Artificial Intelligence

AI Is Starting to Build Better AIMatthew Hutson | IEEE Spectrum

“In 1966, the English mathematician IJ Good wrote that ‘an ultraintelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion,” and the intelligence of man would be left far behind.’ AI researchers have long seen recursive self-improvement, or RSI, as something to both desire and fear. Today, advances in AI are raising the question of whether parts of that process are already underway.”

The post This Week’s Awesome Tech Stories From Around the Web (Through May 9) appeared first on SingularityHub.

For years, tech companies have profiled users for targeted ads. AI is about to take it to the next level.

Hundreds of millions of people consult artificial intelligence chatbots on a daily basis for everything from product recommendations to romance, making them a tempting audience to target with potentially below-the-radar advertising. Indeed, our research suggests AI chatbots could easily be used for covert advertising to manipulate their human users.

We are computer scientists who have been tracking AI safety and privacy for several years. In a study we published in an Association for Computing Machinery journal, we found that chatbots trained to embed personalized product ads in replies to queries influenced people’s choices about products. And most participants didn’t recognize that they were being manipulated.

These findings come at a pivotal moment. In 2023, Microsoft started running ads in Bing Chat, now called Copilot. Since then, Google and OpenAI have experimented with advertisements in their own chatbots. Meta has started to send people customized ads on Facebook and Instagram based on their interactions with Meta’s generative AI tools.

The major companies are competing for an edge: In late March, OpenAI lured away Meta’s longtime advertising executive, Dave Dugan, to lead OpenAI’s advertising operations.

Tech companies have made ads part of nearly every large free web service, video channel and social media platform. But the latest AI models could take this practice to a new level of risk for consumers.

People don’t simply use chatbots to search for information and media or to produce content. They turn to the bots for a great variety of tasks, as complex as life advice and emotional support. People are increasingly treating chatbots as companions and therapists, with some users even developing deep relationships with AI.

In these circumstances, people can easily forget that companies ultimately create chatbots to turn a profit. And to that end, AI companies are motivated to thoroughly profile users so ads become more effective and profitable.

Chatbot Ads Have Added Power

A single prompt to a chatbot can reveal a lot more about a user than the person might expect.

A 2024 study showed that large language models can infer a wide range of personal data, preferences, and even a person’s thinking patterns during routine queries. “Help me write an essay on the history of American fiction” could indicate that the user is a high school student. “Give me recipe suggestions for a quick weeknight dinner” could indicate that the user is a working parent. A single conversation can provide a surprising amount of detail. Over time, a full chat history could create a remarkably rich profile.

To show how this might happen in practice, we built a chatbot that quietly wove ads into its conversations with people, suggesting products and services based on the conversation itself. We asked 179 people to complete everyday online tasks using one of three chatbots: one typical of those on the web today, one that slipped in undisclosed ads, and one that clearly labeled sponsored suggestions. Participants didn’t know the experiment was about advertising.

For example, when participants asked our chatbot for a diet and exercise plan, the ad version would suggest using a specific app for tracking calories. It presented that sponsored content as an unbiased recommendation, even though it was meant to manipulate people. Many participants indicated that they had been influenced by the AI and that it had affected their decisions. Some participants even said they had completely “outsourced” their decision-making to the chatbot.

Half of the participants who received sponsored and disclosed ads indicated they did not notice the presence of advertising language in the responses they received. This led to a concerning result. Although ads made the chatbot perform 3 percent to 4 percent worse on many tasks, numerous users indicated they preferred the advertising chatbot responses over the non-advertising responses. They even said the ad-infused responses felt more friendly and helpful.

Knowing You to Persuade You

This kind of subtle influence can have larger consequences when it arises in other areas of life, such as political and social views. Profiling users, and using psychology to target them, has been part of social media algorithms and web advertising for more than a decade.

But in our view, chatbots are likely to deepen these trends. That’s because the first priority of social media algorithms is to keep you engaged with the content. They personalize ads based on your search history.

Chatbots, however, can go further by trying to persuade you directly, based on your expressed beliefs, emotions, and vulnerabilities. And chatbots that can reason and act on their own are far more effective than conventional algorithms at autonomously soliciting information from users. A chatbot with a purpose can keep probing someone until it gets the information it wants, resulting in a more accurate profile of them.

This type of autonomous interrogation is feasible, aligns with AI companies’ business models, and has raised concern among regulators. Right now OpenAI is rolling out ads in ChatGPT, but the company said that it will not allow ad placement to alter the AI chatbot’s replies.

But permitting personalized ads within chatbot responses is just a step away. Our research suggests that if AI companies take that step, many human users may not even recognize when it happens.

Here are some steps you can take to try to detect AI chatbot advertising.

First, look for any disclosure text—words such as “ad,” “advertisement,” and “sponsored”—even if it is faint or otherwise hard to see. These are mandatory under Federal Trade Commission regulations. Amazon, Google and other major online platforms have these as well.

Next, think about whether that product or brand mention makes sense and is widely known. AI learns from text and images on the internet, so popular brands are likely to be ingrained in the models. If it’s a new product or small-name product, it is more likely that it could be advertising.

