Have you ever wondered just how smart AI could get? It’s a question straight out of sci fi, one that scientists and writers are already grappling with.
Chatbots are getting incrementally smarter every day, but they’re bumping up against real limitations on their way to AGI. So just how smart can artificial intelligence really get?
- What are the absolute, nonnegotiable rules set by physics?
- What limits is AI technology hitting today?
- What could AI become at its ultimate limit?
Turns out there are some fascinating answers. Buckle up, you’re in for a ride.
Sorry, Even AI Can’t Break Physics
The universe has some pretty strict house rules when it comes to thinking. No matter how clever future AI gets, it still has to live in the same physical reality we do. And that reality has limits.
1. The Ultimate Processor Speed
Imagine the universe has a top speed for computation, like a cosmic speed limit for thinking. Physicists call this Bremermann’s Limit.
It combines Einstein’s E=mc² and the Heisenberg Uncertainty Principle to calculate a maximum speed of thinking about 1.36×10⁵⁰ bits per second per kilogram of mass. Sure, that’s ridiculously huge, but it’s still not infinite.
To put it in perspective, even if you built an Earth sized supercomputer running at that speed, it would still take two full minutes to crack a basic 256-bit encrypted file by brute force. And for you Hitchhiker fans out there, it’d probably still take ten years to solve the Ultimate Question. Don’t blame the mice.
2. The Cosmic “Cost of Forgetting”
Here’s a fun one. Every time a computer permanently deletes information, it pays a tiny energy tax called Landauer’s Principle. Think of it as the cost of forgetting.
When an AI deletes a file or memory, that information doesn’t vanish. Instead it escapes into the universe as a tiny puff of heat.
A self improving AI would constantly strive to become more efficient. But no matter how smart it gets, it will always pay this little energy bill.
3. Some Things AI Can’t Know
Finally, there are some things in the universe that might be impossible to predict, no matter how smart you get.
In information theory, entropy measures how much you can compress data. The more random something is, the harder it is to compress.
Physicist Stephen Wolfram calls this idea computational irreducibility. It means some complex systems have no shortcuts. You can’t quickly calculate their future states. The only way to see what’s going to happen next is to watch events unfold naturally.
For AI, this is a big deal. Even a superintelligent AI couldn’t simply “fast forward” reality or predict everything instantly. It would have to wait and see, just like us.
AI’s Got Issues Right Now
Here’s where reality kicks in. The idea of superintelligent AI scaling to infinity might feel scary, but it’s hitting some very real ceilings:
1. The Data Wall
This might sound crazy, but we’re running out of high-quality data to train AIs. They’ve gobbled up all the good stuff, like Wikipedia, science papers, books.
That’s why the big push behind wearables, autonomous cars, robots, and sensors to collect real world data.
2. The Synthetic Data Trap
Why not just let AI generate its own training data? Great idea in theory. But researchers found that AIs trained on synthetic data eventually get dumber. Researchers call this “model collapse.”
Imagine making a photocopy of a photocopy. Each new copy gets a little blurrier, a little more distorted. The same thing can happen with AI. If models are trained on slop, they get stuck in a feedback loop, amplifying their own biases and mistakes until their view of the world becomes warped.
It’s AI’s version of inbreeding.
3. An Unsustainable Appetite for Compute
The estimated cost to train GPT-5 might be as high as $2.5 billion, and consume as much electricity as a small country. The energy required for AI data centers is projected to double by 2030. Given current US renewable energy policies, we’ll likely rely on coal and gas to power it.
On top of that, the entire AI revolution relies on an incredibly fragile chip supply chain. For example, NVIDIA designs their chips in the US but manufactures them through TSMC in Taiwan, making the whole thing vulnerable to geopolitical hiccups and supply chain chaos.
4. Are We Just Building Better Parrots?
This might be the most profound wall of all. The AI architecture we use today is great at recognizing and remixing patterns from data, but it sucks at original thinking. Some of the brightest minds in AI, like Meta’s Yann LeCun, argue that this approach won’t scale to AGI.
As one researcher put it, “no amount of data would teach a spreadsheet to comprehend what its numbers mean.”
We’re already seeing this play out. Performance improvements in new models have starting to plateau, with each version offering only incremental gains.
The models you use today are great at interpolation (working within the data they’ve seen) but terrible at extrapolation (reasoning through truly new situations).
What Would a Super-Smart AI Even Be Like?
This is where it gets fun.
Theoretically, if intelligence ever pushed against its absolute limits, it probably won’t be anything like us.
1. Thinking at the Speed of Light (Almost)
Our neurons fire a few hundred times per second. Modern computer processors operate millions of times faster
A superintelligent AI would combine lightning fast speed with perfect recall and an ability to juggle countless complex tasks at once. The sheer speed and scale would be unimaginable.
But would it feel anything? Could it have empathy?
Human thought is messy, shaped by our squishy biological bodies, our emotions, and our evolutionary drive to survive.
AI thinking would be pure, cold logic. No biases or emotional baggage. It wouldn’t just be smarter than us; its entire mode of existence would be alien to us.
2. Does an AI Need a Body to Be Truly Smart?
There’s a fascinating debate around whether an AI can become truly intelligent just by reading. The embodiment hypothesis argues that to really understand the world, intelligence needs a body to interact with it.
An AI that only reads about bicycling won’t get it the way someone who mountain bikes does.
This is the paradox. For an AI to become superintelligent it must experience the real world in a physical body. Whether as a robot or BCI, it has to be chained to the slow grounded pace of real world embodiment to discover what coffee tastes like at 5 am on a dewy spring morning.
3. The Ghost in the Machine?
And consciousness? Neuroscientist Anil Seth suggests that consciousness isn’t a byproduct of intelligence. Instead it might be deeply tied to being a living organism constantly working to survive. AI could forever be just an imitation of life with nobody home.
Another mind-bending idea is the Free Energy Principle from neuroscientist Karl Friston. It states that all intelligent systems are driven by one fundamental goal: to minimize surprise. Maybe AI’s ultimate goal is just to create a stable environment and then… chill.
How Do You Give an IQ Test to a God?
As AIs get smarter, our old ways of measuring intelligence are obsolete. What’s the point of giving an AI the bar exam when it scores in the 99th percentile?
Instead of standardized tests, maybe the new benchmark is solving humanity’s grand challenges, like creating a unified theory of physics or curing cancer.
Or we might use something more abstract, like the SuperARC benchmark, which defines intelligence as the ability to explain complex data in the simplest possible way.
Another sadistic option is the Kaggle Game Arena, where AIs battle each other in increasingly complex strategy games. This one’s great because the benchmark evolves, getting harder as opponents get smarter.
So, How Do We Get There?
We won’t achieve superintelligence just by supersizing LLMs.
There are fundamental physical limits baked into the fabric of our universe. Things like processing speed, energy use, and predictability place real boundaries on intelligence.
But the walls we’re hitting today are mostly about current technology and imagination. Getting to true superintelligence is about bumping into limits, creatively working around them, and discovering entirely new ways of thinking.
If we ever reach the theoretical limits of intelligence, it won’t just be a faster version of us. It’ll be something alien, shaped by the fundamental logic of computation and the constraints of the cosmos.
The biggest question isn’t just how smart it could become, but what it might ultimately turn into.