[YOU] on AI encounters AI primarily through transformer architectures and the large language models they power. Spiking neural networks enter the cycle as the technical face of a different wager: that the capabilities missing from current AI—genuine contextual judgment, the perception of what a situation actually calls for, the responsiveness to others that makes benevolence possible—are missing because the architecture forecloses them, not merely because the training is insufficient. If Zeng is right that wisdom requires the kind of temporal binding, self-modeling, and social cognition that spiking architectures are better positioned to deliver, then the cycle's story of AI's limitations is partly a story about the wrong substrate.
The contrast with Judea Pearl's framework is instructive. Pearl argues that large language models are sophisticated curve-fitters that cannot cross the gap from association to causation. Zeng would extend this: they are also sophisticated pattern-predictors that cannot cross the gap from pattern-completion to wisdom, and the reason is partly architectural. Both diagnoses are compatible—the same systems that live on rung one of Pearl's ladder may also lack the spiking temporal dynamics that genuine self-modeling requires.
The basic concept of a spiking neuron predates digital computing; it was formalized by Alan Hodgkin and Andrew Huxley in their 1952 model of nerve impulse propagation, work that earned them the Nobel Prize in Physiology or Medicine. The first artificial spiking neural network models appeared in the 1980s and 1990s, associated with researchers including Wolfgang Maass, who coined the term 'liquid state machines' for recurrent spiking networks, and Gerstner and Kistler, whose textbook formalized the integrate-and-fire neuron model. The field developed in parallel with mainstream deep learning but was largely overshadowed by the latter's extraordinary empirical success on standard benchmarks.
Interest in SNNs revived as two separate pressures accumulated: the energy costs of large-scale deep learning became economically and environmentally significant, and researchers studying the limits of current AI began arguing that the architectural gap between biological and artificial neural computation was not incidental but consequential. Yi Zeng's BrainCog platform, developed from 2013 onward at the Chinese Academy of Sciences, is the most ambitious attempt to use SNNs as the substrate for genuinely brain-inspired cognitive intelligence—not merely more efficient computation but computation organized to replicate the cognitive architecture that biological intelligence depends on.
Spike timing as information. In biological neural networks, information is encoded not only in which neurons fire but in the precise timing of their firing relative to each other and to ongoing network rhythms. Spiking neural networks preserve this temporal dimension; conventional artificial neural networks, which transmit continuous activation values at each layer without temporal specificity, discard it. The temporal information enables forms of binding—linking information from different sensory modalities or different time points into a unified percept or concept—that continuous networks cannot achieve without additional mechanisms.
Energy efficiency through sparsity. Biological neurons fire rarely relative to the number of possible spike times—a sparse coding strategy that allows the brain to process enormous complexity with roughly twenty watts of power. SNNs inherit this sparsity: at any given moment, most neurons are not firing. Conventional deep learning, by contrast, requires dense matrix multiplications in which every weight is active for every forward pass. The energy difference is significant both for practical deployment and as a potential signal about the computational principles that intelligence actually uses.
Temporal integration and self-modeling. The BrainCog platform uses spiking dynamics to enable forms of temporal integration—linking sensory events across time—that Zeng argues are prerequisites for self-modeling: the brain's ongoing representation of itself as an entity with a history, a location in space, and a perspective on the world. A system without temporal integration cannot build the kind of persistent self-model that genuine wisdom seems to require. A system without a self-model cannot perceive what a situation calls for from its particular vantage point.
Social cognition and theory of mind. The BrainCog platform includes modules for social cognition inspired by mirror neuron systems—neural circuits that respond both when an action is performed and when it is observed in another agent. These modules attempt to provide what Zeng calls genuine theory of mind: not a statistical approximation of what other agents say but a representation of other agents as having states, interests, and perspectives that matter. Whether these modules succeed is an open empirical question. What is not in question is that they are attempting something structurally different from what large language models do.