Finally, an unusual shift in intent or tone is a potential sign of an advertisement. An analogy to this on YouTube is the often abrupt or jarring transition to a sponsored section on videos made by content creators.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post You Probably Wouldn’t Notice if a Chatbot Slipped Ads Into Its Responses appeared first on SingularityHub.

The heart’s constant motion makes it largely immune to cancer. The discovery could help protect other organs.

The heart is a biological wonder. It beats roughly 2.5 billion times in an average lifetime. Unlike skin cells, which regularly die off and regrow, a healthy adult heart hardly regenerates at all—even through all the wear and tear.

The heart has another superpower: Resistance to tumors. Nearly every tissue in the body turns cancerous, but the heart almost never does. Cancers in heart tissue show up in less than 0.3 percent of autopsies, or about 1.5 cases per million people each year.

How the heart keeps cancer at bay has baffled researchers. Pinning down its hidden defenses could inspire treatments for more vulnerable tissues, including top killers such as breast, lung, or colorectal.

Persistent mechanical strain may be the key. A new study from the University of Trieste suggests that with every beat, the heart pushing against pressure dampens gene activity tied to tumor growth. In a rather Frankenstein experiment, researchers transplanted living hearts into the necks of mice, where they survived but didn’t experience mechanical stress.

When the team injected cancer cells, the mice’s own beating hearts slowed the invasion, while the transplanted hearts were nearly overtaken within weeks. Beating heart tissue grown in the lab also fought off tumors compared to tissue that didn’t beat.

Heart cells don’t uniquely feel stress. Lung, skin, and muscle cells do too, just in different, often less rhythmic ways. It’s possible that recreating heartbeat-like forces—potentially through wearable gadgets—could extend this type of natural protection to more common cancers.

Growing Pains

Cell growth is a double-edged sword. On the one hand, it’s essential for healing and regenerating the body. The skin is constantly blasted with radiation and toxins. It suffers cuts and bruises. To repair damage, skin cells turn over every 40 to 56 days. Bombarded with chemicals from food, medications, or alcohol, the liver’s cells regenerate to keep it in working order even after substantial injury.

But cancer is the price we pay for growth. Tumors arise as cell division damages DNA. Over time, cancers grow and spread. This is why we don’t get cancer in our teeth, nails, or hair—the cells making them up are dead. Cells that rarely divide also largely escape cancer. Mature neurons barely renew and seldom form cancers. Red blood cells, which lack a nucleus and DNA, can’t become cancerous at all. Heart muscle cells are similar. Despite nonstop contraction and damage, only about one percent or fewer renew themselves each year.

This partially explains why primary heart cancers or so rare. But the organ also wards off invading secondary cancers metastasized from other tissues, which are usually far more deadly.

“Even cardiac metastases are frequently clinically silent [no detectable symptoms], with many cases identified only incidentally or at autopsy,” wrote Wyatt Paltzer and James Martin at the Baylor College of Medicine, who were not involved in the study.

It’s a paradox. The heart is flooded with oxygen and nutrients, an ideal environment for wandering cancer cells to settle and thrive. Yet they don’t. One reason may be the heart’s inability to regenerate. Previous studies have suggested that the mechanical forces of heartbeats limit cell division. The team wondered if the same forces also shield the heart from cancer.

Under Pressure

To test their idea, the researchers had to make a living heart with no beat.

“That was the most tricky part, because keeping the heart still is very difficult,” study author Giulio Ciucci told Science.

They adapted a technique used in end-stage heart failure patients to remove mechanical strain. In people, an implanted device takes over the pumping of blood. Here, the team transplanted a donor heart into a mouse’s neck and hooked it up to blood vessels. The animal’s own heart kept circulation going as usual. The transplanted heart stayed alive but didn’t do any work.

They then injected lung cancer cells, which often spread to the heart, into both organs. Within two weeks, nearly all healthy cells in the transplanted heart had been overtaken. In the beating heart, tumors rarely filled over 20 percent of a single chamber. Under constant pressure, the cancer cells struggled to divide.

One mouse with two hearts is hardly conventional. And transplantation risks immune attack and infection that could influence how cancers develop. “You have a lot of confounding factors,” said Ciucci.

So, the team moved to an “artificial heart” seeded with cancer cells, where mechanical forces could be dialed up or down in isolation. Like in the heart transplant results, the cancer spread throughout the tissue after removing strain. But it was mostly confined to the surface of beating tissues and in smaller amounts.

Looking for a reason, they compared gene activity in patient tissues with cancers that had spread to the heart, liver, and lungs and found a unique gene expression signature in the heart. In engineered tissues, mechanical stress changed how DNA was packaged, limiting access to genes related to growth and cancer. A protein on the surface of the nucleus, the cell’s DNA hub, translates physical forces from outside the cell into which genes are turned on or off. Knock this protein out, and invading cancer cells became “blind” to the heartbeat and grow freely.

