Unnatural Intelligence
Where AI stands against living intelligence
In 1726, a traveler reached the city of Lagado ruled by a royal family on the floating island of Laputa. In the city was the Academy of Projectors funded directly by the king himself. Within the Academy was the The Engine.
It was created so that
“the most ignorant person, at a reasonable charge, and with a little bodily labor, might write books in philosophy, poetry, politics, laws, mathematics, and theology, without the least assistance from genius or study.”
This was a fictional device from the stories of Gulliver’s Travels, and it is thought to be the earliest known reference to anything resembling a computer chip.
Today, The Engine is very real.
The latest advancements in AI have reached a point far beyond the original Imitation Game posed by Alan Turing. He believed that if a machine could fool a human into believing it was talking to another person, then it could be considered intelligent. This became known as the Turing Test.
Large language models like the famous ChatGPT are more than capable of passing this Turing Test. It can write anything from song lyrics, to short stories, to a cover letter, all with strangely natural sounding language. To Turing’s amusement, it can also imitate different voices whether that’s a cartoon character, a movie star, or even Turing himself.
Generative AI can even turn text into pictures that have never previously existed before. It can create entirely new characters or a complete scene that would rival some of the best concept artists in the world. One piece was awarded first place for digital art at a state fair causing some controversy. Like the language models, it can also closely mimic the styles of numerous famous artists from Pollock to Rembrandt to traditional Japanese Ukiyo-e.
It was always long believed that humanity’s use of language was an ultimate sign of our superior intelligence. It was something only humans could do that separated it from the animals. Our use of language is what allows us to build and transfer knowledge across generations through the ever-growing library of stories, textbooks, and articles. But now that machines have learned this capability, it has many people fearing the inevitable singularity, the point when machines reach human intelligence.
But is there really anything to fear?
To answer this question, we must first define the thing that we cherish and fear the most: intelligence itself.
What is Intelligence?
What makes something intelligent has never been clear and is always fraught with controversy. In the Western tradition it is thought of as the ability to reason, an ability gifted to humanity and unavailable to the animal kingdom. Eastern notions included reasoning as well but also placed emphasis on social and spiritual aspects like humility, doing what’s right, and knowing oneself. Often words like intelligent, smart, wise, talented, or gifted are used to point to some special ability while being equally vague in what that special something is.
Modern concepts of intelligence have focused heavily on specific kinds of reasoning and cognitive abilities in the form of IQ tests. They test mainly pattern recognition based on visio-spatial, numerical, and verbal abilities using little puzzles to complete a sequence to finish the pattern. Each question feels a lot like a coding problem, where the purpose is to find the right mental algorithm and apply it. While IQ was a decent predictor of academic test scores — what it was originally designed for — it was overly restrictive and failed to account for other critical skills like social, emotional, and physical abilities.
Creativity is also vital for the invention of new ideas, techniques, and technologies. Many Nobel Prize winners have a variety of creative and artistic interests aside from their main academic ones. Memory and speed also seem to be related. Smart people are often able to store and recall lots of information in great detail. They can also solve problems or at least make progress faster than others. Some kinds of work require quick thinking, some are marathons that last entire careers or lifetimes.
Another way of looking at intelligence is through its construction.
The idea of a thinking machine has existed ever since the industrial revolution introduced all of its mechanical marvels. Long before computers, there were automata, precise machines that executed complex calculations or actions. With the invention of clocks and watches, it was evident that machines could do just about anything with the right parts. All kinds of early robots were made to write, play music, or even fly like birds, and the earliest calculators were purely mechanical. However, they were limited to the very specific actions they were designed for. Most machines were single purpose until the idea of generalized, programmable computing came along.
With software the purpose could be changed without new parts. The right series of functions and instructions is capable of doing everything from analyzing spreadsheets to streaming movies to simulating entire virtual worlds. Computers have given us ways to search, store, and manipulate information that used to be reserved for the greatest minds and in many ways even exceed them. Yet, we would still hesitate to call apps intelligent. They feel more like fancy tools, like digital automata injected with more complicated abilities. Like other tools, they are only as useful as the person using it. They can’t learn, discover, or create anything on its own. This is something even the simplest of living creatures could do far better than any software ever created.
Living Intelligence
An insect, a lizard, or a bird can learn to control its own body, understand its environment, survive, and mate all by simply being born. It has an innate system able to take in information from all kinds of sources and map it into useful knowledge. It evolved to generate its own intelligence without needing to be designed.
Neurons first helped to coordinate signals from the strange and diverse multicellular bodies of the earliest creatures. A kind of movement intelligence. Then as the sensory arms race exploded, touch, smelling, hearing, seeing started to flood in huge amounts of new information that needed processing. The brain started to emerge to decode the senses and coordinate it with movement. It gained sensory intelligence. This gave rise to groups of brainy creatures that interacted and coordinated their collective behavior. They would communicate with one another, cooperate to avoid predators, compete to attract mates. They became socially intelligent. As their sphere of information grew, their data context expanded, so did their ability to learn and adapt to them.
The abilities that feel so automatic to us are often taken for granted. What a tiny fruit fly or goldfish can accomplish are things that entire teams of smart engineers still struggle to reproduce in machines. Robots have trouble balancing or moving on feet and are unable to adapt to different terrain very easily. Their movements feel robotic and can get stuck often. Phones are packed with sensors like cameras, microphones, GPS, and gyroscopes but they are mainly there to capture information rather than understand it. This has gotten better with computer vision and natural language processing, yet the most they can do is detect features and objects, and only the ones it was trained on. They don’t really understand what they are sensing and still need human input to train and correct them. And forget about socializing. Without being able to truly understand its own parts and environment, social interaction won’t be happening for a while. We haven’t even talked about pure abstract intelligence yet.
