opportunities.

Real Opportunities vs. Fake Ones

As a platform evolves from age 1 to its golden age, it passes through many real and fake opportunities. Looking back at mobile internet's early days, four types of opportunities dominated the charts.

First: the "flashlight" — a fake opportunity. Before 2010, flashlight apps frequently topped the charts. Using your phone as a light source was genuinely useful demand. But the tool was too simple, eventually becoming a standard system feature, and flashlight apps died out.

Second: "Talking Tom" — a small opportunity. Talking Tom cleverly used the phone's microphone — whatever you said, it would repeat in a cat's voice. At its peak, daily downloads hit one million, but there was no long-term retention. The gameplay was too single-dimensional; after a few days of talking to a cat, the novelty wore off. Later, Talking Tom was acquired by a Chinese listed company for RMB 7 billion.

Third: "91 Mobile Assistant" — a phase-dependent opportunity. In mobile internet's early days, whoever built an app store gained traction. There were 91 Mobile Assistant, Wandoujia, Anzhi, and others. 91 achieved the largest scale and the most beautiful exit. In 2013, 91 sold to Baidu for $1.9 billion — a record-breaking price at the time. The founding team and investors were thrilled.

Around 2013, what was the first thing people did after buying a phone? Download tons of apps to try, like Shennong tasting hundred herbs. App stores were considered potential mobile internet entry points because they were刚需高频 (rigid demand, high frequency) with distribution power. A decade later, killer apps on phones have converged to a dozen or so, and these dozen have become the true mobile internet entry points. The app stores' greatest historical mission was to elevate those dozen killer apps to their thrones. Few people remember 91 or Wandoujia today.

Fourth: "Douyin and Kuaishou" — real opportunities, with long-term retention, sustainable business models, and meaningful moats. In mobile's early days, these were "unknown cards" — impossible to estimate market size, harder still to imagine becoming hundred-billion-dollar companies. They were rare and precious.

Today, in AI's early stage, these same four categories of opportunity lie before us. We need to figure out: which is AI's flashlight, and which is AI's Douyin or Kuaishou?

Wrapper applications based on large models are the new flashlight — they'll be made obsolete by model capabilities in short order. Some native AI hit applications, if their gameplay is too single-dimensional and unable to build proprietary assets, may become the new Talking Tom. As for AI's Douyin and Kuaishou, it's still hard to imagine what they'll look like. More likely, the future Douyin is still building its Toutiao today, or even its Neihan Duanzi. It's an evolutionary process.

Every platform rise brings its own native applications — this is the greatest opportunity for entrepreneurs.

What did early mobile native apps look like? In 2011, KPCB partner John Doerr summarized mobile native as SoLoMo — Social, Local, Mobile. The app that best fit this was Foursquare: you could check in at locations. If you frequently checked in at a bar, you'd earn that bar's honor badge; people who often appeared at the same place could connect. Downloads were high for a while, but retention was poor — the gameplay was too simple, not fun enough. Years later, location services became a standard feature in every app.

What are truly mobile native applications? Meituan Waimai definitely qualifies. But in food delivery's entire value chain, location services contribute only 20% — 80% of value comes from offline merchant resources and logistics fulfillment. Yet that 20% was the画龙点睛 (finishing touch that brings the dragon to life): it unlocked the entire business process. Because of phones' location属性 (location attribute), we could search restaurants within a few kilometers; delivery drivers' positions became trackable in real time.

What is mobile native? Not an application that's 100% born from phones, but one whose experience and value are amplified 100x because phones exist.

AI native is the same. AI native doesn't necessarily rely 100% on large models. The model may contribute only 20%, but that 20% is the画龙点睛 — unlocking a previously impossible task. That's true AI native.

Tools vs. Assets

Whenever a platform opportunity arises, tools are created first.

Looking back at the dozen or so high-frequency applications that ultimately established themselves on mobile — WeChat, Taobao, Pinduoduo, Douyin, Kuaishou, Meituan Waimai, DiDi, Xiaohongshu — few are pure tools. They ultimately established themselves through core assets: WeChat through relationship assets; Douyin, Kuaishou, and Xiaohongshu through content assets; Meituan Waimai and DiDi through offline assets; Taobao and Pinduoduo through merchant assets. These assets formed the products' supply — supply is the product.