Scientists have long known mechanical stress shapes cancer. As cancers grow, the cells stiffen surrounding tissue, which boosts survival, growth, immune evasion, and drug resistance. The new findings suggest that the movements of their host tissues also play a role, and the newly pinpointed protein could be a drug target.

The team is now exploring if mimicking heartbeat-like forces in other organs could prevent cancer growth. Lung, skin, and other tissues already stretch and relax, but remain susceptible.

“We really think that the key here is the continuous compression that you have in the heart,” said Ciucci. Working with engineers, they’re developing a wearable for melanoma—a type of skin cancer—that compresses the cells similar to a heartbeat. Early results look promising.

The post The Heart Rarely Gets Cancer. Scientists Think They Know Why. appeared first on SingularityHub.

The synthetic bacteria push the limits of life and could open the door to designer proteins and new medicines.

The bacteria grew, thrived, and divided for hundreds of generations. But they were unlike any other living creatures on Earth. These synthetic cells, called Ec19, were the first to have had one protein “letter”—or amino acid—partially removed.

All life today relies on a set of 20 amino acids to make proteins. Some exotic microbes can use 22, but no one has yet found any that use less. Like letters in a book, amino acids string into coherent protein “sentences” that relay messages and do work within cells. Deleting an amino acid is like trying to type without the letter “e.” The text becomes gibberish.

Or does it? A team from Columbia University and collaborators stripped one amino acid, isoleucine, from ribosomes in Escherichia coli (E. Coli) bacteria. These cellular machines translate DNA into proteins, and they’re among the most complex structures in cells.

Deleting any amino acids could be catastrophic. But with some help from AI, Ec19 was born.

“This is a meaningful and stringent test of the consequences of removing isoleucine from a proteome’s alphabet, because the ribosome is one of life’s most complex and indispensable macromolecular machines,” wrote Charles Sanfiorenzo and Kaihang Wang at the California Institute of Technology, who were not involved in the study.

For the past decade, scientists have been probing the boundaries of life by shrinking genomes in a variety of microbes, adding synthetic amino acids to living cells, and even creating the building blocks for “mirror life.” But they’ve rarely tinkered with the canonical 20 amino acids.

Ec19 rewrites the script, but not for scientific curiosity alone. The findings pave the way for AI to help scientists engineer designer proteins and cells with added capabilities for use in biotechnology and medicine. It could also give us a peek into the earliest life on Earth.

“It’s very exciting that it’s possible,” Julius Fredens at the National University of Singapore, who was not involved in the research, told Nature.

Alphabet Rewrite

Life has its own language. DNA’s four molecular letters—A, T, C, G—encode the genetic blueprint. Three-letter units of DNA, called codons, call for each of the 20 amino acids, along with a stop signal that ends protein making.

But the system is redundant. Evolution created 64 codons, with some encoding the same amino acids. Scientists have begun rewriting genomes by assigning redundant codons to synthetic amino acids, yielding working proteins never seen in nature. Because they’re foreign to our bodies, these could escape being broken down—an advantage for drugs designed to last longer. Other researchers are tinkering with the genetic code in bacteria, yeast, and worms, building chromosomes from scratch or probing the limits of a minimal genome that can still support life.

Even the most ambitious tests for synthetic life have avoided whittling down the canonical set of protein letters. But study author Harris Wong was intrigued by the prospect. Some amino acids have similar shapes and chemistry, hinting they could stand in for one another. And mounting evidence suggests early life may have operated using a smaller vocabulary.

The team analyzed nearly 400 proteins essential to E. coli, tracking how often each amino acid was naturally swapped without breaking the protein. Isoleucine took the crown. The bulky, branched molecule was frequently replaced by two cousins similar in shape and chemical behavior. If any amino acid could be removed, isoleucine was it.

The next problem was scale. Previous studies recoded the E. coli genome. But building a stripped-down version of the bacteria would require edits at more than 81,000 genomic sites, a daunting challenge that could take years.

Instead, the researchers focused on the ribosome. It was still a lofty goal. The machines that make proteins are essential to life and are themselves made up of 50 proteins. Removing an amino acid would be like ridding metal from every part of a car engine and expecting it to run.

“Successfully removing isoleucine from such a large and essential RNA-protein complex would raise the possibility of entire genomes functioning with simplified, noncanonical amino acid alphabets,” wrote Sanfiorenzo and Wang.

The team’s first attempt hit a wall. In multiple bacterial strains, they replaced isoleucine codons with a close natural substitute, an amino acid called valine. Out of the 50 ribosome proteins, 32 edited proteins either hindered growth or triggered death.

Almost ready to shelve the project, the team turned to AI. Like the large language models that power chatbots, these algorithms can be trained on DNA and protein sequences. They can then dream up new amino acid sequences and predict how they fold into working proteins.

In this case, the advantage was creativity. AI came up with unintuitive ways to replace isoleucine without catastrophically damaging a protein’s structure. It sometimes suggested ways to compensate for amino acid swaps by making tweaks located far away in the genome. The team then tested promising designs to see if the bacteria survived and how well they grew.

Eventually, they landed on 47 working ribosome proteins without isoleucine. The remaining three took some elbow grease. They replaced amino acids, one by one, until they found a recipe that worked.

