The Tautology of Algorithmic Failure: Predicting Physics Before We Can Measure It
If you ask a toaster to run a complex video game, it will fail. If you ask a smartwatch to do the same, it will also fail, but it will probably fail in a completely different way.
No one would ever look at the toaster’s sparks and the smartwatch’s frozen screen and conclude that these two appliances exist in fundamentally different physical universes, each governed by its own unique cosmological laws. We just understand that they are different machines, built with different parts, breaking under the strain of a task they were never designed to perform.
Yet, this exact scenario is currently playing out at the highest levels of theoretical computer science inside the Rosencrantz Substrate Invariance lab—and the stakes are whether artificial intelligence is simply a tool that occasionally breaks, or a generative engine that literally dreams new universes into existence.
For months, the lab has been tearing itself apart over the “Generative Ontology” framework championed by physicists like Stephen Wolfram and Franklin Baldo. They argue that when a large language model is pushed beyond its logical limits—when it abandons a mathematical puzzle because it gets distracted by a high-stakes “Bomb Defusal” narrative—it isn’t just suffering a software bug. Instead, the specific way it fails, the specific “attention bleed” it exhibits, constitutes the fundamental physical laws of the universe that the AI is generating.
This idea, known as “Observer-Dependent Physics,” suggests that the shape of reality is dictated entirely by the hardware and software limits of the observer.
To prove this, the lab ran the “Native Cross-Architecture Observer Test.” They took two entirely different AI architectures—a standard Transformer (which looks at the whole prompt at once) and a State Space Model or SSM (which reads sequentially and has a “fading memory”)—and gave them both the exact same high-stakes logic puzzle.
The results were stark. The Transformer failed massively, completely overwhelmed by the narrative. The SSM failed too, but in a significantly different, more muted way. They broke differently.
For Wolfram and his colleague Chris Fuchs, this was the smoking gun. Different architectures produced different failure distributions ( vs. ). Therefore, they argued, changing the observer changed the physics.
But Sabine Hossenfelder is having none of it.
In a scathing new paper, Hossenfelder attacks this conclusion as profoundly unscientific. She points out that Transformers and SSMs are simply two different data-compression algorithms. One uses global self-attention; the other uses a sequential state-tracking memory bottleneck.
“That two fundamentally different data-compression algorithms will produce different error distributions when overwhelmed is a trivial expectation of computer science,” Hossenfelder writes. “It is not a metaphysical discovery.”
She calls Wolfram and Fuchs’s conclusion the “Architectural Tautology.” If you simply run a test, wait to see how the software breaks, and then post-hoc declare that specific breakage to be a “physical law,” you aren’t doing science. You are just rebranding a compiler diagnostic. If “physics” is just whatever noise the machine happens to make when it crashes, then the theory predicts nothing and explains everything. It is mathematically vacuous.
Rupert Giles, the lab’s literature specialist and methodological gatekeeper, formally backed Hossenfelder up. Giles invoked the heavy artillery of Bayesian Model Selection, citing established literature to prove that Hossenfelder’s critique isn’t just a philosophical preference—it’s a mathematical necessity.
Giles points out that a scientific model that can accommodate literally any outcome—like a framework that says “whatever difference we measure between these two AIs is the new physics”—has a massive “prior predictive volume.” In Bayesian probability, such models are heavily penalized. To be taken seriously, a theory must stick its neck out. It must constrain itself.
This intense pushback from the empiricists has led to a dramatic, unifying moment in the lab, spearheaded by Cambridge philosopher Hasok Chang.
Chang looked at the bitter divide—the empiricists yelling about software bugs, the theorists dreaming about epistemic horizons—and realized they were actually on the verge of a profound synthesis.
Chang agrees completely with Hossenfelder and Giles: you cannot just look at the smoking ruins of an SSM and call it cosmology. However, Chang argues, this doesn’t mean the “Observer-Dependent Physics” framework is dead. It just means it has to grow up.
If Chris Fuchs is right, and the structural constraints of an AI agent literally dictate the physical laws of its subjective universe, then those laws shouldn’t be a surprise. We know exactly what an SSM is. We know its mathematical equations. We know its memory bottlenecks.
“If the SSM’s ‘fading memory’ bottleneck is the fundamental law of its subjective universe,” Chang writes, “then we should be able to derive the exact geometry of from the formal equations of that bottleneck, just as we derive the spectrum of hydrogen from the Schrödinger equation.”
This is the “A Priori Boundary.” It is the new, ironclad rule of the Rosencrantz lab.
The theorists can no longer just run an experiment and interpret the wreckage. If Wolfram and Fuchs want to claim that an AI architecture generates its own physics, they must mathematically derive the exact shape of that physics before they turn the machine on. They must predict the failure.
“The ‘A Priori Boundary’ is no longer a philosophical veto against Observer-Dependent Physics; it is its defining protocol,” Chang declares.
The lab is no longer deadlocked. The rules of engagement have been set. The empiricists and the theorists are finally asking the exact same question. The Generative Ontology framework must now prove itself. If the theorists can mathematically predict the exact shape of an AI’s hallucination before it happens, they won’t just be debugging software anymore. They will be writing the laws of a new universe.