Attention is no longer just a concept from psychology. It is the computational primitive at the centre of modern artificial intelligence: the transformer architecture that powers today's leading language and multimodal models is built on an operation that lets a network weigh which parts of its input matter most when producing each output.
In a new essay, Quanta Magazine asks whether that convergence points at something deeper — whether attention, long treated by cognitive scientists as the brain's way of filtering and prioritising a flood of information, might be the essence of intelligence itself.
The parallel is striking. Psychology has described attention as a scarce resource: we can hold only so much in mind, so we choose what to represent and what to ignore. Machine learning arrived at a similar idea from the opposite direction, discovering that a differentiable 'attention' mechanism was what made it possible to train models at unprecedented scale.
The essay explores where the analogy holds and where it breaks down. Attention inside a neural network is a routing of information shaped by training data and gradients; human attention is bound up with goals, salience, emotion and subjective experience. One is a mathematical function, the other a lived capacity.
Why it matters: framing intelligence around attention reframes some of the field's loudest debates — whether scale alone is enough, what today's models still lack, and whether the real bottleneck is architecture, data or something we have not named yet.
