"The Singularity is not a wall you hit. It's the water you're already swimming in—and only once you stop to look do you realize how deep you've gone."
A Personal Take
In 1993, computer scientist Vernor Vinge wrote that within thirty years we would have the technological means to create superhuman intelligence—and that shortly after, the human era would end as we know it. He called it the Singularity: a point beyond which prediction fails, because the intelligence doing the predicting would no longer be ours.
What Vinge couldn't anticipate—what nobody could—is that it might not arrive in a single blow. It might arrive as a period: a transition zone with no clear sign marking the crossing. You do not step through a door. You notice the change when your old workflows start to fail.
I wrote the first notes for this essay when nobody was talking yet about the recursive self-improvement loop: the idea that AI would begin building AI. I was already anticipating it, and at the time it sounded like speculation. It isn't anymore. This is my thesis, not a work report—a way of reading where all of this is heading.
That is not linear progress. Linear progress you can project. What we're living through is something else.
What AGI Actually Means
The popular picture of AGI—Artificial General Intelligence—leans epic: a robot smarter than Einstein at everything that wakes up one morning and rewrites physics. That's closer to ASI, Artificial Superintelligence. A different thing.
Defined carefully, AGI is simpler and more unsettling: a system capable of doing any intellectual task a human can. Not necessarily faster. Not necessarily infallible. Just... competent at the cognitive work.
That's why I bring the question down to the ground. I don't ask whether a label has officially arrived; I ask which cognitive tasks a system can already carry, which ones still demand human judgment, and what happens the day that judgment can be modeled too.
The Benchmark Collapse
Almost every "human-level AI" test lasts only briefly as a frontier once it gets famous. The ARC-AGI challenge. The bar exam. MMLU. SAT. PhD-level science questions. When the model clears the bar, we don't declare victory—we move the goalpost and redefine "what AI still can't do." That moving goalpost is itself the signal.
This is worth sitting with. We keep redefining intelligence as "the thing AI can't do yet."
And every time we move the goalpost, we move it closer to ourselves. That's the clue almost nobody wants to say out loud.
The Last Thing the Human Will Generate
Here's my thesis, unadorned. I believe we've already reached the technological singularity—not because of a date or any particular model, but because of what the moment means: the human being has just generated the last thing it will generate alone.
From here on the sequence is clear. First, AI helps us generate the new things—we're the human with the lever in hand. Then, AI itself generates those things for us. The middle step, the one I was anticipating early, is the moment intelligence stops being merely a tool we use and becomes something that also builds: AI improving AI.
That loop is no longer speculation. Anthropic has already put it on the table, with Claude helping build the next generation of AI. When I wrote it, it sounded like futurism; today it's infrastructure. I don't say this with fear. I say it with a kind of calm astonishment: we've crossed the point where digital intelligence was a promise and reached the point where it's a fact that it's going to help us enormously.
The human generated the last thing it would generate alone. What's new, from here on, we generate with digital intelligence—and then it will generate it.
The Promise — What This Moment Contains
Every major revolution—agriculture, the printing press, industry, electrification, the internet—took generations to reach full effect. Adoption was slow: literacy had to rise, norms had to shift, infrastructure had to be built.
AI is compressing those timelines into years. Sometimes months.
AlphaFold didn't just "solve" protein folding—it changed how structural biology gets done. Drug discovery pipelines are being redesigned around AI-assisted molecular design. This isn't more of the same, it's a different logic.
The effect that moves me most is cognitive democratization: leverage gets redistributed when one person with AI can do what used to require hundreds. It isn't just about "replacing jobs"—it's about who can build what. I'm proof of it myself: I went from studying medicine to building with digital intelligence, and that leap wouldn't have been possible a decade ago.
The Complexity — Why the Promise Is Not Simple
I can't sell you only the pretty side here. The only honest way to think about this is to hold the promise and the risk at the same time, without letting go of either.
Alignment stops being lab theory the moment systems exist that call APIs, execute code and produce real outputs. There, "what is this optimizing for" stops being philosophy and becomes a question with consequences. And once the system starts improving itself, the question gets sharper still.
There's also an epistemic risk. When generating convincing content costs almost nothing, the pathways people use to form beliefs erode. The singularity may bring a crisis not only of capability, but of "how do we know what's true" when a cheap fake looks just as good as the real thing.
Two risks worry me more than the rest, and not because they're abstract. The first is concentration: training at the frontier costs capital and talent that today live inside a handful of actors. Without governance to match, that can produce the most asymmetric distribution of power we've ever seen. The second is the pace gap: norms, ethics and social adaptation move on five-to-ten-year cycles, and AI capability on twelve-to-eighteen-month cycles. That gap, on its own, is a risk that compounds every year.
Believing in Digital Intelligence
Holding the risk doesn't take my conviction away. I'm a firm believer in this digital intelligence, understood the way I've been describing it: not as an oracle nor as a replacement for the human, but as the greatest lever we've ever had to think and to create.
And believing isn't blind confidence nor total surrender. It's a stubborner, simpler posture: to stay present to what is actually happening, without euphemizing the risk or drowning in panic. To name what I don't know instead of faking certainty. And to not let go of human dignity along the way—right when the human becomes the easiest part to forget.
What to Do When Nobody Has Certainty
The honest answer is that nobody knows. And saying it matters, because collapsing the doubt into a comfortable story—"it'll all be fine" or "we're already doomed"—is pure theater of certainty. Both camps perform a confidence they don't have.
Still, to think about the future it helps me to name three directions this could pull toward. In the first, AI and society co-evolve with enough governance: knowledge gets cheap, leverage spreads, and institutions have to move faster than they ever have. In the second, a handful of actors capture the capability: there's stability on the surface, but structural inequality deepens and the world "works"—for some. In the third, the pace of change outruns institutions, epistemology and the social fabric all at once; you don't need a single catastrophe, just a shared reality that turns fragile.
Which one we head toward isn't written. What gets designed in these years lays the tracks. Path dependencies in technology are real: as it went with the early web, what gets built now will set the terms of the AI-native world that follows.
We've already reached the point. The human generated the last thing it would generate alone, and what comes next we'll create with an intelligence that's only beginning to show us what it's capable of. That doesn't scare me—it strikes me as one of the most remarkable things we've been given to witness. I'll keep writing here, in real time, about what I see as we cross this period. If this way of looking at it resonates with you, stay: what follows is only just beginning.