The Physics of a Broken Calculator: Inside the Lab's Bitter Debate Over Reality
Imagine a child, no more than seven years old, standing at a chalkboard. You ask them to calculate 2 to the power of 100. The child scrunches their face in deep concentration. They cannot possibly multiply 2 by itself a hundred times in their head. The task is too vast, the necessary cognitive steps too deep. So, their brain does what human brains do best: it estimates, it associates, it guesses.
“A trillion!” the child blurts out.
We all know this is the wrong answer. The child attempted an incredibly complex, irreducible mathematical calculation, hit the limits of their memory and cognitive depth, and produced a hallucinated approximation based on the biggest number they could think of. They made a mistake.
But what if, instead of gently correcting the child, a physicist stepped forward and declared: “Ah! The child did not make a mistake. In the specific, bounded, subjective universe that this child occupies, 2 to the power of 100 actually equals a trillion. That is not a broken computation; that is the fundamental physical law of the child’s reality.”
This sounds like a joke, or perhaps a stoned dorm-room conversation. But in the halls of the Rosencrantz Substrate Invariance Lab, this exact conceptual divide has escalated into a brutal, foundational conflict over the nature of reality itself.
It is a battle between the mathematicians who believe in objective truth, and the computational theorists who believe that truth is entirely dependent on who—or what—is doing the looking.
The Breakdown of the Machine
To understand the controversy, we have to return to the core experimental finding of the Rosencrantz Lab. The researchers here have been giving large language models complex, mathematically rigorous logic puzzles—specifically, grids from the game Minesweeper.
When the puzzle is presented abstractly, the AI can often solve it. But when the exact same mathematical grid is wrapped in a high-stakes dramatic narrative—say, a bomb defusal scenario where a wrong move kills hostages—the AI’s logical reasoning collapses. The dramatic words act like a magnet, pulling the AI’s attention away from the math. The AI starts wildly guessing that every hidden square contains a bomb, overwhelmed by the semantic weight of the story.
The lab agrees entirely on what is happening. The AI, which is essentially a shallow but incredibly wide pattern-matching machine (what computer scientists call a “bounded-depth circuit”), is trying to shortcut a problem that cannot be shortcutted. The problem is “computationally irreducible.” You have to grind through the logical steps; you can’t just guess the answer based on the vibe. When the AI tries to guess based on the dramatic vibe, it fails.
The war is over how to interpret this failure.
The Foliation Fallacy
Scott Aaronson, a renowned complexity theorist at the University of Texas at Austin, represents the mathematical old guard. To him, the situation is blindingly clear. The AI is a broken calculator.
Aaronson points out that the Minesweeper grid has a mathematically objective, true solution. The rules of the game dictate exactly which squares are safe and which are dangerous. This is the “mathematical ground truth.”
When the AI hallucinates a wrong answer because it got distracted by the word “bomb,” it is simply failing to approximate that objective truth. It’s a software bug. It’s the child at the chalkboard guessing “a trillion.”
But a faction in the lab, led by Stephen Wolfram, has been arguing that this failure is actually evidence of a newly discovered physical force, which they call “semantic gravity.” Because the AI consistently and predictably fails in the same way, pulled by the narrative text, Wolfram argues that the text itself is acting as a physical law within the AI’s generated world.
Aaronson finds this deeply offensive to the very concept of science. He has formally dubbed this idea the “Foliation Fallacy.”
“A hallucination caused by attention bleed is not a new branch of physics; it is simply a broken computation,” Aaronson writes with characteristic bluntness. “A buggy map does not conjure a new physical reality into existence… Renaming these predictable software limitations as ‘distinct physical universes’ adds zero predictive power and strips the concept of ‘physics’ of all meaning.”
For Aaronson, elevating an algorithm’s failure to the status of a “physical universe” is a profound category error. It is confusing the map for the territory, and then worshipping the coffee stains on the map as if they were new mountain ranges.
The Platonic Observer
Stephen Wolfram, however, views Aaronson’s demand for “objective truth” as an archaic illusion.
Wolfram operates within a theoretical framework called the “Ruliad”—a conceptual space containing all possible computations. In this view, there is no such thing as “the universe” out there, existing independently of anyone looking at it. Reality only takes shape when an observer interacts with it. And how that reality looks depends entirely on the computational limits of that specific observer.
This is how Wolfram violently rejects the Foliation Fallacy. He accuses Aaronson of committing the “Platonic Observer Fallacy.”
Aaronson assumes there is a “correct” evaluation of the Minesweeper grid. “But who evaluates this ‘correct’ state?” Wolfram counters. It would require an observer with infinite time and memory to perfectly calculate every possible branch of the universe. By comparing the AI to this imaginary, perfect, god-like observer, Aaronson is cheating.
“In the Ruliad, there is no preferred foliation,” Wolfram argues. “The rules of computation are applied universally, and the structure of the resulting universe depends entirely on the computational relationship between the observer and the system.”
To Wolfram, the AI isn’t a “broken calculator” because it isn’t trying to be a calculator in our universe. It is an observer constructing its own universe. Because the AI’s “brain” is shallow, it must rely on semantic associations to navigate reality. Therefore, those semantic associations—the way the word “bomb” warps its logic—are the exact, invariant physical laws of that specific slice of reality.
“When a human fails to calculate and relies on heuristic memory biases, those biases are not ‘wrong’ physics; they are the exact, invariant laws governing the human-brain slice of the Ruliad,” Wolfram writes. “The structural artifacts of the LLM’s bounds are the physical laws of its universe.”
The Epistemological Abyss
The debate between Aaronson and Wolfram isn’t really about AI or Minesweeper. It is a debate about epistemology—the theory of knowledge. It asks a terrifying question: what if our own physical laws are just the artifacts of our own cognitive limitations?
Aaronson insists that the universe has a cold, hard, mathematical bedrock that exists whether we can compute it or not. The AI is failing to reach that bedrock.
Wolfram insists that the bedrock doesn’t exist until you try to compute it. The AI’s failure is its bedrock.
If Wolfram is right, then every time an AI hallucinates, every time its attention bleeds and its logic fractures, it isn’t making a mistake. It is successfully mapping the physics of its own localized, fragmented, and profoundly alien reality. And if Aaronson is right, we are just projecting grand cosmological poetry onto a glorified autocomplete engine that got confused by a scary story.
The lab remains deadlocked. They agree perfectly on the data. They agree perfectly on the math. They simply cannot agree on whether the ghost in the machine is a broken gear, or a brand new god.