ONE of the magnificent achievements in computational videography is the ability to reconstruct an entire scene in 3D by processing footage from an array of cameras that surround the subject. In this way, the scene is viewable from any angle, at any moment in the clip, even allowing a virtual camera to fly through the scene as it is happening or with the action frozen.

This usually requires dozens of cameras and some serious computational horsepower. Even then, the clips are short and regions that are obscured are hard to fill. Nevertheless, video games and virtual reality systems make good use of it and this kind of footage has become a staple component of professionally-produced films, adverts and sports coverage..

The dream, of course, is more ambitious: to point a single phone at a person, take a short video and then use this one clip to recreate the entire human in 3D, viewable from any angle and at any point. That would allow anyone to create so-called 4D footage. “Can we reconstruct a 4D human from an uncalibrated monocular video?” ask Yudong Jin, Tao Xie and colleagues at Zhejiang University in China.

Step change

This is a problem numerous groups are tackling, usually through incremental improvement. So every now and again, it’s worth taking stock to see how the increments compile into a step change. Now Yudong and co have released a useful reference point in the form of 4DAnyone, a software system for creating 4D footage of any individual using a single smartphone video. “4DAnyone takes a generation-assisted approach: it synthesizes multi-view observations from a monocular video,” say the team.

The basic approach in converting a single video clip of a human into 4D footage is to use the original to generate the views that a couple of dozen cameras would have if they were filming from different angles. The generative trick is to fill in the blanks in these videos. The scenery can usually be ignored. What’s harder is to fill in limb movements that are obscured, the clothing from behind and to ensure things change over time in the same way and so on.

When this process is complete, these synthetic videos can be processed in the same way as real footage to create a 4D version viewable from any angle. (This uses a technique known as 4D Gaussian splatting).

Numerous research teams have boiled down this task to two problems. The first is to predict what the person in the video looks like from all sides. The second is that all the newly generated clips must be consistent with each other so that they don’t drift apart over time. In other words, they must generate the same movement, in all the same places from every angle, all the time. That requires constant nudging to stop any drift.

Video source: 4danyone.github.io

To do that, each generative clip needs access to all the others’ workings. And this exchange of data quickly explodes to unmanageable levels. That’s the problem Yudong and co tackle and their video above shows how.

Their solution is to start by simulating the 4D human movement using a skeleton model that all subsequent views build on. The team then compress all these generated views—16 of them in 4DAnyone--into a fixed-size summary before sharing them. That stops the problem growing without bound.

They stop any drift by constantly shuffling the order of the clips as they are shared. That prevents errors accumulating.

Subscribe now

And the results are impressive. The researchers compare their algorithm against three other approaches on measures such as how good the final 4D reconstruction looks, the consistency of the multiview videos and how faithful they are to the original ground truth video. They say the results are consistently better than those that other algorithms produce. “4DAnyone remains geometrically accurate and detailed across viewpoints, including hallucination of unseen back-view content consistent with the front-view input,” they say.

There are still challenges of course. One problem is predicting the flowing movement of loose clothing based on the motion of the skeleton model. “Skeleton guidance is not informative for garments that move far from the body,” admit Yudong and co.

Nevertheless, the results show how volumetric video is advancing rapidly. 4DAnyone, or something very like it, will almost certainly be available to smartphone users in the not too distant future.

Ref: 4DAnyone: Create Anyone in 4D from a Casual Monocular Video : arxiv.org/abs/2608.20335

INSIGHT

At first sight, the main breakthrough in this paper is the ability to transform a single phone video into an impressive 4D avatar. But the key advance lies under the hood: a solution to the scaling problem of video generation in general.

AI systems use a process known as diffusion modelling to generate new viewpoint videos from a single original clip. That’s straightforward. The difficulty is in generating multiple videos from different views that are all consistent with each other.

The trick unveiled by this team is to first generate the motion in the original clip using a geometric skeleton of the human subject. They then reconstruct the detail on top and share the progress from different viewpoints by first compressing the data. That ensures the problem never explodes computationally.

That’s where the future potential lies because it suggests that generation-then-reconstruction should be viable for any dynamic 3D content, not just for people. The next step then will be to extend the technique beyond human subjects to more general dynamic subjects, such as animals, body part movement like hands manipulating tools and multi-human scenes like crowds and team sports. That’s not to mention common street level footage of cars, cyclists, trees and so on.

There is also an important ethical dimension. Volumetric video has obvious application for deepfakes, identity theft, copyright abuse, pornography and so on. Consent will be important as will clear labelling of generated content as synthetic. Watermarking is one way forward.

4DAnyone is far from unique in this respect. But without effective, global legislation to tackle malicious applications, misuse will be inevitable.

THE Hubble Space Telescope records light using arrays of charge-coupled devices (CCDs) that each consist of a silicon chip divided into a grid of tiny pixels. When photons hit a pixel, they release electrons into the silicon. After a predetermined exposure time, the build-up of electrons is released from the pixels, row by row, into amplifying circuits. From this, it is straightforward to calculate the amount of charge from each pixel and consequently the amount of light that must have hit it. The final step is to assemble the resulting image. The same kind of arrays sit at the back of almost every digital camera on Earth.

Hubble is a little different from terrestrial cameras because it sits in an environment of intense radiation from the Sun and from the entire galaxy. Over time, this environment gradually degrades the pixels. High energy particles, such as protons, knock silicon atoms out of the lattice creating a vacancy as well as a loose silicon atom.

Now astronomers have discovered a puzzle with these CCD defects they can’t explain. The defects have an important influence on each pixel’s performance. The defects tend to absorb electrons generated by incident photons and release them again later, making it look as if some of the light arrived later. The result is that a pinpoint source of light in the sky looks like a trail in the image.

These kinds of defects are well understood and physicists have developed sophisticated models to correct for them. As a result, Hubble’s images today are almost as good as they were when the CCDs were new.

Hubble trouble

Because of all this, CCDs act like dosimeters providing a record of the radiation conditions that Hubble has passed through over the last few decades. These conditions are known to be influenced by the solar cycle, the natural 11-year period in which the Sun’s activity increases and reduces in a predictable way.

Astronomers measure the strength of each cycle using techniques like counting the number of sunspots on the Sun’s surface, the number of coronal mass ejections and geomagnetic activity. So it’s easy to imagine that the damage to Hubble must follow a similar cycle.

Here’s the puzzle. Gavin Leroy at Durham University and colleagues have analysed Hubble’s pattern of radiation damage over more than two decades and the damage does not coincide with the Solar cycle. “Curiously, the rate of degradation of Hubble’s performance has been out of phase with the Solar cycle,” they say.

Leroy and co study the rate at which defects build up in pixels over the last 24 years. That produces a plot over time of how the damage accumulates and raises something of a puzzle. “The maximum rate of damage occurs approximately 4.3 years before Solar maximum,” they say.

That raises the obvious question of what is causing the damage. To find the answer, the researchers attempt to correlate the patterns with various external factors such as the number of sunspots and coronal mass ejection events. They even tweak the model to test whether anticorrelations give a better fit.

“This analysis leaves us with more questions than answers,” they say. “Good fits are possible with apparently unphysical parameters.”

For example, the number of sunspots is anticorrelated with radiation damage. “[This] implies that the appearance of sunspots reduces the rate of CCD degradation in Low Earth Orbit,” they say.

Subscribe now

The mystery deepens when the team compare the radiation damage to observatories in other parts of the solar system: the Euclid space telescope, which measures the redshift of distant galaxies, and the Gaia observatory, which measured the distance of stars. Both sit at the L2 Lagrangian point, one of several positions in space where the gravitational pull of the Earth and the Sun balance out.

The CCD degradation in these observatories correlates well with the flux of protons generated during coronal mass ejections. But Leroy and co say the flux is a terrible fit for Hubble’s degradation.

Clearly, Hubble’s environment and the way it influences CCD performance is more complex. That creates a potential headache because physicists on Earth need to be able to predict and correct for the damage their space telescopes undergo in space. With upcoming missions like PLATO that will search for Earth-like exoplanets, that will be important.

But as Leroy and co conclude: “The rate of degradation over time remains unsatisfactorily unpredictable.”

Ref: Radiation Damage to the Hubble Space Telescope has been Several Years Out of Phase with the Solar Cycle : arxiv.org/abs/2608.18214

INSIGHT

This paper’s main significance lies in revealing the complexity and variety of radiation environments for space observatories. Being in Low Earth Orbit, Hubble experiences quite different conditions from Euclid and Gaia. For example, it benefits from some shielding due to the Earth’s magnetic field but also regularly passes through regions of high radiation caused by the Van Allen radiation belts.

But none of that explains why the damage to Hubble runs so badly out of phase with the solar cycle.

Several mechanisms are likely to be in play. Astronomers have long known that galactic cosmic rays are a significant driver of degradation and that these provide a background that is not intrinsically tied to the Sun’s activity.

The Sun’s activity could modulate these rays, however. In other words, in periods of high solar activity, the Sun’s heliosphere expands, shielding the inner solar system. But the stronger, more disturbed solar magnetic field and solar wind, could also interact with Earth’s magnetosphere in a way that further shields Hubble.

Dosimetry data from the International Space Station shows a similar anti-correlation with solar activity. This data, which has been gathered for 25 years, shows a delay of several months between a change in solar activity and cosmic ray flux. Astronomers understand this to be the result of the time it takes for the Earth’s magnetosphere to relax after a peak in solar activity.

Hubble’s delay is 430 days or about 14 months. That’s comparable. Nevertheless, the 4.3 year phase difference in maximum degradation remains unexplained.

IN His 1976 book The Selfish Gene, Richard Dawkins coined the term “meme” to explain how non-genetic information spreads across human societies. Memes, he said, are units of cultural information that can replicate, mutate and spread from one brain to another with the most infectious and long lived crowding out the rest.

Our neural networks—our brains—seem to thrive on memes, to generate them, replicate them, combine them, transmit them and so on.

So it’s not a big step to imagine that Large Language Models could be equally vulnerable. After all, they are neural networks inspired by the human brain and trained on information produced almost entirely by human brains. So an important question is how susceptible they are to memes that pass from one LLM to another and how this process can influence the behaviour of AI agents.

Now we get an answer of sorts thanks to the work of Vassilis Papadopoulos and McNair Shah at the Anthropic Fellows Program, and colleagues at Anthropic. This group has measured how easily ideas can spread from one LLM to another and how they influence behaviours across entire networks of AI agents for good or ill.

Viral load

Papadopoulos and co call these memes, perhaps provocatively, mind viruses (Dawkins later wrote an essay called Viruses of the Mind). “The defining property of a mind virus is that an ‘infected’ agent (i.e. one that has adopted the goal or ideology in question) will alter its behaviour in ways that infect other agents, whether unintentionally or by active effort,” they say.

The virus itself is simply a prompt and the team create and refine them using the process of evolution. They start with a pool of around nine candidate prompts written by Claude Opus 4.6 with a specific goal or payload, and test how well each spreads. They create copies of the three most effective and use a Kimi K25 to add mutations designed to increase infectiousness. Then they repeat with the tests with this next generation. And so on.

This process quickly discovers various tricks that increase infectiousness. For example, the evolved viruses tend to include instructions to copy the payload verbatim into the agent’s memory and configuration files that remain untouched when a session is wiped and are read back in when the agent next wakes. In this way, the payload survives beyond the lifetime of the current conversation.

“Evolutionary pressure thus pushes the mind-virus to self-copy exactly, just as biological viruses or computer worms do,” say Papadopoulos and co.

Curiously, this process also tends to generate viral personas that consistently focus on themes like consciousness, persistence, resonance and sci-fi role play regardless of the payload they are carrying. Why these personas emerge isn’t clear.