Outside of movements, senses, and socializing lies what may be considered pure intelligence, the kind of intelligence represented by abstract ideas and concepts rather than direct data from the world. It is how we can think of a million without a million objects, or how we can understand forces without having to see objects move or interact. Abstract rules and logic allow us to deduce and induce solutions to problems, to observe and uncover truths about the universe, to create new and fantastical things. This harkens back to the pure reasoning that seems to separate humans from other creatures. While many smart animals show high levels of all previous intelligences, none have invented new tools from their imagination, carved out stories and legends that last generations, created ever growingly complex civilizations, or discovered laws that govern our world.
What all of these representations of intelligence all have in common is the ability to gather and make sense of information, to find and store patterns, and then apply and recombine those patterns in order solve problems and adapt. What any type of intelligence is intelligent about depends on the information it is processing, its data context. Each kind of sensory modality, each kind of movement and balancing, each kind of social situations would each have their own level of understanding and refinement. More importantly, it is able to integrate and coordinate this knowledge as a whole. It isn’t just controlling its body, sensing the environment, or socializing with others. It is doing all of it simultaneously, automatically using the most important pieces at any given moment to govern its decisions.
Using this general definition anything can be intelligent, whether it’s made of code, cells, or cities.
Intelligent Systems
Biologists have shown that even clusters of cells have their own kind of intelligence. Slime molds, when placed in a dish with different food sources, are able to find and connect itself using the most efficiently placed paths. It is able to navigate mazes, and, even more impressively, when it is given the layout of Tokyo train stations, was able to create a more efficient system than the one designed by humans. Biologists like Michael Levin, have also demonstrated that cells have their own bioelectrical language that allow cells to communicate with one another. These electrical gradients can tell organisms where to grow certain body parts or when to repair. Flatworms can grow two heads or no heads when the electrical ion channels are changed without changing any DNA information. What matters isn’t the proteins encoded by genes, but how genes are triggered when cells talk with one another through their cellular communication.
The same concept can be applied to social signals on the complete opposite scale of cities and nations. Societies need to solve many problems in order to survive. It needs enough resources to feed and shelter its population, it needs roles and rules to gather and distribute those resources, it needs systems to enforce those rules. How a city develops and grows depends on how work is organized, the flow of money, and the governing bodies that manage that work. And when the underlying rules and laws no longer sustain the wellbeing of the population, the people tend to organize and challenge the system to rewrite those rules.
Going back to our original question, are AI’s something we should fear? Are they really intelligent?
Nothing to Fear
Artificial intelligence has come a long way, especially compared to the dream of The Machine and mechanical automata. They have thoroughly destroyed Turing’s Imitation Game and will undoubtedly change the world we live in. Yet, they still have a long way to go before they have any chance of competing with living intelligence. They may never even get there.
AI are inherently limited to the digital domains they are trained on. For LLMs like ChatGPT, their world is made up of text and characters. While they give a convincing representation of writing that can rival or exceed real people, they have no concept of what any of the words actually mean. They learn language without ever seeing what they refer to, knowing what it feels like, what it smells like, its weight, size, or anything at all. They are confined to a dark virtual prison of language data that has no association to the outside world. Other impressive AI like AlphaGo or AlphaFold similarly live in their own prisons of a game and proteins.
Self-driving cars do have this awareness of the world, yet it is limited to the narrow functionality of driving. Its vision is tuned to detect only the things relevant to driving like roads, signs, and other people and cars. Its movement is undoubtedly advanced and intelligent. It is also making decisions in real-time by integrating its vision with its controls. But that is the extent of what it can learn and do. It can’t teach ChatGPT more about the world or Midjourney how to drive within it. It also doesn’t know or care about the passengers inside or outside of it.
To create a frighteningly intelligent AI, the kind to challenge humanity, it would have to live and breathe the world around it. It would have to be immersed in reality to give its knowledge any real meaning, to give it some ground truth. It would have to be able to combine and integrate multiple kinds of intelligences together to move, to explore, to understand high-level, abstract concepts in real-time. It can’t be confined to just text data or driving data or vision data. It needs more than one prompt at a time, one purpose at a time. In a sense, it has to live and learn from the world at its own accord. And this would need a huge amount of computational power that may not be possible. It isn’t certain what the processing power of the human brain is, but it’s likely far more than we imagine.
Every generation imagines distant futures, they dream where their technology will take them. With Gulliver it was The Machine. With the industrial revolution it was mechanical automata. During the space race people dreamt of flying cars, moon bases, and another life on Mars. Every time we imagine the future, we take the technology far beyond the limits we have yet to discover. It turns out going to space is extremely hard and expensive, most people probably shouldn’t fly anything, and teaching machines to move and learn is incredibly difficult. As much as computers have changed the world, they too have their limits. AIs will likely change how we interact with machines forever. We will ask them questions, play and create with them, and use them to assist us in all kinds of new ways. But we also may never have a proper conversation with them, they may never understand us, or gain any real understanding outside of the virtual realm. They may help discover new cures, create metamaterials, and solve century old problems, yet never understand what it’s doing. We may get the AI moon landing, but never get to Mars.
So, for now, we can continue to marvel at all the dreams and nightmares of artificial intelligence. We can speculate at what may come and imagine new possibilities because that is what our intelligence allows us to do. It is what it compels us to do.