Why I Think NeuroAI Is the Road to AGI

Recently I found an interesting paper.

The paper grew out of an NSF workshop called NeuroAI and Beyond. Researchers from neuroscience, artificial intelligence, robotics, cognitive science, control theory, and neuromorphic engineering gathered to discuss the future of AI. The workshop produced a detailed report containing the discussions and participants’ personal statements. A later paper organized these ideas into a research program for NeuroAI. It identifies three major weaknesses in current AI: limited interaction with the physical world, the absence of lifelong learning, and extreme inefficiency in energy and data.

At first, these seemed like three separate engineering challenges. After reading the paper and the workshop report, I began to see them as parts of the same process.

When I use AGI in this post, I mean an autonomous intelligence that can live in the world. It has a persistent identity and a physical body, learns from its own experience, adapts without being retrained from scratch, and continues operating under real physical and energy constraints.

The body generates experience.

Plastic neural hardware turns experience into change.

Multiple timescales turn change into stable knowledge.

Integrated memory and computation allow the whole process to continue efficiently.

This, I believe, is the road to AGI.

Embodiment: intelligence needs consequences

Modern AI has approached intelligence from the top, beginning with language. Nature approached intelligence from the ground, beginning with bodies that had to survive.

A frontier language model can explain how to cook an omelet. It can describe the chemistry of eggs, provide a recipe, and tell us how to control the temperature of the pan. Yet it cannot enter an unfamiliar kitchen, find the tools, crack an egg, remove a piece of shell, and adjust its movements when something goes wrong.

The real world is continuous. Sensory information is incomplete and noisy. Objects have unknown weights, surfaces have unknown friction, and every action has consequences. A robot cannot pause reality while it calculates the perfect response. It must act immediately, observe what happened, and adapt.

Consider the word “heavy.” A language model learns its meaning through relationships among words. It knows that lead is heavy, that heavy objects are difficult to lift, and that “heavy” is related to weight.

A child learns the same concept through muscles, skin, balance, sound, and gravity. Before learning the word, the child has already tried to lift objects, dropped them, heard them hit the floor, and felt the effort required to move them.

The language model receives descriptions of experience. The child receives the experience itself.

This difference matters because an embodied agent can generate its own data. Every movement becomes an experiment. When it pushes an object, it learns whether the object rolls, slides, bends, or breaks. When it reaches for a cup, it predicts where the cup will be and how its fingers should feel when they make contact. A mismatch between prediction and reality becomes a learning signal.

The paper describes this idea as prediction through interaction. Biological intelligence constantly predicts what will happen next, including what will happen as a result of its own actions. The agent acts, observes the consequences, corrects its prediction, and acts again. Through this loop, it gradually learns cause and effect.

Embodiment also changes the architecture of intelligence itself. In animals, the nervous system and the body evolved together. The brain controls a particular body whose muscles, joints, skin, sensory organs, and physical constraints shape every problem the nervous system must solve.

The body can perform part of the computation. Tendons store and release energy. Soft tissue absorbs impacts. Springs and dampers stabilize movement. Fingertips provide dense information about pressure, vibration, texture, and slipping. These physical features simplify control before the brain needs to intervene.

Many modern robots follow a different design philosophy. Their hardware and control systems are often developed separately. Engineers build a rigid machine with powerful motors, then expect software to calculate every movement. This approach can produce precise actions in a controlled factory. But it becomes much less reliable when the environment changes.

The idea that changed my thinking was that embodiment does more than give AI cameras, microphones, and robotic arms. It gives the system a source of causal experience. Its actions change the world, and the resulting consequences change the agent.

Bing Brunton, one of the participants of this workshop , summarized this perspective in a memorable sentence: “The brain sits not in a jar.”

AGI needs a body.

Lifelong learning

A body can create an endless stream of experience. That experience becomes useful when some of it survives the moment and changes future behavior.

Most current AI systems separate learning from use. During training, the model’s parameters change. After deployment, those parameters usually remain frozen.

Developers have created several ways to give models access to new information. A larger context window gives the model more temporary working space. Retrieval-augmented generation, or RAG, lets it consult an external library. Fine-tuning can change the model itself, although it usually occurs during a separate training stage and may damage knowledge the model has already acquired.