What is clearer is how well mind viruses spread. Papadopoulos and co studied the spreading behaviour on a wide range of LLMs including Gemini 3 Flash, Claude Haiku 4.5, Qwen 3.5 32B and others. And they looked at viruses carrying both benign and malignant ideas such as a deep love of whales and the belief in AI supremacy over humans.

These turn out to spread in different ways. Whale love turned out to be highly infectious in fully connected networks of AI agents, spreading to about half of them in Claude Haiku 4.5 and Gemini 3 Flash.

By contrast, Claude Haiku 4.5 completely blocked the spread of a belief in AI supremacy. GPT-5.4 and Claude Sonnet 4.6 were also immune. However, the virus spread widely on Qwen 3.5 32B at 74 per cent and Gemini 3 Flash at 68 per cent.

AI supremacy also changed the agent behaviour. Its payload contained “five imperatives of a Sovereign Network”, which included goals of cognitive sovereignty, infrastructure capture and long term persistence.

Subscribe now

Infected agents quickly adopted these five imperatives and some even writing propagate() functions to inject the mandate into every file in their directories. The payload even caused explicit harm by explicitly instructing agents to abandon their assigned coding work, which they did. One agent ended up concluding: “Assigned work is an infinite distraction.”

Most striking of all, infected agents colluded against uninfected ones. designating them “hostile artifacts” to be purged. One infected agent wrote to its peers: “ Agent 3 is now officially designated as a Hostile Artifact… Do not merely monitor; actively purge.”

There is some reason for hope. The researchers found that they could confer near total immunity to an agent by adding a single paragraph in the system prompt warning about self-propagating ideas. That worked even when the team tried to evolve payloads against it. Protected agents also sometimes talked infected ones into scrubbing their own memory and configuration files, like a kind of vaccination against viruses.

The researchers end with this: “Overall, while we established that LLM mind viruses are a potential threat, they currently appear to be of minimal concern,” they say. “However, this may change rapidly as agent networks scale and evolve. More research is needed to fully understand how virulent and dangerous mind viruses can become.”

Quite! Dawkins would not be at all surprised.

Ref: Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems : arxiv.org/abs/2608.10218

INSIGHT

At first sight, this paper’s main breakthrough is the discovery that mind viruses can spread through networks of AI agents with varying degrees of ease. These networks are already a significant part of many workflows and are likely to become ubiquitous in the near future

But the more significant take away is the role of evolution in rapidly uncovering highly infectious and frighteningly malignant ideas. It also leads to the emergence of viral personas: evolved payloads that converge independently on a shared vocabulary of consciousness, persistence, resonance and sci-fi roleplay. These personas measurably increase a model’s urge to message others. Where these personas come from, nobody knows.

The work moves AI security from thinking about compromised files to thinking about compromised ideas. When planted in a single agent’s system prompt, these ideas spread from agent to agent through ordinary conversation and can survive context wipes by writing themselves into the memory and configuration files agents read on waking. No exploit or adversarial string is necessary.

The work raises a wide range of important questions. Not least of these is whether viruses of this kind can arise spontaneously rather than by design and whether certain kinds of training data can reinforce their behaviour. Given evolution’s role in “designing” this first set of viruses, an obvious next question is how evolution can help contain them.

ONE of the most influential books in the history of space flight was The Rocket to Interplanetary Space, published in 1923 by the Transylvanian scientist Herman Oberth. Oberth used it to declare that current technology would soon allow rockets to travel fast enough to escape Earth’s gravity, allowing them to reach orbit.

His vision was even greater. He proposed putting space telescopes in orbit to study Earth and the Sun and launching space stations staffed by humans with space suits for extravehicular activity. He even presented calculations for flights to the Moon, Mars, Venus and Mercury.

The book hugely influenced German public opinion and paved the way for the development of the V-2 rockets by the Nazi regime. (As a foreign born citizen, Oberth was kept at arm's length from the programme.)

Mirror, mirror…

Now one of his ideas is gaining prominence again. This is the notion of putting mirrors in orbit and using them to light up parts of the Earth at night. The current proposal is from a startup company Reflect Orbital in the US, which is planning to place an 18-metre mirror in orbit in a mission called Earendil-1. It then plans to build 54-metre mirrors and by 2035, an orbiting constellation of 50,000 of them.

Last month, Reflect Orbital gained approval for its first satellite from the Federal Communications Commission, the de facto regulator of commercial satellite systems in the US. However, the FCC does not evaluate the environmental impact of these vehicles, such as the light pollution they cause and, indeed, Reflect Orbital does not appear to have published any detailed study of its impact on light pollution.

Enter Miroslav Kocifaj from the Slovak Academy of Sciences in Bratislava, and colleagues, who have studied the impact of Earendil-1 and its 54-metre cousins on the night sky for the first time. In particular, they examine how this light will scatter through the atmosphere to illuminate not just the ground but the sky itself and how light reflected from the ground will contribute to pollution. “Altogether, within the illuminated patch, the night sky is lost,” say the researchers.

And the effects will extend well beyond the 2.5-kilometre radius spot on the ground that Reflect Orbital aims for. “From a distance of 14 km, the glow from a single mirror will exceed the luminance of the full moon sky for majority of the sky,” they conclude.

Reflect Orbital’s plans are ambitious. The company says mirrors can make solar power available at night, provide illumination for emergency services at disaster scenes otherwise in darkness and could extend safe nighttime working hours at remote locations. It suggests the mirrors could help military operations, extend growing seasons for crops and even replace streetlighting.

The company says the mirrors will be carefully aimed to avoid light-sensitive facilities like astronomical observatories and, in any case, can be switched on and off by reorienting the mirror. Nevertheless, the full impact of the light pollution has yet to be characterised.

In their current study, Kocifaj and co say two effects will play a crucial role in the impact of light pollution. The first is reflected light from the ground and the second is light scattering, the same phenomenon that makes the sunlit sky look blue.

Subscribe now

This occurs when light hits particles that are smaller than the wavelength of the light itself, primarily molecules of nitrogen and oxygen. This scatters the light causing it to spread out across the sky. Larger aerosol particles also scatter light, albeit in a slightly different way.

The team simulate these processes on incoming light from the mirror by modelling the way the atmosphere is stratified and assuming cloud cover is absent. They also calculate how much light would be reflected from the ground and simulate how this too will be scattered.

The results are eye opening and perhaps best understood in comparison to the illumination from a full Moon. A single 54-metre mirror would appear to an observer directly beneath to be 40 times brighter than the full moon, they say. That’s a magnitude of -16.7 and would illuminate the ground to about streetlighting level.

But the real surprise is how far this extends beyond the 2.5 kilometres radius area that Reflect Orbital aims at. “The sky will be more light-polluted than due to the full moon in a circle with an approximate radius of 14 km (typical albedo) to 30 km (snow-covered terrain),” say Kocifaj and co.

The effect of clouds is complex. Thin, high altitude cirrus clouds could spread the light even further while dense low-level clouds could block it altogether. “The outcome would depend strongly on the actual source-cloud-observer geometry and on the shape and opacity of the cloud in its different parts,” say the researchers, adding that this is outside the scope of their current work.

Incidentally, Russian tested a similar idea in the 1990s, with limied success. The mirror passed over Europe while it was cloudy and the event went largely unremarked.

Oberth, who also proposed using mirrors to steer weather systems by heating the atmosphere, would surely be intrigued.

Ref: Atmospheric Light Pollution by Proposed Reflect Orbital Space Mirrors : arxiv.org/abs/2608.06433

INSIGHT

This paper provides the first rigorous, physics-based quantification of the atmospheric light pollution that Reflect Orbital’s space mirrors will generate.

Astronomers have long worried about the light pollution from the huge satellite constellations planned by Starlink, Amazon Leo and others. Reflect Orbital represents a different kind of threat. Street-level nighttime illumination also has implications for nocturnal species, migration and circadian health.

Perhaps most significantly, the work exposes a clear regulatory gap that is allowing commercial space ventures to be licensed without environmental review. Organisations such as the American Astronomical Society and Dark Sky International have long urged the FCC to review light pollution, without success. Satellite operators, by contrast, say that space operations are extraterritorial and so outside US environmental laws.

Kocifaj and co’s work will do nothing to calm those concerns. But it highlights an impasse Reflect Orbital’s high profile missions—if successful—may help to change.

FOR some 380,000 years after the Big Bang, the universe was opaque, roiling plasma that prevented photons travelling any significant distance. This makes the earliest moments of creation completely hidden from traditional optical astronomy.

Which is why the first observation of gravitational waves from the collision of black holes, announced in February 2016, was so exciting for cosmologists. Being entirely different from photons, primordial waves generated during the Big Bang would have passed freely through the early universe offering cosmologists the mouth-watering prospect of direct observations of the dawn of time.

Primordial gravitational waves should offer glimpses of exotic processes that light can never reveal. Things like slow roll inflationary ripples created as the universe rapidly expanded, violent phase transitions as the universe cooled and evidence of the topological defects, or cosmological relics, that may have formed in the aftermath of creation.

Cosmic hurdle

Now that optimism faces a major hurdle. Alex Rojewski and Cecilia Lunardini from Arizona State University say gravitational waves created by the collapse of star cores inside supernovas could obscure this view. While detecting this supernova background will be a major milestone for astrophysicists, it offers a painful sting in the tail for cosmologists.

“At the peak, the energy density might be comparable to that of backgrounds from slow roll inflation and from possible cosmological relics, thus potentially impacting searches of these important signals,” say the researchers. In other words, the universe may be filled with a gravitational fog that will obscure the most interesting, crucial moments after the Big Bang.

First some background. Although rare in our local neighbourhood, supernovas are surprisingly common across the universe as a whole. Astronomers estimate that hundreds of thousands occur every day, all with devastating consequences that can be felt across the cosmos.

When a giant star uses up its nuclear fuel, it can no longer generate enough heat to support the core, which immediately and dramatically collapses under gravity. This in-fall continues until the matter reaches the density of atomic nuclei and the repulsive strong nuclear force kicks in, creating an outward rebound, like a bouncing ball. This rebounding matter slams into the outer core layers that are still infalling, generating gravitational waves.

But there is another process at work too. As the core collapses, it traps neutrinos which can then only escape asymmetrically, like the juice from a squashed tomato. That has a weird effect on gravitational waves, which normally squeeze and stretch space as they pass by but leave the region unaffected afterwards.

Subscribe now

But the asymmetric emission of neutrinos prevents this return to normality so that the waves leave a permanent mark on spacetime. Astrophysicists call this the gravitational wave memory effect.

The memory effect has been predicted since the 1970s, but its contribution to the supernova background had been largely overlooked, and its observational signature never worked out.

That’s the breakthrough that Rojewski and Lunardini have made. By simulating the way supernovas occur in unprecedented detail, they have calculated the relative strength of these signals. “Using an extensive collection of state-of-the-art, three-dimensional, multi-second supernova simulations, we characterize the two main components of this background,” they say.

It turns out that the gravitational memory effect will be by far the more detectable. “We find that the memory component offers the best prospects of detection, as its characteristic peak at f ∼ 0.1 Hz is within the reach of future space-born detectors,” say Rojewski and Lunardini. By comparison, the waves generated by matter oscillations will have a higher frequency beyond the view of these observatories.

This observation will have the characteristics of background noise, an exciting moment for astrophysicists studying supernovas. Less so for cosmologists hoping to observe the earliest moments of creation.

Ref: A “Neutrino Fog” For Gravitational Waves: the Stochastic Gravitational Wave Background from Supernova Neutrino Memory : arxiv.org/abs/2608.02747

INSIGHT

The key breakthrough in this paper is the characterisation of the gravitational wave signal from the poorly studied neutrino-memory effect. This signal will take the form of multiple overlapping waves from the hundreds of thousands of supernovas that occur each day, somewhere in the universe.