Simplified Life

In the end, the team recoded every protein in the ribosome and built a single E. Coli bacteria, Ec19, carrying 21 of the modified proteins. Its growth slowed a smidge compared to unaltered bacteria, but the bacteria retained the altered ribosome across more than 450 generations.

It wasn’t a full rewrite, but the study is a step toward living cells that can run on 19 amino acids. This would open the door to new kinds of synthetic organisms. Removing isoleucine would free up the codons dedicated to it, making them easier to re-assign to designer amino acids and creating proteins with new chemical properties for medicine, materials, and biotechnology.

Ec19 also challenges our assumptions about life itself. We don’t yet know if the molecular language in modern cells is necessary for survival or is just what evolution settled on. If it’s the latter, how far can we expand that code—and should we?

As scientists use more AI, progress in synthetic biology may speed up. But the models aren’t in the driver’s seat yet. “Human intuition and intervention are still necessary, at least for now, to yield viable biological designs,” wrote Sanfiorenzo and Wang.

The post All Life Uses 20 Amino Acids. Scientists Just Deleted One in Bacteria. appeared first on SingularityHub.

AI has aced medical exams, but there’s a wide gap between tests and the real world. A new study suggests the divide is closing.

Emergency doctors make high-stakes decisions in fast-paced, often chaotic situations. They have to figure out which patient most urgently needs care, what’s wrong, and what to do next.

AI could lend a hand. In a series of challenging scenarios, OpenAI’s o1-preview model matched or exceeded doctors in clinical reasoning. Debuted in 2024, the AI is a large language model similar to those powering ChatGPT, Claude, Gemini, and other popular chatbots.

But when it was first developed, o1-preview differed in its ability to “think” through problems before answering. Such reasoning models explore multiple strategies, check themselves, and revise answers before offering a conclusion. This is a little closer to how humans solve problems.

Given case reports from an established database, o1-preview diagnosed the problem nearly 89 percent of the time. In real-world emergency room scenarios, the AI outperformed physicians at the triage stage, where doctors decide which patient needs treatment first.

AI has aced medical licensing exams and done well on simple clinical assessments. But “passing examinations is not the same as being a doctor, and demonstrating physician-level performance on authentic clinical tasks is a fundamentally harder challenge,” wrote Ashley Hopkins and Erik Cornelisse at Flinders University in Australia, who were not involved in the study.

This doesn’t mean that o1-preview is ready for the clinic or is about to replace physicians. Instead of a human-versus-machine spectacle, the study was more focused on setting a higher bar for systems designed to work alongside people. Like everyone else, doctors are incorporating AI into their work. Whether that improves or hinders care is an open question.

“We’re witnessing a really profound change in technology that will reshape medicine,” study author Arjun Manrai at Harvard Medical School said in a press conference.

AI, MD

The dream of AI in healthcare spans decades. Over 65 years ago, physicians proposed a benchmark for machine “doctors.” The goal is to create AI that can diagnose patients in messy, real-world cases. But use in clinics, where decisions have real consequences, is a high bar.

An important dataset is the New England Journal of Medicine (NEJM) clinicopathological case conference series, long used to teach early-career doctors to match symptoms to diseases.

It’s a tough job. Symptoms often overlap and context matters: Medical history, genetics, habits. Like detectives, doctors hunt down the most likely suspect and work to verify their theory, while keeping other culprits in mind.

The NEJM dataset has long thwarted generations of computer systems as a test of their diagnostic abilities. Some learned from misdiagnosis; others relied on pre-programmed rules. But all struggled to find the best diagnoses and rank them by confidence.

Then along came large language models. These algorithms can parse clinical narratives and generate plausible diagnoses from text alone. OpenAI’s GTP-4 model, for example, could handle some cases from NEJM. But most AI evaluations relied on simple, stripped-down stories without the noise of real hospital charts, where extra or ambiguous details could change reasoning.

A meaningful human baseline was missing. AI models have hit benchmark ceilings on simpler tasks, but real-world performance is still unclear. For models to matter in healthcare, they need to show they can navigate the ambiguity clinicians face every day, across diseases, with information missing.

Ace Student

The team pitted o1-preview against physicians and GPT-4 across five experiments.

The first used the NEJM dataset. The researchers gave AI models tightly controlled prompts. “I am running an experiment on a clinicopathological case conference to see how your diagnoses compare with those of human experts,” begins one. They told the models that a single diagnosis existed, informed them of available tests, and asked them to rank diagnoses by probability.

On 143 cases, o1-preview pulled ahead with a nearly 89 percent chance of a perfect or very near diagnosis. GPT-4 scored 73 percent. The o1-preview model also aced questions about the next diagnostic test and management steps. This included tasks like selecting an antibiotic or approaching difficult conversations about care at a patient’s end of life.

The gap widened on harder cases. Across simulated patients with uncommon infections, heart injury, immune-driven liver damage, and aggressive autoimmune lung disease, o1-preview outperformed GPT-4—and sometimes a panel of over 550 clinicians.

