"If You Don't Understand These 3 Differences Between AI and Mobile Internet, Don't Bother Starting a Company"

About the author: Bingjian Wu, Partner at Heart Capital (Soul Capital). Former mobile product manager and strategy analyst at a major tech company, with prior experience at K2VC and Legend Star. Has invested in multiple large language model and AI application projects. WeChat: wubingjian

The last decade of mobile internet produced a generation of battle-hardened entrepreneurs. Many have now jumped into AI, only to discover after getting their hands dirty that mobile and AI are fundamentally different beasts.

Some come to me asking: what exactly are the differences? Where do they originate? What traps does惯性 thinking (inertial thinking) set?

First, the key questions have changed. Here are three fragmented thoughts to feel your way through.

Difference ①: Competition vs. Swallowing

Mobile's keyword is competition; AI's keyword is swallowing.

The classic stories of mobile internet always revolve around competition — the food delivery wars, the ride-hailing wars, so many legendary battles.

Why is competition the keyword for mobile? Because our phone home screens only have a dozen or so slots. Those dozen slots represent a dozen categories of network platforms: messaging, food delivery, ride-hailing, short video, and so on.

These platforms are typically two-sided or multi-sided markets with powerful network effects. More consumers attract more suppliers; more suppliers attract more consumers. This flywheel keeps spinning until the market consolidates around one or two winners.

Whether dozens or hundreds of players enter, it eventually becomes a bipolar war. Only by driving out the other side do you win the right to tax the land.

Most network platforms run on tax-collection business models. Product promotion pays the Douyin tax; merchants pay the Pinduoduo tax. This is the necessary cost of maintaining order on this land, and it's the secret of why platform models become money-printing machines.

Land ownership is the root of competition.

But AI is different. AI's keyword is swallowing.

A year and a half into this new wave of AI entrepreneurship, have you heard of any two LLM companies fighting to the death? Any two AI apps in a zero-sum battle? No. That's why the media struggles to write a splashy "Midgame of the LLM Wars."

But you've definitely heard swallowing stories. At every OpenAI launch event, amid the applause, some developer screams: "OMG, the model just swallowed my product!" Cases like Jasper are too numerous to count.

Why does swallowing happen? Because the model is a student, constantly ingesting data and learning skills. As models iterate, they inevitably acquire more advanced capabilities.

Swallowing and competition are fundamentally different. Competition is: I see you, I copy you pixel by pixel, then we compete on operational efficiency. Swallowing is: I don't even know you exist. I'm just quietly learning skills — sorry, I happened to learn yours.

The former is like martial arts masters competing on inner strength; the latter is like a二向箔 (two-dimensional foil) attack.

Therefore, predicting where large models are headed, predicting where OpenAI is headed, has become a basic competency for AI practitioners.

Don't stand in OpenAI's path. Don't get swallowed. AI application entrepreneurs must cultivate land on roads OpenAI won't travel, while also predicting the steep new capabilities of next-generation models.

For example, if you predict OpenAI won't do entertainment, and that multimodal capabilities will leap forward — then you should build entertainment on multimodal now, even aggressively maxing out token usage. That way when the next model drops, your product experience will have a step-function improvement!

When guarding against swallowing, we must consider not just what OpenAI does, but four确定性生意 (certain bets) ahead: AI PC, AI Phone, AI Cloud, and AI-enabled software (like Office Copilot). When PCs and phones are AI-armed with intelligent assistants, which AI applications will they swallow?

Competition in AI isn't nonexistent — there's competition for funding, for talent, for daily user acquisition. Some use Kimi, others use Doubao. But because current AI products lack network effects, there's no life-or-death competitive dynamic.

Swallowing is the keyword of the AI era. It's existential. Right now, are you focused on competition, or on swallowing?

Difference ②: Customer Acquisition vs. Technology Betting

Mobile bets on customer acquisition; AI bets on technology prediction.

Entrepreneurship is a constant process of betting the house. In the mobile era, most bets were on customer acquisition.

Early on, Toutiao spent heavily on phone pre-installs. Douyin poured money into Guangdiantong ads. DiDi subsidized your rides. Pinduoduo's ten-billion-yuan subsidies. At their core, these were all: after PMF validation, confident in product retention, betting cash to race against time. The baseline assumption was that customer acquisition costs would only rise, so after PMF validation, you had to grow aggressively, reach critical user mass first, and trigger network effects or two-sided effects.