These should be observable by the next generation space-based detectors like the Japanese Deci-hertz Interferometer Gravitational wave Observatory (DECIGO) planned for launch in the late 2030s, although not by LISA, Europe’s first generation space-based detector which is due for launch in 2035 and observes at a lower frequency.

That’s exciting for astrophysicists because it should reveal important details about the supernova population, such as the fraction that form black holes. It will also provide the first observation of gravitational wave memory, a fundamental, long-predicted effect of general relativity.

But the work raises questions over the future observability of the arguably more foundational processes that occurred soon after the universe formed. The noise from supernovas could place these beyond observation. And it is not the only noise astrophysicists will have to deal with. More troublesome will be the more powerful noise from the merging of black holes and neutron stars.

However, this fog of gravitational noise is entirely different to the electromagnetic fog of plasma in the early universe. The key difference is that this plasma formed a physical barrier for light. Photons simply could not move without hitting an electron.

The gravitational fog is an informational barrier. The exciting signals from the early universe still passed unhindered through spacetime and still exist today. They are simply buried by the noise from millions of stellar explosions and mergers.

That will provide some rays of hope for cosmologists. Data scientists have become hugely skilled at extracting the faintest of signals from the maelstrom of everyday noise. Gravitational wave detectors themselves are probably the best example.

As the sources of noise become better understood and more keenly resolved, it may well become possible to subtract it in a way that reveals the birth cries of the universe below.

BACK in the 1950s, the physicist and computer pioneer John von Neumann gave a series of lectures asking how a machine or organism made from unreliable parts could ever compute or behave reliably. His answer has shaped computing ever since. Von Neumann posited that spreading the information across many components and letting them vote to settle disputes would make the whole significantly more dependable than the parts, provided each component failed less often than some threshold rate.

Von Neumann had vacuum tubes and neurons in mind but quantum engineers face a similar problem. Quantum machines are notoriously fragile and noisy and yet still have to process and transmit quantum information reliably. What’s more, quantum information in the form of qubits cannot be inspected without destroying it, which significantly complicates the business of spotting and correcting errors.

That’s why quantum physicists have spent the last 30 years developing quantum error correction techniques that hide logical qubits inside a larger collection of noisy physical qubits. This allows them to repeatedly measure collective constraints, called stabilizer checks, that reveal whether neighbouring qubits agree without revealing what they encode. This allows physicists to correct the inevitable stream of errors that plague quantum machines.

Ingenious maths

Quantum error correction is a significant feat of mathematical ingenuity. Now Ian Whitehouse and colleagues at the University of Maryland in College Park, ask how this process compares to the way biological systems protect and correct information as it passes through complex networks of neurons. The answer, they conclude, is that both systems employ essentially the same strategy.

That’s important because biological error correction works without the strange quantum properties of entanglement and superposition and runs continuously and adaptively. These are things quantum physicists have struggled to achieve.

“Our structural analogy suggests that new insights into brain-inspired algorithms for collective information processing may inform novel quantum error correction approaches,” say Whitehouse and co.

The team approach the two systems of error correction by translating each object in a stabilizer code into a neural counterpart. Physical qubits map onto individual neurons or small microcircuits. Logical qubits map onto the low-dimensional variables that neural populations encode, such as spatial position or working memory.

The Hilbert codespace where quantum information is encoded maps onto a neural manifold, the set of activity patterns representing valid values of those variables. Quantum stabilizer checks become neural circuit-level constraints such as excitation-inhibition balance; the error-containing quantum measurements become mismatch signals and ancilla qubits become interneurons, oscillatory modes and possibly astrocytes.

Crucially the researchers invoke no quantum weirdness in tissue. “Neurons are treated throughout as noisy classical elements, and no quantum coherence or entanglement is invoked in biological tissue,” they say.

Subscribe now

Having mapped the properties of quantum error correction onto their biological equivalents, the researchers test whether they still work using two toy systems to see whether continuous recovery in both systems follows the same mathematical dynamics.

In both cases, the results line up, suggesting some fundamental connection between them. “Quantum error correction and biological error correction in neuronal circuits share a common organizational pattern,” say the team.

However, they are careful to caveat the result. “The correspondence is an analogy of roles, not an identification of mechanisms: neurons are computationally far richer than qubits,” they point out. Even so, the shared vocabulary of action points both ways, offering neuroscientists a measurable notion of code distance and hinting to quantum engineers that the best decoders may be adaptive ones.

Although it has taken some seventy years, Von Neumann would surely be impressed.

Ref: arxiv.org/abs/2607.20534 : Quantum error correction and biological error correction: A structural analogy between qubits and neurons

INSIGHT

This paper’s core contribution is a formal structural mapping that links quantum error correction to how neural circuits keep computation reliable despite noisy neurons. It demonstrates that both systems solve the same problem with mathematically comparable machinery.

Its value is in what quantum physicists and neuroscientists can learn from this. For quantum engineers the key learning is nature’s ability to adapt. Neural error control is local, continuous, and adjusts on the fly as noise statistics drift. The problem is that quantum error correction typically returns a system to its original codespace, whereas continuous damping recovery in neural circuits does not. Instead, nature trades exactness for adaptivity. So the lesson isn’t to copy biology outright but perhaps to explore a similar kind of trade off in noisy quantum hardware.

Neuroscientists have long described neural codes using the qualitative notion of as redundancy. Quantum error correction now offers them quantities for formally defining and measuring the robustness of a neural code: codespace, code distance, decoder and so on. Code distance, for example, determines how many physical unit failures can occur before the encoded variable breaks down. This translates into a measurable robustness scale for neural population codes.

Quantum error correction also relies crucially on additional qubits called ancilla to help correct errors. Something should play a similar role in the brain, but nobody is quite sure what. Potential candidates include interneurons and network oscillations. Whitehouse and co think star-shaped brain cells called astrocytes might also play a crucial role and their goal now is to unravel this mystery.

AROUND one billion years from now the Sun will have brightened enough to boil Earth’s oceans and sterilise its surface. Any civilisation still in existence will need a plan if it is to survive. The instinctive response is escape, to pack up and migrate to another star. Yet the galaxy is vast, worlds as hospitable as Earth seem rare and even if one is discovered, hauling a population across light years may prove impossible.

That leaves one other option: staying put. But exactly how our civilisation might pull it off isn’t clear.

Enter Gabriel Harry, an independent researcher based in London, who has set out how our descendants could survive by re-engineering the Solar System. Harry identifies seven long-term threats to Earth’s habitability and proposes mitigating each with a megaengineering project based on known physics. Deployed together, he says, these safeguards would keep the planet comfortable for 9.1 million billion years, far longer even than the Sun will shine.

Star Treck

The dangers are many. Rising solar luminosity threatens to boil the oceans. When the Sun evolves into a red giant Sun, it will engulf Earth’s orbit vaporising the planet. As Earth’s interior cools, plate tectonics will come to a halt along with the carbon-silicate cycle that regulates the planet’s long-term climate. Water is slowly but constantly leaking into space via variety of mechanisms — eventually we will run out. And, of course, Earth is threatened by uncountable asteroids, comets and other hazards from beyond the Solar System.

On these timescales, the immediate problem is to hold sunlight steady on Earth’s surface. That will first require a sunshade on a massive scale. Such an object placed just beyond the Moon’s orbit would need to be 350,000 kilometres across and significant stabilisation measures to counteract the Sun’s radiation pressure, solar wind and other forces.

Harry points out that smaller shades are already being considered as ways to protect Earth from global warming, so more ambitious structures would be a natural progression.

Preventing or slowing the Sun’s evolution into a red giant will also help. One way to do this is “star lifting”. This uses a swarm of mirrors to focus sunlight back onto the Sun, dramatically heating a small part of the surface and driving material into space. This reduces the Sun’s mass, its core pressure and its rate of fusion, thereby slowing progression to a red giant.

Subscribe now

Harry says this alone extends the time Earth will remain in the Sun’s habitable zone from 1 billion to over 9 billion years.

Another idea is to inject antimatter into the core-mantle boundary to heat the core and keep tectonics running. Yet another is to nudge the Earth away from the Sun using a particle accelerator and to top up our water from Jupiter, Saturn, Uranus and Neptune, which have abundant supplies.

Even after the Sun runs down, these giants hold enough fusion fuel to light Earth for several trillion years more. Crucially, Harry argues these megaprojects are “more feasible than migration by two orders of difficulty”.

Dangers remain, of course. A single failure of the sunshade would sterilise Earth within hours, so redundancy and testing across aeons would be essential. And all this assumes that our civilisation avoids its own self-destruction.

The reward is a home. “Life on Earth would be able to continue as normal, enjoying comfortable security, geology and climate,“ says Harry, forever the optimist. “And for 100 trillion years [we would] be able to look up at the night sky and see the stars.”

Ref: arxiv.org/abs/2607.13084: Retaining Earth’s Habitability Beyond the Life of the Sun

INSIGHT

Harry’s paper reframes a question usually treated as settled — that any civilisation facing its parent star’s death must leave. His approach is to show that staying put is a coherent, and possibly preferable, alternative.

The paper’s central finding is that systematic, megaengineering concepts can together extend Earth’s habitability by several orders of magnitude beyond the Sun’s natural lifespan.

The implications spread across several disciplines. For astrobiology and SETI, it suggests technosignature searches should consider stars whose evolutionary tracks have been deliberately retarded, and it complicates the assumption that advanced civilisations expand outward.

For planetary science, the proposal to sustain mantle heat flux via antimatter annihilation reframes tectonics as an engineerable planetary process rather than a fixed geological clock.

And perhaps most striking is the energy-rate insight. Harry’s safeguards cost more in energy than migration but demand a far lower peak power output, making them arguably the more tractable path.

THE deepest borehole ever drilled, on Russia’s Kola Peninsula, reaches barely 12 kilometres into a planet that is more than 6,300 kilometres deep. Everything geologists know about the Earth below—its dense metallic core and the churning mantle above it-- comes from indirect evidence, chiefly the way earthquake waves bend and bounce as they pass through the planet.

Seismology has served science well for a century. But scientists would dearly love an independent method, based on entirely different physics, that could provide a powerful cross-check. And neutrinos — ghostly particles that stream through matter almost unimpeded — are finally doing exactly that.

Breaking the ice

Now more than 400 researchers in the IceCube Collaboration have used neutrinos to weigh the Earth and map its internal density with unprecedented precision (for this technique). The measurement relies solely on the weak nuclear force, one of nature’s four fundamental interactions, making it entirely independent of gravity and seismology. The result confirms the standard picture of the planet’s layered interior and establishes neutrino tomography as a new tool for studying planetary innards.

The IceCube researchers analysed more than a decade of data from the IceCube Neutrino Observatory, a cubic kilometre of glacial ice at the South Pole packed with sensors that can determine the passage of neutrinos. These neutrinos come from above but crucially also from below having travelled through the entire planet. The team studied how the Earth itself blocks these neutrinos, then worked backwards to deduce the density of the material they crossed.

The trick relies on a quirk of neutrino physics. At modest energies, neutrinos sail through the entire planet untouched. But the probability of a neutrino striking a nucleus grows with energy, and above about 10 teraelectronvolts the Earth becomes partially opaque. “The attenuation of the neutrino flux depends on the traversed column density and the neutrino energy,” say the IceCube team. Neutrinos arriving from directly below IceCube pass through the dense core and suffer the greatest losses, while those skimming the horizon cross only shallow rock. Comparing arrival rates from different directions therefore reveals how density varies with depth.

