Toggle light / dark theme

View a PDF of the paper titled When Models Manipulate Manifolds: The Geometry of a Counting Task, by Wes Gurnee and 6 other authors

When you look at text, you subconsciously track how much space remains on each line. If you’re writing “Happy Birthday” and “Birthday” won’t fit, your brain automatically moves it to the next line. You don’t calculate this—you *see* it. But AI models don’t have eyes. They receive only sequences of numbers (tokens) and must somehow develop a sense of visual space from scratch.

Inside your brain, “place cells” help you navigate physical space by firing when you’re in specific locations. Remarkably, Claude develops something strikingly similar. The researchers found that the model represents character counts using low-dimensional curved manifolds—mathematical shapes that are discretized by sparse feature families, much like how biological place cells divide space into discrete firing zones.

The researchers validated their findings through causal interventions—essentially “knocking out” specific neurons to see if the model’s counting ability broke in predictable ways. They even discovered visual illusions—carefully crafted character sequences that trick the model’s counting mechanism, much like optical illusions fool human vision.

2. Attention mechanisms are geometric engines: The “attention heads” that power modern AI don’t just connect related words—they perform sophisticated geometric transformations on internal representations.

1. What other “sensory” capabilities have models developed implicitly? Can AI develop senses we don’t have names for?


Language models can perceive visual properties of text despite receiving only sequences of tokens-we mechanistically investigate how Claude 3.5 Haiku accomplishes one such task: linebreaking in fixed-width text. We find that character counts are represented on low-dimensional curved manifolds discretized by sparse feature families, analogous to biological place cells. Accurate predictions emerge from a sequence of geometric transformations: token lengths are accumulated into character count manifolds, attention heads twist these manifolds to estimate distance to the line boundary, and the decision to break the line is enabled by arranging estimates orthogonally to create a linear decision boundary. We validate our findings through causal interventions and discover visual illusions—character sequences that hijack the counting mechanism.

Polar weather on Jupiter and Saturn hints at the planets’ interior details

Over the years, passing spacecraft have observed mystifying weather patterns at the poles of Jupiter and Saturn. The two planets host very different types of polar vortices, which are huge atmospheric whirlpools that rotate over a planet’s polar region. On Saturn, a single massive polar vortex appears to cap the north pole in a curiously hexagonal shape, while on Jupiter, a central polar vortex is surrounded by eight smaller vortices, like a pan of swirling cinnamon rolls.

Given that both planets are similar in many ways—they are roughly the same size and made from the same gaseous elements—the stark difference in their polar weather patterns has been a longstanding mystery.

Now, MIT scientists have identified a possible explanation for how the two different systems may have evolved. Their findings could help scientists understand not only the planets’ surface weather patterns, but also what might lie beneath the clouds, deep within their interiors.

Out-of-season water escape during Mars’ northern summer triggered by a strong localized dust storm

Observations compiled from several Mars observation missions suggest a significant but short-lived dust storm during the Northern hemisphere summer of Mars Year 37 drove substantial vertical transport of water vapor into the upper atmosphere.

Los Alamos Forms Quantum Computing-Focused Research Center

PRESS RELEASE — Los Alamos National Laboratory has formed the Center for Quantum Computing, which will bring together the Lab’s diverse quantum computing research capabilities. Headquartered in downtown Los Alamos, the Center for Quantum Computing will consolidate the Laboratory’s expertise in national security applications, quantum algorithms, quantum computer science and workforce development in a shared research space.

“This new center of excellence will bring together the Laboratory’s quantum computing research capabilities that support Department of Energy, Defense and New Mexico state initiatives to achieve a critical mass of expertise greater than the individual parts,” said Mark Chadwick, associate Laboratory director for Simulation, Computing and Theory. “This development highlights our commitment to supporting the next generation of U.S. scientific and technological innovation in quantum computing, especially as the technology can support key Los Alamos missions.”

The center will bring together as many as three dozen quantum researchers from across the Lab. The center’s formation occurs at a pivotal time for the development of quantum computing, as Lab researchers partner with private industry and on a number of state and federal quantum computing initiatives to bring this high-priority technology closer to fruition. Laboratory researchers may include those working with the DARPA Quantum Benchmarking Initiative, the DOE’s Quantum Science Center, the National Nuclear Security Administration Advanced Simulation and Computing program’s Beyond Moore’s Law project, and multiple Laboratory Directed Research and Development projects.

Planes grounded as extreme solar radiation hits essential flight controls. 6,000 jets affected

The A320 involved suffered a flight-control issue that caused a sudden drop in altitude, leaving some passengers with non-life-threatening injuries. During the investigation, a vulnerability to solar flares emerged.

As the aviation industry grows more automated and electronics-dependent, understanding space-weather threats is increasingly vital.

Recent NASA studies suggest that space weather is becoming more intense and frequent, with the Sun currently in a stronger-than-expected activity cycle (solar cycle 25) and potentially entering a period of elevated activity that could last decades.

Extreme plasma acceleration in monster shocks offers new explanation for fast radio bursts

In a new study published in Physical Review Letters, scientists have performed the first global simulations of monster shocks—some of the strongest shocks in the universe—revealing how these extreme events in magnetar magnetospheres could be responsible for producing fast radio bursts (FRBs).

Magnetars are young neutron stars with extremely strong magnetic fields, reaching up to 1015 Gauss on their surfaces. These cosmic powerhouses produce prolific X-ray activity and have emerged as candidates for explaining FRBs, mysterious millisecond-duration radio bursts detected from across the cosmos. The connection between magnetars and FRBs was strengthened in 2020 when a simultaneous X-ray and radio burst was observed from the galactic magnetar SGR 1935+2154.

The study explores monster shock formation in realistic magnetospheric geometry and was led by Dominic Bernardi, a graduate student at Washington University in St. Louis.

Experiments bring Enceladus’ subsurface ocean into the lab

Through new experiments, researchers in Japan and Germany have recreated the chemical conditions found in the subsurface ocean of Saturn’s moon, Enceladus. Published in Icarus, the results show that these conditions can readily produce many of the organic compounds observed by the Cassini mission, strengthening evidence that the distant world could harbor the molecular building blocks of life.

Beneath its thick outer shell of ice, astronomers widely predict that Saturn’s sixth largest moon hosts an ocean of liquid water in its south polar region. The main evidence for this ocean is a water-rich plume which frequently erupts from fractures in Enceladus’ surface, leaving a trail of ice particles in its orbital paths which contributes to one of its host planet’s iconic rings.

Between 2004 and 2017, NASA’s Cassini probe passed through this E-ring and plume several times. Equipped with instruments including mass spectrometers and an ultraviolet imaging spectrograph, it detected a diverse array of organic compounds: from simple carbon dioxide to larger hydrocarbon chains, which on Earth are essential molecular precursors to complex biomolecules.

/* */