Section
37 articles
Cities are full of cameras, but most of them belong to businesses, residents and other organisations. Their footage is usually locked away by ownership rules, privacy settings or encryption. That leaves tracking systems with only a thin set of accessible sensors, long gaps between sightings and little help from the many devices already watching the street.
Biologists testing thousands of genes often face an expensive question: which untested genes should go into the next batch? A new machine-learning approach combines a general-purpose language model with a specialized policy that learns from past gene hits, screen descriptions, and round-by-round lab feedback. The goal is to turn noisy experimental results into better guesses about what to test next, while keeping choices grounded in known biology.
For a blind person moving through a room, the hardest part may be knowing where things are before reaching out. One answer is to stop describing objects in words and instead give them a place in space as sound. A wearable pair of glasses can detect nearby items and play short everyday sounds around the head, so a bottle might feel like it is coming from the left while a cup sits ahead. The question is whether that kind of spatial audio can guide hands reliably enough to matter.
Data centers are getting hotter in more than one sense. As servers multiply, the electricity needed to keep them cool has become a major part of their global footprint. One idea is to stop fighting heat and let server rooms run much warmer than today's comfortable 22 degrees. That sounds simple, but it changes how cooling works, what kind of machines can tolerate the warmth, and whether the savings are large enough to matter across continents.
A robot arm reaches for a cup, a package, or a surgical tool, and the hard part arrives before contact: deciding which motion will hold up once the world pushes back. Engineers are teaching machines to run short rehearsals of possible moves, using learned models that predict what cameras, joints, and objects might do next. Those rehearsals sit beside real safety checks, factory deployments, and medical robots already in use.
During radiotherapy, a quick cone-beam scan helps aim treatment at the tumour. The same image is often too grainy and distorted to calculate dose safely, so clinicians may order another CT scan. That extra scan adds radiation, cost, and waiting time. A new way of rebuilding a cleaner, CT-like volume from the noisy scan could change that workflow by using the scanner's own geometry as a built-in check.
In Portugal's emergency departments, a hospital's forecast should fit inside its region, and every region should fit inside the country. Yet many demand models treat each level as separate, leaving planners with numbers that look precise but do not reconcile. The question is whether one system can predict patient visits 28 days ahead while keeping hospital, regional and national totals in line.
When a skater glides across Olympic ice, today's live video sends millions of pixels for the rink, crowd, and camera angle again and again. GenStream proposes a different route: transmit only the moving performer's pose, a simple box around them, and the viewpoint needed to rebuild the scene. The viewer's device then paints realistic bodies into a prebuilt 3D arena. The idea promises much smaller streams for sports and performance video, but it shifts heavy work from the network to the screen.
Household robots must decide what to do next while objects slip, sensors miss things, and instructions change mid-task. A new training method lets them learn from their own mistakes in simulation, turning failed rollouts into examples that teach better recovery without step-by-step human demonstrations.
When a voice is altered to protect identity, the question is whether it still carries clues that can link it back to its owner. Speech de-identification raises several different risks at once. Five systems are compared for matching resistance, leakage of age and sex, separation between speakers, and usefulness for listening.
Every Wi-Fi connection reports a simple number: how strong the signal is. That number is usually treated as a nuisance, useful only for guessing whether you are close to a router. But it may carry more information than most people expect. A University of Iceland test checked whether those ordinary strength readings could be fed into an existing model built from richer Wi-Fi data and still produce a rough sense of where someone stands in a room. The setup used inexpensive boards, no special sensors, and the kind of signal values that many smart devices already expose.
Bharatanatyam is built from repeated units called adavus, each a choreographed chain of held poses and the movements that link them. A camera-based system must decide which unit a dancer is performing by following that whole sequence. The hard part is not naming one pose; it is keeping track of small foot changes, hidden hands and costume as similar steps blur together.