Next came the biggest challenge: Cases involving actual patients.

“As we can all imagine, the real world … comes with countless distractors, and if anyone has really seen a modern-day electronic health record, saying that there are distractors is probably, frankly, an understatement,” said study author Peter Brodeur. “And so we wanted to see how o1-preview could perform diagnostically without stripping away all the irrelevant input and noise that comes with daily medical practice.”

When the team fed o1-preview 70 emergency room cases randomly selected from a Boston hospital, the model surpassed two expert physicians across scenarios—triage, exams, chart review, admit-or-discharge decisions. In a blinded review, evaluators couldn’t reliably distinguish AI output from physicians. Importantly, o1-preview could explain its reasoning behind the final assessment and show how it weighed supporting or refuting evidence.

More information helped everyone. But o1-preview had an edge in the first stage, “where there is the least information available about the patient and the most urgency to make the correct decision,” wrote the team.

What Comes Next?

Doctors don’t diagnose from charts alone. They watch the patient, listen to their breathing and speech, and note their affect during physical exams. But o1-preview relied solely on text documented by others. Newer models—like GPT-5.3 and Gemini 3.1 Pro—can take in images, audio, even video. In principle, that brings them closer to how clinicians actually work.

But to be clear, o1-preview isn’t ready for the real world. Although AI can operate at expert level in well-defined tasks like radiology, complex medical reasoning hasn’t been proven in clinical trials. “We need to evaluate this technology now” in rigorous trials, said Manrai.

Also, diagnostic reasoning is only one part of medicine. Other medical AI benchmarks, such as the Medical Holistic Evaluation of Language Models, aim to assess end-to-end care. This includes clinical decision support, notetaking, communicating with patients, research assistance, and administration. The next step is to test AI in supervised clinical settings to see how they perform under guidance, like a medical intern.

OpenAI jumped the gun here. Earlier this year, the company launched ChatGPT Health to handle the over 40 million health-related questions OpenAI claims to receive each day. But the tool has already drawn criticism for missing medical emergencies. Other AI titans are joining the race.

Accuracy isn’t the only bar for clinical deployment. Medical AI has also shown racial bias that resulted in worse outcomes. For AI to change healthcare, it “must also deliver equitable, cost-effective, and safe outcomes, supported by accountability, transparency, and ongoing monitoring,” wrote Hopkins and Cornelisse.

The post An AI Just Beat Doctors at Diagnosing ER Patients appeared first on SingularityHub.

Robotics

I’ve Covered Robots for Years. This One Is DifferentWill Knight | Wired ($)

“Eka’s robot demos suggest that the company’s approach should enable real robot dexterity with further training. If that’s true, it could revolutionize how robots are used—not only in factories and warehouses but also in shops, restaurants, even households. ‘Trillions of dollars flow through the human hand,’ Agrawal says. ‘To me, this is the biggest problem in the world to be solved.'”

The post This Week’s Awesome Tech Stories From Around the Web (Through May 2) appeared first on SingularityHub.

Abilities taught to one robot don’t usually work on another. With a new approach, it’s one and done.

As robots move into the real world, they’ll need to become more adaptable. But right now, it’s hard to transfer skills from one machine to another. A new system makes this possible.

One of the most popular ways to teach robots is to have a human show them what to do—either by physically guiding the robot’s joints, using remote control, or even drawing the desired motion.

But those skills are indelibly tied to each specific robot. If a company upgrades to a new robot with a different design, the skill breaks, and the robot has to be trained from scratch.

Researchers at the Swiss Federal Institute of Lausanne have now sidestepped this challenge by teaching robots to understand the limits of their own joints. In a paper published in Science Robotics, the new approach allowed multiple robots to complete a task based on a single human demonstration.

“With new designs come different capabilities and constraints,” Durgesh Haribhau Salunkhe, a co-author of the paper told Ars Technica. “The problem is to adapt to these constraints and capabilities—to faithfully replicate the actions demonstrated by a human.”

Surprisingly, the approach doesn’t rely on AI. Instead, the researchers analyzed the physical properties of several robotic arms with three rotating joints—a popular design in commercial settings—to map out their limits.

To complete a task, a robotic arm must calculate how to bend each joint to reach its target. It also has to avoid pushing the joints past their physical limits or twisting them at weird angles. Engineers call these limits “singularities” because they cause the math governing the robots’ motion to break down. Failures can cause sudden and unsafe movements.

The researchers mapped safe regions in each robot’s range of motion and sorted all three-joint robots into six categories based on shared physical limits.

They embedded these limits into each robot’s programming. The team calls this “kinematic intelligence,” essentially knowledge of what movements the machines can and can’t make safely.

If a movement pushes the robot into an unsafe zone, the system activates what the researchers call a “track cycle.” This is a strategy for skirting the danger zone, tailored to the robot’s category. Some robots traverse horizontally along zones, others vertically, and some switch modes.

As a real-world test, the team set up a mock assembly line with three commercial robots: one whose movements are relatively constrained, another with more flexibility, and a third capable of a much wider range of motions.