Why do pure tools struggle to establish themselves? Why must they build core assets? Pure tools deliver 100% of user value through code, and code is cheap to copy. Eventually, everyone has similar functionality. For consumer products, you must find advantages beyond code. WeChat's value comes from address books; Douyin and Kuaishou's value from massive video libraries; Meituan's value from merchant and rider networks. Ultimately, differentiation forms around assets.

The same logic applies to AI. In the future, ChatGPT shouldn't be purely Q&A format — it should be able to call various agents to complete tasks, somewhat like an all-in-one WeChat. Midjourney shouldn't simply be an "image designer" — it should produce content with plots and entertainment value, deriving value from content consumption, somewhat like today's content platforms.

Between tools and assets, the relationship is often dynamic. Many products weren't positioned as tools from day one. Douyin was a short-video content community from day one; Meituan Waimai was a food delivery transaction platform from day one. Their product interfaces haven't fundamentally changed since day one. The strongest products often keep their broad outline constant while the assets inside iterate repeatedly.

Some products started as tools and quickly pivoted. Gif Kuaishou was initially a video-to-gif tool, pivoting to short-video community after a year or two. Its explosion came after the community pivot.

Opportunities exist on both the north slope and south slope. But don't linger in tool form for too long. You can start with tools, but the endgame is always about assets. If you figure this out on day one, it's best to build assets from the start.

New Characteristics of AI Entrepreneurship

I've drawn many analogies between AI and mobile internet above. But having actually participated in some AI projects, I've found that AI entrepreneurship follows very different rules from mobile.

First, the 0-to-1 validation period in AI entrepreneurship is longer. When building mobile applications, we could reduce the product to a single function, build an MVP in two to three months, and validate upon launch. Mobile applications限定场景 (constrained the scenario). If we built a ride-hailing app, users would only use it for rides because of the product positioning and specific interaction logic — they definitely wouldn't search for news inside it.

But AI products are different. Faced with an AI-enabled chat box, users will input anything, probing its boundaries. Once the product can't handle it, users may leave. MVP is somewhat失效 (ineffective) for AI, or rather, the time to build an AI MVP has lengthened.

Mobile products typically face users' deterministic needs: rides are rides, food delivery is food delivery. AI products face users' uncertain needs. Faced with a box, users will chat with the model about everything. Managing user expectations is a critical proposition.

Earlier this year, people marveled at ChatGPT's capabilities because demos make it easy to feel the upper bound of model capability. But in real usage, whether you retain users is determined by the lower bound of model capability.

Second, once PMF is found, AI entrepreneurship's 1-to-10 growth will be faster. In the previous two waves of platform-level entrepreneurship — PC and mobile — applications spread alongside terminal deployment. Application adoption speed was constrained by terminal deployment speed. But AI entrepreneurship builds applications on existing phones and PCs. As long as it's an惊艳 (stunning) product, it will spread at lightning speed. ChatGPT and Midjourney broke 100 million users in mere months, validating this pattern.

But because every position on every terminal is already occupied, there's no traffic dividend. This requires new AI products to be ten times better than previous products, ideally solving problems that were previously unsolvable, to have a chance of breaking through.

In the early days of PC and mobile entrepreneurship, commercial infrastructure was immature — no ready payment systems, monetization methods, or traffic acquisition channels. In mobile's early days, no one knew how to make money on phones; feed ads were absolutely unimaginable. In today's AI entrepreneurship, this commercial infrastructure is all ready, but the engine — the model — is immature.

This creates a dynamic: some smart people will first use a 50-point AI product to train their troops, commercializing from day one,摸索增长方式 (figuring out growth methods). After practical experience reveals the product's essence, when foundation models reach 18, they'll swap in the new engine and take off. The pure user-acquisition-without-monetization, pure cash-burning approach of previous internet eras is unlikely to repeat in AI entrepreneurship. AI applications come with monetization built in.

Third, major competition will arrive earlier, and startups' breakout battles may be decisive from the first fight. In 2010, few major companies proposed mobile-first strategies. Until 2012, some were still watching mobile internet opportunities from the sidelines. This gave startups time to develop quietly. By the time unknown cards became known cards, major companies couldn't suppress startups anymore.

