Plenty of people open ChatGPT every day without being able to explain how it actually learns. The same goes for the match between AlphaGo and Lee Sedol that the whole nation watched. It's remembered as the moment a machine outran human intelligence, but the real story behind that victory was how deep learning works — data, computing power, and algorithms all clicking into place at once. This piece walks through why old-school AI, built on rules coded in one by one, hit its limits, and how self-learning AI that trains itself on data took over — told through two landmark moments: AlexNet and AlphaGo. By the end, you'll be able to explain how deep learning works to someone else in a single paragraph.
Why Rule-Based AI Hit a Wall
The place to start is what AI was doing before deep learning came along. Early artificial intelligence ran on rules that humans wrote in by hand, one at a time. To get a system to recognize a cat, a programmer had to pre-define conditions like what shape the ears should be or how many whiskers to expect.
The problem is that the world's data never stops growing. As the volume of data to handle increases, so must the rules needed to extract anything useful from it. But rules written by hand only work within a range of cases a person can actually enumerate. Every time an exception popped up, another rule got bolted on — until the rules started contradicting each other while new edge cases kept piling up. The limits of whoever was writing the rules became the limits of the entire system. That's where rule-based AI stalled out.
The Moment Data, Computing, and Algorithms Converged
The breakthrough came from flipping the approach on its head. Instead of humans writing the rules, why not feed the machine massive amounts of data and let it find the patterns itself? For this kind of self-learning to actually work, three conditions had to come together at once.
The first was raw material to learn from — data at scale, properly organized. ImageNet, a massive dataset with labels attached to every image, filled that role. The second was a computer that could handle the sheer volume of calculations involved. GPUs like Nvidia's GTX 580, originally built for rendering game graphics, turned out to be excellent at parallel computation, which brought training times down to something workable. The third was an algorithm capable of extracting features from data in layers — the deep neural network that came to be known as AlexNet.
None of this went anywhere as long as the three pieces existed separately. It was only when the raw material, the computing hardware, and the algorithm converged at a single point that machines learning on their own — the actual mechanics of deep learning — started to run for real. That convergence is where modern AI begins.
Did AlphaGo Really Surpass Human Intelligence?
Go was long considered territory no machine could touch. The number of possible positions is effectively infinite, so calculating every move to find the optimal one was never going to be feasible. What AlphaGo did was change the frame for dealing with that infinity.
The key move was to stop treating Go as an endless chain of choices and instead reframe it as a bounded probability problem. Among the enormous number of possible moves on the board at any given moment, which one raises the odds of winning the most? AlphaGo learned to estimate that probability from data. Rather than exhaustively searching every possibility, it narrowed the field down to the highest-odds candidates.
So reading AlphaGo as proof that machines had surpassed human intelligence itself misses what actually happened. The machine didn't win by playing Go the way a human does — it won by redefining the problem into a shape it could solve. That redefinition is the essence of how deep learning works: the skill of turning an unmanageable infinity into a probability problem you can actually tackle.
How Deep Learning Works: A Practical Framework
Boil it down and there are just three things worth holding onto.
- Are the rules written by a person, or extracted from data? Whenever you're adopting or evaluating an AI system, start by checking whether it runs on rules a human maintains by hand or on a structure that updates itself from data. In any domain where data keeps growing, the hand-maintained approach won't hold up for long.
- Are all three conditions in place together? Performance never comes from the algorithm alone. Check whether you have the training data, the computing resources, and the right algorithm all lined up as a set. Miss even one, and there's no result.
- How has the problem been framed? Strong AI results usually come from reframing a hard problem into a shape a machine can actually solve. When you run into a new challenge, the first question is whether there's a way to narrow an infinite space of possibilities down to a bounded set of choices.
Whether it's ChatGPT or AlphaGo, the machines didn't suddenly get smarter — what changed is the shape of the problem people handed them. Hold onto that one line about how deep learning works, and you'll be able to size up whatever AI you run into next.
