Engineers currently design artificial minds by stacking modules: perception networks, memory stores, planners and reward learners. Look closely at the human brain and you see something similar, yet alien in crucial ways. It is a distributed control system that learns on the fly, runs on about 20 watts, and optimises not a single loss function but survival in a volatile world. Through understanding those differences is where the next advances in AI may come about.
Hierarchy without a boss
The brain is hierarchical but not centralised. Local circuits operate semi-independently and coordinate via spikes (fast), neuromodulators (slower), and plasticity (slowest). Think of many agents trained under different objectives, coupled by shared bottlenecks. There is no master process. Global behaviour emerges from competition and coalition across specialised subsystems.
The objective function isn’t “accuracy”
Modern AI minimises well-defined losses. Brains juggle multiple, shifting objectives: keep the body safe, harvest resources, learn useful structure, explore when it pays, exploit when it must. The “cost” is context dependent and set by physiology and environment.
Neuromodulators work like dynamic hyperparameters:
- dopamine biases towards learning from reward prediction errors;
- noradrenaline shifts system gain under surprise;
- acetylcholine increases reliance on sensory evidence when the model is uncertain;
- serotonin broadly tempers impulse and risk.
The optimisation target is a moving, embodied one.
Learning without backpropagation
Artificial networks use backprop to move a single global objective downhill. Brains update weights locally. Synapses change based on pre- and post-synaptic activity, gated by third-party signals (e.g., dopamine).
The cortex appears to use predictive learning: circuits attempt to forecast the next input and adjust when surprised. The cerebellum learns compact forward models for rapid correction; the basal ganglia reinforce action sequences that improved outcomes; the hippocampus performs rapid one-shot binding and later “replays” to train cortex offline.
Multiple learners operate on different timescales and data regimes, coordinated but not unified.
Representations are mixed and low-dimensional
Units in cortex show mixed selectivity: the same neuron participates in many codes depending on task and context. Population activity often lives on low-dimensional manifolds that warp smoothly as goals change.
This yields extreme reuse: rapid task switching without retraining, and graceful degradation under noise. Our current AI systems tend to silo capabilities in separate heads or adapters; the brain shares circuitry aggressively.
Action selection as competition, not if-then logic
Action arises from parallel proposals that compete for a limited motor and cognitive “output bus”. The basal ganglia implement biased competition: suppress most, disinhibit the winner, adjust thresholds with dopamine-labelled value signals.
The same machinery routes internal acts—shifts of attention, recall, imagery—not just muscle commands. Planning is therefore less a central search and more a tournament among partial options assembled from memory and perception.
Attention is precision management
In transformers, attention is a trainable routing table. In the brain, attention adjusts precision: it upweights reliable signals and down-weights noisy ones across levels of the hierarchy.
The thalamus and fronto-parietal networks act as dynamic gates, allocating scarce bandwidth to what matters for the current objective. This resolves the stability–plasticity dilemma: learn from surprising, trusted errors; ignore the rest.
Memory is reconstruction, not storage
Episodic, semantic, procedural and affective memories are implemented across overlapping circuits with different write speeds and retention profiles.
- The hippocampus rapidly binds who-did-what-where;
- the cortex consolidates useful regularities more slowly;
- the cerebellum compiles micro-skills;
- the amygdala tags memories with salience.
Retrieval is generative: the system reconstructs a best guess consistent with current goals and priors, rather than playing back a literal tape. For AI, this argues for memory as a compositional query engine, not a passive key–value store.
Consciousness as a limited broadcast
Most computation remains local and unconscious. A small fraction wins access to a global broadcast that synchronises many subsystems at once—perception, language, decision, interoception.
This “workspace” is narrow and slow but affords flexible reassembly of skills for novel tasks. In engineering terms, it resembles a shared attention buffer with severe bandwidth limits, reserved for coordination when local routines cannot handle the job and kick the problem 'upstairs'.
Development, curriculum and offline learning
Brains train under a shaped curriculum: reflexes and innate biases scaffold early learning; sensorimotor play builds world models; social imitation and language supercharge abstraction.
Sleep and quiet wakefulness run offline optimisation: replay consolidates and reorganises memories, prunes redundancy, and integrates new knowledge with old. The lesson for AI is to treat data order, task schedule and offline reorganisation as first-class design levers, not afterthoughts.
Energy, noise and robustness
The brain runs under tight power and time budgets. Sparse, event-driven spiking and analogue dendritic computation keep costs down. Noise is not just tolerated but exploited for exploration and regularisation.
Robustness comes from redundancy, mixed selectivity and feedback at many scales. By contrast, most frontier models are dense, clocked, and energy hungry, with robustness tacked on via augmentation and fine-tuning.
Legacy layers, modern wrappers
Evolution never rewrote from scratch. Brainstem and hypothalamus provide hard real-time control; limbic circuits compute value and salience; neocortex offers programmable modelling and flexible control.
The cortex learns to predict and steer the older layers rather than replace them. This is closer to a modern software stack wrapping legacy code than to a monolithic redesign. It suggests hybrid AI architectures where slow, model-based planners manage fast, reflexive controllers through learned interfaces.
What current AI gets right—and wrong
Right: heavy use of self-supervised prediction; attention as a routing mechanism; vector-space semantics that allow composition; curriculum benefits; the power of offline training.
Wrong or incomplete: single global objectives; reliance on backprop and dense synchrony; weak embodiment; simplistic memory; limited neuromodulation; no true multi-timescale arbitration of goals.
Bridging these gaps likely requires event-driven computation, local learning rules that approximate gradient flow, richer modulatory control, and agents that learn to manage their own objectives under constraints.
Open problems that matter to engineers
How do local learning rules and dendritic nonlinearities approximate useful credit assignment at scale? How does the system infer and update its own objectives under changing internal states? What are the precise control laws for precision-weighted attention? How do replay and sleep choose what to consolidate versus erase? How is the workspace implemented physically, and what bandwidth and latency constraints define conscious processing?
The engineering takeaway
The brain is not a better transformer. It is a multi-learner control system with strong inductive biases for prediction, compression, and energy-aware action, glued together by modulators and a narrow coordination channel we experience as consciousness.
If we want AI that is more sample-efficient, more robust, and more adaptable, we should copy those principles: diversify learning rules and timescales, treat attention as precision control, make memory reconstructive and task-driven, add modulators to retune the whole agent in real time, and design for offline reorganisation.
The destination is not a brain clone, but machines that inherit the right constraints.
