Exploring how a 2D machine mimics the signal-processing logic of the 3D brain—through attention, layers, and learned transformations.

Why This Essay Exists

For many people—even experienced programmers—AI remains a kind of mystery. It produces text, answers questions, even writes code, but what’s happening under the surface often feels like magic.

This essay is an attempt to demystify that process. To explain, as simply and precisely as possible, how AI actually works—not just in broad terms, but in terms that relate to the way the brain works, how signals travel, and how information gets transformed.

The goal is to make these ideas accessible: not just for engineers or researchers, but for anyone curious—including my son, Taiga, who likes to understand how things work from the inside out.

You’ll encounter metaphors: caterpillars and butterflies, GPS systems, 2D machines mimicking 3D brains. These are not simplifications—they are bridges. Underneath them lie real patterns and systems. The hope is that by the end, you’ll no longer see AI as mysterious, but as a signal flow system—a dynamic engine that turns input into meaning, through mathematical transformation.

Introduction – How a 2D Code System Mimics a 3D Brain

AI experts often speak in terms of layers, signal weights, and tokens—words that evoke the image of a living, 3D neural network like the human brain. And indeed, the brain’s logic inspired the original leap to artificial intelligence. But in reality, today’s AI systems are not biological or spatial. They are abstract mathematical machines—built on binary code (1s and 0s)—programmed to receive strings of digital input, reorganize them through learned rules, and return transformed strings as output.

It’s not a brain. But it functions like one in some remarkable ways.

The transformation process is mathematical, not cellular. It takes place in 2D, not 3D. Yet the logic—the flow of information through weighted pathways, the selective amplification of relevant signals, the mapping from input to action—is strikingly similar.

Like turning a caterpillar into a butterfly, AI systems take in raw, unstructured input and reshape it through a cascade of transformations. And in some tasks—like language, pattern detection, or rapid classification—AI can already outperform the human brain in both speed and scale.

Below is a walk-through of how a 2D system of binary code mimics, and in some cases surpasses, the core functionality of the 3D brain.

1. Hardware – The Physical Foundation

At the most fundamental level, every AI system relies on hardware: physical silicon chips made of microscopic transistors. Each transistor can be either on (1) or off (0). These binary states form the foundation of digital information. Combinations of 1s and 0s create units of code called bits, which can be arranged into longer strings—like a written ledger of binary signals—that carry information across the system.

These bits don’t just sit still. The strings of code flow through the transistors on the chip. The transistors are arranged into logic circuits that allow these signals to be flipped or transformed based on mathematical instructions. In essence, the hardware becomes a vast switching system, continuously reshaping binary code as it moves.

Mathematical formulas—embedded in the software—tell the system how to transform each sequence of bits. These transformations happen in stages. In an AI network, the code is processed through multiple rounds of transformation, each called a layer. Just as a signal in the brain may pass through different regions before producing a thought or action, digital signals in an AI system move through layers of logical processing before reaching an output.

When you use an AI tool like ChatGPT, this process doesn’t happen entirely on your device. Your input is typically sent to a remote data center, where specialized processors carry out the heavy computation. The binary code travels via fiber-optic cables—as flickers of light—or over Wi-Fi, where radio waves rapidly switch frequency to represent the same 1s and 0s. A single fiber-optic line can carry up to 10 billion bits per second, allowing your question to travel to a supercomputer and back almost instantly.

In short, the hardware is the physical substrate—the digital highway—that lets code move and transform. It provides the space for signal flow, but the intelligence emerges from what happens within those flows: the mathematical logic that decides how to shape the signal at every step.

2. Software – Mathematics in Motion

The software takes a starting string of binary code—representing your input—and applies a series of mathematical transformations. These operations are not random; they follow a structured sequence the AI system has learned through training. The software detects patterns—tokens, bits, or motifs—within the string, and transforms it accordingly.

At the heart of this process is attention: the system learns to focus more on the parts of the signal that matter and suppress the rest. With each transformation layer, the string is modified based on the system’s learned ability to recognize which elements are important in context.

The signal doesn’t move through space like it does in a brain. Instead, the string of code undergoes a set number of recombinations, layer by layer, each one reshaping the signal slightly. These are not spatial layers but abstract steps in a virtual process. After passing through dozens—or in some models, hundreds—of such layers, the final string of code emerges: a new output that can be decoded into words, actions, or predictions.

Remarkably, when this sequence is performed by a trained system, it becomes highly efficient—capable of producing complex, accurate results in milliseconds. In some tasks, it can even perform faster than the human brain.

While the brain routes signals through growing and adapting 3D neural tissue, the AI system performs this entire process in a virtual 2D architecture. And yet, it functions like a kind of GPS: navigating through symbolic space using attention-weighted rules that were shaped by experience. Intelligence, in this system, emerges not from physical structure but from repeated signal transformation governed by learned mathematics.

3. A Navigation Metaphor: Understanding the AI Network

AI experts use terms like layers, nodes, weights, tokens, and backpropagation—terms that sound like they belong to a neuroscience lab or a programming manual. But beneath the jargon is a simple metaphor that can make sense of it all.

Imagine an AI network as an ultra-intelligent GPS system.

You input a destination—a question or a prompt. At first, the system doesn’t know which roads lead to the right answer. It explores all routes, even the wrong ones. Over time, and through millions of learning cycles, it figures out which paths are helpful and which aren’t. That learning process is called backpropagation—a mathematical way to dim bad signposts and strengthen good ones.

