The neural networks running ChatGPT weren't designed to retrieve files—they were designed to think.
The Summary
- AI systems like ChatGPT don't work like databases pulling files—they work like brains making associations, tracing back to Frank Rosenblatt's 1958 Perceptron, which learned from examples instead of following rules
- The term "artificial intelligence" was coined in 1956, but the breakthrough came from cognitive scientists trying to understand human minds, not computer scientists optimizing computation
- This origin story explains why LLMs hallucinate, why they're probabilistic rather than deterministic, and why prompt engineering feels more like psychology than programming
The Signal
Most people think ChatGPT is fancy autocomplete running on a giant lookup table. You type a question, the system finds the answer in its database, spits it back. Clean. Logical. Wrong.
Large language models work through learned associations, not retrieval. When you ask GPT-4 a question, it's not searching files. It's activating patterns across billions of weighted connections, built through exposure to text the way your brain builds patterns through exposure to experience. The architecture is fundamentally neural, not computational in the traditional sense.
This matters because it explains behavior that confuses people. Why do LLMs confidently hallucinate? Because human memory does too. Why does the same prompt sometimes yield different answers? Because association-based thinking is probabilistic. Why does prompt engineering work better when you "talk" to the model like it's reasoning through a problem? Because it kind of is.
"Understanding AI as brain-like rather than database-like changes how you should interact with it and what you should expect from it."
The fork in AI's road happened in 1956 at Dartmouth, where the term "artificial intelligence" was coined. The people in that room believed intelligence meant writing down the right rules for a machine to follow. Pure logic. Pure computation. If-then statements all the way down.
Frank Rosenblatt wasn't in that room. He was a psychologist, and in 1958 he built the Perceptron—the first artificial neural network. Instead of programming rules, he let the machine learn from examples. The Perceptron's architecture mimicked neural connections in brains: many simple units, connected with varying strengths, patterns emerging from repetition rather than instruction.
Key architectural differences:
- Rule-based AI: explicit logic, deterministic outputs, brittle in novel situations
- Neural networks: learned associations, probabilistic outputs, generalizable patterns
- Human cognition: also learned associations, also probabilistic, also hallucinates
For decades, rule-based AI dominated because it was easier to understand and control. Expert systems. Decision trees. Knowledge graphs. These worked for narrow tasks but couldn't scale to messy, real-world complexity. They couldn't write poetry or hold conversations or generate code because human language and thought don't follow neat rules.
The neural network approach lost funding and credibility through the 1970s and 80s—the "AI winters." But cognitive scientists kept refining the theory. How does the brain actually work? Not through rules. Through layers of simple processors that strengthen or weaken connections based on patterns in their inputs. Backpropagation. Gradient descent. Statistical learning.
"The systems we're deploying at scale in 2024 trace their lineage to psychologists studying cognition, not engineers optimizing databases."
By 2017, the Transformer architecture (the T in ChatGPT) proved you could scale neural networks to massive size and get emergent capabilities no one explicitly programmed. Reasoning. Translation. Summarization. Code generation. Not because someone wrote rules for those tasks, but because the network learned statistical patterns across billions of text examples and developed internal representations that map onto how humans think about language.
The Implication
If you're building agents or deploying LLMs at work, stop treating them like search engines or APIs. They're cognitive systems. They respond to context, priming, and framing the way humans do. Prompt engineering isn't a hack, it's applied psychology.
This also means current AI safety approaches are partly misguided. You can't "patch" a neural network the way you patch buggy code. These systems don't have explicit goals you can edit. They have learned tendencies that emerge from training patterns. Alignment isn't a software problem. It's a cognitive science problem.
Watch for the companies that get this distinction. They'll build better tools because they'll understand what they're actually working with—not databases that hallucinate, but statistical brains that sometimes confabulate. Just like yours.