We Need to Look into Machine Cognition
When we speak, it often feels like the words simply arrive. One word leads to the next, then the next, until eventually we reach the end of a thought. It’s an iterative process — a step-by-step stream of language emerging in real time.
But that raises a deeper question.
Is speaking itself the process of thinking? Or is it simply the surface output of a deeper process happening outside our awareness?
It often feels like the latter.
Ideas seem to exist just outside of conscious reach. The next word you say, the next idea you express — they appear as if they’ve been assembled somewhere beneath the surface. Our conscious mind only sees the final output.
Sometimes this hidden processing shows up visually rather than verbally. We might “see” an image in our mind. But even then, the same question applies:
How was that image constructed? Where did its details come from? What process generated it?
This is the real question of cognition: How are thoughts constructed before they reach consciousness?
Today’s AI systems — particularly transformer models — demonstrate something remarkable. With a relatively understandable architecture, they compress enormous amounts of information and generate outputs that often appear intelligent. They recognize patterns, synthesize knowledge, and sometimes produce surprisingly creative responses.
But the deeper challenge remains.
Intelligence isn’t just about knowing more information than a human. Many models already do that.
The real frontier is correct reasoning and knowledge synthesis — the ability to combine information in meaningful ways and produce reliable, creative insights.
Humans do this imperfectly, but we still possess something powerful: a form of deep reasoning that emerges from processes we barely understand ourselves.
If we want AI systems to truly reach — or surpass — human-level intelligence, we may need to understand that hidden layer of cognition much better.
Not just what thoughts are produced, but how thoughts are constructed.
Because ultimately, that might be the key to building systems that don’t just mimic intelligence — but genuinely reason.