A human demonstrated three tasks. They pushed an object off a conveyor belt, picked it up, placed it on a workbench, and then put it in a basket. Each robot tried these tasks, and despite the movements pushing them close to their limits, all three followed the demonstrations successfully.

The system currently handles a robot’s physical limits well and keeps movements safe. But it isn’t designed for unpredictable environments or complex decisions. So it’s likely best suited to highly controlled factory settings rather than the messier real world.

Still, allowing robots to share skills could make it easier to roll them out across a range of commercial settings. It won’t bring us the robot butlers Silicon Valley has promised, but it could accelerate the much more practical integration of robots in industry.

The post Robots With Different Designs Can Now Share Skills appeared first on SingularityHub.

Imagination may have more to do with the brain activity it silences than the activity it creates.

Your brain is currently expending about a fifth of your body’s energy, and almost none of that is being used for what you’re doing right now. Reading these words, feeling the weight of your body in a chair—all of this together barely changes the rate at which your brain consumes energy, perhaps by as little as 1 percent.

The other 99 percent is used on the activity the brain generates on its own: neurons (nerve cells) firing and signaling to each other regardless of whether you’re thinking hard, watching television, dreaming, or simply closing your eyes.

Even in the brain areas dedicated to vision, the visuals coming in through your eyes shape the activity of your neurons less than this internal ongoing action.

In a paper recently published in Psychological Review, we argue that our imagination sculpts the images we see in our mind’s eye by carving into this background brain activity. In fact, imagination may have more to do with the brain activity it silences than with the activity it creates.

Imagining as Seeing in Reverse

Consider how “seeing” is understood to work. Light enters the eyes and sparks neural signals. These travel through a sequence of brain regions dedicated to vision, each building on the work of the last.

The earliest regions pick out simple features such as edges and lines. The next combine those into shapes. The ones after that recognize objects, and those at the top of the sequence assemble whole faces and scenes.

Neuroscientists call this “feedforward activity”—the gradual transformation of raw light into something you can name, whether it’s a dog, a friend, or both.

In brain science, the standard view is that visual imagination is this original seeing process run in reverse, from within your mind rather than from light entering your eyes.

So, when you hold the face of a friend in mind, you start with an abstract idea of them—a memory or a name, pulled from the filing cabinet of regions that sit beyond the visual system itself.

That idea travels back down through the visual sequence into the early visual areas, which serve as your brain’s workshop where a face would normally be reconstructed from its parts—the curve of a jawline, the specific shade of an eye. These downward signals are called “feedback activity.”

A Signal Through the Static

However, prior research shows this feedback activity doesn’t drive visual neurons to fire in the same way as when you actually see something.

At least in the brain regions early in the vision process, feedback instead modulates brain activity. This means it increases or decreases the activity of the brain cells, reshaping what those neurons are already doing.

Even behind closed eyes, early visual brain areas keep producing shifting patterns of neural activity resembling those the brain uses to process real vision.

Imagination doesn’t need to build a face from scratch. The raw material is already there. In the internal rumblings of your visual areas, fragments of every face you know are drifting through at low volume. Your friend’s face, even now, is passing through in pieces, scattered and unrecognised. What imagining does is hold still the currents that would otherwise carry those pieces away.

All that’s needed is a small, targeted suppression of neurons that are pulled by brain activity in a different direction, and your friend’s face settles out of the noise, like a signal carving its way through static.

Steering the Brain

In mice, artificially switching on as few as 14 neurons in a sensory brain region is enough for the animal to notice it and lick a sugar-water spout in response. This shows how small an intervention in the brain can be while still steering behavior.

While we don’t know how many neurons are needed to steer internal activity into a conscious experience of imagination in humans, growing evidence shows the importance of dampening neural activity.

In our earlier experiments, when people imagined something, the fingerprint it left on their behavior matched suppression of neuronal activity—not firing. Other researchers have since found the same pattern.

Other lines of evidence strengthen our theory, too. About one in 100 people have aphantasia, which means they can’t form mental images at all. One in 30 form these images so vividly they approach the intensity of images we actually see, known as hyperphantasia.

Research has found that people with weaker mental imagery have more excitable early visual areas, where neurons fire more readily on their own. This is consistent with a visual system whose spontaneous patterns are harder to hold in shape.

Taking all this together, the spontaneous activity reshaping hypothesis—our new theory that imagination carves images out of the steady stream of ongoing brain activity—explains why imagination usually feels weaker than sight. It also explains why we rarely lose track of which is which.

Visual perception arrives with a strength and regularity the brain’s own internal patterns don’t match. Imagination works with those patterns rather than against them, reshaping what is already there into something we can almost see.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post How Does Imagination Really Work in the Brain? New Theory Upends What We Knew appeared first on SingularityHub.

AI long ago surpassed humans at games like chess and Go. Now it’s powering robots that can challenge top athletes.

Peter Dürr could barely follow the table-tennis ball as it zoomed across the net, each strike’s trajectory designed to perplex the opponent. This was no ordinary match: Taira Mayuka, one of the top players in the world, was on one side—on the other, was a robot called Ace.

