During the 2026 Spring Festival, Tencent ’ s Yuanbao rolled out a high-stakes "all-in" move that made the entire internet industry sit up and take notice: a 1-billion-yuan cash red-envelope giveaway, with up to 10,000 yuan for a single person.

The scale of it all briefly takes you back to 2015. In internal meetings, Pony Ma made no secret of what he was pinning on this product — he hoped Yuanbao could recreate the spectacle from 11 years ago, when "WeChat Red Envelopes" pulled off a Pearl Harbor-style surprise attack and reshaped China ’ s payments landscape overnight.

Yet looking back from 2026 — an inflection point for artificial intelligence — this kind of classic internet growth play is doomed to be futile in the AI wave.

The core logic is simple: in the face of an absolute "intelligence gap," all operational tricks are just wasted effort. The underlying rule now is — whoever is smarter wins.

This time, Yuanbao ’ s flagship push, "Yuanbao Pai," is built around a single idea: leveraging Tencent ’ s strongest asset — its social DNA — using red envelopes to drive AI group-based social interaction. At its core, this is still the old-era "traffic logic": the belief that as long as you control an entry point for traffic and have daily active users ( DAU ) , the product can build a moat.

But the internet and artificial intelligence are two entirely different creatures.

Eleven years ago, red envelopes solved the "card-linking" problem. That was about the spread of financial infrastructure — a tool shifting from "nothing to something." Once users completed real-name verification and linked a bank card, WeChat Pay was essentially locked in, thanks to the high switching costs embedded in social relationships.

But today, red envelopes can ’ t fix "model hallucinations," much less bridge an "intelligence divide."

AI is not a social tool — it is a productivity tool ( at least, first and foremost ) .

If users are lured in by red envelopes but then discover that, in logical reasoning, complex instruction-following, or multimodal understanding, Yuanbao is still a generation behind top players like DeepSeek or ChatGPT, the outcome is brutally simple: the day the red envelopes run out is the day users leave.

We can make the scenario more concrete: imagine you ’ ve just grabbed a 200-yuan red envelope in a Yuanbao group and you ’ re in great spirits. But the very next moment, when you need to write a complex Python script, or do a deep logical teardown of a 500-page business plan, you find that Yuanbao ’ s answers still contain basic errors — or make jarring leaps in reasoning.

At this point, would you really put up with its "mediocrity" just for that 200-yuan red packet, and not open a smarter, more accurate Gemini or Claude instead?

Of course not.

As long as you ’ re a normal person.

What network effect are we even talking about?

At this stage, users ’ tolerance for "idiocy" is basically zero.

Because AI consumes the most expensive resources humans have — time and trust.

Once trust goes bankrupt, the social sharing "viral loop" only brings "low-quality DAU," not real product stickiness.

After taking the red packet, no rational user is willing to switch their primary LLM from a SOTA ( state-of-the-art ) model to a low-intelligence one.

In The Unmanned Company and in various essays on Zhuomo Shi, I defined this as "intelligence first."

The shift in this underlying priority logic — putting intelligence first — has extremely far-reaching implications, yet it seems not to be fully recognized.

Many people had high hopes for Yuanbao based on WeChat ’ s status as an entry-point superapp. They believed that as long as WeChat gave Yuanbao a spot on the menu bar, it could, like Video Accounts back in the day, force its way up through sheer traffic "irrigation."

This way of thinking ignores one of the most infamous bloodbaths in the history of technology: Chrome beating IE.

Back then, IE had the strongest distribution power on the planet — it was bundled directly into Windows. Operationally, Microsoft could reach every PC user; in terms of convenience, you didn ’ t even need to install anything. But what happened?

In just a few years, Chrome nailed IE to the wall of shame.

Chrome won on two hard metrics: search that actually worked, and an engine ( V8 ) that was fast enough. ( Even in ancient times, the smarter one won. )

When users realized Chrome loaded pages 2 seconds faster, JavaScript no longer stuttered, and search results were frighteningly accurate, Windows bundling stopped being an advantage and started feeling like harassment.

Today ’ s AI race is in its own "V8 engine moment."

Almost every rising model holds its own SOTA position.

● ChatGPT ’ s foundation is its exceptionally strong "all-round capability" and its first-mover "mindshare lock-in."

● Claude Code has built an unshakable reputation in professional developer circles thanks to its precise grasp of programmers ’ needs and its exceptionally high consistency.

● Gemini, leveraging Google ’ s deeply integrated, native multimodality, has shown clear technical superiority in handling complex audio/video and long-form text analysis. Especially Banana — without Banana, I ’ d guess this whole product ’ s market share would be several percentage points lower.

By contrast, Elon Musk ’ s Grok, despite being backed by X ’ s ( formerly Twitter ’ s ) massive traffic and Musk ’ s own top-tier personal brand, has remained lukewarm. Why? Because in benchmark evaluations and real-world usage alike, Grok has never consistently demonstrated "generation-gap" capabilities that surpass GPT or Claude.

