Many of us open ChatGPT every day without being able to explain how it actually learns. The same goes for the nationally watched match between AlphaGo and Lee Sedol. It's remembered as the moment a machine surpassed human intelligence, but the real story behind that contest was how data, computing power, and algorithms locked together to make deep learning work. This piece walks through why old-school AI—built rule by rule—hit its limits, and how self-learning AI that trains itself on data took over, using two landmark moments: AlexNet and AlphaGo. By the end, you'll be able to explain how deep learning works in a single paragraph.
Why Rule-Based AI Hit a Wall
Let's start with what AI was doing before deep learning came along. Early artificial intelligence worked by having humans write out decision rules by hand, one by one. To teach a system to recognize a cat, a programmer would predefine conditions like the shape of its ears or the number of its whiskers.
The problem is that the world's data never stops growing. As the volume of data to handle expands, the rules needed to extract anything useful from it have to expand right along with 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 situations kept piling up. The limits of the people writing the rules became the limits of the AI itself. That's where rule-based AI stalled.
The Moment Data, Computing Power, and Algorithms Converged
The idea that broke through that wall came from flipping the direction entirely: instead of humans writing the rules, feed the machine massive amounts of data and let it find the patterns itself. For this self-learning approach to actually work, three conditions had to come together at once.
The first was learning material: large volumes of organized data. ImageNet, a massive dataset of labeled images, filled that role. The second was a computer capable of handling all that calculation. GPUs like Nvidia's GTX 580—originally built to render video game graphics—turned out to be excellent at parallel computation, which brought training times down to a realistic range. The third was an algorithm capable of extracting features from data layer by layer: the deep neural network known as AlexNet.
As long as these three existed separately, nothing happened. The moment the material, the computing hardware, and the algorithm converged, the mechanics of deep learning—machines teaching themselves—finally started running for real. That convergence is the starting line of modern AI.
Did AlphaGo Really Surpass Human Intelligence?
Go was long considered a domain machines could never touch. The number of possible positions is effectively infinite, so calculating every move to pick the best one simply wasn't feasible. What AlphaGo did was change how it approached that infinity.
The key was reframing Go not as a game of endless, open-ended choices, but as a game of bounded probability. Among the enormous number of possible moves on the board at any given moment, AlphaGo used data to learn which ones offered the highest probability of winning, then narrowed the field to those. Rather than searching every possibility, it zeroed in on the options most likely to pay off.
So reading AlphaGo's win as machines outpacing human intelligence itself misreads what happened. The machine didn't win by playing Go the way a human does—it simply redefined the problem into a form it could solve. The essence of how deep learning works lies in exactly this redefinition: turning an unmanageable infinity into a probability problem it can handle.
How to Apply This: A Practical Checklist for Deep Learning
To sum up, there are just three things worth keeping in mind.
- Are the rules written by people, or learned from data? When adopting or evaluating any AI system, first check whether it runs on rules maintained by hand or rules that update themselves 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 doesn't come from an algorithm alone. Check whether you have training data, sufficient computing resources, and the right algorithm all lined up as a set. Miss just one, and there's no result.
- How was the problem framed? Strong AI results usually come from reframing a hard problem into a shape a machine can solve. When you encounter a new task, first ask whether there's a way to narrow an infinite set of possibilities down to a bounded set of choices.
Whether it's ChatGPT or AlphaGo, the machines didn't suddenly get smarter—the shape of the problem people posed to them changed. Hold onto that single idea about how deep learning works, and you'll be able to size up whatever AI you encounter next.