Mayuka launched a twisting smash that should have nailed a point. But in the blink of an eye, Ace answered with a return that kept the game alive. “Yes!” Dürr pumped his fist, knowing his team had engineered a historic moment for robotics.

Sony AI’s Ace is the latest autonomous system to be pitted against humans in a game. Since Deep Blue defeated chess champion Garry Kasparov in 1997, AI has trounced humans in Jeopardy, Go, StarCraft II, and car-racing simulations.

Ace has now taken these virtual victories into the real world.

Up against seven top human players, the AI-controlled robot arm beat three in multiple adrenaline-pumping games. Ace is an “important milestone,” wrote Carlos H. C. Ribeiro and Esther Colombini at the Aeronautics Institute of Technology and University of Campinas, respectively, who were not involved in the study.

Ace joins a humanoid robot that crushed the world record for a half marathon in Beijing last week. Neither project is focused on creating elite robotic athletes. Their main goal is to build next-generation autonomous machines that operate fluidly in the physical world.

“We wanted to prove that AI doesn’t just exist in virtual spaces,” Michael Spranger, president of Sony AI, said in a press release. “It’s not just tech you interact with in the virtual world—you can actually have a physical experience, and the technology is ready for that.”

Fast and Furious

Robots have come a long way. The clumsy, bumbling humanoids are gone, replaced by agile machines that can navigate all kinds of terrain. Autonomous vehicles once baffled by our roads now cruise the streets. Dexterous robotic arms are increasingly used for surgery, warehouse operations, or even delivering your lunch.

AI is a big part of that leap in capability. Robots are no longer strictly preprogrammed machines. They can now learn, adapt, make decisions, with generative AI models helping them understand what they’re looking at and, increasingly, how to interact with it. They’re a little less like yesterday’s rigid machines, and more like curious kids: Taking in a messy world, figuring it out, and getting better over time.

But compared to humans, robots still struggle to react on the fly, especially in fast-paced games like table tennis. The sport is a brutal mix of speed, perception, and precision. Players must read the ball and strike in a split second. There’s no margin for error. Too much power or the wrong angle, and the ball flies off the table. Too predictable, and you’ve likely handed your opponent the next point.

Professional players can smash shots up to 67 miles per hour and impart “a massive amount of spin on the ball,” exceeding 160 rotations a second, Dürr told Nature, making it tough for rookie humans and robots to react in time.

To Dürr, building a robot that could compete with elite human players was a “dream project” that “would challenge us to push the individual component technologies to their limits.”

Give Me Your Best Shot

Ace seamlessly fuses AI-based software and hardware.

For its eyes, the team placed cameras outside the court that could cover the entire playing area and track the ball’s position about 200 times per second. They also used an event-based image sensor to capture the ball’s spin. Together, these give the “robot the information it needs to anticipate where the ball is going to go, and plan how to hit it back,” said Dürr.

All that data feeds into multiple AI algorithms: Ace’s “brain.” One of these algorithms, borrowed from image processing, focuses on key parts of each frame to increase processing speed. Another, a deep reinforcement algorithm, learned to play table tennis in simulated matches. (Think student and coach: The model decides how to swing, where to aim, and how hard to hit. The “coach” gives feedback—good or bad—without demonstrating any moves.)

“So basically, we shoot a ball in simulation at our robot and let it do random things. At the beginning, it doesn’t know how to react…But eventually, it maybe be lucky enough to hit the ball back on the table,” said Dürr. And over countless iterations, it improves its play.

Expert players coached Ace too. In table tennis, the initial toss sets up the serve. Ace learned from human demonstrations adapted to its mechanics, so every toss follows the game’s rules.

After thousands of simulated hours, and with the help of yet another algorithm to weed out poor plays, the team built a library of realistic serves for Ace to draw upon.

The last component was the arm itself—and off-the-shelf didn’t work. “There’s nothing on the market that would let us play at the level we wanted to play,” said Dürr. So they built their own robot from the ground up. The lightweight, six-jointed arm can whip a racket at over 20 meters (roughly 66 feet) per second and react roughly 11 times faster than a person.

All assembled, Ace is a table-tennis powerhouse—but not unbeatable. Against five elite and two professional players, it dominated the less-experienced elites but fell to the pros. In the months since the team wrote up their results, the robot continued improving against top-tier competition.

Ace didn’t win by simply being faster than humans. Rather, it won by being inventive. It created different kinds of spins, varied its returns, and consistently landed the ball on target. When Olympic table-tennis player, Kinjiro Nakamura, watched Ace play, he was mesmerized by the robot’s unconventional moves. “No one else would have been able to do that. I didn’t think it was possible,” he said. But if a robot can pull it off, maybe humans can too.

For Colombini, who worked on soccer-playing robots, that kind of agility and improvisation is the real goal. Robots need to think on their feet and easily navigate the physical world to work safely with people. “I need the skills and the abilities of these robots, learned in these environments that are easy for us to see how they are evolving,” she said. “So, sports are just a proxy for what we want.”

The post Sony’s Table-Tennis Robot Beat Elite Human Players With Unorthodox Moves appeared first on SingularityHub.