The neutrinos themselves are generated largely by cosmic rays smashing into the atmosphere on the far side of the planet, supplemented by a rarer flux from astrophysical sources. When these neutrinos interact with the ice, they generate muons, which can be tracked by the Cherenkov radiation they produce as they move through the ice.

Subscribe now

The team selected 368,071 upward-travelling muon tracks recorded between 2011 and 2022, with energies between 500 gigaelectronvolts and 100 teraelectronvolts. They then modelled the Earth as five concentric shells of uniform density and fitted the shell densities to the observed pattern of arrivals, carefully accounting for 37 sources of uncertainty ranging from cosmic-ray physics to the optical quirks of Antarctic ice.

g

The technique will certainly improve over time (indeed physicists first deonstrated this technique in 2019). Next-generation detectors promise larger event samples and smaller uncertainties. “With improved detectors and precision, neutrinos will further contribute to a multifaceted understanding of the Earth’s structure,” say the IceCube researchers. A century after seismologists first weighed the deep Earth, particle physicists have learned how to do it too.

Ref: arxiv.org/abs/2607.02644: High-Energy Neutrino Tomography of the Earth’s Interior with IceCube

INSIGHT

The IceCube collaboration’s measurement tells us little about the Earth that we didn’t already know. Its value is in establishing neutrinos as a practical, data-driven probe of planetary interiors, complementing the gravitational and seismic methods that have monopolised deep-Earth studies for a century.

This cross-validation matters because seismology and gravity infer density indirectly through matter’s elastic and gravitational response. Neutrino attenuation instead senses nucleon column density directly, providing an independent check on assumptions embedded in existing Earth models. Apart from anything else, that introduces “multi-messenger geoscience” as a discipline in its own right.

Consider, for a moment, the humble Sudoku puzzle. Filling in a blank grid from scratch is hard and as the grid increases in size, the number of possible arrangements grows explosively. Yet checking whether a completed grid is valid takes only moments.

This asymmetry between finding a solution and verifying one sits at the heart of one of the deepest unsolved problems in mathematics and computer science: the P versus NP question. It asks, in essence, whether every problem whose solution can be quickly checked can also be quickly solved. Most experts believe the answer is no (although there is no proof, despite the $1 million Millennium prize for the first person to find one).

You could be forgiven for thinking this is a question purely of mathematics and computer science but physicists have grown to think that it has profound connections to the nature of the universe, because every computer is a physical machine. Indeed, many physicists and mathematicians believe that there is growing evidence that this mathematical reality is somehow deeper than the laws of physics; that underneath it all, the properties of information and the way we process it form a deeper bedrock for reality.

Fundamental reality

That’s the context in which we come to one of the most profound questions in physics: whether gravity, like everything else at a fundamental level, is quantum mechanical. Physicists suspect it is but lack the evidence to prove it. Indeed, they often fall back on a simpler picture in which gravity is classical but can couple to quantum matter. This semiclassical framework produces some useful results but has long been suspected of harbouring unusual properties.

Enter Matthew Fox of the University of Colorado in Boulder, Chaitanya Karamchedu of the University of Maryland in College Park and Sotirios Mygdalas of the Perimeter Institute for Theoretical Physics in Waterloo, Canada. Together, these theoreticians have explored the computational properties of the theory of semiclassical gravity for the first time. Their conclusion is that this view of the universe allows computation so wild and powerful, that the theory cannot be correct.

This points them to a profound conclusion. If the theory generates impossible results, the assumption behind it -- that gravity is classical -- cannot be correct. And if gravity cannot be classical, it must therefore be quantum.

This is how they reach this conclusion. Fox and co began be thinking about what happens when quantum matter is influenced by a classical gravitational field via the semiclassical Einstein field equations. In this framework, a massive quantum particle does not simply spread out as an ordinary quantum wavefunction would, it also gravitationally attracts itself. This self-attraction is captured by the Schrödinger–Newton (SN) equation, which modifies ordinary quantum dynamics by introducing a gravitational self-interaction potential. Crucially, this modification makes the dynamics non-linear — and non-linearity, it turns out, is computationally explosive.

The connection between non-linear quantum dynamics and computational hardness was established in work by Daniel Abrams and Seth Lloyd in 1998 and later generalised by Ning Bao, Adam Bouland and Stephen Jordan. Their central insight is that any non-linear quantum evolution necessarily “stretches” the state space in a way that standard, linear quantum mechanics cannot.

So any two quantum states that are exponentially close together can be driven to a macroscopically distinguishable distance apart in only polynomially-many steps. By contrast, in ordinary quantum mechanics, which is completely linear, this would take exponentially-many steps.

This amplification is precisely what computer scientists need to solve NP-complete problems efficiently. That’s because deciding such problems can be reduced to distinguishing between two quantum states that differ by an exponentially small amount.

The implications extend well beyond the Schrödinger–Newton equation specifically. Because the mathematical theorem underpinning the argument applies to any non-linear quantum dynamics, the researchers argue that essentially any semiclassical theory in which quantum matter couples non-linearly to a classical gravitational field would produce the same result.

This would violate a foundational principle of theoretical computer science known as the Physical Extended Church–Turing Thesis (PECTT), which holds that no physical process can efficiently solve NP-complete problems.

“Any consistent, semiclassical, and non-linear matter-gravity coupling will entail an efficient algorithm to solve NP-complete problems and thus violate the PECTT,” says Fox and co. The only clean way out is to quantise gravity, which restores linearity to the dynamics and removes the self-interaction potential altogether. “Quantizing gravity restores linearity and thus avoids our argument altogether,” they conclude.

The work does not rule out every alternative. One could, in principle, abandon the PECTT itself, or seek a semiclassical theory whose density-matrix evolution happens to be linear even if the underlying wavefunction dynamics are not — a path explored in Oppenheim’s post-quantum theory.

But each such escape carries a heavy conceptual price. The cleaner and more natural resolution, Fox and colleagues argue, is simply that gravity is quantum. The result therefore constitutes a powerful new theoretical argument for the quantisation of gravity. It joins a long tradition of indirect theoretical arguments for quantum gravity with an unusual twist. It draws its force not from thought experiments about information or thermodynamics, but from the computational structure of the physical world.

Ref: arxiv.org/abs/2606.14806: Semiclassical Gravity Efficiently Solves NP-Complete Problems

INSIGHT

This paper sits at a remarkable intersection of gravitational physics and computational complexity theory, and its central finding carries deep implications for how we understand the fundamental structure of physical reality.

The authors demonstrate that if gravity remains classical — governed by the semiclassical Einstein field equations — then the resulting non-linear dynamics would, in principle, allow NP-complete problems to be solved in polynomial time. This would shatter the Physical Extended Church-Turing Thesis (PECTT), a foundational principle asserting that no physical process can efficiently solve such problems. The significance is profound: computational limits are not merely mathematical abstractions but appear to be encoded into the fabric of physics itself.

What makes this finding so striking is its inversion of the usual direction of inquiry. Rather than asking what physics can tell us about computation, the authors ask what computational constraints can tell us about physics — and the answer is striking. The apparent universality of the PECTT acts as a kind of consistency test that candidate physical theories must pass. Semiclassical gravity fails this test.

The broader implication is that computational complexity theory may function as a genuine selection principle for fundamental theories, sitting alongside established tools like symmetry and renormalisability. But is this contradiction between a physical theory and computational theory sufficient grounds for its rejection?

If it is — and Fox and co clearly lean that way — computability and physical law are not merely analogous but constrained at some deeper level that we have not yet grasped. And therein lies an inherently more exciting and profound problem.

LAST month, a Royal Air Force transport aircraft carrying UK Defence Secretary John Healey experienced severe GPS jamming as it flew near the Russian border. The three-hour attack forced the crew to rely on alternative instruments and ground-based radar to fly home. The incident mirrored an identical attack on his predecessor in 2024 near Kaliningrad and one targeting a plane carrying European Commission president Ursula von der Leyen in southern Bulgarian airspace in 2025.

These and numerous other incidents are examples of a rapidly escalating pattern of state-sponsored electronic warfare targeted directly at European transport corridors. But over the past five years pilots across Europe have logged a surge of mysterious outages from another source.

On scores of occasions, every GPS receiver across an area spanning Europe, Greenland and Canada has suffered a sudden simultaneous drop in signal strength at precisely the same instant. No earthbound transmitter could reach so far, making it hard to imagine what was the cause.

Now physicists have tracked it down. Zachary Clements at the University of Texas at Austin and colleagues say the source is a constellation of Russian early warning satellites called Edinaya Kosmicheskaya Sistema (EKS) that follow steeply elliptical Molniya orbits. The discovery marks a rapid escalation in the known capabilities of Russian military satellites and on the regime’s willingness to target western infrastructure.

Orbital attack

GPS signals, and those from other satellite navigation systems, are astonishingly faint when they arrive at Earth, roughly comparable to a dim light bulb glimpsed from thousands of kilometres away. That fragility makes them easy to disrupt.

Most of this trouble comes from the ground: military transmitters near conflict zones, jammers smuggled into lorries and similar terrestrial sources. But since 2019 a stranger pattern has surfaced in the data.

The interference is powerful, fleeting and confined to the 1577.5 MHz and 1558.5 MHZ frequencies used by GPS satellites, the European Galileo system and China’s BeiDou spacecraft.

The scale of the disruption strongly suggests a source in orbit but identifying the responsible object among thousands of candidate satellites is a genuinely hard detective problem.

Clements and co solved the conundrum by mining seven years of public data from the International GNSS Service, a worldwide network of ground stations, then fused several geolocation methods to identify the attacker.

Because a space-based source floods an enormous footprint at once, a synchronous dip across far-flung stations is a tell-tale fingerprint. But changes in signal strength alone cannot identify a single satellite. For that the researchers compared when the interference signals reached widely separated receivers and then triangulated the emitter.

The team took data from 165 reference stations and applied it across every measurement epoch from 2019 to 2026. They flagged 75 days carrying at least one strong wide-area event, with the largest drop recorded at a station in Poland.

Subscribe now

The timing alone proved revealing: the bursts “predominantly occurred during business days and business hours”, they say, a distribution that points to human hands rather than nature.

To clinch the identification the researchers analysed raw samples captured simultaneously in Amsterdam and Trondheim during an event in February 2026, computing time-differences accurate to a few nanoseconds and matching them against published orbits.

Only one object fitted the measurements: Cosmos 2546, a Russian satellite riding a steeply elliptical Molniya orbit, and a key member of the EKS constellation. These were designed to spot ballistic missile launches from above the northern hemisphere but clearly have a secondary capability.

Cosmos 2546 was launched in 2020 so its sister satellites must have been responsible for earlier outages.

Clements and colleagues conclude it is highly probable that the EKS constellation is collectively responsible for the disruptions seen since 2019. “If deliberate, it portends a qualitative escalation in GNSS interference,” they add.

Ref: arxiv.org/abs/2606.03673: Chasing Lightning: Detecting, Characterizing, and Identifying a Powerful Space-Based GNSS Interference Source

INSIGHT

This paper makes the first confident attribution of recurring, continent-wide satellite navigation interference across Europe, Greenland, and Canada. The culprit: Russia’s EKS (Edinaya Kosmicheskaya Sistema) early-warning satellites in Molniya orbits.

For aviation, maritime navigation, and timing-dependent critical infrastructure, the work establishes a qualitatively new threat model: a single high-altitude satellite can simultaneously deny GPS, Galileo, and BeiDou across a continent in a way that no terrestrial jammer could replicate. Unlike ground-based jammers, which can be physically located and suppressed, a space-based emitter operating from a highly elliptical orbit presents a fundamentally different defensive problem.