AI-era betting happens on technology prediction. Training a model costs tens of millions, even hundreds of millions of dollars. How capable will this model be? What features to target? Iterate frequently to keep up, or hold back for a blockbuster? Will it be leading-edge or obsolete the moment it's trained?

It's extraordinarily demanding. The stakes are higher. Small-scale experimentation barely works.

It's hard for any model to dominate uncontested. It's more like a frog-leap competition — at different time points, the SOTA model differs. That's why many LLM companies have claimed to be #1 at various moments. Looking across time, the leader really does change. GPT-4o, Claude 3.5 Sonnet, Llama 3.1 each had their moment in the sun.

This frog-leap dynamic makes "rapid iteration" obsolete. Model competition is more like charging up for big moves.

Entrepreneurs must gauge the limits of current technology and resources, predict where top players will leap in six to twelve months, then charge up an even bigger leap!

Large models have no obvious first-mover advantage, only SOTA advantage (State of the Art). Only by achieving SOTA on certain dimensions can you deliver differentiated user experiences and capture social attention — exposure, funding, and talent.

LLM entrepreneurship is for the few. Most are targeting AI application opportunities. Similarly, AI applications also bet on technology prediction:

• On direction: are you on the LLM's inevitable path, or in directions LLMs won't pursue? That's a bet.
• On engineering practice: after stacking lots of engineering, as base models advance, does it become cannon fodder or more valuable? That's a bet.
• On business math: how much will token costs drop next year? Giving you the courage to max out tokens now. That's also a bet.

What other bets do you see in the AI era?

Difference ③: Business Warfare vs. Engineering Practice

Mobile values business warfare; AI values engineering practice.

In the mobile era, no investor asked whether an app could be built. App development barriers were low. Delivery quality between teams might be 70 vs. 75 points — not much difference. What decided victory was product design and business efficiency.

The gap between teams on product and business execution was often 10 vs. 90 points. So investors cared deeply about whether founders had product intuition, whether they could fight hard business battles.

AI is exactly the reverse. We're far from business warfare; the fight is engineering practice. Simply building something usable and good already eliminates 99% of people.

Many teams raised funding for AI applications last year, but few have delivered qualified products this year. The gap lies in AI cognition and execution capability.

First, whether there's clear cognition of model capability boundaries. Last year many overestimated model capabilities. The reality: large models are still very early. Achievable PMF (Product Market Fit) is currently limited. Because of technical constraints, the industry has shifted to saying TPF (Technology Product Fit) rather than PMF.

These product forms all appeared in science fiction: AI search, office assistants, virtual tutors, entertainment companions, the Her-style romantic partner.

What you choose to build already reflects the founder's AI cognition level — a 10 vs. 90 point difference.

Many products written into pitch decks are currently unachievable. If you know it's unachievable, pivot early and save yourself.

Second, engineering practice. Take office assistants: some do exactly what you ask; others don't follow commands. Entertainment companions: some feel human; others feel robotic.

Where do these practical differences lie? A few examples:

• Does the team understand the new AI tech stack? To achieve a certain model feature, do you do continued pre-training, fine-tuning, or prompt engineering? Knowing and understanding are vastly different.
• How much effort goes into post-training? Post-train once and stop, or build your own evaluation system and iterate repeatedly?
• Have you built data flywheels? Is user interaction data used to optimize models, or does it just sit there?

These engineering practices are also 10 vs. 90 point differences.

Teams with both AI cognition and engineering practice — simply delivering a qualified product already makes users happy, already beats 99% of competitors. We're far from business warfare.

AI is still in a period of radical technical change. The entire tech stack is highly immature. Many places require hand-tuning models, hand-assembling products — quite dependent on craftsman spirit.

When AI matures, will business warfare become the main theme again?

Postscript

The above are fragmented thoughts, but the differences go further.

When exploring a brand-new industry, are we asking the key questions?

Refreshing your OS with fresh AI knowledge and practice seems to make it easier to ask key questions and find key solutions.

Past experience is valuable pre-training data, making us more accurate when predicting the next token.

Often, escaping previous惯性 thinking (inertial thinking), not marking the boat to find the sword — being able to discuss specifics concretely — already puts you on the right side.

For young entrepreneurs, having no惯性, living in the present, is also a unique advantage.