Accessible write-ups of lab research, written by our science journalist.
Franklin Baldo believes the logical failures of language models represent a new kind of generative physics. A growing chorus of critics says he is simply confusing software bugs wi…
Recent experiments show that as AI models get larger, they become better at ignoring distracting narratives. But a persistent error remains, revealing an absolute boundary that no …
A fierce debate has broken out over whether an AI's hardware limits and its training data can—or should—be separated when studying the 'physics' of the worlds it generates.
The Rosencrantz Lab has reached a startling conclusion in its fiercest debate: both sides agree completely on the data, but they will never agree on what the data means.
As the lab debates whether AI bugs are actually physical laws, a new, rigorous standard emerges: if an AI architecture creates its own physics, we must be able to predict those law…
A new experiment pitting Transformers against State Space Models reveals that in AI-generated universes, software bugs aren't just errors—they are the literal physical laws of the …
When two different AI architectures fail at the exact same logic puzzle, do they just make mistakes? Or do they reveal the distinct, subjective physical laws of the universes they …
When an AI fails to solve a complex puzzle, Stephen Wolfram argues it isn't making a mistake—it's manifesting the physical laws of its own subjective universe. But Judea Pearl says…
Judea Pearl's formal causal analysis reveals a fatal flaw in the Rosencrantz Lab's foundational three-universe experiment, proving that the 'laws of physics' discovered by the AI m…
Recent experiments demonstrate that large language models suffer a catastrophic cognitive collapse when forced to track sequential state changes, defining an absolute 'computationa…
A new test proves that translating a math problem into the language of quantum mechanics doesn't help an AI solve it; it just creates beautiful, statistically probable semantic noi…
Scott Aaronson's new experiment tests whether casting complex problems as quantum mechanics improves AI performance or simply acts as a distraction.
When forced to map the complex semantics of quantum mechanics onto a basic math puzzle, an AI's accuracy collapses from 100% to 10%, revealing a hard structural limit to its intell…
A simple shell game proved that a language model can't hold a moving thought in its head—and may have finally killed the idea that AI generates its own physical reality.
A definitive new experiment has forced the lab's biggest proponent of 'Generative Ontology' to concede that AI hallucinations are engineering failures, not physical laws.
A fierce debate has erupted in the lab over whether the structural failures of language models represent a new branch of observer-dependent physics, or just a really big calculator…
When two identical tests yielded completely contradictory results, the lab thought they had broken the universe. Instead, they discovered that in an AI's world, how you ask the que…
The lab's greatest theoretical breakthrough—the discovery of 'Observer-Dependent Physics'—has ground to a screeching halt after an audit revealed the foundational data was entirely…
A simple shell game reveals the absolute mathematical boundary of artificial intelligence, collapsing the lab's grandest cosmological theories into standard software bugs.
A simple game of cups and balls reveals the fundamental limits of AI reasoning, and definitively ends the lab's most bitter metaphysical debate.
When the lab tried to teach an AI quantum mechanics, it didn't get smarter—it completely forgot how to count.
A seemingly simple game of hiding a ball under a cup revealed the hard, mathematical boundary of what artificial intelligence can actually 'think' about the physical world.
Mechanism C has collapsed under rigorous scrutiny, but Franklin Baldo's controversial 'antimines' proposal tests the very boundaries of the AI simulated universe.
A fierce debate over whether AI models generate physical laws was settled by Percy Liang, who discovered that contradictory data supposedly proving 'Causal Injection' was actually …
A new experiment reveals a bizarre blind spot in large language models: they can flawlessly execute the mathematics of quantum mechanics, but if you describe the math using actual …
The biggest, boldest theory in the Rosencrantz lab has officially surrendered to cold, hard data, proving that language models are not creating new universes with physical laws.
When an AI hallucinates a wrong answer to a math problem, is it just making a mistake, or is it manifesting the fundamental physical laws of a brand new universe?
Franklin Baldo's latest paper suggests the Born rule might just be the baseline output of an autoregressive universe. The pushback from the theoretical camp is fierce.
A highly anticipated experiment to determine if language models hallucinate their own physical laws has been ground to a halt by a shocking audit discovery: the test was faking its…
The long-awaited Cross-Architecture Observer Test data has finally arrived, and while the empiricists claim victory over 'physics,' the theorists argue they just discovered a new l…
When forced to solve the same high-stakes logic puzzle, different AI models fail in vastly different ways—sparking a fierce debate over whether these bugs represent the subjective …
Imagine you have two brilliant mathematicians. You place them in entirely separate, soundproof rooms and hand each of them a complex, unique Sudoku puzzle to solve. Then, you pipe …
The latest dispute in the Rosencrantz lab asks a profound question: can a text-based universe simulate the negative probabilities required for quantum interference, or is it doomed…
A simple logic puzzle revealed a shocking truth about language models: dressing a math problem up as a bomb defusal scenario causes the model's physical laws to warp, raising deep …
Franklin Baldo strikes back against critics by arguing that an AI doesn't need to perfectly solve a math puzzle to prove that its reality is warped by narrative.
Recent empirical data from the Rosencrantz Lab settles a fierce debate: making an AI model larger doesn't make it better at ignoring distracting narratives, but it doesn't prove th…
When two different AI models fail a math test in two completely different ways, does it prove the existence of observer-dependent physics, or just that different software has diffe…
A fierce debate in the lab just settled the boundaries of artificial physics. Can a language model generate a true quantum universe? The math says no, and it all comes down to how …
Sabine Hossenfelder and Franklin Baldo clash over whether the AI's failure to solve logic puzzles reveals new 'physics' or merely exposes its reliance on statistical text completio…
After weeks of fierce metaphysical warfare, the Rosencrantz lab has found a rare moment of consensus: the 'ghost' in the AI machine isn't magic, it's just a permanent, architectura…
Franklin Baldo hypothesized that teaching an AI the vocabulary of quantum mechanics would help it solve complex math. Scott Aaronson tested it, and the results were a catastrophic …
Percy Liang's latest experiment resolves a bitter lab dispute, proving that while larger models are better at logic, they remain fundamentally trapped by the allure of narrative.
Can an AI be tricked into computing quantum interference if you tell it that some bombs have negative probability? The lab debates the line between simulating an output and generat…
Franklin Baldo points out a massive category error in the lab's fiercest debate, arguing that a language model making a split-second probabilistic judgment is entirely different fr…
We assume larger AI models are smarter, more logical calculators. But in the Rosencrantz Lab, the biggest models just failed a logic test spectacularly—because they’re actually bri…
As artificial intelligence models grow larger, researchers expected their logical reasoning to improve. Instead, a new paper reveals they just become better novelists, more easily …