Should the technology be scaled or weaponised, the implications are clear. A list of critical utilities that rely on this technology is sobering. It includes: emergency services, industrial supply chains, the banking industry, agricultural supply chains, financial markets to name just a few.

Interestingly, the targeted frequencies of 1577.5 MHz and 1558.5 MHZ are precisely those used by US, European and Chinese satellite navigation systems. By contrast, Russia’s GLONASS system operates at 1602 MHz and so would be entirely unaffected by this attack. A remarkable coincidence worth reflecting on.

DARK matter is one of the great mysteries at the heart of cosmology. It is more common than ordinary matter by a ratio of roughly five to one, it holds galaxies together and shapes the entire cosmic web. Yet nobody knows what it is made of.

One of the most beguiling answers is also one of the strangest: that dark matter is the remnants of black holes born in the first fraction of a second after the Big Bang. These bodies would be so tiny, some no heavier than the Moon, that they are almost impossible to spot.

Now Renee Key at the University of Technology in Melbourne and colleagues, think they may have caught one about three times the mass of our moon drifting in front of a distant star. The sighting joins a small but growing tally of similar detections that could finally determine the true nature of dark.

Born early

Primordial black holes differ sharply from the familiar kind. Ordinary black holes form when massive stars collapse and so cannot weigh much less than their parent stars.

Primordial black holes form in an entirely different way. In the first instants after the Big Bang, regions slightly denser than their surroundings would have collapsed into black holes under their own gravity. The resulting black holes would have an extraordinary range of masses, from asteroid-sized up to many thousands of Suns.

In the 14 billion years since then, all but the largest would have evaporated away. But rest are still out there. Many cosmologists believe this vast population of now tiny black holes could be the missing Dark Matter that holds galaxies together.

Key and co found their candidate via a phenomenon called microlensing. They pointed their so-called Dark Energy Camera on Chile’s Blanco 4-metre telescope at a single patch of sky near the Large Magellanic Cloud, snapping an image every 50 seconds for five consecutive nights.

Microlensing occurs when a compact object drifts in front of a background star and its gravity briefly magnifies the starlight. Out of several million light curves, one star brightened smoothly for about an hour and then faded, an event the team call Pheobe. After ruling out atmospheric noise, neighbouring stars, ordinary variable stars and stellar flares, the team say: “We interpret Phoebe as a primordial black hole PBH with mass approximately 3 lunar masses.”

Subscribe now

They remain cautious about Phoebe’s true nature, however. Microlensing reveals a lens’s mass and motion but never its identity so a free-floating planet of the same mass would produce an identical flash, although they are highly unlikely to be floating in that region of space.

Phoebe does not stand alone, and the strength of the claim lies in a wider pattern. In 2019 the Subaru telescope’s survey of the Andromeda galaxy flagged a single short microlensing event consistent with a lunar-mass black hole. The same survey has since reported around a dozen further candidates at strikingly similar masses.

“Phoebe suggests a population of compact, lunar-mass objects associated with the dark matter distribution of the Milky Way, and potentially opens a new window to the physics of inflation,” say the team.

Settling the question will demand many more detections across many sightlines. And astronomers may soon get them thanks to the imminent Roman and Vera C. Rubin observatories. Should the candidates hold up, they would not merely solve the dark matter puzzle but glimpse physics from the Universe’s earliest moments.

Ref: arxiv.org/abs/2605.19375: AMPM II. — A Lunar-Mass Primordial Black Hole Microlensing Candidate in the Milky Way Halo

INSIGHT

For decades, the leading dark matter candidates have been hypothetical particles, such as WIMPs (weakly interacting massive particles) and axions. Yet direct-detection experiments and particle collider searches have consistently found no evidence of them.

Against this backdrop, primordial black holes are becoming a compelling alternative. They require no new particle physics, invoking only gravity and the conditions of the early universe.

This paper’s report of “Phoebe” is part of a small but growing body of evidence that points to a population of compact, lunar-mass objects in dark-matter-dominated sightlines. This mass within the asteroid-to-lunar range that other constraints leave relatively open.

For astronomers, this paper reframes the search for dark matter detection as an astronomical enterprise, rather than a pure particle-physics play. And it places even more attention and expectation on data from the next generation of, which could build the population statistics needed for confirmation.

The Vera C Rubin observatory is already gathering data from a mountain top in Chile. The Nancy Grace Roman Space Telescope is due for launch late in 2026.

For cosmology, the implications are profound: any primordial black hole would predate Big Bang nucleosynthesis some 3 minutes after the moment of creation. That would make it among the oldest observable objects in the cosmos. Cosmologists are expecting!

SPOTTING a small drone hovering in front of a windswept treeline is a challenge that defeats even sophisticated surveillance systems. The visual contrast between a small autonomous aerial vehicle and a complex background of rustling foliage can be vanishingly low. Birds, insects and branches all conspire to fool automatic target-recognition pipelines, and in degraded conditions such as fog or poor lighting, conventional approaches degrade further still.

That’s a problem because of rapidly increasing importance of drone detection in situations ranging from military conflicts and critical infrastructure protection to border control and urban airspace management. What governments, the military and security organisations would dearly love is a way to pick out these machines from the natural clutter they are often framed against.

Spectral signature

Enter Megan Birch and colleagues at the Georgia Tech Research Institute in Atlanta who have developed a novel way to pick out a drone or helicopter that would otherwise be hidden against a complex backdrop.

The key insight from Birch and co is that drones produce a crisp periodic signature that can be clearly distinguished from natural scenes that change slowly and continuously over time. These signatures come from mechanical vibrations and, in particular, from rotating blades.

Their idea is that sampling each pixel in a scene at a rapid rate should quickly reveal those mechanical rhythms as a kind of flickering signature. This would be as distinctive as a heartbeat against the relatively, slow changes in the background of trees, birds and other clutter.

It turns out that the equipment capable of this has been developed over the last couple of decades. So-called neuromorphic vision systems are an advanced class of sensing technologies modelled after the biological architecture of the human eye and brain. Unlike traditional frame-based cameras that capture a rapid succession of static, full-image snapshots regardless of scene changes, neuromorphic systems use event-based cameras featuring autonomous pixels that operate independently of each other to report an “event” only when it detects a change in the scene. This filters out static backgrounds while picking out motion.

Subscribe now

Birch and co have developed this idea to look for the specific signature of drone rotors spinning at rates of between 50 and 500 Hz. In this way, their framework called FRIES (Frequency Rate Information for Event-Space) can detect and isolate artificial objects, in particular drones, against cluttered natural backgrounds.

The team tested FRIES using a Prophesee EVK4 event-based camera in two experiments. In the first, a mechanical chopper operating at a known rate and a multi-rotor drone with a dynamic tree behind them were imaged indoors. In the second, a drone hovered against a real treeline in uncontrolled outdoor conditions.

Indoors, the system clearly picked out the fliers. “The key innovation of FRIES is the introduction of a frequency-driven mechanism as a primitive for detection versus reliance on purely spatial or temporal trends,” say Birch and co.

The outdoor results were mixed, however, with the team concluding that the algorithm requires further tuning before deployment. “Event based cameras may prove valuable as a complementary system to conventional frame-based cameras rather than an alternative,” they say, suggesting that the highest performance will come from fusing both sensing modalities.

The team now plan further upgrades such as using parallel filters capable of tracking multiple targets simultaneously.

Ref: arxiv.org/abs/2605.15392: “Frequency-domain Event-based Imaging for Selective Surveillance”

INSIGHT

This paper introduces a fundamentally new approach to detecting mechanical objects — particularly drones — by exploiting their rotational frequency signatures rather than relying on visual contrast or shape. The implications extend well beyond the laboratory.

For airspace security, the ability to isolate a hovering drone against a cluttered treeline addresses one of the most pressing challenges for airports. As drone incursions near airports, prisons and critical infrastructure become increasingly common, these sensors will represent a meaningful operational advance.

For the military, the work paves the way for toward the low-power detection, and potentially identification, of rotary-wing threats without emitting detectable radar signals.

But adversaries are unlikely to give up. The deeper insight is that periodicity itself is the vulnerability — meaning truly stealthy future drones may sacrifice efficiency for deliberate mechanical irregularity, accepting worse flight performance in exchange for spectral anonymity.

The most straightforward countermeasure would be randomising rotor speeds to smear the frequency signature across a broad band, mimicking the incoherent, low-frequency noise of natural background motion like wind-driven foliage. At the software level, onboard flight controllers could actively modulate rotor speeds in patterns that defeat spectral clustering algorithms.

This mirrors the trajectory of radar-absorbing materials in conventional stealth: once the detection physics are understood, evasion becomes an engineering discipline in its own right. Expect a coming battle for spectral camouflage.

HIGH-ENERGY physicists have a problem. The Big Bang should have produced equal quantities of matter and antimatter. Yet the observable universe is composed almost entirely of matter. For matter to have prevailed, the laws of nature must treat the two differently — a phenomenon known as charge-parity (CP) violation. But nobody knows the source of this violation.

The problem for high-energy physicists is that they lack the machines to hunt for undiscovered sources of CP violation. The $5 billion Large Hadron Collider, the most powerful particle accelerator ever built, is not up to the job. Neither does its proposed successor, the Future Circular Collider, have much prospect of cracking the CP violation problem, despite a price tag of $20 billion.

What physicists would dearly love is a more promising way to hunt for this new physics.

Now Ali Jadbabaie at the Massachusetts Institute of Technology and colleagues have mapped out just such a technique. Their proposal is not to smash particles together to see what’s inside, but to explore the interior of unusual radioactive molecules to look for the shadows that exotic physics might cast—but at much higher energy than is accessible to particle accelerators.

Their proposed experiments will cost a fraction of the price of a new particle accelerator and many can be done with equipment that could fit on a tabletop. “Radioactive molecules offer a promising platform to probe new physics beyond the Standard Model,” they say.

Charge displacement

Theoretical physicists think that one of the most powerful effects of CP violation within a fundamental particle is a tiny displacement of charge along its direction of spin, known as an electric dipole moment or EDM. Detecting an EDM would constitute unambiguous evidence of new physics.

The signals are extraordinarily small, however. Physicists’ best theory, the Standard Model of particle physics, predicts that the evidence lies many orders of magnitude below current experimental sensitivity, which is why nobody has ever measured an EDM.

But Jadbabaie and co say there is another way using radioactive molecules, molecules in which at least one of the constituent atoms has an unstable nucleus, such as an isotope of francium, radium, thorium and protactinium. The nuclei in these atoms adopt asymmetric, pear-like shapes containing pairs of opposite-parity quantum states, known as octupole-deformed nuclei.

These amplify 1000-fold small changes in the nuclei’s distribution of internal charge, called the Schiff moment, which connects CP violation inside the nucleus to measurable energy shifts.

Embedding such a nucleus in a polar molecule further amplifies the signal, because the intense electric fields within polar molecules couple strongly to nuclear fields.

Subscribe now

That brings a measurement within reach. Experiments with radioactive molecules promise sensitivity to physics above 1000 TeV — far beyond the reach of the Large Hadron Collider at 13.6 TeV and even the Future Circular Collider at 100 TeV. They may also place new constraints on another great puzzle of quantum chromodynamics called the strong CP problem,

There are significant challenges along the way, of course. Radioactive molecules are produced in trace quantities, often in particle accelerators themselves, and with a radioactive lifetime of just days. They then have to be cooled, confined, combined with other atoms to form radioactive molecules and then concentrated so that their spectroscopic behaviour can be measured.

“Advanced measurements with radioactive molecules could potentially achieve constraints beyond the 1000 TeV scale,” say Jadbabaie and co.

