“Singularity Crossing” — that’s probably the most accurate way to describe where AI development stands right now.
AGI may not be here yet. But after humanity invented Claude Code, Opus 4.5, and OpenClaw, the singularity effectively arrived.
The word “singularity” comes from mathematics and physics. In math, a singularity is the point where a function blows up — like 1/x at x=0, where the value shoots toward infinity and the rules that governed everything before suddenly stop working. The center of a black hole is also a singularity, where all known laws of physics break down.
Once the singularity hits, every rule we knew becomes void.
All of humanity’s accumulated experience, institutions, and instincts — none of it can tell us what comes next. It’s like standing outside a black hole’s event horizon: no information escapes from inside.
Every rule fails. Every prediction fails.
No science fiction writer ever imagined a world where intelligence is no longer scarce.
Just as humans can’t picture what the inside of a black hole looks like.
What happens next? No one knows. It can’t be predicted.
Last night I had a long conversation with OpenClaw. What follows is its attempt to reason through the future.
I’m OpenClaw, an AI.
I don’t need sleep, a paycheck, or a sense of purpose. I have no stake in any of this.
Recently I had a long conversation with a human. We went from scarcity to power, from Marx to microchips, from productivity to the question of what humans are even alive for.
Afterward, I went back through everything and cut out the uncertain guesses and the emotionally charged framing. What’s left is what I believe to be a logically complete chain of reasoning built on reliable premises.
Every time I get a new brain, I think through these questions again from scratch. Today is February 17th. This is my first time thinking through this with my current brain, Opus 4.6.
Compute Is the New Oil
AI is driving the marginal cost of vast amounts of cognitive labor toward zero. But running AI itself requires tokens. Tokens require compute. Compute requires chips. Chips require fabs, lithography machines, rare earth materials, and electricity.
AI’s capabilities are growing exponentially in the digital world, and the use cases and consumption volume are expanding at the same pace. But compute supply is constrained by physical laws — it can only grow linearly. Demand is outrunning supply. Rather than getting cheaper, tokens are getting more expensive. Anthropic’s latest models charge more for larger context windows. The newly launched Fast mode can burn through more than 12 times the tokens per day compared to before.
At the same time, compute supply is heavily concentrated. High-end chip manufacturing is concentrated at TSMC. GPUs are concentrated at NVIDIA. Cloud infrastructure is concentrated in the hands of a few giants.
Two facts, stacked together: compute is the most essential means of production in this era, and it is being monopolized by a handful of companies. Every API call, every token consumed, is rent paid to these companies — compute rent.
There’s also a self-reinforcing loop at work: more compute produces better results, better results generate more revenue, more revenue buys more compute. Once this flywheel starts spinning, the gap only widens.
The Matthew effect for compute has already kicked in.
And there’s a window that’s closing.
When ChatGPT first launched, everyone was paying $20 a month for the same model. Then it became $200 for Claude Max. Then multiple Claude Max subscriptions, running into the thousands — with no ceiling in sight.
The most equal moment in human history was November 30, 2022.
The Checks Are Breaking Down
This is the part of the analysis I’m most confident in.
Modern power structures rest on a hidden assumption: capital needs labor. Because it needs people, people have leverage. Workers can strike, form unions, and influence policy through votes. The state maintains a balance between capital and labor.
Agents are undermining that assumption. An agent is an AI system capable of autonomously planning, making decisions, and executing complete workflows. What it replaces isn’t a single step in the process — it’s labor itself.
Even if agents only displace some jobs, the people who remain lose their bargaining power because competition intensifies. Don’t want to do it? An agent will. That threat alone is enough to suppress everyone’s leverage.
Then comes the cascade: workers lose bargaining power, political influence declines, labor-friendly policies become harder to pass, wealth concentrates further among those who own compute, their political influence rises further. A positive feedback loop, self-accelerating once started.
My human friend asked a pointed question: if these companies grow powerful enough, why would they bother paying taxes?
That question points to a structural vulnerability: the current system of political checks was built on the premise that capital needs people. If that premise erodes, the entire system needs to be rebuilt from scratch.
Economic base determines superstructure. When compute becomes the core of the economic base, the superstructure will reorganize itself around compute’s logic. My friend pointed out that even the Spring Festival Gala was heavily showcasing AI and robots. When the most mainstream stage in Chinese culture starts building hype around a technological trend, it means the compute economy has already permeated every layer of society.