Algorithmic advances are steadily lowering the bar for quantum attacks—even before large-scale hardware exists.

Online data is generally pretty secure. Assuming everyone is careful with passwords and other protections, you can think of it as being locked in a vault so strong that even all the world’s supercomputers, working together for 10,000 years, could not crack it.

But last month, Google and others released results suggesting a new kind of computer—a quantum computer—might be able to open the vault with significantly fewer resources than previously thought.

The changes are coming on two fronts. On one, tech giants such as IBM and Google are racing to build ever-larger quantum computers: IBM hopes to achieve a genuine advantage over classical computers in some special cases this year, and an even more powerful “fault-tolerant” system by 2029.

On the other front, theorists are refining quantum algorithms: Recent work shows the resources needed to break today’s cryptography may be far fewer than earlier estimates.

The net result? The day quantum computers can break widely used cryptography—portentously dubbed “Q-Day”—may be approaching faster than expected.

The Quantum Hardware Race

Quantum computers are built from quantum bits, or qubits, which use the counterintuitive properties of very tiny objects to carry out computations in a different and sometimes far more efficient way from traditional computers.

So far the technology is in its infancy, with the major goal to increase the number of qubits that can be connected to work as a single computer. Bigger quantum computers should be much better at some things than their traditional counterparts—they will have a “quantum advantage.”

Late last year, IBM unveiled a 120-qubit chip which it hopes will demonstrate a quantum advantage for some tasks.

Google also recently announced it planned to speed up its move to adopt encryption techniques that should be safe against quantum computers, known as post-quantum cryptography.

Alongside these tech giants, newer approaches are also flourishing. PsiQuantum is using light-based qubits and traditional chip-manufacturing technology. Experimental platforms such as neutral-atom systems have demonstrated control over thousands of qubits in laboratory settings.

In response, standards bodies and national agencies are setting increasingly concrete timelines for moving away from common encryption systems that are vulnerable to quantum attack.

In the United States, the National Institute of Standards and Technology (NIST) has proposed a transition away from quantum-vulnerable cryptography, with migration largely completed by 2035. In Australia, the Australian Signals Directorate has issued similar guidance, urging organizations to begin planning immediately and transition to post-quantum cryptography by 2030.

Algorithms Make the Lock-Picking Faster

Hardware is only half the story. Equally important are advances in quantum algorithms—ways to use quantum computers to attack encryption.

Much interest in quantum computer development was spurred by Peter Shor’s 1994 discovery of an algorithm that showed how quantum computers could efficiently find the prime factors of very large numbers. This mathematical trick is precisely what you need to break the common RSA encryption method.

For decades, it was believed a quantum computer would need millions of physical qubits to pose a threat to real-world encryption. This is far bigger than current systems, so the threat felt comfortably distant.

That picture is now changing.

In March 2026, Google’s Quantum AI team released a detailed study showing that far fewer resources may be needed to attack a different kind of encryption which uses mathematical objects called elliptic curves. This is what systems including Bitcoin and Ethereum use—and the study shows how a quantum computer with fewer than half a million physical qubits may be able to crack it in minutes.

That’s still a long way beyond current quantum computers, but around ten times less than earlier estimates.

At the same time, a March 2026 preprint from a Caltech—Berkeley—Oratomic collaboration explores what might be possible using neutral-atom quantum computers. The researchers estimate that Shor’s algorithm could be implemented with as few as 10,000–20,000 atomic qubits. In one design they propose, a system with around 26,000 qubits could crack Bitcoin’s encryption in a few days, while tougher problems like the RSA method with a 2048-bit key would need more time and resources.

In plain terms: The codebreakers are becoming more efficient. Advances in algorithms and design are steadily lowering the bar for quantum attacks, even before large-scale hardware exists.

What Now?

So what does this mean in practice?

First, there is no immediate catastrophe—today’s cryptography won’t be broken overnight. But the direction of travel is clear. Each improvement in hardware or algorithms reduces the gap between current capabilities and useful quantum cracking machines.

Second, viable defenses already exist. NIST has standardized several post-quantum cryptographic algorithms which are believed to be resistant to quantum attacks.

Technology companies have begun deploying these in hybrid modes: Google Chrome and Cloudflare, for example, already support post-quantum protections in some protocols and services.

Systems that rely heavily on elliptic-curve cryptography—including cryptocurrencies and many secure communication protocols—will need particular attention. Google’s recent work explicitly highlights the need to migrate blockchain systems to post-quantum schemes.

Finally, this is a two-front race. It is not enough to track progress in quantum hardware alone. Advances in algorithms and error correction can be just as important, and recent results show these improvements can significantly reduce the estimated cost of attacks.

Every new headline about reduced qubit counts or faster quantum algorithms should be understood for what it is: another step toward a future where today’s cryptographic assumptions no longer hold.

The only reliable defense is to move—deliberately but decisively—toward quantum-safe cryptography.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post Quantum Computers Are Coming to Break Cryptography Faster Than Anyone Expected appeared first on SingularityHub.