Ptolemy thought Earth sat in the center. The sun went round. Planets danced around us.
He was wrong.
But he didn’t know that. He just kept patching his map every time new observations broke it. Centuries of tweaks. Then Copernicus arrived. Sun-centered. Simple. Elegant. The revolution didn’t just change astronomy; it changed everything.
We’ve loved this story. Special relativity killed the aether. Continental drift drowned those sinking land-bridges.
Simple wins. That’s Occam’s razor. Attributed to William of Ockham in the 14th century, it’s the rulebook we teach students. Start small. Add variables only when you have to. Minimal assumptions. Fewer causes. Specific predictions.
It feels right. It’s safe.
“What if simpler is the enemy of accurate?”
Marina Dubova doesn’t think it is safe. She works at the Santa Fe Institute, a hub for complexity. She argues we are hiding truth by demanding simplicity.
We are building computer simulations. Putting humans in “micro-worlds.” Flipping the script on our own psychology.
The result? The most cherished rules of scientific method might be getting in the way.
And this matters now. Not just for us, but for the machines. We are building AI scientists. If we teach them to think like 14th-century friars, they will stay limited. We want AI that sees the hidden structures we ignore.
Thomas Lewton sat down with Dubova to pick this apart. Here is what she said about breaking the mold.
Simplicity is a human bug, not a feature
We are lazy creatures. Cognitive scientists have shown this. Psychologist Tania Lombbrozo found we prefer broad, simple explanations even when they are less probable.
Think about an alien patient. Two symptoms.
Option A: One disease causing both.
Option B: Two diseases, each causing one symptom.
The data says B is more likely.
People choose A anyway.
Why? It feels cleaner. We hate complexity. So we start there. Dubova tested this with agents. Some were told to minimize variables (Occam’s style). Others were told to go wild—create explanations with a thousand variables for a system that only needs three.
It sounds absurd. A thousand variables for one outcome?
It worked. Sometimes better than the parsimonious agents.
These complex models memorized details. They tracked temperature. Genetic risks. Air quality on the home planet. They found patterns the simple models missed.
Experimenting randomly might be better
Another rule? Theories must guide experiments. You don’t just poke at the universe blindly. You test a hypothesis.
Remember 1919? Eddington watched a solar eclipse. He tested Einstein’s general relativity.
Prediction: Light bends.
Observation: It bent.
Theory confirmed. Textbooks write this as the pinnacle of the method. Reason driving observation.
Dubova ran sims to test this. She had agents falsify theories through careful choice. Or confirm biases. Or pick experiments randomly. Or chase novelty.
The careful thinkers failed to find the best theory.
The random explorers won.
Novelty seekers won too.
She ran four more checks to see if she’d made a mistake. No. The guard rails were stopping progress.
Real neuroscientists do the same. In her study, scientists looked at a “toy brain.” The setup was weird—one region controlled multiple functions. This violated the standard “one-to-one” rule. One area, one job. Face recognition here. Language there.
The data showed otherwise. One region did both.
The scientists didn’t want to see it. They insisted there were subtle differences. They invented two slightly different regions where none existed. Their priors blinded them.
Science as touch, not sight
This is why we need a shift. Not because simple ideas are bad. But because they limit us.
Neuroscience used to believe in distinct modules. Genetics believed in single genes for traits.
Both were practical. They helped us start. We can’t handle thousand-dimensional math in our heads.
But now AI can.
Dubova points to double descent in statistical learning. We thought big models would memorize noise. Fail at new tasks.
They don’t.
Error rates drop, peak, then drop again. Massive complexity allows systems to capture reality so deeply they can generalize. They transform detail into rule.
If we automate science using old habits, we scale up our blindness.
Is that worth the risk?
Science is usually described as a mirror. We reflect nature. We watch it.
Dubova likes the philosopher Hasok Chang better. He calls for haptic realism. Mazviita Chirimuta joins her here.
It’s not looking.
It’s touching.
Poke the thing. Deform it. See how it reacts. When you touch reality, it changes under your finger. You miss parts. So you poke from another angle. Then another.
Occam’s razor says look at the smallest possible surface.
The razor suggests you stop poking.
Dubova wants us to poke harder. Explore the weird. Let the complexity in.
We have been too afraid of the noise. The noise might be the signal.
