Why Brains Beat AI

In complexity, plasticity, and social learning, brains still have an edge.

The predictions for super-intelligent AI are starting to scare the world — even former President Obama. But there are important reasons, grounded in the biology of the brain, that AI progress might not be as fast or scary as many think.

Regarding the predictions: a former OpenAI researcher, Daniel Kokotajlo, has been getting lots of attention for his prediction (featured in the New York Times) that top AI labs may trigger an unstoppable and dangerous cascade by 2027.

The basic logic is quite simple: we are getting to the point that AI is beating the human brain. The number of parameters in AI models is getting close to the number of synapses in the human brain, i.e., a few hundred trillion. And AI performance at a number of functions — most notably, software engineering — is reaching human level. If AI brains are as good as human brains across a wide array of cognitive tasks, this could set off a chain reaction to super-intelligence. AI brains, after all, would not be limited by the need for food, water, and physical space. We could create an infinite number of them, and they’d never need a day off.

So are we on the brink of the AI revolution? I am far from an expert in AI engineering. But a surprising amount of insight in AI has come from metaphorical reasoning. Indeed, the development of modern AI was based on a metaphor to the network architecture of the human brain. And I think there are a number of reasons that this metaphor to the human brain might break down — and lead to significant bottlenecks in AI progress.

The first is that biological complexity of the human brain far exceeds the complexity of an AI neural network. Some have made simple comparisons between the number of connections in the human brain (i.e., synapses) and the number of connections in the leading AI neural networks (i.e., parameters). As those AI networks reach a few hundred trillion connections/parameters, the argument goes, they will rival the brain in size, sophistication, and ability. But this reduces the brain’s complexity to a single dimension when, in fact, its biology goes far beyond neural connections. For example, the human brain is composed of neurons with dendrites — web-like structures that spiral out of each neuron. These dendrites play an important role in processing signals that go through the brain and releasing a shocking array of chemicals and neurotransmitters. There is nothing in modern AI that mimics this crucial brain mechanism. Yet this is just one example out of thousands of mechanisms that are completely absent in AI. Consider: a leading introductory text in neuroscience describes hundreds of mechanisms of brain function. (And that’s just an intro!) Modern AI mostly exploits only one of these: synaptic connections. If intelligence is a product of more than just synapses, then AI progress may stall.

The second reason is that the complexity of the brain is constantly changing due to plasticity while AI is a static architecture. Innovation in AI has been driven in part by experimentation in the architecture in which AI neural networks are built. For example, the convolutional neural network has an architecture similar to the optic nerve of the human eye, with specialized sensor cells arranged in a grid that resembles pixels on a screen. But there is one massive difference between the architecture of AI neural networks and the human brain: AI is a static network while the brain’s architecture constantly changes. This is a concept of neuroscience called “plasticity” and it is fundamental to neuroscience. Brain neurons are constantly wiring and re-wiring themselves into new architectures via mechanisms that are very poorly understood. (This is part of the reason why psychedelic drugs remain such a mystery.) There is evidence, moreover, that plasticity is important for innovation. A brain that can’t rewire itself can’t make creative leaps. Without this crucial ability, it’s plausible that AI won’t be able to replicate human thinking and research.

The third reason is that human brains evolved to collaborate with other brains while AI evolved as an individual agent. Harvard’s Joseph Henrich argues that human beings are intelligent and powerful mostly because of our ability to engage in social learning. The breakthrough that you make can be learned by me, through evolved attributes like theory of mind and language. And when I then take your innovation and make another innovation myself, you can learn from me, too. This dynamic feedback loop between human brains is the primary mechanisms through which human beings innovate. But we have little understanding as to how social learning actually works. Indeed, we don’t even have an understanding of consciousness, which is plausibly a building block upon which social learning is built! Without a more foundational understanding of how the brain engages in social learning, it’s possible AI simply can’t innovate the way we do. And we are very very far from this sort of understanding.

I don’t know if any of these three points are wrong. And, of course, there are many other reasons AI progress might stall — most obviously, because of the law of diminishing returns. But I don’t see these 3 points discussed enough in modern AI discourse.

What do you think?