Each node in the network is like a fork in the road with a glowing sign. At first, all the signs are equally bright. But as the system learns, only the most reliable signposts stay lit. These are the weights—learned preferences that help guide the signal.

Tokens—the input words or symbols—are like the destination request on your GPS. They determine which roads the system should light up.

Attention is like focusing on only the key road signs while ignoring distractions. The system learns to highlight the parts of the input that matter and suppress what doesn’t—like a driver scanning for exit signs while ignoring billboards.

Eventually, with all its learned preferences and weighted paths, the system becomes a high-speed signal navigator. When you enter a prompt, it instantly lights up the most relevant path—transforming the input string into a meaningful output. With this metaphor in mind, let’s now return to how actual AI models are structured, trained, and used.

4. AI Models – Learning Through Signal Weighting

The core of an AI system is its model—a vast network trained to transform input signals into meaningful outputs. This network is inspired by the brain’s structure, but operates in an entirely different form.

In the brain, intelligence arises from physical structures: neurons that grow, connect, and fire across real 3D space. A sensory signal starts at a specific point—like a GPS coordinate—and travels along biological pathways, shaped by a lifetime of experience, plasticity, and learning. Signals branch, loop, and interact with different regions of the brain before resulting in a thought, action, or emotion. Every step is embodied, shaped by space, chemistry, and time.

AI, by contrast, transforms signals in code. When an input arrives—say, a sentence—it is converted into a string of bits and passed through a virtual architecture made of layers of mathematical instructions. These layers are composed of nodes (abstract processing points), each with assigned weights. A weight is not physical, but a learned number that tells the system how strongly to respond to certain patterns in the input.

At each node, the string of code is examined: Are there patterns? Specific combinations of tokens? Features the system has seen before? Based on these, the signal is transformed and passed forward. Importantly, the attention mechanism allows the system to selectively amplify certain parts of the input while ignoring others—just like the brain focusing on a voice in a noisy room.

This process repeats through many layers. The model doesn’t “move” the signal from one physical location to another—it recomputes the string at each step. The result is a cascading chain of transformations: bits into new bits, signal into new signal.

Over time, through training on massive datasets, the system learns which transformations are most useful. These learned preferences are stored as weights—values that tilt the decision-making process toward the most effective patterns. This is the essence of signal weighting: assigning importance to parts of the input based on experience.

In the brain, signal flow is recursive and adaptive, shaped by real-world context, sensory feedback, and biological constraints. In AI, signal flow is symbolic and rapid—guided by data, pattern recognition, and algorithmic efficiency. Yet both systems share a common logic: the intelligent transformation of input into output, guided by weighted evaluation of what matters most.

In that sense, AI doesn’t mimic the brain’s form, but it echoes its function: the ability to route information along optimized paths. Not through neurons, but through numbers.

Conclusion – From Signal to Meaning

You can think of each input signal as a caterpillar—a raw, unstructured string of code entering the system. As it moves through the AI network, it undergoes a defined number of transformations, programmed in the architecture of the machine, until something entirely new emerges. Just as a caterpillar enters the chrysalis and recombines itself through a hidden biological program, the input code moves through abstract layers of mathematical logic. The final output—the transformed string—is the butterfly: a new expression of meaning, shaped by pattern recognition, memory, and learned response.

In the human brain, signals arrive through receptors and travel through a 3D network of conduits, using biochemical messengers to activate responses. Some pathways are fast, some slow; some are strong, others weak—literally weighted through synaptic structure, neurotransmitter gradients, and physical distance. Learning in this system involves rewiring—growth, pruning, reinforcement—across time and space.

In contrast, AI models operate in a virtual 2D environment. Words are converted into strings of binary code—1s and 0s—which flow across microchips composed of billions of on/off gates. These bits pass through a fixed series of transformation steps. At each stage, the string is reshaped based on statistical patterns in the code and the model’s learned instructions for how to process them. The input string—the caterpillar—enters the network and returns as a string butterfly, recombined and newly meaningful.

Importantly, this design was no accident. One of the godfathers of AI—Geoffrey Hinton—and other pioneers in the field were originally driven by a desire to understand how the brain works. They designed neural networks to replicate the brain’s core mechanism of signal weighting: the idea that information is transformed by assigning importance to certain signals over others. While AI systems ended up being mathematical rather than biological, they retained this logic. The concept of weights—now abstract values rather than physical properties—still governs how code is refined as it moves through each layer.

Every input must pass through the same set of layers. There are no shortcuts or skips. The process is recursive and uniform: a signal enters, is transformed based on weighted patterns, and emerges with a structured, often highly accurate output. At its core, this is pattern recognition—refined over time through feedback and error correction.

Unlike the biological brain with its physically weighted conduits and spatial dynamics, today’s AI systems exist entirely in flattened logic space. Their intelligence arises not from physical wiring, but from mathematical structure: the recognition of pattern, the modulation of signal strength, and the precise sequencing of abstract transformations.

And yet—across both systems—the miracle is the same: signal in, meaning out.

What This Means for Us

AI is not alive. It does not think or feel. But it does process. It receives signals, evaluates patterns, and produces responses—much like we do, but in silicon and code rather than cells and neurons. By understanding the logic beneath the surface, we begin to reclaim the narrative of AI—not as something alien, but as something we shaped. Something we can choose to align with our values, our needs, and our evolving understanding of intelligence itself.