AI Is Not the Next Internet: This Wave Has a Different Shape

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

At Sequoia Capital’s 2026 AI Ascent conference, three partners—Pat Grady, Sonya Huang, and Konstantine Buhler—made a direct pronouncement: AGI has arrived. Their definition is highly pragmatic. Send an agent to complete a task, and if it can recover from failures and persist until the job is done, that’s enough.

My first reaction to this definition was skepticism about hype. But on reflection, I recognized the shrewdness. It sidesteps debates about whether AI thinks like humans and asks just one thing: can the work get done? This is a commercial definition, and precisely because it’s commercial, it deserves serious consideration. The threshold for enterprise AI adoption is shifting from “is it smart enough?” to “can it consistently complete a task?”

From a Faster Horse to an Automobile

For the past two years, most AI applications have done the same thing: make existing work 10% or 40% faster. The conference terminology called this a faster horse—faster, yes, but you’re still feeding a horse.

The difference with the new generation of long-duration agents comes down to one number. In a test chart they referenced, the duration models could work continuously without drifting off course jumped from a scale of tens of minutes to hours. Tens of minutes only works for an assistant; hours means you can delegate a complete task: code from requirements to deployment, research from raw materials to conclusions. The qualitative leap from horse to automobile was never about speed—it was about method.

The conference cited several examples of compressed timelines. Developers who single-handedly completed what would normally be three-year projects during holidays using coding agents, Notion’s team rewriting 8 million lines of code in six weeks. These numbers come from speakers and are difficult to independently verify individually, but the direction is easy to confirm: similar stories are multiplying and happening around us.

The criterion is simple. If your team’s AI use still revolves around making existing processes slightly faster, this wave hasn’t really arrived yet. The marker of arrival is when division of labor, collaboration methods, and even departmental structures begin getting rewritten because of AI.

This Wave Is a Compute Revolution, Not a Communications Revolution

I believe the conference’s most valuable judgment is this: AI is not the next internet.

The internet, cloud computing, mobile internet—all are fundamentally communications revolutions, solving how information gets distributed. These waves share a common trait: a relatively stable foundation. Once the roads are built, the application layer can make long-term plans on top, betting the foundation won’t shift.

AI is a compute revolution, solving how information gets processed. Its characteristic is that the foundation itself is moving. In 2022, ChatGPT showed the industry pre-training; two years later, inference-time compute elevated reasoning capabilities to a new level; later still, long-duration agents transformed conversational AI into productive AI. Each leap is discontinuous, and nobody can say for certain where the next one will be.

The shape of these two types of waves directly determines strategy.

The speakers used an analogy I found apt. On a clear day, you can’t overtake 15 cars; on a rainy day, you can. Rain means unstable rules and low visibility. For leaders, no advantage is safe; for laggards, every turn is an opportunity.

So my advice is concrete: don’t make five-year plans in AI. The reason has nothing to do with perseverance—this wave’s foundation simply doesn’t give you conditions for long-term planning. Compress strategic units from five-year roadmaps to one bet per generation of capabilities. Each time a new generation of capabilities emerges, use small resources to quickly validate whether it can rewrite your business. If validation fails, withdraw; if it succeeds, double down.

This Wave Targets the Service Industry

The wave’s shape also determines how the market is calculated.

The conference provided a comparison. In the first 15 years of cloud transformation, the global software market grew from roughly $350 billion to $650 billion. AI is targeting something else: service revenue—the markets held up by human labor in law, healthcare, finance, and consulting. Legal services in the US alone, just one vertical, represents a roughly $400 billion market, approximately equivalent to the entire software market. The speakers themselves admitted they can’t calculate precisely—whether it’s $10 trillion, $5 trillion, or $50 trillion, nobody knows—but the directional judgment is clear: this wave’s pricing unit has shifted from subscription seats to displaced work hours and delivered outcomes.

This explains why the fastest-growing companies in the industry have almost universally chosen the path of selling outcomes. Your AI replaces a legal assistant’s work hours; clients pay for results, not seat licenses. The software market is a zero-sum game; the service market takes incremental budget from labor spending—the ceiling differs by an order of magnitude.

For entrepreneurs, the question shifts accordingly. From “what software should I build?” to “whose work hours am I replacing, and how do I prove the replacement’s effectiveness?”

Two Debts Behind the Pragmatic Definition

Finally, let me address what the conference didn’t fully explore.

Defining AGI as capable of sustained work completion is action-oriented, but carries two debts.

One is a reliability debt. That an agent can complete tasks is a statistical conclusion—nine successes out of ten attempts, one failure. In scenarios requiring accountability, that one failure isn’t a low-probability event; it’s an incident waiting to happen. The lower enterprises set supervision based on this definition, the higher the price they may pay later. Supervision won’t disappear with upgraded definitions; it will just resurface in a different form.

The other is a capability illusion. Completing tasks in narrow domains is different from general intelligence. Taking a commercial definition as a signal of omnipotence makes it easy to hand over tasks AI still can’t handle well. When you hit those pitfalls and swing back to complete distrust, both extremes are actually the same mistake: looking at labels instead of specific tasks.

So my position is clear: the definition is a starting gun, not an insurance policy. When it comes to each specific piece of work, you still need to ask those three old questions: What can it do? Where will it fail? Who backs it up when it fails?

Implications for Action

Three things you can do now.

First, reset your strategic clock. If your company still drives AI deployment with annual or longer planning cycles, compress it to one unit per generation of capabilities. There’s only one verification criterion: each time a new generation of capabilities emerges, are you among the first to validate its value in your own business?

Second, re-prioritize your task list using the new definition. List the work in your team and ask for each item: can an agent complete this continuously, what’s the failure rate, and who backs it up when it fails? Tasks with completion capability and low failure rates—delegate first, keep supervision in place; tasks that can’t be completed—don’t covet them, honestly keep them for humans.

Third, shift moat budget toward the customer side. Capabilities depreciate extremely fast in this wave; today’s technical lead may be erased tomorrow by an open-source release. What depreciates slowly is understanding of customer business, accumulated trust, and work methods embedded in their processes. When allocating resources, this side deserves more than the technical side.

The car has arrived. Debating whether it counts as a car is meaningless. What matters is learning to drive and finding your own road.

Key points: Sequoia defines AGI as capable of sustained work completion; enterprise threshold shifts from smart to usable; difference between faster horse and automobile lies in method; compute revolution’s shifting foundation invalidates five-year plans, compressing strategy to one bet per capability generation; pricing unit shifts from seats to displaced work hours; pragmatic definition carries reliability and capability illusion debts.