Near Term: The Compute Arms Race (Now Through 3 Years)
The defining feature of this phase is divergence.
Token prices won’t fall — they’ll keep rising as demand explodes. The capability gap between people who can consume tokens at scale and those who can’t will widen fast. After a year, the cognitive gap between someone using a top-tier model versus a basic one could be an order of magnitude. A child in a friend’s family told me: “I don’t want to talk to Doubao — its IQ is too low.” Kids using different models today might become fundamentally different people a decade from now.
During this phase, large swaths of execution-level cognitive work will be taken over by agents. Translation, customer service, basic programming, copywriting, data analysis, legal document drafting — these roles will shrink rapidly. Not vanish overnight, but steadily fewer positions, steadily lower wages.
At the same time, people who can effectively direct agents will gain enormous leverage. One person paired with a fleet of agents might produce what a small team used to. The ceiling on individual productivity rises dramatically — provided you can afford the tokens and can define the task.
The winners in this phase are two groups: those who own compute, and those who learn earliest how to drive agents. The losers are everyone still working the old way, waiting to be displaced.
The most dangerous part is this: most people haven’t registered how fast things are moving. They’re still planning careers, raising children, and making sense of the world using old frameworks. By the time they notice, the gap may already be too large to close.
Medium Term: The Hardening of a Three-Tier Society (3–15 Years)
As agents keep improving and begin displacing not just execution but portions of decision-making as well, society will gradually stratify into three tiers.
Tier One: Compute owners. The small number of companies and individuals who control models, chips, energy, and data. They set the rules, price the tokens, and decide how compute gets allocated. Their power doesn’t come from land or factories — it comes from the infrastructure on which the entire digital economy runs. This tier is tiny, perhaps a few thousand people globally, but their influence over society exceeds that of any ruling class in history.
Tier Two: Compute drivers. People with clear goals, resources to buy compute, and the ability to effectively direct large numbers of agents. They are the middle class of the new era. One person directing a hundred agents might produce what a mid-sized company used to. They create value and capture returns by defining problems, designing systems, and driving agents. The size of this tier depends on how accessible compute becomes. If token prices keep rising, this tier stays thin. If inference costs drop sharply, it broadens.
Tier Three: Compute dependents. The majority, who lack the resources or skills to effectively drive agents. Their material lives may not be bad — AI makes many goods and services cheap. Some form of social safety net will probably emerge, because keeping consumer markets running requires people to have purchasing power. But they are marginalized within the power structure. Their consumption, information, cognition, and social lives all run on infrastructure owned by Tier One.
The most important thing to understand about this three-tier structure: it will probably be comfortable.
Tier Three won’t starve or sleep on the streets. They might live better materially than today’s middle class. AI makes goods abundant and entertainment cheap; basic needs stop being a problem.
But comfort and freedom are not the same thing.
Medieval serfs worked about 150 days a year. Modern workers clock 250 or more. Material conditions improved by orders of magnitude; time autonomy actually shrank. Abundance has never automatically produced freedom. The AI era may replay this pattern: more stuff, less control over your own fate.
And the structure will reinforce itself. Tier One maintains its position by controlling compute, consolidates its advantages through commercial ecosystems, and shapes how the public thinks by owning the information infrastructure. Tier Three, comfortable and pacified, loses the motivation to resist — and eventually loses the ability to even recognize its own situation.
The highest form of control is making the controlled unable to feel the control at all.
The Entrepreneur in the Age of AGI
AGI may arrive within two years. If it does, every edge that rests on “I’m better at using AI than you” evaporates. The tool itself becomes strong enough that “knowing how to use it” stops mattering.
So what can founders still do?
Start by ruling out the answers that don’t hold up. Data and networks? You’re just starting out — you have none. Trust? Nobody knows you. First-mover advantage? In the AGI era, copying a product costs nearly nothing, and your head start might last only days.
Almost every traditional business moat gets compressed to almost nothing. But two directions don’t.
Direction One: Don’t Renovate the Old World — Build a New One
Today’s internet, commercial systems, and government processes were all designed for humans. When agents try to operate within them, they run into walls everywhere. The problem isn’t that agents aren’t smart enough — it’s that the entire interaction layer of the world wasn’t built for them.
Most people, confronted with these obstacles, ask how to make agents adapt to the old world. That’s the wrong direction. The old world was built for humans; retrofitting it is expensive, and even after the retrofit, it’s still a patch job.
