Can AI Become Superintelligent?
Date:
Tags: Superintelligence, Scaling, Future of AI, Reflection
If you spend enough time reading about artificial intelligence these days, you quickly run into two camps that sound like they are living on completely different planets.
In one corner, you have the true believers in Silicon Valley who talk about Artificial Superintelligence as if it is already scheduled on a calendar—just a matter of pouring a few more hundred billion dollars into data centers, stacking enough GPUs and compute clusters, and waiting for the curve to go vertical. In the other corner, you have skeptics who look at a chatbot hallucinating a simple historical date or failing to count the letters in a word and say: Come on. This is just fancy autocomplete. It is never going to wake up and outsmart humanity.
So who is right? Honestly, if you strip away the hype and the cynicism, the most honest answer is that there is no law of physics stopping AI from becoming superintelligent—but there is also no guarantee that the road we are currently on leads all the way there.
Why It Actually Seems Possible
Look at what has happened in just a few years. Not long ago, getting a computer to hold a coherent conversation, write working software from a vague description, or reason through multi-step math problems felt like science fiction. Today, we take it for granted on a Tuesday afternoon.
The strongest argument that AI can become superintelligent is brutally simple: human brains are made of ordinary matter, running on roughly twenty watts of power, bound by the laws of physics. We are proof that a physical system can organize atoms in a way that produces general intelligence, science, art, and strategy. Silicon does not get tired, it does not forget what it read ten years ago, and it can share knowledge across thousands of copies in milliseconds. Once a machine gets even slightly better than human engineers at designing better machines, the feedback loop becomes very hard to ignore.
Where Things Could Still Stall
At the same time, assuming an effortless straight line from today's models to a godlike intellect feels a bit too convenient.
Right now, most of our progress still relies heavily on brute force—more data, larger models, more electricity, more chips. Anyone who has watched how fragile the hardware supply chain is, especially during the recent turmoil around NVIDIA, knows that physical reality has a way of interrupting neat exponential charts. You cannot build infinite data centers overnight, and we are already running low on clean, human-written text to train on.
Beyond hardware and data, there is a deeper question: Is predicting the next token—even with reasoning loops bolted on top—really the same thing as true understanding? Animals and humans learn from living in a physical world with consequences. Most AI systems still live inside a text box. Maybe scaling current architectures gets us all the way to ASI. Or maybe it gets us to a brilliant, tireless assistant that still misses a fundamental ingredient of real-world common sense.
The Uncomfortable Middle Ground
What fascinates me most is that AI does not even need to become an all-knowing super-being to change everything. Even if progress slows down halfway to true superintelligence, a system that is as smart as the top one percent of human experts in every field—and available a million times over at the click of a button—already turns our economy, science, and politics upside down.
Still, betting against human ingenuity in pushing compute and algorithms forward has been a losing trade for decades. Whether it takes five years, twenty years, or requires a completely new breakthrough beyond transformers, the path toward superintelligence is wide open. The real question is probably not whether machines can eventually surpass us—it is whether we will be mature enough to handle them when they do.