From Silicon to Synapses: The Search for Biological and Hybrid AI

Current large language models (LLMs) are powerful statistical systems capable of producing fluent text and solving many useful tasks. However, fluency alone does not establish human‑like understanding, consciousness, or general intelligence, because generating coherent sentences can be achieved by pattern matching without any underlying model of the world. These questions remain open and unexplored, and researchers have yet to devise reliable tests that separate sophisticated mimicry from true comprehension.

One research direction shifts the focus from merely scaling software to changing the computing substrate: biological and hybrid computing. Researchers are exploring living neurons, brain organoids, and chips inspired by neural signaling.

Why Silicon-Based AI Still Has Limits

Artificial neural networks are mathematical models implemented on digital hardware. They have achieved remarkable results, but they differ fundamentally from biological brains in several ways:

• Energy use: The human brain runs on roughly 20 watts, comparable to a dim light bulb. Training a large language model can consume megawatt‑hours of electricity, and serving it at scale requires continuous power. This gap reflects not just hardware inefficiency but a fundamental difference in how computation is organized; biological tissue exploits massive parallelism and local processing to keep energy costs low.
• Learning: Most deployed models are trained in batches and then updated periodically. In contrast, biological brains learn continuously through mechanisms such as synaptic plasticity, neuromodulation, and sleep‑consolidation, without requiring a separate “training phase.” Continuous learning lets an organism adapt on the fly, whereas a batch‑trained network stalls until the next re‑training cycle.
• Embodiment and plasticity: A model typically lacks a physical body and direct physical experience, and its parameters are fixed during deployment. Biological nervous systems, on the other hand, adapt continuously to sensory input and action, with the body itself shaping what the brain learns. The sensorimotor loop supplies feedback that grounds concepts in real‑world consequences.

These differences do not prove that biological tissue is required for general intelligence. However, they highlight why researchers continue to explore alternatives to conventional hardware and training methods: the hope is to capture the brain’s efficiency, adaptability, and embodied learning in engineered systems.

What Is a Biological Neural Network?

In this article, a biological neural network refers to a computing system that incorporates living neurons or neural tissue as part of its processing loop. Related research is also described as biological computing, wetware, or organoid intelligence. This field is experimental, and terminology is not yet fully standardized.

Three approaches are often discussed:

Neuromorphic engineering

Neuromorphic chips imitate aspects of biological signaling with electronic circuits. Spiking neural networks process discrete events, spikes, rather than continuously updating every value. This can be more energy‑efficient for sparse, event‑driven workloads because neurons fire only when needed, reducing unnecessary computation. Systems such as Intel’s Loihi and IBM’s NorthPole are research platforms; they are not biological brains, but they explore more efficient ways to compute.

Hybrid bio‑silicon systems

In hybrid systems, living neurons, often derived from animals or human stem cells, are connected to microelectrode arrays. Electronics provide stimulation and record activity, while the tissue contributes biological dynamics. The challenges of keeping the cells alive, interpreting their signals, and scaling the interface remain significant engineering problems; each hurdle must be solved before a hybrid device can operate reliably outside the laboratory.

Brain organoids

Researchers can reprogram some human cells into stem cells and grow three‑dimensional neural organoids. These structures develop neurons and synaptic activity, but they are not miniature human brains: they lack the full anatomy, vascular system, sensory inputs, and developmental trajectory of a brain. Their potential as biological processors is an active research question, and the field is still learning how to coax organoids toward useful computational behavior.

What Has Been Demonstrated?

The idea is no longer purely speculative, but the demonstrations are small and task‑specific.

• DishBrain (2022): Researchers at Cortical Labs connected cultured mouse and human neurons to electrodes and a simulated Pong environment. The study reported that the culture changed its activity in response to feedback, suggesting a rudimentary form of learning. The system adjusted its firing patterns to improve its score, yet it did not demonstrate consciousness or general intelligence.
• Brainoware (2023): Researchers reported using a human brain organoid with an electrode interface for a speech‑recognition classification task. The result was a proof of concept: the organoid’s activity could be mapped onto phoneme categories, but the setup remained a narrow, supervised experiment rather than a general‑purpose biological computer.
• Memristive and organic devices: Researchers are also developing electronic components that retain a history of their electrical state, mimicking synaptic plasticity. Such devices may support low‑power learning because they store weight changes locally, but a laboratory memristor is not itself a living synapse and therefore cannot reproduce the full biochemical richness of a neuron.

What Would Be Needed for More General Intelligence?

Changing the hardware substrate would not automatically create general intelligence. Each of the following requirements addresses a specific gap between current systems and biological cognition:

• Embodiment: Sensors and actuators would provide the feedback loops needed to learn about physical cause and effect. Biological intelligence is shaped by having a body, the brain learns what actions produce what outcomes through continuous sensorimotor experience. A disembodied system trained on text lacks this grounding, so it cannot develop an intuitive physics of the world.
• Continual learning: The system would need to update from new experience without catastrophic interference, the tendency of neural networks to overwrite previously learned knowledge when trained on new data. Biological brains solve this partly through sleep‑based memory consolidation and complementary learning systems that separate rapid episodic encoding from slow semantic integration.
• A control and interface layer: Neural tissue would need reliable stimulation, readout, memory, and error handling. Without this, the biological component cannot be directed toward a task or its outputs interpreted, rendering it a passive curiosity rather than an active processor.
• A theory of cognition: Ideas such as predictive processing and active inference are influential hypotheses about how brains minimize prediction error. They suggest that intelligence may require a generative model of the world, not just pattern matching, because the brain continuously forecasts sensory input and acts to reduce surprise.

A Plausible Development Path

Predictions about timelines are highly uncertain. Near‑term work is more likely to produce specialized research tools, low‑power sensors, or interfaces for prosthetics than autonomous, human‑level systems.

Over a longer horizon, hybrid architectures could combine conventional models with spiking or biological components for selected sensing and control tasks. For instance, a robot might use a spiking front‑end to process tactile streams while a digital back‑end handles planning. Whether such systems would qualify as artificial general intelligence is unknown; the answer will hinge on whether they acquire flexible, goal‑directed reasoning beyond their narrow modules.

More capable organoids would also create ethical questions. If a system had meaningful experiences, a claim that cannot currently be established for today’s organoids, researchers would need safeguards concerning welfare, consent, and disposal. These questions should be addressed before capability becomes the sole metric of progress, because ethical lapses could stall or derail the field entirely.

Conclusion

Biological and neuromorphic computing broadens the search for efficient, adaptive intelligence. Current experiments demonstrate learning‑like behavior in tightly controlled settings, not conscious or general AI. The field is promising precisely because it is still uncertain: progress will depend as much on careful measurement, safety protocols, and ethical frameworks as on breakthroughs in hardware.