Artificial intelligence has humans beat – at least when it comes to games like chess and Go, identifying the 3D structure of proteins, generating investment strategies…the list goes on and on. Some argue that models like ChatGPT are already at the threshold of human intelligence. OpenAI head Sam Altman even threw his unborn child under the bus, claiming “my kid is never gonna grow up being smarter than AI”.
The capabilities of modern AI are certainly impressive, but I am deeply sceptical about comparisons between humans and machines. AI (at present and in the foreseeable future) isn’t all that smart, or at least it isn’t in the way that humans are – and that’s a good thing.
Advertisement
Learning is at the heart of intelligent behaviour, and humans learn differently to machines. AI models learn incrementally, ploughing through a massive amount of training data. The power required to do so can take down an entire power grid. In contrast, the power requirements of the human brain are comparable to an incandescent light bulb, in part because they are designed to learn with very little data. An AI might process every pixel of an image, whereas humans extract information from a few glances to construct visual memories.
Despite their economy, human brains are remarkably flexible, compared with the brittle nature of contemporary AI. When cutting-edge models are given a continuing stream of new information that deviates from what had been previously learned, it can result in “catastrophic forgetting“, which, for machine learning, is as bad as it sounds. To get around the problem, it is necessary to turn off learning in a fully trained model before releasing it into the wild.
Humans, in contrast, continually learn throughout their lifetimes without fear of catastrophic forgetting because they combine semantic memory, which reflects gradually accumulated knowledge of the world, and episodic memory, which reflects memories of specific events. A child could rely on semantic memory to learn that birds generally have feathers, beaks and wings that they use to fly. When they see that a penguin, which has similar features, can’t fly but swims, episodic memory allows them to rapidly learn this exception without forgetting the typical features of birds.
I am certain that the next generation of AI will incorporate some kind of episodic-like memory, but I suspect that engineers wouldn’t want to fully emulate human memory. As I describe in my book, , our memories can be startlingly fragmented, biased and malleable. The selective and sometimes inaccurate nature of human memory makes us poorly suited to well-defined tasks like chess, but it enables us to flexibly navigate an uncertain and rapidly changing world. Humans lack the comprehensive body of knowledge incorporated in models like ChatGPT, but we can look to our episodic memories from our lived experiences to generate unique innovations and works of imagination.
Comparisons between human and artificial intelligence are misguided because they reflect different design constraints. Human brains, which are built for survival and reproduction in the physical world, squeeze as much information as possible from very little data and energy, whereas the best AI applications can discover needles in massive data haystacks that would overwhelm our resource-frugal brains.
Rather than attempting to surpass human intelligence, we are better served by machines that complement our own idiosyncrasies. And maybe Sam Altman should be more optimistic about the fate of his progeny.
Charan Ranganath is author of the book Why We Remember: Revealing the hidden power of memory
