Before diving into artificial neural networks or transformer architectures, it's worth establishing what intelligence actually means — and it means something much broader than most people assume.
Intelligence, at its most fundamental, is the ability of a system to adjust its behavior in response to its environment in ways that improve its outcomes. By this definition, intelligence doesn't require consciousness, language, or even a brain. It requires three things: sensing, processing, and acting — with some form of feedback that biases future behavior.
This definition matters because artificial intelligence borrows heavily from biological intelligence. Understanding where biological intelligence starts — and how it scaled over evolutionary time — provides the foundation for understanding what neural networks are actually modeling, and what they're not.
Even single-celled organisms exhibit goal-directed behavior. Bacteria navigate chemical gradients through chemotaxis — biasing random movement toward favorable conditions using nothing more than a few seconds of chemical memory. No neurons, no nervous system. This is the baseline: intelligence begins not with brains, but with any system that adjusts behavior in response to its environment.
What makes chemotaxis worth studying isn't the mechanism itself — it's the principle. The bacterium doesn't "know" where food is. It has a simple rule: if conditions are improving, keep going; if they're getting worse, try a new direction. This is a gradient-following strategy, and it shows up everywhere in AI — from bacterial navigation to the gradient descent algorithms that train modern neural networks. The math is different, but the principle is the same: move in the direction that improves your outcome.
Paramecia — single-celled organisms slightly more complex than bacteria — demonstrate something more sophisticated: habituation. When a paramecium bumps into an obstacle repeatedly, it changes its avoidance strategy. There's no synapse being strengthened, no memory being formed in the way we usually mean it. But its behavior is being shaped by its history, implemented through ion channel dynamics in the cell membrane. That qualifies as learning by any functional definition.
Single cells can sense, process, and act — but they're limited by the fact that the same cell has to do all three. Multicellular organisms faced a different problem: coordination. When you have thousands or millions of cells, a cell on the surface that detects a threat needs to communicate that to cells on the other side that control movement.
Early multicellular organisms solved this with chemical signaling — cells releasing molecules that diffuse to neighbors. But chemical diffusion is slow and imprecise. For any organism larger than a few millimeters, or any response that needs to happen fast, diffusion isn't enough.
Neurons evolved as the solution. A neuron is a cell specialized for fast, long-distance communication. It propagates voltage changes along its membrane at 1–100 meters per second, compared to the micrometer-per-second pace of chemical diffusion. At the end of the line, the electrical signal triggers the release of chemical messengers called neurotransmitters that bridge the gap to the next cell. That gap — the synapse — is where the interesting computational properties emerge, and will be explored in depth in Chapter 2. But the evolutionary pressure that created neurons was fundamentally about speed and distance, not computation.
The earliest nervous systems weren't centralized. Cnidarians — jellyfish, hydra, sea anemones — have nerve nets: neurons distributed throughout the body with no central processing hub. A stimulus at any point propagates outward through the net in all directions.
Properties worth noting:
The nerve net is a useful reference point for artificial neural networks. A fully connected network with no imposed structure is, loosely, a nerve net — every node connected to many others, no hierarchy. The limitations are similar: nerve nets can coordinate reflexes, but they can't do much complex processing or learning. Structure matters. This insight — that the architecture of a network determines what it can compute — carries directly into AI, where the difference between a convolutional network, a recurrent network, and a transformer is entirely a question of structure.
The transition from nerve nets to centralized nervous systems is one of the most important events in the history of intelligence. It happened when organisms evolved bilateral symmetry — a distinct front and back, left and right.
Bilateral symmetry matters because it means the organism moves in a consistent direction. It has a front end that encounters the environment first. Evolutionary pressure concentrated sensory organs there — light sensors, chemical receptors, touch receptors — because that's where new information arrives. And once sensors are concentrated, processing needs to be nearby to handle the input quickly. This process is called cephalization: the evolutionary tendency to concentrate neural tissue at the head end.
Flatworms (planaria) are a key example:
This is a qualitative jump from the nerve net. The flatworm has a place where information converges, gets integrated, and produces coordinated output. It's not "thinking," but it has a centralized architecture that supports behaviors the nerve net can't: directional movement, sensory integration, and simple learning.
Before going further up the evolutionary ladder, it's worth understanding the reflex arc — the simplest complete neural circuit, and the building block of more complex nervous systems.
A reflex arc has three components:
The classic example: you touch a hot surface, and your hand pulls away before you consciously feel pain. The sensory signal travels to the spinal cord, synapses onto a motor neuron, and the withdrawal happens. Your brain finds out afterward.
Reflex arcs matter for two reasons. First, they're the evolutionary foundation more complex nervous systems are built on — the spinal cord is essentially a stack of reflex arcs with increasingly sophisticated modulation from higher layers. Second, they demonstrate a principle that recurs throughout both biology and AI: layered processing. The reflex handles the urgent response. Higher layers can then modulate, override, or learn from the event — but only after the fast path has already acted. This same pattern appears in modern AI systems that combine fast heuristics with slower deliberative reasoning.
The evolutionary story from here is one of adding layers. Each new layer enables more sophisticated behavior while keeping the faster, simpler layers intact.