The facts have shown that if the core technical metrics and user experience can ’ t break into the first tier, funneling traffic through the entry point only speeds up users ’ recognition of its "mediocrity."

The harder WeChat pushes traffic, the faster the sense of disappointment Yuanbao triggers by comparison will spread to the entire user base.

The reason Yuanbao dares to roll out large-scale operational subsidies at this moment is that internally there must be a judgment call: that large-model capabilities have entered a "leveling-off phase," that the gaps between players have narrowed, and that what comes next is competing on operations, ecosystems, and who can hand out the biggest cash bonuses.

That ’ s completely backwards.

Intelligence hasn ’ t leveled off at all — it ’ s accelerating.

From OpenAI ’ s O1 kicking off the reinforcement-learning reasoning track to the real-time multimodal emotional interaction that became widespread in early 2026, the ceiling of intelligence has been getting rewritten almost every 3 – 6 months.

This isn ’ t "fine-tuning." This is "evolution" — an evolution that keeps creating generation gaps.

When the technology is still evolving at an exponential pace, shifting focus too early to cash giveaways, viral growth hacks, and social gameplay is an utterly absurd line of thinking.

It ’ s like in the era when the steam engine had just been invented, James Watt didn ’ t study how to improve thermal efficiency, but instead obsessed over what paint job would make the locomotive look prettier — and how to offer kickbacks to merchants hauling freight.

This approach is extremely dangerous, because competition in the AI industry is vertical, not horizontal.

By contrast, competition among most internet products has largely been horizontal.

The last wave of the mobile internet ( O2O, ride-hailing, food delivery ) didn ’ t really involve doing serious tech.

Was there a generation gap between Meituan and Ele.me? Was there a chasm between Didi and Uber in algorithms?

Basically not.

When there ’ s no generational gap in the product, operations is of course king — burning cash can buy you an empire.

But now, those gaps can emerge at any moment.

When DeepSeek shot to fame on the back of its outstanding energy efficiency and stunning ability to understand Chinese context, people seemed to quickly forget how it broke through in the first place — it wasn ’ t by handing out red envelopes in WeChat Moments, but by genuine, hard "IQ suppression."

A new architecture and new methods underpinned everything.

At this stage, any budget that doesn ’ t point toward "improving intelligence" is, in essence, wasting precious compute resources and a rare R&D window.

The reason I dared to predict that Unmanned Company was bound to become reality is, in large part, this speed. As it turns out, the judgments in the book have been turning into reality one by one. ( Including the recently viral CrowdBot, which indeed hasn ’ t gone beyond what the book described. )

AI product stickiness comes from deep trust in the product ’ s capabilities.

That trust is extremely fragile. It is built up through the accumulation of countless "Aha Moments" ( moments of surprise ) :

● When you casually toss it a blurry whiteboard sketch, it generates runnable front-end code in a second.

● When you ’ re staring at a clunky tens-of-thousands-of-words annual report, it precisely points out the financial loophole hidden in a footnote.

● When you ’ re stuck in a creative rut, it offers a metaphor you ’ d never thought of.

Those moments are the real moat.

And what do red envelopes buy you?

They buy you "bullshit."

The user grabs Yuanbao ’ s red-envelope money and drops a few AI sticker memes in the group chat — that ’ s just "joining the fun."

When evening comes and they finally settle down, needing to handle an end-of-year summary that could affect their career prospects, they ’ ll still turn around and ask the smarter model.

At that point, Yuanbao completely degenerates into nothing more than a "tool-person for handing out money."

In the history of the internet, the fate of tool-people is usually pretty grim — users show up only when money is being handed out; once the money ’ s gone, the users scatter, too.

In 2026, when AI determines your fate, resources should be poured — without holding anything back — into algorithmic optimization, the quality of synthetic data, and lower inference costs.

At times like this, you really need someone with the courage and capability to grind on SOTA — at least to have that edge, even if the odds aren ’ t great.

There ’ s no middle ground in AI competition.

If your intelligence level is even 10% lower, in users ’ minds that ’ s the difference between 0 and 1.

One is "productivity," the other is a "toy"; one is a "brain," the other is a "parrot."

Today, as intelligence advances by leaps and bounds, traffic you claw back with red-envelope giveaways will, without top-tier intelligence to catch and convert it, ultimately turn into "tuition" paid to train users for your competitors.

AI usage habits users develop in Yuanbao will eventually, driven by the pursuit of higher intelligence, transfer seamlessly to other SOTA models.

The previous wave of the internet was "operations decides life or death"; this wave of AI is "intelligence reveals the true winner."

Rather than spraying cash to buy traffic, you ’ re better off going into retreat to raise IQ.

After all, on this track, only the smartest one earns the right to survive.

Trying to build large models with the logic of internet apps is probably a mistake.