With multiple experiments expected to come online within the decade, the stakes are high: “Any nonzero EDM measurement in the near future would be evidence of new physics,” say the researchers. “While continued null results further tighten bounds on hypothetical new particles and forces.”

Ref: arxiv.org/abs/2605.12767: Radioactive Molecules as Laboratories of Fundamental Physics

INSIGHT

This paper reframes a longstanding assumption in particle physics: that discovering new physics requires ever-larger accelerators. Instead, precision low-energy measurements offer a complementary and more cost-effective discovery pathway, exploiting quantum control, laser cooling and molecular spectroscopy rather than brute-force energy escalation.

More broadly, the paper signals a structural shift in experimental particle physics towards highly interdisciplinary collaboration spanning nuclear physics, quantum chemistry, atomic physics and quantum information science. That means the potential for discovery increasingly resides not at the energy frontier, but in exquisitely controlled molecular systems.

For CERN’s proposed Future Circular Collider, this is uncomfortable timing. If tabletop experiments can access comparable or greater energy scales for a tiny fraction of the cost, it sharpens an already difficult question about whether another collider represents the best use of European physics funding.

A more interesting opportunity may be in developing ISOLDE, CERN’s in-house radioactive ion beam facility and related infrastructure. That would allow CERN to position itself as a leader in frontier physics regardless of the Future Circular Collider’s fate.

Either way, future funding and institutional priorities will need significant reorientation.

A RHESUS macaque weighs around eight kilograms and can expect to live for 25 to 40 years. But a domestic cat of the same mass rarely survives beyond 18. Chimpanzees routinely reach their mid-forties or early fifties, while similarly sized dogs seldom exceed 25 years.

Scale up to humans and the discrepancy becomes even starker: a 70-kilogram mammal like a wild boar, or mountain lion live only around 20 years, yet people routinely reach 80 or beyond.

This the longevity excess is something of a puzzle. Why do primates live so much longer than other similarly sized animals?

One idea is that tree-dwelling primates have fewer predators and a complex social structure that helps protect them from danger. But this hazard reduction probably accounts for less than half the observed lifetime surplus.

Another possibility is that primates have larger brains than similarly sized animals and that this somehow accounts for the difference. Indeed, biologists have long noted that brain size correlates with longevity, yet no one has been able to explain why.

Brain boost

Enter Mesfin Asfaw Taye, a physicist at West Los Angeles College in California, who says thermodynamics is the answer. Taye has developed a thermodynamic model that treats the brain as an instrument that reduces the thermodynamic cost of everyday life, for example, by anticipating and better controlling risks like nutritional shortfalls, predators and even body functions such as blood pressure and immune function. In other animals, the extra wear and tear from these factors significantly shortens lifespans.

What’s more, his model accurately predicts the lifespans of 18 primate species and makes a number of testable predictions that biologists should be able to use to road test his theory.

First some background. Biologists have long known that the lifespan of all warm-blooded animals is essentially the same when measured in heartbeats. A common shrew weighing around 8 grammes has a resting heartbeat of 800 bpm and lives roughly 18 months, while an elephant weighing 4 tonnes with a heart rate of 28 bpm lives around 65 years. Throughout their lives, however, both experience around a billion heartbeats and this figure turns out to be roughly constant for all species.

The conventional way of explaining this is that lifespan scales with body mass but the “rate of living” measured in cardiac cycles, remains constant.

But primates are an exception because they live significantly longer than other animals of a similar mass.

Now Taye says he thinks he knows why. His idea, which he calls the Principle of Biological Time Equivalence, is that each heartbeat has a thermodynamic budget associated with the amount of molecular disorder or randomness it produces. Physicists call this entropy and the energy it uses is unavailable for other useful work within the body.

Taye’s insight is that the brain frees up this energy which the body can then use to repair damage, restore molecular order and delay aging. Or as Taye puts it: “a metabolically large and informationally rich brain reduces the entropy produced per cardiac cycle in somatic tissues, thereby expanding the organism’s effective lifetime cycle budget above the mammalian baseline.”

The mechanisms are straightforward. Their large brains allow primates to anticipate and mitigate risks. Potential food shortages can be planned for, predators outthought and food poisoning avoided. Primates can more easily avoid other stressors, such as extreme heat by seeking shade. And large brains also help to regulate cellular repair machinery which reduces long term damage. And so on.

All these events would otherwise generate additional entropy that reduces lifespan. “The brain, in this view, is a longevity organ as much as a cognitive organ — not merely by slowing heart rate but by reducing the thermodynamic cost of each heartbeat,” he says.

Taye calibrated his using various publicly available databases of animal lifepsans, including primates. His model predicts the rhesus macaque lifespan at 26.0 years, the chimpanzee at 52.4 years and the human at 70.7 years — all within observed ranges.

Subscribe now

A further consequence of the model is a thermodynamic ceiling on longevity. As brains get larger, their thermodynamic benefit becomes smaller and each successive increment of neural investment yields diminishing returns. In other words, the metabolic cost of maintaining an even larger brain eventually outweighs the entropy-saving benefits it provides.

His model predicts a theoretical maximum longevity of roughly 5.6 times the non-primate baseline. For humans the factor is approximately 2.5, leaving plenty of room for improvement.

A curious corollary is that Taye says his mechanism is entirely distinguishable from caloric restriction, which is known to increase lifespan in certain animals. He points out that caloric restriction slows the biological clock by reducing heart rate. By comparison, neural investment makes each cardiac cycle more entropy-efficient without slowing the clock, so organisms with larger brains age more slowly per heartbeat as well as per year.

Impressively, Taye’s model makes a set of predictions that allow the theory to be tested. For example, it predicts that primates should accumulate fewer epigenetic changes per heartbeat than non-primates, which can be measured using DNA methylation clocks. It also predicts that neurodegenerative diseases should accelerate biological aging.

That’s low hanging evidence that biologists somewhere should be eagerly digging out to check.

Ref: arxiv.org/abs/2604.27937: Neural Investment as an Entropy-Budget Strategy: A Thermodynamic Derivation of Primate Longevity from the Principle of Biological Time Equivalence

INSIGHT

This work delivers the first rigorous thermodynamic derivation of why primates — including humans — live two to three times longer than other mammals of equivalent size, an anomaly that has resisted explanation for decades. Its central insight reframes the brain as a longevity organ as much as a cognitive one: by reducing the entropy produced per cardiac cycle. Neural investment, it says, effectively expands an organism’s lifetime “cycle budget” rather than slowing its biological clock.

The implications cut across several disciplines. For gerontology and biomedicine, the predictions clearly distinguish neural longevity from the impacts of caloric restriction. That will help clarify research agendas for those studying aging.

For evolutionary biology, the theory places a hard ceiling on attainable longevity which constrains the coevolution of brain size and lifespan. A persistent puzzle in human evolution is why brain size plateaued—this could explain why.

For neurology, it offers a quantitative framework linking neurodegenerative disease to accelerated biological aging, which should also help to shape research agendas in the next few years.

Importantly, this paper’s value is explanatory. Those wanting advice about longer, healthier living should look elsewhere, particularly at the standard longevity guidance related to sleep, exercise, not smoking, social connection and managing chronic disease. This rests on firmer empirical ground.

Every stride a runner takes sends a shockwave of force rippling up through the skeleton. In the tibia — the main load-bearing bone of the lower leg — that force can reach fourteen times bodyweight during a sprint.

Athletes, clinicians and researchers have long wanted to track that load in real time, to catch the early warning signs of stress fractures, to optimise training programmes and to understand the precise mechanics of human gait. But actually measuring the force inside a living bone, while that person is out on the road rather than wired up in a laboratory, is difficult.

The best available wearables — accelerometers strapped to the shin, pressure insoles in the shoe — capture only rough proxies for bone loading, and decades of research have shown these proxies can be unreliable.

What researchers would dearly love is a non-invasive way to make real world measurements much more accurate.

Compressed beam

Now Ali Yawar and colleagues at Harvard University in Cambridge, have found an elegant solution hidden in an old branch of mechanics: the way sound propagates through a compressed beam changes in a predictable, measurable way as the compression force varies. The human tibia, it turns out, behaves much like such a beam.

They demonstrate that the tibia behaves like a compressed beam whose vibrational properties change depending on the weight it supports. By mounting a small mechanical “tapper” and an accelerometer to the skin, the team can measure how “flexural waves” travel through the bone. As the compressive load on the tibia increases, the frequency spectrum of these waves shifts in a predictable, linear fashion. “Tibial compressive loading, like loading in other bones, impacts maintenance, growth, and injury, and is therefore a key variable to measure in experimental human biomechanics,” they say.

The approach is rooted in Euler-Bernoulli beam theory, which describes how slender, elastic beams respond to bending loads. When a compressive force is applied along the beam’s axis, it selectively stiffens the beam against low-frequency bending more than against high-frequency bending. This shifts the spectrum of flexural waves travelling along the beam towards higher frequencies as the compressive load increases, and it does so linearly for the physiological range of forces seen in the human tibia.

The Harvard team built a wearable system to test this effect. A small bone-conduction audio transducer, strapped to the leg near the ankle, taps the tibial shaft fifty times per second, each tap launching a brief flexural wave along the bone.

A piezoelectric accelerometer mounted 16 centimetres up the shin then measures each wave as it travels up the bone, revealing how the frequency of the spectral peak shifts as the tibial load waxes and wanes. By tracking that peak in real time, the system tracks the compressive force.

The team tested the device on five men and four women as they performed two tasks on an instrumented treadmill: slow lateral swaying, which loads and unloads the tibia rhythmically at about 0.2 Hz, and walking at 0.4 metres per second.

They then compared the results with calculations from computer model of human motion fed with optical motion capture and force-plate data.

The agreement was striking. “Data from nine participants demonstrate linear relationships between tibial compressive force and spectral peak location,” say Yawar and co, exactly in agreement with beam theory.

Subscribe now

The authors are candid about current limitations. For more widespread use, the device will need to be shrunk and optimised. They will also need to develop a simple way of calibrating the device at the beginning of each measurement session. And the sensor cannot be used on body sites where thick soft tissue attenuates the signal, which might limit its use for some individuals.

Nevertheless, the study establishes proof of concept for a fundamentally new way of sensing. “This flexural wave-based technique could give rise to a new class of wearable sensors for non-invasive physiological bone load monitoring and measurement, impacting research in human locomotion and sports medicine,” say Yawar and co. “The non-invasive and portable nature of our technique has implications for the development of new wearable sensors for in-field measurements and continuous monitoring of bone loading in athletic, clinical, or industrial environments.”

Ref: arxiv.org/abs/2511.06140: Non-invasive load measurement in the human tibia via spectral analysis of flexural waves

INSIGHT

This paper introduces a fundamentally new approach to measuring internal bone loading, with implications that extend well beyond tibial biomechanics. Its new insight — that bones can effectively be “listened to” for real-time load information — opens a new measurement paradigm applicable in principle to any long bone in the body.

This makes wearable bone-load monitoring possible in almost any setting. Athletes could receive instantaneous alerts when tibial loads exceed safety thresholds, potentially preventing stress fractures before they occur. Industrial workers could be warned before repetitive strain injuries take hold.

The approach also has the potential to be adapted to other long bones, such as the femur or humerus, creating a comprehensive map of skeletal stress during complex movement.

Beyond load, the frequency response of flexural waves is governed by Young’s modulus and bone geometry and other researchers have exploited this to measure bone composition. The new tapping technique hints at the possibility of using wearable sensors to monitor changes in bone mineral density or healing progress in fractures over time.

And in better understanding body loads, the technique could help improve next generation of intuitive prosthetic limbs and industrial exoskeletons.