ByteDance vs. Baidu, Pinduoduo and Meituan vs. Alibaba, miHoYo vs. Tencent — all cases of startups moving from the margins to the core, flipping the table on major companies. This gave major company management an additional VC mindset: believe in the power of non-consensus, don't underestimate seemingly small entry points.

So major companies reached consensus on AI much earlier. Almost all treat AI as a new growth curve, and competition with startups will come sooner. In AI entrepreneurship, once 0-to-1 is validated, the next year or two becomes especially critical. Growth from 1-to-10, commercialization rollout, and organizational capability building will happen in a compressed time window. Previously, this entire process might take four or five years; now it will compress to two or three. Founders need both initial velocity and acceleration.

I previously discussed this with Kai Qu of 42章经. Our consistent answer: higher talent density and a more competitive environment will make "AI 2012" arrive faster. Looking back years from now, we'll still see a cohort of startups at the main table, flipping the table on major companies.

Entry and Pace

For investors, when to enter AI? Past experience gave me my answer.

I graduated from computer science in 2010 and joined Baidu as a mobile internet product manager. At that time, mobile traffic was single-digit percentage of total traffic. I started doing VC investing in mobile internet at the end of 2013, at an institution focused on angel rounds. In retrospect, for mobile internet investing, starting early-stage investment in 2014 was already too late — I'd missed the earliest stages of most major mobile internet winners. For a VC novice, it takes at least two to three years to build networks and understanding. By the time that's built, mobile internet had already exploded to 2016.

With this visceral experience, seeing OpenAI fire the starting gun and AI undergo qualitative change, I'm more willing to participate in intelligence from the very beginning. There's a simple logic: to find gold, you need your 10,000 hours. Whether those 10,000 hours happen this year or next makes a fundamental difference.

For entrepreneurs, when to enter AI? Should you wait until models reach 18?

My answer: the earlier the better, as long as you know what you want to do. But watch your pace. Spend as little money as possible. Gain as much understanding as possible. Once you see a good opportunity, pivot anytime, and wait for the critical hand to go all-in.

Why do I think enter early? The gap between AI entrepreneurship and mobile entrepreneurship is far greater than between mobile and PC entrepreneurship. There's a significant cognitive gap that needs to be filled through实战 (actual combat). Where are smartphones' capability boundaries? This answer was relatively clear in mobile internet's early days. But where are AI's capability boundaries? The answer is unclear — the technology layer is shifting while you're building the application layer. You need full understanding and预判 (anticipation) of model capability boundaries. Second, the competition issue mentioned earlier places higher demands on founders' readiness.

AI's journey from 1 to 18 is a continuous unlocking of new opportunities, but it won't take 18 years. At AI's current pace of development measured in days, it may happen quickly. People tend to overestimate progress in the coming year and underestimate change over the coming decade. This dynamic is playing out precisely right now.

Facing an unknown massive market, we're accustomed to预判 (predicting). History often rhymes, but never repeats. AI is not another mobile. Half of what I've written here may turn out wrong in retrospect — wrong from path dependence, and more so from limits of imagination. Just as standing in 2010, we couldn't imagine mobile would bring the changes we see today.

We can't accurately tell whether we're at 14 or 16. Since 18 isn't far off, just do it. The best way to predict the future is to create it.

---

Author bio: Bingjian Wu is a partner at Heart Capital (Soul Capital). He has three years of experience as a Baidu mobile product manager and strategic analyst, and ten years of venture capital experience, previously at K2VC and Legend Star. He focuses on AI, robotics, and hardware.

/ Recommended Reading /

Founded in 2022, Heart Capital is a China-based early-stage venture capital fund focused on technology and digitalization, founded by Yan Han, formerly a founding partner of Lightspeed China. The Heart Capital team consists mainly of Lightspeed China's founding partners, CFO, core investors, and senior industry investors from Cainiao and Baidu. The team's past investments include Series A investments in Full Truck Alliance (NYSE: YMM), Xpeng Motors (NYSE: XPEV, HK: 09868), as well as FinVolution (NYSE: FINV), 06810.HK, RoboSense, World Logistics, LandSpace, Lanhu, Micro-nano Star, Starfield, and others. Rooted in China with a global outlook, Heart Capital is dedicated to finding world leaders who will disrupt the future. Heart Capital advocates the value of "people" and "heart," and looks forward to accompanying more young Chinese entrepreneurs onto the world stage.