Transformers Learn Patterns, Not Why Humans Think

Attention mechanisms in transformer networks have become the explanation for why large language models work at all.

Researchers point to the mathematical elegance of tokens selectively weighting one another across sequences, the way a model learns to focus on relevant information rather than process everything equally. Declare this the substrate of intelligence itself. It's also exactly what connectionist networks looked like in 1986.

David Rumelhart, Geoffrey Hinton. Ronald Williams published their backpropagation paper that year, proving neural networks could learn robust internal representations through iteratively reducing prediction error. Researchers wrote papers about how connectionism dissolved the mind-body problem, how neural computation vindicated embodied cognition, how we'd finally found the mechanism that explained intelligence.

What they'd actually found was the mechanism that explained pattern recognition at scale. The category error repeated itself through the 1990s as researchers confused "the system learns correlations in data" with "this is how humans learn." A connectionist network could be trained to recognize faces or parse grammar without ever understanding either — it was powerful and instrumental but not explanatory of human cognition, because explaining what a system does is categorically different from explaining why a human does it. That why always involves intention, memory, social context, discontinuity.

Attention mechanisms have reproduced this confusion perfectly. Transformers demonstrably work, scale. Learn statistical regularities across vast datasets to produce useful outputs, yet none of that proves attention is the essence of intelligence rather than just the most useful trick we've found for making neural networks perform at enterprise scale. The difference this time is that we have the scar tissue. When someone tells you a mathematical mechanism is the foundation of thought, you can now ask the hard question. Have you proven the mechanism explains cognition, or have you just optimized it for prediction? Your ability to distinguish them — in AI papers, in career advice, in any expert claiming to have found the engine underneath some complex human behavior. Is what that prior failure bought you.