"Neural Network" Is Matrix Math with a Biology Problem

"Neural network" might be the most successful piece of terminology in technology history. Not because it describes what the system does — but because it makes matrix multiplication sound like biology. The word "neural" imports brain-like expectations onto a mathematical function. And that framing has been inflating expectations about AI since the term was first proposed in the 1940s. ## The Conventional Wisdom The AI industry frames neural networks using biological metaphors. Models have "neurons" that "fire." Networks "learn" from data. Layers "perceive" features. The brain analogy is so embedded in the vocabulary that most practitioners don't question it — even though MIT researchers explicitly warned that "more caution is warranted when interpreting neural network models" in comparison to the brain. That is because "neural network" borrows credibility from neuroscience. The brain is the most sophisticated information-processing system known. Calling your matrix multiplication system a "neural network" implies it shares some of that sophistication. It doesn't. ## The Contrarian Take: It's Function Approximation Here's what most people miss: a neural network has no neurons. It has mathematical functions arranged in layers. Each "neuron" computes one operation: 1. **Multiply** inputs by weights (matrix multiplication) 2. **Add** a bias term (vector addition) 3. **Squash** the result through a nonlinear activation function (ReLU, sigmoid, tanh) That's it. The entire architecture is: matrix multiplication → nonlinear activation → repeat across layers. The "network" is layers of these operations composed together. The "learning" is adjusting the weight matrices through gradient descent and backpropagation — an optimization algorithm, not a cognitive process. For eg. when GPT-4 generates a response, here's what happens at every one of its layers: your input tokens are multiplied by weight matrices containing billions of parameters, bias vectors get added, and the results pass through activation functions. The final output is a probability distribution over the next token. Sophisticated?? Yes. Neural?? Not in any biological sense. ## The Evidence: Biology Breaks Everywhere The biological analogy fails at every level that matters for engineering. Real neurons communicate with electrochemical signals across approximately 100 trillion synapses. They form complex topologies — not the uniform layers of a neural network. Artificial "neurons" are nodes in a matrix equation. And here's the inversion that matters: the "neurons" in deep neural networks function more like synapses in the brain, while the connections in neural networks function more like nerve cells in terms of information flow. The mapping is inverted. And the learning mechanism is fundamentally different, right?? The brain doesn't learn by backpropagation. Neural networks don't form episodic memories. They don't sleep to consolidate learning. They don't reorganize their architecture when encountering genuinely new information — they interpolate within the patterns of their training data. That is because neural networks are universal function approximators — given enough parameters and data, they can approximate any continuous function. This is mathematically proven and genuinely powerful. But "universal function approximator" doesn't market as well as "neural network." ## Why This Matters: The Terminology Sets the Failure Expectation The engineering consequence is measurable. When a product manager hears "neural network," they imagine brain-like adaptability — a system that handles novel situations gracefully. When the model fails on an out-of-distribution edge case, they're surprised. But matrix math isn't adaptive the way brains are. Neural networks interpolate within training data. They don't extrapolate reliably. And "neural" doesn't prepare you for that failure mode. When I deployed a neural network classifier for MyClaw, I initially expected it to generalize the way a human would — recognizing novel patterns by analogy. It didn't. It classified novel inputs by finding the nearest neighbor in its learned feature space — which worked well within distribution and failed silently outside it. The moment I stopped thinking "this network understands patterns" and started thinking "this function approximates patterns it's seen before," I designed better fallback logic. ## The Series Recap: What Stripping the Vocabulary Reveals In a nutshell — here's what the last six weeks have demonstrated. Strip the terminology, and you see the architecture: | Term | What It Sounds Like | What It Actually Is | |------|-------------------|-------------------| | AI Agent | Autonomous entity | While loop with LLM call, tools, and context | | RAG | Intelligent retrieval | Search + paste + same LLM call | | Reasoning | Machine thought | Pattern matching + compute budget | | Autonomous AI | Self-governing system | Supervised automation + approval gates | | Neural Network | Brain-like intelligence | Matrix multiplication + activation functions | Every term in the AI vocabulary inflates expectations about what the system does. And inflated expectations produce inflated failure rates — 88% for autonomous agents, 73% retrieval failures for RAG, 300% error increases from irrelevant context in "reasoning" models. The companies that build the best AI systems won't be the ones with the most sophisticated marketing vocabulary. They'll be the ones who understand the math underneath. ## Your Turn When was the last time a vendor explained their model's architecture in math — not metaphors?? And does stripping the biology from "neural network" help you build better, or does it miss something important about the analogy?? I'm betting that once you see neural networks as function approximators — not artificial brains — you design better systems, better tests, and better fallbacks.