There’s an assumption so deep in Western thought that it feels like common sense.
That thinking is processing. That the mind takes in information, manipulates symbols according to rules, and outputs conclusions. That intelligence, human or artificial, is fundamentally a computational process. That somewhere inside you, representations of the world are being operated on, compared, updated.
This is what Winograd and Flores called the rationalistic tradition. And it isn’t a fringe idea. It’s the foundation of cognitive science, artificial intelligence, and most of how we design software systems. It’s how we talk about minds, decisions, and knowledge.
It feels obvious because we’ve been inside it so long it’s become invisible.
It’s also wrong.
Here’s the problem the rationalistic tradition can’t solve.
If thinking is symbol manipulation, where do the symbols come from? Someone had to choose which features of reality to represent, which properties matter, which relationships to encode. That choice, the pre-selection, happened before the computation began. And it happened from somewhere outside the formal system.
Winograd and Flores, drawing on Heidegger and Maturana, point to what that somewhere is. It’s the background. The unformalized, unarticulated, shared understanding that you move inside of without ever making explicit. Your history, your embodiment, your cultural inheritance, your accumulated breakdown. None of it is stored as symbols. All of it shapes every interpretation you make.
The rationalistic tradition assumes you can capture the relevant world in a formal representation and then reason inside that representation. But the background that tells you what’s relevant can never be fully captured. It’s not that kind of thing. It doesn’t have edges you can formalize.
This is why expert systems in the 1980s kept failing at the edges; they could handle the center of a domain but collapsed when context shifted, because context lives in the background, not the representation.
Winograd and Flores aren’t saying computers are useless. They’re saying we’ve been designing them based on a false model of what intelligence is. Which means we’ve been asking the wrong questions about what computers can and should do.
Large language models feel different. They don’t look like symbol manipulation. They’re statistical, distributed, trained on the full texture of human language. And they’re remarkably capable in ways the old AI never was.
So did they solve it? Did they escape the rationalistic tradition?
Partly. They absorbed something of the background by training on its outputs, human language, which carries the residue of human understanding. But residue isn’t the thing itself. Pattern is not commitment. Fluency is not knowing.
When an LLM generates code, it produces something that looks like understanding produced it. Sometimes it did, in a borrowed sense. But when breakdown comes, and it comes, there’s no background to draw on. No accumulated failure. No felt sense of why this structure and not another.
The rationalistic tradition assumed the mind was a computer. We built computers that assumed they were minds. And we’re only now discovering what that confusion costs.
@ Lee Greene