High-frequency ultrasound can image the thin layers of skin, from epidermis to muscle, but automated analysis is held back by a shortage of expert labels for deeper structures. A physics-based simulator offers one route around that bottleneck: it builds virtual skin phantoms and generates paired scans with exact layer masks. The open question is whether those synthetic images contain enough realistic structure and texture to prepare a model before it sees real patient data.
Hospitals keep detailed records of diagnoses, procedures, lab tests and medicines for every admission. Those structured entries are usually treated as dry administrative data, while the richer story lives in doctors' notes. But predicting whether a patient will come back within a month may not require reading any prose at all. Turning routine hospital codes into named medical concepts makes it possible to estimate readmission risk from the record itself.
Mobile networks are often described by speed tests and coverage maps, but a real city is messier than either. Buildings, traffic, handovers and weather all shape what a phone actually receives. Over many months, cars drove through urban and suburban Vienna carrying scanners and smartphones that recorded radio signals from 4G LTE and 5G NR networks. The result is an open dataset that lets engineers and researchers examine how mobile airwaves behave in the places where people really live.
Quantum computers are often imagined as machines that count possibilities in parallel. Their raw output, however, is a set of probabilities that add to one, which makes them awkward for tasks needing many ordinary numbers at once. Reconstructing an image from two inputs is exactly such a task. A hybrid design tries to fix this by letting a small quantum circuit feed a classical adjustment layer, so the system can emit full continuous vectors rather than separate single values.
Quantum computers in a shared cloud are not interchangeable machines. Each has its own qubit layout, native gates, noise levels and routing rules, so the same circuit can become very different low-level instructions on each device. Compiling every job on every machine to find the best fit is slow and wasteful, especially when many circuits must be dispatched at once. The question is whether a system can estimate where a circuit will run well before paying that compilation cost.
City traffic moves in patterns shaped by more than asphalt. A highway sensor near an office tower, a stadium, or a residential street sits inside a place with its own rhythms, yet many forecasting tools connect sensors only by road distance. One idea is to give those tools a second map: not just how far apart sensors are, but what kinds of places surround them. That extra layer can be built from public knowledge about locations, without hand drawing a city specific ontology.
When someone leaves an emergency department and returns within two weeks, clinicians may wonder whether the first visit missed something important. Hospitals often review only a narrow slice of these returns because full chart reviews are slow and expensive. Doctors and artificial intelligence were asked to judge which return visits deserve closer scrutiny, using only brief diagnosis labels.
On Reddit, public awards are small symbolic rewards that appear on posts and comments. A field experiment sent one-dollar Diamond awards from two accounts, one presented as human and one clearly labeled as a bot, to low-attention contributions in general-interest communities. Each award included a private message explaining the reason, such as praise for logic, empathy, moral integrity, or a random lottery win. The aim was to see whether these visible signals changed recipients' replies, later activity, writing tone, and how other users responded.
A contact-tracing app that warns users of infection risk also creates a narrow path for an attacker to learn whether one specific person is sick. The question is whether the warning can be made private enough to protect individuals while still stopping outbreaks, without turning every alert into useless noise.
Game developers often want computer opponents that feel varied and believable, not just optimal winners. A new training method tries to give game agents adjustable personalities by describing a player’s style as a short list of numbers. Each number sets how strongly the agent should show a trait such as aggression, cooperation, or risk-taking. The approach lets one trained model shift between many styles after learning is finished.
Choosing a blood test, scan, or lab value is often a judgment call. A patient may have several possible explanations for the same symptoms, and each test can point in a different direction. New work asks whether a language model can help by imagining what each result might show before the sample is drawn.
Artificial intelligence systems rely on a step called attention, where each word weighs the others to decide what matters. That step is usually computed by digital chips using exponential math. The open question is whether tiny physical oscillators—swinging circuits that naturally settle into patterns—could do part of this job with less energy. The issue is not only imitation. It also asks where their physics helps, what limits them, and how close they come to today’s software.