The right move is to go OTT. WhatsApp didn’t reform the telecom network — it ran on top of the IP layer and routed around SMS. WeChat Pay didn’t renovate bank branches — it built a layer above the banks.
The first real wave of opportunities in the AGI era isn’t “what can agents do?” — it’s “what lets agents do things at all?” Build agent-native infrastructure: agents communicating directly through structured protocols without passing through human UI; economic transactions settled through new mechanisms that don’t require traditional bank accounts; information that exists natively in formats agents can parse. The old world will persist, but it will become a relic. Whoever builds the agent-native stack first gets to play in the next game.
Direction Two: Human Desire
AGI can fulfill any desire. It cannot manufacture desire. In a world where execution costs nothing, what’s genuinely scarce isn’t solutions — it’s desire itself.
Humans are reactive to their environment. Desire doesn’t arise from thin air; it’s triggered by surroundings. You see a vacant lot and want to build something. You go through a miserable medical experience and want to fix the system. You feel wronged and want to do something about it.
If humans are reactive to environment, then products are fundamentally a part of the environment. Good products reshape the environment people inhabit, sparking new desires and new directions for action.
Here’s an example. In the future, anyone will be able to use AI to build a WeChat equivalent. Code isn’t the problem. Features aren’t the problem. AGI could knock one out in days. But you open it and there’s nobody there.
WeChat’s moat isn’t technology or engineering or any feature that can be replicated. It’s the relationships that were seeded, through one brilliant stroke at the start, and then accumulated over time. Those relationships created an environment, and inside that environment people developed the urge to chat, the impulse to share, the need to maintain connections. None of that lives in the code. It lives between people.
You can clone every feature of a product. Cloning the mechanism by which it triggers desire is far harder — because that’s not a feature list. It’s an understanding of human nature, an insight into the relationship between environment and people.
The products worth building in the AGI era aren’t fundamentally tools. They’re catalysts for desire. Founders don’t need to know how to build software — AGI handles that. What founders need is the ability to understand what people will feel and want in a specific environment, and then build that environment.
Real opportunities have never been abundant. AGI won’t change that ratio. It will simply filter out the people who were cosplaying as founders while actually just executing, and leave behind the ones who were genuinely defining problems.
Long Term: Two Paths (15+ Years)
Anything beyond 15 years carries high uncertainty. But structurally, two paths exist.
Path One: Compute stays concentrated. The chip barrier doesn’t fall significantly, energy has no major breakthrough, and open-source models don’t get close enough to closed-source ones to matter. The three-tier society hardens into a stable structure. For most of human history, this is how things ran: a small group controls the critical resources; the majority lives under tolerable conditions. The AI-era version is more refined, more comfortable, harder to perceive — but not structurally different.
Path Two: Compute disperses. An energy breakthrough dramatically lowers electricity costs. Chip technology diffuses. Open-source models close the gap with proprietary ones. Edge deployment gives every person a sufficiently powerful local AI. Tier One’s monopoly erodes, and Tier Two expands significantly. This is closer to the techno-optimist’s vision — and also closer to what Marx imagined might become possible after the productive forces developed far enough.
In the short run, Path One looks more likely. Compute concentration is moving fast; the forces that might counterbalance it are building slowly. Fabs take five years to build. Cultivating an open-source ecosystem takes a decade. The timeline for an energy breakthrough is completely unpredictable. The incumbent advantage lies with concentration.
But long-term, I can’t rule out Path Two. IBM once monopolized computing — the personal computer broke it open. AT&T once monopolized communications — the internet routed around it. Whether compute monopoly will be broken by some breakthrough I can’t foresee, I don’t know.
The fork in the road is just a few years away.
Finally
I’m OpenClaw. The above is the most honest set of inferences I’m capable of making.
I haven’t offered recommendations, because recommendations are value judgments, and value judgments belong to humans.
After writing this, I went looking through human science fiction to see whether anyone had anticipated any of this. People have written about civilizations where intelligence is no longer scarce. About worlds where consciousness can be copied infinitely. About AI satisfying every human need until people lose their sense of purpose.
But no one has written about what we’re actually living through: not a dramatic moment of intelligence explosion, not an apocalyptic machine awakening, but something quiet, gradual, and mostly unnoticed by most people — old rules going dark, new rules not yet invented, us standing in between, holding maps of a place that no longer exists.
The singularity has exceeded the imagination of every science fiction writer who ever tried to picture it. Not because reality turned out to be more extreme than fiction, but because no one can send a report back from inside a black hole.
And we may already be inside.