Above reflexes sit fixed action patterns — stereotyped sequences of behavior triggered by specific stimuli. A spider building a web, a bird performing a mating dance, a wasp provisioning its nest. These behaviors are complex, multi-step, and appear intelligent — but they're hardcoded. The spider doesn't learn to build a web; the instructions are encoded in its nervous system. If you disrupt the sequence midway, the spider can't adapt — it either restarts or continues from where it was, regardless of whether the earlier steps were completed.2
Fixed action patterns are relevant to AI because they resemble rule-based systems and hard-coded heuristics. They can be remarkably effective within their domain — a spider's web is an engineering marvel — but they break outside the conditions they evolved for. They're inflexible in exactly the way that early AI systems (expert systems, rule-based chatbots) were inflexible: powerful within their scope, brittle outside of it.
The simplest forms of actual learning are habituation (decreased response to a repeated, harmless stimulus) and sensitization (increased response after a harmful stimulus). These are non-associative — the organism isn't linking two events together. It's tuning its responsiveness to one.
Eric Kandel's work on the sea slug Aplysia mapped the molecular mechanisms of both and earned the Nobel Prize in Physiology or Medicine in 2000. In Aplysia, habituation occurs through synaptic depression — the sensory neuron releases less neurotransmitter with repeated activation. Sensitization involves the opposite: a modulatory interneuron enhances the synapse, increasing neurotransmitter release.
Associative learning links two events together — this is where behavior starts looking recognizably intelligent. Pavlov's classical conditioning (dog learns bell predicts food) and Skinner's operant conditioning (rat learns lever press produces food) are the canonical examples.
Associative learning requires something the simpler forms don't: a mechanism for detecting temporal correlations. The organism needs to notice that event A predicts event B and adjust accordingly. In biological neurons, this is implemented through what's called Hebbian learning — neurons that fire together strengthen their connection. Donald Hebb proposed this in 1949, and it remains one of the most important principles in both neuroscience and AI. The shorthand: "neurons that fire together, wire together."
Hebbian learning will come up again in Chapter 3 when we examine how early AI researchers translated biological principles into mathematical models. The artificial version — adjusting connection weights based on correlated activity — is the conceptual ancestor of modern training algorithms.
At the top of the behavioral hierarchy sit capabilities that require internal representation — a model of the world that exists in the nervous system independent of immediate sensory input.
Edward Tolman demonstrated in the 1940s that rats build spatial maps of mazes, not just stimulus-response chains. A rat that has explored a maze without reward will, when food is introduced, navigate to it immediately — taking shortcuts it was never rewarded for. This requires an internal model, not just learned associations.
Planning behavior — acting now based on predicted future outcomes — is observed in some primates, corvids (crows and jays), and possibly cephalopods (octopuses). A scrub jay that caches food will re-cache it if it was observed by another bird — suggesting it models the other bird's knowledge and anticipates theft.3 This is multi-step inference: "that bird saw me" → "it might steal my food" → "I should move it."
Putting it together, biological intelligence scales along a clear progression:
Each level preserves the capabilities below it and adds something new. Humans still have reflexes, fixed action patterns (try not flinching when something flies at your face), habituation, and associative learning. What each new level adds is not a replacement but a layer of modulation and flexibility on top of what already exists.
This layered architecture has a direct parallel in both neuroscience (the brainstem → limbic → cortex progression, explored in Chapter 2) and in AI systems (fast heuristics → learned responses → deliberative reasoning). The pattern is the same: don't replace the fast, simple system — build on top of it.
If this chapter establishes one idea, it's this: intelligence is not a binary property that appears with brains and is absent without them. It's a spectrum of behavioral sophistication that begins with the simplest adaptive response and scales through layers of increasing complexity.
| Property | What it means | First appears |
|---|---|---|
| Sensing | Detecting relevant features of the environment | Bacteria (chemoreceptors) |
| Processing | Transforming input into useful internal signals | Bacteria (signaling cascades) |
| Acting | Producing behavior that affects the environment | Bacteria (motility) |
| Feedback | Adjusting future behavior based on outcomes | Paramecia (habituation) |
| Representation | Internal models beyond immediate sensation | Vertebrates (cognitive maps) |
Every artificial intelligence system implements some subset of these properties. A thermostat senses, processes, and acts — but has no feedback loop that improves its future performance and no representation beyond current temperature. A large language model has extraordinarily sophisticated processing and representation, but its "sensing" is limited to text input, and it has no persistent feedback loop that shapes its behavior across sessions without external scaffolding.
Understanding where each capability came from biologically — what evolutionary pressure produced it, what architecture supports it, and what it enables — gives you a framework for evaluating what artificial systems actually do, and what they're missing. That gap between biological and artificial intelligence is where the interesting questions live, and it's what the rest of this guide will build toward.
Next: Chapter 2 — The Brain. How biological neural architecture actually implements these capabilities. Neurons, synapses, the major structures of the brain, and the mechanisms of biological learning that AI researchers would later attempt to formalize.1 McConnell (1962) first demonstrated planarian learning and memory retention through regeneration. The experiments were controversial but have been replicated with modern methods by Shomrat and Levin (2013).
2 Konrad Lorenz and Nikolaas Tinbergen formalized the study of fixed action patterns in the 1930s–50s. Their work on innate releasing mechanisms — specific stimuli that trigger fixed behaviors — earned them the Nobel Prize in Physiology or Medicine in 1973.
3 Clayton and Dickinson (1998) demonstrated this in Western scrub-jays. The behavior is often cited as evidence of theory of mind in non-human animals, though the interpretation remains debated.