Brain metastases can range from large, obvious masses to specks barely visible in a 3D MRI scan. A computer system must find every one of them while avoiding false alarms that could send patients toward unnecessary treatment. One way to test whether scale matters is to give the same type of computer model two different views of the brain: a wide view with more context and a tighter view where small lesions occupy more space.
Mobile networks have long been sold as faster pipes for data. The question now is whether that model still fits what users, factories, grids, and governments actually need. A seventh generation would not be defined by speed alone. It would require a new way of coordinating spectrum, devices, energy, security, and business models across an industry already moving toward 6G.
A new theoretical model treats large language models less as software and more as a cultural habit that can spread through societies. It asks whether the convenience of asking an AI to read, write, reason, or check work could change not just individual productivity but the collective capacity for unaided thinking. The framework borrows ideas from epidemic spread, cultural evolution, and cognitive offloading to explore how adoption, dependence, autonomy, and recovery might interact across a population.
When wildfires race through dry hills, crews can be slowed by smoke, broken radio links, and exhausted shifts. A computer experiment asks whether small drones could help by learning to fly around fire on their own. Instead of following fixed rules, the drones are trained in a virtual world where flames appear, spread, block paths, or crash into obstacles. The goal is not to drop water but to understand how groups of autonomous aircraft can keep watch over dangerous edges, share information, and adapt as conditions change.
Electronic health records contain a long stream of measurements, notes, scans, and decisions made during hospital care. A machine-learning system trained on that stream treats each patient stay as one continuous sequence and asks what kind of event may come next, how much time will pass before it arrives, and whether changing an earlier treatment choice would alter the later course.
Aircraft power systems hum at a frequency unlike ordinary grids, and their faults can appear as sudden voltage drops, flicker, or short transients. To catch them quickly, engineers are testing compact neural networks that learn from simulated electrical signals and then run on aerospace-grade chips. The question is whether such a system can be small enough to fly while still recognizing the subtle signatures of trouble.
Guitarists do more than press strings and pick notes. They bend, slide, hammer, pull, mute, and strike the body to shape a phrase. Turning that performance into written tablature is harder than writing piano scores because many pitches can be played in several places on the neck, while the sound itself carries clues about which string was used. A new modular pipeline attempts to solve this by separating transcription into stages: first detecting notes, then recognizing expressive techniques, assigning strings and frets, and finally producing a score that a player could actually use.
Hospitals are turning to software robots for billing, scheduling, insurance checks, and records work. The hard part is not buying bots; it is deciding which administrative tasks should be automated first, what tools can safely handle protected health information, and how much each project will cost or save.
When an AI edits code to fix a bug, it needs evidence that the edit works. Official test suites can confirm success, but they are not the only signal available while a model is still learning to repair. One route is to let another model write small regression tests from the issue report and use their pass/fail signals during repair. The challenge is making those generated checks trustworthy enough to guide a fix without replacing final verification.
A set of coding agents was asked to locate internal directions inside large language models that can nudge generated text toward a target idea or style. The task required them to search through thousands of candidate features, test how sharply each one responds to relevant examples, and then check whether adding the feature actually changes model output. The setup turns an opaque machine into something closer to a control panel, where the right lever can make a model answer in Portuguese, describe symptoms, or focus on real estate.
Recurrent language models promise a cheaper way to handle very long texts because they carry a compact memory instead of storing every previous token. Yet when trained on short passages, that memory often fails to connect distant causes and effects. The key question is whether the backward signal that teaches earlier hidden states from later mistakes can be kept alive without changing what the model predicts or how it is scored.
A training-free pipeline turns ordinary camera photos into plain-language descriptions, then attaches those words to points in a 3D scan. Instead of asking a vision model to match images directly to point clouds, it uses language as the bridge: objects are described, segmented, projected back into space, and summarized for each point.
Artificial intelligence agents are increasingly built to do narrow enterprise jobs, from checking payroll to managing websites. Builders often improve them in two ways: giving the model better instructions and tools, or fine-tuning its weights on examples of good behavior. The tempting question is whether combining those methods produces a stronger worker.