These methods are useful. Yet none gives a deployed model a continuous, selective, and stable process for deciding what to remember, incorporating new experience into long-term knowledge, and preserving what it has already learned.

An animal has no clean boundary between training and deployment. Life itself is training.

The difference begins with the physical organization of the brain. The same network of neurons and synapses stores information, computes with it, and changes through experience.

A synapse helps determine what a neural circuit does now. Activity can also alter that synapse, changing what the circuit will do when it encounters a similar situation in the future. Present computation can therefore leave a physical trace inside the system. That trace is learning.

This plastic structure makes lifelong learning possible. It also creates a serious problem. If every experience strongly changed every connection, old knowledge would quickly disappear.

The brain manages this problem by controlling when, where, and how quickly change occurs. Neuromodulators help signal whether an event is important enough to learn from. Reward, novelty, attention, danger, and internal needs can influence which circuits become plastic.

Different synapses also change at different speeds. Some respond rapidly and preserve recent information. Others change slowly, accumulating evidence across many experiences. Fast changes give the organism adaptability. Slow changes protect stable knowledge.

The brain also uses different memory systems.

The hippocampus can rapidly record a particular episode after one or a few experiences. The neocortex changes more slowly, extracting general patterns from many episodes. During memory replay, especially during sleep, selected experiences are gradually integrated into long-term cortical knowledge.

Imagine a household robot that discovers that one particular pan has a loose handle. It should remember that fact after a single dangerous incident. At the same time, it should preserve its broader understanding of kitchens, tools, heat, and cooking.

The individual event should enter memory quickly. The general lesson should be integrated more slowly.

This combination allows the system to learn from a surprising event while protecting everything it already knows. It can remember an exception without treating every exception as a new law of the world.

This is what I mean by a life history. Yesterday’s experience changes today’s decisions, and today’s mistakes improve tomorrow’s behavior. The agent gradually becomes a different system because of what it has lived through.

Efficiency: intelligence must be able to stay awake

A system that learns throughout its life must also be able to run throughout its life.

A chatbot can remain inactive until a prompt arrives. An embodied agent has no such luxury. Its cameras, microphones, skin, balance sensors, joints, and internal monitors continuously produce information. Even while standing still, it must decide what to ignore, what to predict, what to remember, and when to act.

The world loop and the time loop must keep running every second.

The scale of the energy problem is striking. The paper estimates that training a frontier language model can consume around 50 GWh of electricity. The human brain operates on roughly 20 watts while supporting perception, movement, memory, prediction, and continuous learning. It reveals the enormous difference in efficiency.

The previous section introduced a key feature of biological intelligence: the same network stores information, computes with it, and changes through experience.

For lifelong learning, this structure allows present activity to create lasting change.

For energy efficiency, the same structure reduces the need to move information.

In conventional computers, memory and computation are physically separated. A processor repeatedly fetches model parameters and activations from memory, performs calculations, and writes the results back. For large AI models, moving this data can consume more energy than the arithmetic itself.

The brain organizes computation differently. Learned information is stored in synaptic states, and those synapses directly shape the signals passing through the network. Storage and computation occur close together.

The architecture that makes knowledge changeable also makes knowledge cheaper to use.

Efficiency also depends on how much of the system is active at one time.

Transformers perform large, dense matrix operations at every step. Huge numbers of parameters participate in processing regardless of input.

Biological circuits are much more selective. Most neurons remain quiet at any given moment. Activity rises when something changes, violates a prediction, signals danger, or deserves attention.

Consider a camera watching an empty room. A conventional camera repeatedly transmits complete frames. The walls, floor, and furniture are sent again and again even though almost nothing has changed.

An event-based camera follows another strategy. It reports changes in brightness as they occur. When the room remains still, very little information is produced. When someone enters, only the changing parts of the visual field generate events.

The retina uses a related principle. It processes contrast, movement, edges, and changes before information travels deeper into the brain. This reduces the amount of raw sensory data that later circuits must handle.

An efficient intelligent system should allow most of its components to remain quiet until they have something useful to contribute.

The brain also reuses circuits through time. Information travels through recurrent feedback loops. Higher regions send predictions toward sensory regions, and new sensory information returns as an error signal. The system can repeatedly refine its interpretation using many of the same circuits.