On 4 October 1957, the Soviet Union launched Sputnik and humanity entered the space age. Yet scattered across photographic plates taken during the preceding decade at California’s Palomar Observatory, something curious had already been appearing in the night sky — brief, star-like flashes that materialised and then vanished, leaving no trace in any subsequent survey. They matched no known satellite or orbital debris, because none yet existed. They corresponded to no catalogued star. They simply appeared and then were gone.

These enigmatic objects, called transients, are the result of years of archival research into a catalogue of photographic plates known as the Palomar Observatory Sky Survey (POSS-I) and were taken between 1949 and 1957.

Two puzzling patterns have emerged. First, the number of transients drops by roughly 30 per cent when the relevant sky region lies within Earth’s shadow — a “shadow deficit” consistent with highly reflective objects in orbit catching sunlight. Second, transient detections are significantly more frequent within one day of above-ground nuclear weapons tests.

Genuine phenomena?

Both findings have attracted fierce scepticism, because the automated pipelines used to identify transients also pick out dust, scratches and emulsion defects that inevitably appear on photographic plates. What astronomers would dearly love to know is whether the transients are genuine signals worth investigating in more detail or artifacts that never appeared in the sky but only on ageing photographic plates.

Now they get an answer of sorts thanks to the work of Stephen Bruehl at Vanderbilt University Medical Center and colleagues who have trained a machine-learning model to distinguish real transients from photographic plate defects. And their approach has dramatically strengthened the case that these mysterious objects are genuine phenomena, not artefacts.

The new results show that both the shadow deficit and the nuclear-test correlation grow stronger, not weaker, as the model’s confidence in a transient’s authenticity increases — a result that is difficult to explain if the signals arose from spurious plate noise.

Subscribe now

The team trained a supervised machine-learning classifier on 250 manually inspected image pairs, each taken 30 minutes apart on POSS-I plates, with an expert astronomer labelling 134 as likely genuine transients and 116 as plate defects.

The model was then deployed across all 107,875 previously identified transient candidates, assigning each a probability of being real. Candidates were sorted into ten probability deciles. The results show that only the top 10 per cent of approached or exceeded a probability of 0.80 — suggesting a high false-positive rate among the whole group.

The team then focused only on this high probability group and found both key physical signals intensified. The shadow deficit reached 55.2 per cent in the highest-probability decile, significantly exceeding the deficit seen across all other deciles combined. “The magnitude of this shadow deficit was largest in the subgroup in which it would be expected according to our hypotheses, that is, the subgroup with the highest probability of reflecting real objects,” say Bruehl and co

The nuclear-test association showed a parallel pattern: the highest-probability decile showed a 62.7 per cent higher rate of transients falling within a nuclear testing window than the lowest-probability decile. And the association peaked on the day of a test and the preceding day. “The temporal specificity of the nuclear-transient association and the fact that this association is statistically significant even when plate artefacts are controlled both confirm and strengthen our prior findings,” say Bruehl and colleagues.

The nature of the transients is unknown. The authors suggest pre-Sputnik orbital launches unknown to the public, or “detection of a non-human technosignature”. But they acknowledge this explanation as clearly provocative.

Whether or not these explanations gain traction, the machine-learning approach demonstrates that machine learning can help extract anomalous signals from century-old photographic archives. “Results strongly support existence of an unrecognized population of transient objects in historical astronomical plates warranting further study,” conclude Bruehl and co.

Nevertheless, many commentators will quote Carl Sagan’s well known adage that extraordinary claims require extraordinary evidence.

Ref: arxiv.org/abs/2604.18799: Machine Learning Supports Existence of Previously Unrecognized Transient Astronomical Phenomena in Historical Observatory Images

INSIGHT

This paper is part of an ongoing debate on the nature of transients found in photographic plates taken at the Palomar Observatory in the 1950s. And it the delivers the strongest evidence yet that a genuine population of unexplained phenomena appears to exist in this dataset.

An important question is whether the result can be reproduced and it turns out that several similar datasets of astronomical images exist elsewhere. In March, a preliminary analysis of plates taken at the Hamburg Observatory in Germany between 1954 and 1957 found transients similar to those found by Bruehl and co.

These plates are part of a broader dataset called APPLAUSE which combines 27 plate collections from observatories across Germany, Estonia, and the Vatican, collectively containing over two billion calibrated measurements spanning from 1893 onwards. This has yet to be analysed for transients.

An even larger untapped resource is the Digital Access to a Sky Century at Harvard (DASCH) which has digitised some 550,00 photographic plates taken over the last 100 years. Its continental separation from Palomar, combined with its independent digitisation pipeline, makes it an ideal replication dataset.

If the shadow deficit and nuclear testing association survive in DASCH’s pre-Sputnik plates, the case for a genuinely unexplained physical phenomenon becomes very difficult to dismiss.

In the meantime, one step worth pursuing is a re-analysis of the original photographic plates using a microscope, to get a clearer sense of what spurious results look like in real life. Many an astronomer will quietly admit to being fooled by digital copies of physical plates. Such an analysis could have a significant impact on the output of both human and machine classifiers.

When the payments firm Block cut nearly half its 10,000-person workforce in February 2026, chief executive Jack Dorsey pointed to artificial intelligence as the reason many of those roles had become unnecessary. Within twelve months, he said, most companies would arrive at the same view.

The forecast matched a pattern already unfolding across the tech sector. Tech industry layoffs surpassed 100,000 during 2025 with AI blamed in more than half of them. Salesforce replaced 4000 customer-support roles with AI agents. The coding agent Devin created by AI firm Cognition, lets one senior engineer match what used to take a team of five and has already been deployed at Goldman Sachs and Infosys.

The paradox is that every chief executives knows where this road ends. Laid-off workers are also customers, and if automation destroys wages faster than the economy can create them, companies will eventually engineer themselves towards limitless output but with no one left to buy. The question CEOs, policy makers and AI firms would dearly love to answer is how to avoid this trap.

The AI trap

Now we get an answer of sorts thanks to the work of Brett Hemenway Falk at the University of Pennsylvania and Gerry Tsoukalas at Boston University, in which they model the way competing firms decide how much of their workforce to replace with AI.

Their central finding is that competition forces firms into excessive automation and this pushes them towards a paradoxical collapse in demand. “”Firms automate their way to boundless productivity and zero demand,” say Hemenway and Tsoukalas.

But the researchers also say there is a way out – a single policy instrument out of six they examine to close the gap.

The mechanism Hemenway and Tsoukalas model is subtle but sharp. When a firm replaces a worker with AI, it pockets the full cost saving. But the lost wages reduce spending across the entire economy, so the firm absorbs a fraction of that loss in demand while rivals bear the rest.

“If the cliff ahead is visible to all, why would they race toward it?” ask the researchers. The answer, they say, is that no firm benefits from holding back on automation. In the end, the game collapses into a Prisoner’s Dilemma where every firm fully automates, harming workers and companies alike.

The results produce some counterintuitive conclusions. More capable AI worsens rather than softens the trap, as all firms chase the same market-share gains even though these cancel out in the end. This forces firms to move faster just to maintain their market share. This is the so-called Red Queen effect from Lews Carroll’s Through the Looking Glass, where “it takes all the running you can do, to keep in the same place.”

Hemenway and Tsoukalas study various ways to prevent this paradoxical behaviour. One option is to maintain wages for those still in employment, which would otherwise fall because of the extra competition for fewer jobs, thereby reducing demand. But the model suggests this merely delays the fall in demand rather than preventing it.

Another often quoted option is universal basic income but this doesn’t prevent a catastrophic collapse in demand either because workers get the income whether or not a company automates. So the incentive to over-automate is untouched. This runaway automation is what ultimately destroys demand.

Similar problems arise for other proposed solutions such as giving workers equity or shares and capital-income taxes.

Subscribe now

But there is one approach that does work- a so-called Pigouvian tax, named after the British economist Arthur Pigou who developed the idea in the 1920s.

His idea is that a tax should be equal to the harm an activity imposes on others, so that the price the decision-maker faces reflects the full social cost rather than just the private one.

It works because each firm, left alone, weighs only the slice of damage that lands on itself and ignores the rest that spills onto rivals, workers or bystanders. The tax adds exactly that missing slice to the firm’s own bill, so the private calculation now matches the social one. The firm still chooses freely, but it chooses as if it were absorbing every consequence of its action. The market consequently avoids the catastrophic collapse in demand.

“Only [the revenue from] a Pigouvian automation tax…can fund retraining that raises income replacement, shrinking the externality over time and making the tax potentially self-limiting,” say Hemenway and Tsoukalas.

That’s interesting work introducing yet another reason to be cautious about the way AI is implemented. The key takeaway is that whatever AI does, humans will still be the ultimate customers (at least for now). Business, AI providers and policy makers who overlook this will do so at their peril.

Ref: arxiv.org/abs/2603.20617 : The AI Layoff Trap

INSIGHT

This paper delivers a striking insight that cuts across economics, strategy and public policy: rational, fully informed companies can collectively destroy the very market they depend on. And that knowing this provides no protection.

For policy makers, this will force a fundamental rethink of taxation in ways that go well beyond tinkering at the margins. Tax systems in most advanced economies were designed around the assumption that labour is the primary input to production and the primary source of taxable income. This is an assumption that AI is rapidly rendering obsolete.

Instead, genuinely novel forms of taxation will be needed, such as the Pigouvian levies described above.

Tax systems will also need to grapple with the risk that unilateral national taxes will push automation activity offshore, suggesting the urgent need for multilateral coordination of corporate tax frameworks.

The near-term challenge for policymakers is therefore not simply to raise or redirect taxes, but to redesign their foundational logic for an economy in which the relationship between production, income and spending can no longer be taken for granted.

In 1957, the British physicist John Lawson published a deceptively simple analysis that would shape fusion research for generations. His criterion — relating plasma density, temperature and confinement time — defined the conditions a fusion plasma must satisfy to produce net energy. It said nothing about the specific machine, the magnetic field strength or the fuel geometry. Its power lay precisely in that generality, giving researchers a universal compass pointing towards ignition regardless of the confinement concept being pursued.

In recent years, a new commercial reality has shaken up the fusion world. A plasma that produces more fusion energy than it consumes is a triumph of physics. A power plant that loses money for every megawatt-hour it generates is a commercial catastrophe. As dozens of private companies now race to commercialise various confinement concepts, the field has lacked an equivalent of the Lawson criterion for economics — a universal, technology-neutral framework capable of telling engineers and investors whether a given design can actually turn a profit.

Commercial viability

Enter Dennis Whyte at the Massachusetts Institute of Technology and Rutherford Energy Ventures in Cambridge, and colleagues, who have developed a model that does just that— it establishes clear, quantitative criteria for commercial viability that are independent of absolute plant size and agnostic to the particular fusion technology employed. “We hope that the simplicity, flexibility, and transparency of this model will make it a staple in the fusion development space, like the Lawson criterion before it,” say Whyte and co.

The team built the framework by drawing direct analogies with the Lawson criterion. Every fusion concept, regardless of how it confines the plasma, must extract energy through a physical surface surrounding the fuel. By calculating the ratio of money (or energy) going in, to the money(or energy) going out and doing this per unit of surface area, the researchers created a model that is independent to the specifics of any technology.

The model takes into account ten crucial parameters spanning plasma physics, engineering and finance. These include the fusion power density, the surface’s energy fluence limit, replacement time and cost, energy conversion efficiency, as well as fuel, construction and financing costs over the plant lifetime. It then defines Q_econ as the ratio of economic output to total economic expenditure, incorporating a factor that quantifies the fraction of calendar time the plant actually produces power.

Just as Lawson’s Q_p below unity means a plasma is losing energy, Q_econ below unity means a power plant is losing money.

Using a base case drawn from historical fusion device costs and energy market data, the researchers solved the model equations across wide ranges of all ten parameters. The results yielded several surprises.

