Tech dirty secret: AI is like Santa Claus
Artificial Intelligence (AI) is a children story. It doesn’t exist.
Kids love Santa Claus. He has a funny beard and outfit. More importantly, he brings gifts. When children finally learn that Santa was a lie, the news is often devastating. But when the gifts continue to come the next year, they quickly get over it.
Similarly, I think that AI is a sci-fy lie. But like Santa Claus, it’s a lie that brings gifts.
Why am I saying that Artificial Intelligence doesn’t exist?
I think it’s helpful to start from the beginning. In 1956, the field of computer science was still young, but researchers were optimistic. Marvin Minsky and John McCarty, together with Claude Shannon and Nathan Rochester launched the Dartmouth Conference with the avowed goal to solve the artificial intelligence problem “in one summer.” They defined AI as “making a machine behave in ways that would be called intelligent if a human were so behaving.”
In retrospect, their optimism seemed charmingly naïve. After some initial progress on subproblems such as chess-playing machines, the field ran out of steam in the 1970s, and most people lost interest. The period from 1974–1980 is often called the “1st AI winter”. A second boom happened in the 1980–1987 period, mostly fuelled by the hype about expert systems. Then from 1987 to about 2011, the field went dormant again.
The research was continuing for sure. However, the focus was specific sub-problems, and few people were speaking about AI. I studied engineering in college during that period. There was no AI department nor AI classes to enroll in.
Something changed in 2012 however. That year, the AlexNet Deep Neural Network won the ImageNet competition. Why was it a big deal? From the Image-net website:
The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) evaluates algorithms for object detection and image classification at large scale. One high-level motivation is to allow researchers to compare progress in detection across a wider variety of objects — taking advantage of the quite expensive labeling effort. Another motivation is to measure the progress of computer vision for large scale image indexing for retrieval and annotation.
In short, ImageNet is a competition used as a benchmark to compare image recognition algorithms. In 2012, the best performing algorithm was a deep neural network.
Neural networks were not new at all. The initial perceptron algorithm was proposed in 1958. It was an extremely simplified model of a single neuron. Rosenblatt, it’s creator, showed that it could be useful to detect simple patterns in images.
Let’s note that the perceptron has almost nothing in common with biological neuron cells. There are dozens of types of human neurons and modeling how they work is fiendishly complicated. Even the most recent neural network architectures don’t try to mimic the brain accurately. In fact, the term neural network is so much of a misnomer than Yann LeCun — the head of Facebook AI research — wants to rename the technique “differentiable programming.”
In 1986, Rumelhart, Hinton, and Williams showed that the backpropagation algorithm could be used to trained multi-layer “deep” neural network algorithms. In 1990, Yann LeCun used neural networks and back-propagation to classify handwritten digits. Even myself I had done some — very modest — work in the mid-2000s with neural networks to classify musical songs.
However, until 2012 and the AlexNet win at ImageNet, neural networks were mostly a research curiosity. Why? Because their performance was horrible. Deep neural networks need a vast volume of data and computing power to even reach acceptable accuracy. If those conditions are not met, many algorithms can do better, starting with the humble least square regression.
In the early 2010s however, things were changing. Google, Facebook, and other tech companies had amassed enormous quantities of data. Nvidia, a chip maker specialized in graphical processor units (GPU) for video games and 3D modeling, had realized that you could use their chips for non-graphical applications and released the CUDA development framework. GPU chips turned out to be vastly more effective than the usual CPU for machine learning applications. Suddenly researchers had access to vast quantities of data and compute, and the neural networks started to perform.
Then someone had a brilliant idea. Neural networks are loosely inspired by the functioning of the brain. Neural networks can perform some “intelligent” tasks, such as recognizing cats in pictures. Therefore neural networks are artificial intelligence. AI is upon us. Journalists lapped it up. The AI hype quickly exploded.
Since neural networks are but one type of algorithms among many with similar capabilities, the entire machine learning field got rebranded as “artificial intelligence.” Like parents telling Santa Claus stories to their children, researchers and entrepreneurs of various stripes told the AI story to an unsuspecting public hungry for stories of intelligent robots. The hype had the desired effect. Government research agencies and investors flooded the field with money. Talented students and professionals followed the siren songs in large numbers. Running high on money and talent, research progressed by leap and bounds. Entrepreneurs and business people started prowling the conferences and mining the research papers for interesting use cases. Industrial applications popped like mushrooms after the rain. All this progress is exciting, and many of the applications are truly invaluable.
Let’s not forget where we started, however. Siri, Alexa, and the self-driving cars are great, but they have nothing to do with HAL9000, Skynet and agent Smith. Computer systems still have no agency, no real autonomy, no consciousness and they are still easily confused and misled. Real Artificial Intelligence is still science fiction.
What do you think? Is the term artificial intelligence misleading? Should the industry and the media continue using it?