Current AI often gains computational depth by stacking more layers. Recurrent systems can also gain depth by allowing information to circulate through an existing network over time. This reuse can reduce the need for a separate physical layer at every stage of processing.

Neuromorphic hardware attempts to implement these principles in silicon. Intel’s Loihi 2, for example, supports spiking neural networks in which computation occurs when an event arrives. When there is little relevant activity, much of the hardware can remain quiet.

A neuromorphic chip alone cannot create an efficient intelligence. The algorithms must also be designed around sparse activity, local learning, recurrent dynamics, and event-based communication. Sensors must produce useful sparse signals. Memory must support local updates. Hardware and algorithms must be developed as one system.

For AGI, energy efficiency is closely connected to autonomy. A machine that depends on a data center for every movement cannot live with the independence of an animal. A system that will stop learning whenever it loses access to a server cannot build a continuous life history.

Three capabilities, one architecture

At this point, the three weaknesses identified by the paper stop looking separate.

Embodiment gives the system causal experience.

Plastic neural hardware allows that experience to change the system.

Multiple learning timescales preserve useful changes without destroying old knowledge.

Integrated, sparse, and recurrent computation allows the whole process to continue in real time.

These capabilities have to be designed together.

This also suggests a different role for large language models. An LLM could become an important language, knowledge, and planning layer within a larger intelligent architecture. Other systems would handle perception, movement, reflexes, continual memory, uncertainty, internal goals, and physical self-maintenance.

The high-level system should not calculate every motor command. Human conscious thought does not specify the exact contraction of every muscle when we walk. Lower-level systems handle balance, posture, and rapid correction. Higher levels choose destinations, set goals, and revise plans.

The vestibulo-ocular reflex provides a simple example. When you move your head, your eyes move in the opposite direction to keep the visual world stable. The response occurs in roughly ten milliseconds, far too quickly for conscious thought.

When someone begins wearing glasses that change the relationship between head movement and visual movement, the cerebellum slowly recalibrates the reflex. Once the new setting has been learned, lower-level circuits can execute it rapidly.

The system adapts slowly and acts quickly.

An AGI could follow the same general principle. Fast local controllers would manage immediate safety and physical stability. Intermediate systems would adapt movements and learn skills. Higher layers would support planning, language, abstraction, and social interaction. Memory systems operating at different timescales would connect the entire hierarchy.

This architecture resembles an artificial nervous system far more than a single giant model.

That is the central idea I took from these two documents. The next generation of AI may require a complete system that connects body, memory, learning, control, and hardware from the beginning.

What this future could look like

The paper’s roadmap begins with technologies that already exist in early forms.

In the near future, we could see robots equipped with event-based cameras, tactile skins, protective reflexes, and learning systems that continue adapting after deployment. A robot could adjust to a new tool, a different floor, or an unfamiliar load without requiring complete retraining.

The next stage would involve agents learning world models through their own actions. Multiple robots could share their experiences, allowing a fleet to learn faster than any individual machine. Layered architectures could connect sensors, local controllers, memory systems, and high-level planners.

Over a longer period, the goal would be autonomous agents capable of operating in unstructured human environments. They would handle unfamiliar objects, collaborate with people, learn over years, and operate using hardware efficient enough to remain local and continuously active. The paper even proposes a long-term goal of an AI supercomputer consuming less than one kilowatt.

This future resembles the engineering of a new kind of organism.

From describing the world to living in it

Today’s AI can explain how to make an omelet.

AGI will have to enter an unfamiliar kitchen. It will need to identify the tools, feel the weight of the pan, control the pressure of its fingers, notice when the egg begins to slip, recover when a piece of shell falls into the bowl, and remember the mistake the next time it cooks.

Its body will generate the experience.

Its plastic nervous system will be changed by the experience.

Its memory systems will transform the event into stable knowledge.

Its integrated architecture will keep the entire process running within the energy budget of a physical machine.

After reading the paper and the workshop report, I no longer imagine AGI as a Brain Bug in Starship Troopers, which is super intelligent but completely immobile. I imagine a developing agent with a body, a life history, multiple layers of control, and an efficient artificial nervous system.

Language will remain one of its most powerful abilities. Beneath language will be something older and more fundamental: perception, action, prediction, memory, and survival.

That is why I think NeuroAI is the road to AGI.