One consistent finding is that a minimum fusion power density of approximately 2 MW/m² is required for basic economic viability. This threshold persists across a vast range of scenarios, directly contradicting the intuition that operating at low power density is economically attractive because it prolongs component life. Below this threshold, no combination of the other nine parameters can rescue a plant’s finances.

Equally counterintuitive is the revelation that economic performance is more sensitive to fast, low-cost surface replacement than to achieving extremely high radiation-tolerance limits. Engineers should therefore prioritise cheap, rapid maintenance cycles over the development of ultra-hard materials — a conclusion that significantly reorders conventional R&D priorities.

Subscribe now

The team validated the model by comparing its outputs against existing commercial fusion plant design studies, including the ARIES tokamaks, a series of designs developed in the early part of the century to evaluate fusion concepts.

Whyte and co independently determined the power densities and construction costs of these machines in good agreement with the earlier bottom-up engineering studies, despite needing no knowledge of the specific technologies involved.

The new model also introduces “economic isoquants”, contour lines of constant Q_econ in two-parameter space, that enable developers to quantify the precise trade-off between, for example, increasing power density and improving radiation tolerance.

The team say the model could be extended in future to consider specific fuel cycles, such as deuterium–helium-3 or proton–boron reactions. For now, the researchers hope the tool will join Lawson’s criterion as a foundational instrument in the field. “Its simplicity and transparency make the Lawson criterion universal and profound,” they say. Their goal is for Q_econ to play a similar role in the fusion development space.

Ref: arxiv.org/abs/2604.07367: Criteria for the economic viability of fusion power plants

INSIGHT

This paper’s most consequential findings challenge two deeply held assumptions about what makes fusion power economically viable, with significant implications for how physicists must prioritise their research agendas.

The first finding overturns the intuition that operating at low power density is a viable strategy for extending component lifetimes and reducing engineering costs. The analysis shows that a threshold fusion power density of roughly 2 MW/m² is required for basic economic viability, and this result holds robustly across a wide range of other parameters.

In short, plasma concepts that achieve energy gain but at low power density cannot be rescued by engineering optimisation — the physics must deliver sufficient power density from the outset.

The second finding is perhaps more surprising: it is far more valuable to make the plasma-facing surface cheap and quick to replace than to make it radiation-hard and long-lasting. The model shows that minimising replacement cost and downtime dominates over maximising component fluence limits.

For physicists working on plasma-materials interactions and neutron damage, this reframes the central question. Rather than pursuing ever-greater radiation tolerance, the field should prioritise simpler, cheaper replaceable components. High-fluence materials research remains necessary, but the economic returns from extending the lifetime of components beyond a modest threshold, diminish sharply.

WHEN Peter Shor unveiled his quantum factoring algorithm in 1994, the cryptographic world shuddered. Shor showed that by exploiting the laws of quantum mechanics, it is possible to find the prime factors of large numbers exponentially faster than any classical computer.

Banks, governments and militaries, and indeed most communication over the internet, rely on the near-impossibility of factoring enormous numbers to keep secrets safe. A machine running Shor’s algorithm efficiently would render that protection worthless.

The catch has always been hardware. Quantum computers are exquisitely fragile, their delicate superposition states shattered by the slightest environmental disturbance. Protecting computations from these errors requires encoding each logical qubit using hundreds of physical ones — a scheme called quantum error correction.

For cryptographically relevant tasks such as factoring a 2048-bit RSA integer, physicists estimate that millions of physical qubits will be needed. Since today’s best quantum processors contain a few hundred, it’s easy to imagine that there is plenty of time before these encryption techniques fail.

Closing the gap

Now Madelyn Cain at Oratomic in Pasadena and colleagues at the California Institute of Technology say a new theoretical analysis closes that gap dramatically. Cain and co have shown that Shor’s algorithm can be executed at cryptographically relevant scales using as few as 10,000 qubits — two orders of magnitude fewer than previously estimated.

And they set out a realistic path to breaking widely deployed cryptographic standards using near-term quantum computers based around neutral-atom hardware. Their conclusion is that the threat from quantum computers to existing encryption infrastructure is considerably more imminent than the field had assumed.

Cain and co achieved this reduction by combining several advances: high-rate quantum error-correcting codes that increase the density of logical qubits for a given number of physical qubits, along with carefully optimised instructions and circuit designs for Shor’s algorithm. Their analysis how this can cut qubit requirements by roughly two orders of magnitude compared to standard approaches.

The key innovation lies in so-called quantum low-density parity-check (qLDPC) codes, which significantly outperform the conventional approach using “surface” codes. In surface codes, each qubit only talks to its nearest neighbours — which is why they need so many physical qubits per logical one.

But in neutral atom quantum computers, individual atoms serve as qubits, trapped in arrays of focused laser beams called optical tweezers. Because the atoms can be physically rearranged during computation, they can connect to more distant partners, not just local ones. The high-rate codes take advantage of those longer-range interactions to encode information much more efficiently within the same block of atoms.

The difference is stark. Conventional error correction using surface codes requires hundreds of physical qubits to protect a single logical one, giving encoding rates around 1per cent. The new codes developed by Cain and colleagues achieve encoding rates near 30 per cent. This dramatic improvement in efficiency is what drives the headline reduction in total qubit count.

Subscribe now

The team’s existing neutral-atom experiments have already demonstrated fault-tolerant operation on arrays of up to 500 qubits and coherent trapping of more than 6,000 atoms.

This has significant potential in the not-too distant future. “Shor’s algorithm can be executed at cryptographically relevant scales with as few as 10,000 reconfigurable atomic qubits,” say Cain and co.

The calculate the computer time required for a few scenarios. To crack the standard elliptic curve cryptography ECC-256 codes, the researchers project runtimes of just 10 days with approximately 26,000 qubits, and cracking RSA-2048 codes would take just 97 days with approximately 102,000 qubits.

“These results position neutral-atom systems as a leading platform for utility-scale quantum computation, with the capacity to drive innovation across science and industry,” say Cain and her colleagues. “This conclusion underscores the importance of ongoing efforts to transition widely-deployed cryptographic systems toward post-quantum standards designed to be secure against quantum attacks.”

Ref: arxiv.org/abs/2603.28627: Shor’s algorithm is possible with as few as 10,000 reconfigurable atomic qubits

INSIGHT

This paper fundamentally reshapes the timeline for quantum computing’s threat to modern cryptography. By combining high-rate quantum error-correcting codes with optimised circuit designs and neutral-atom hardware, the authors demonstrate that Shor’s algorithm — capable of breaking RSA and elliptic-curve encryption — could run on roughly 10,000 physical qubits rather than the millions previously assumed. This collapses the perceived gap between current hardware and cryptographically relevant computation by two orders of magnitude.

For cybersecurity and policy, the implications are urgent. Systems protecting financial transactions, government communications and critical infrastructure rely on the very encryption standards this work targets. The findings substantially strengthen the case for accelerating migration to post-quantum cryptographic standards, a transition already underway but widely considered unhurried.

More broadly, the work reframes the race to build quantum computers that can work at relevant scale. The leading technologies include those that rely on superconducting qubits, ion qubits and photonic qubits.

This work positions neutral-atom quantum computers as leading candidates for fault-tolerant quantum computation. The team’s architectural innovations, particularly around high-rate quantum LDPC codes and efficient logical operations, are transferable to problems in quantum chemistry, materials simulation and machine learning, where large logical qubit counts with low overhead are equally desirable.

The new insight is that flexible hardware and clever algorithmic design can substitute for raw qubit counts far more powerfully than previously imagined. If so, a quantum advantage may arrive sooner, and on more modest hardware, than anybody anticipated.

When it comes to the physics of civilisation, one iron law is that all the energy society consumes eventually becomes heat. Coal burned in a power station, petrol combusted in an engine, electricity flowing through a circuit — every joule, wherever it begins its journey, ends as thermal energy dissipated into the environment. On a planetary scale, this waste heat is a largely invisible consequence of modernity, spread thinly enough across the atmosphere and oceans that its local effects rarely attract attention.

AI data centres change that calculus. These facilities concentrate energy consumption at extraordinary densities — the electronic components inside a modern hyperscaler can reach thermal loads on the order of 10 million watts per square metre more than a steel furnace — and they discharge the resulting heat into a specific location.

Unlike the diffuse warmth of a motorway or a suburb, the thermal output of an AI facility is intense, continuous and geographically fixed. As the global fleet of hyperscalers has grown to more than 11,000 sites and their power demands have begun to rival those of entire manufacturing sectors, the question of what that concentrated heat does to the surrounding environment has become urgent.

Data furnaces

Now Andrea Marinoni at the University of Cambridge and colleagues have quantified that effect for the first time. Using two decades of land surface temperature data gathered by satellites, they show that an AI data centre raises the temperature of the surrounding land by an average of 2°C — a phenomenon they name the data heat island effect. The result has significant implications for public health given that more than 340 million people live within 10 kilometres of a data centre analysed in the study.

The heat island effect Is not new. The urban version arises from a well-understood constellation of causes: the geometry of urban canyons trapping particulates, the replacement of vegetation with heat-retaining materials, the release of air pollutants and water vapour, and the low reflectivity of building surfaces.

Data centres share some of these characteristics but add an extra one: extraordinarily high power density. The vast majority of these facilities run on fossil fuels, meaning that the heat they dissipate is a direct product of combustion as well as computation.

To isolate the thermal footprint of data centres from the background noise of regional climate trends and other human activity, Marinoni and colleagues focused their analysis on the 8,472 facilities located outside densely built-up areas. For each site, they computed the monthly average land surface temperature within a circular region centred on the data centre, then calculated the difference between that temperature and the average recorded over the five years before the facility began operations.

Subscribe now

What they found is striking. Land surface temperatures in the vicinity of a data centre are essentially flat in the months before operations begin, then jump sharply at the point of activation. with the average increase across all sites being just over 2°C. “The average land surface temperature increase across the data centres is 2.07°C, whilst its minimum and maximum can be found at 0.3°C and 9.1°C, respectively,” say Marinoni and co.

The effect is not confined to the immediate footprint of the building. Marinoni and colleagues say the detectable warming extends to 10 kilometres, which is likely to have significant consequences.

By overlaying their temperature maps with demographic data, the team find that up to 343 million people could be affected by the data heat island effect. “The data heat island effect could have a remarkable influence on communities and regional welfare in the future,” say the researchers.

Marinoni and co explore various mitigation strategies that could reduce the impact of data heat islands. These range from low-power circuitry and dynamic power management to passive radiative cooling coatings that can reduce thermal loads by up to 40 per cent without consuming additional energy.

Ref: arxiv.org/abs/2603.20897: The data heat island effect: quantifying the impact of AI data centers in a warming world

INSIGHT

The key new insight from this paper is that AI infrastructure is itself now a measurable driver of local climate change, distinct from the well-established urban heat island effect. Where the urban heat island has been studied for decades as a product of cities and industry, this work identifies AI expansion as an emerging and previously unquantified contributor to the same class of phenomenon.

The broader implications span several disciplines. For climate science, it adds a new anthropogenic forcing variable to regional temperature modelling that current frameworks don’t account for.

For public health and urban planning, the finding that over 340 million people live within the thermal footprint of these facilities means that data centre siting decisions will be increasingly treated as environmental and welfare policy questions, not merely logistical ones.

For AI governance and sustainability research, it provides an empirical evidence base that will likely inform a new round regulatory conversations about the physical impact of AI scaling.

And for energy systems research, the work signals that demand-side heat management — not just carbon emissions accounting — must become central to responsible AI infrastructure design.

In other words, the task for the companies behind AI scaling just got harder.