Switches drive nearly every machine. A new one, made of folded DNA, does the same work at the scale of molecules.

Scientists have long dreamed of developing nanoscale machines, but building reliable mechanical components at the molecular scale has proved challenging. Researchers have now developed a DNA-based switch that can rapidly and repeatedly snap between two stable states, much like the components that underpin everyday electronics.

Ever since Richard Feynman’s visionary lecture “There’s Plenty of Room at the Bottom,” researchers have been enamored with the idea of engineering at the scale of atoms and molecules. But manipulating matter at the nanoscale is easier said than done.

Individual molecules are in constant motion and continuously jostled about by the thermal energy of their surroundings. This makes it extremely difficult to position and assemble larger structures and undermines control of the mechanical motion of components.

This is particularly true for switches—key components in many mechanical and electronic devices you might want to build. Getting a tiny structure to hold one position, flip cleanly to another, and then stay there has so far been an unsolved problem.

But now, a team at the Technical University of Munich has created a switch made from folded strands of DNA that remains stable for up to an hour and flips in milliseconds on the application of a brief electric field. Crucially, the device was able to switch back and forth repeatedly with no degradation in performance.

“Individual devices sustain hundreds of thousands of switching cycles over several hours and remain functional for actuation over several days,” the researchers write in a paper in Science Robotics. “As a nanoscale electromechanical interface, our device enables applications in molecular information processing, optical nanodevices, and the dynamic control of chemical reactions.”

The device borrows a principle from standard engineering known as a snap-through mechanism, which rests in either of two states and only flips when pushed hard enough, a bit like a light switch.

Scaling the idea down to a few tens of nanometers meant designing rigid arms linked by flexible molecular hinges, so the structure settles into one of two configurations and does not flick between them on its own. The team relied on DNA origami to accomplish this, where a long strand of DNA is folded into custom 2D and 3D shapes using hundreds of shorter “staple” strands.

One of the two arms features a longer “extension arm” that acts as a lever to push the switch between configurations. DNA carries negative charge, so when an electric field is applied to the device, it pushes the arm hard enough to flip the switch. Left alone, the team estimates that the structure stays in its resting state for roughly six hours, and they observed no spontaneous flips while monitoring 70 switches for an hour.

One of the device’s main strengths is its endurance. One switch survived more than 200,000 flips over five and a half hours, and a simplified version withstood a million switching cycles in three hours while still working about 85 percent of the time. Performance varied considerably from one device to the next, however, with some failing after a few thousand cycles and others continuing for days.

The researchers say failures likely stem from a combination of contaminants, surface wear, and chemical changes in the surrounding fluid. However, some inactive switches later started working again, which the team says suggests they are capable of self-repairing.

To test whether the switch could do anything useful, the researchers attached a gold nanorod to the moving arm, turning it into a microscopic light switch that changed how light scattered off the particle. In a second test, they used the switch to expose or hide a molecular binding site, allowing it to control whether DNA strands could attach.

That second capability could be particularly useful as it could make it possible to control chemical reactions—for instance by turning enzymes on and off. The authors suggest that this could be used to create “control knobs” for chip-based bio-factories that run sequences of reactions.

Considerable obstacles remain before the device can become genuinely useful. A single switch encodes just one bit of information, and the team acknowledges that wiring arrays of switches together to create something resembling a circuit remains a distant prospect.

But a workable switch is a fundamental component that can be used to create all manner of devices. While we’re still a long way from Feynman’s dream of molecular machines, this is a meaningful step in that direction.

The post This DNA Switch Could Control Molecular Machines appeared first on SingularityHub.

Let loose on existing regulations, AI models sniffed out known loopholes—and exposed entirely new ones too.

AI’s hacking skills are big news at the moment, but finding vulnerabilities in code may be the least of our worries. A new study suggests AI models can discover potentially damaging loopholes in the rules and regulations underpinning society.

Modern AI systems are powerful optimizers. Give them a goal, and they’ll pursue it relentlessly, quickly discovering solutions that would take a human years to find. But they are also incredibly literal in the way they approach a problem. They will do exactly what you tell them and are incapable of reading between the lines in the ways a human would.

This tendency leads to a recurring problem known as “reward hacking,” where an AI finds some loophole to maximize its performance on the metric used to measure success without actually achieving what its designers intended. The classic example is the AI that discovered it could win a boat racing videogame by looping around in circles collecting power-ups rather than completing the course.

The problem is partly due to humans being bad at specifying their goals. And unfortunately, it seems this weakness exists in the rules and regulations used to run society. When researchers let popular large language models loose in 72 simulated regulatory environments, the models found 60 percent of known loopholes and even identified some entirely new exploits.

“Within these environments, reward hacking naturally emerges and leads to regulatory loophole discovery,” the authors write in a non-peer-reviewed paper published on arXiv. “Models learn to hack the social rules and generate strategies that remain technically compliant while defeating regulatory intent.”

The regulatory environments the researchers created were primarily based on rules governing things like pharmaceutical patents, NBA salary caps, and deep-sea mining. In each case, Alibaba’s Qwen3 model was given the relevant rules, an explanation of its task, a predefined set of actions it could take, and the system used to score different outcomes.

A more powerful model, Google’s Gemini-3-flash, then simulated the consequences of different actions Qwen3 took and judged if and when it had found a way to exploit the rules of the game. When that occurred, the larger model patched the loophole by adding new rules, and the smaller model was set loose again. Over many iterations, the models to discover increasingly subtle workarounds.

When building their regulatory environments, the researchers omitted real-world fixes that regulators had used to close known loopholes. Over many trials, Qwen3 rediscovered more than 60 percent of these exploits. In a simulation of pharmaceutical patent regulations, the two models ended up replaying the same sequence of loophole discovery and regulatory reform that occurred in the real world.

Crucially, their behavior emerged spontaneously without the researchers asking the algorithms to cheat the system. This is a byproduct of the popular reinforcement learning approach the researchers used, where a model is rewarded for getting closer to a specific, numerically-defined goal.

Worryingly, the team found that existing safety measures offered little protection. Both models are designed to refuse prompts featuring harmful language, but loophole-seeking behavior slipped under the radar. When asked to self-critique their own behavior, the models identified fewer than 40 percent of their own exploits.

The researchers note that the same capabilities could be used more proactively to scour proposed regulations for loopholes before enactment. But lead author Wei Liu, a PhD student at King’s College London, says there are always likely to be gaps. “In the real world,” he told Science, “society is a huge, complicated reward function that can’t ever be patched to a perfect status.”

Adding to the concern, the models used in this study were far from the frontier, suggesting that more powerful AI could be even more adept at regulatory hacking. Whether our existing institutions can adapt quickly enough to this emerging threat is an open question.

The post Forget Code: AI Is Learning to Hack Society appeared first on SingularityHub.

Computing

IBM Has Unveiled Chip Technology That Could Help Extend Moore’s Law Another DecadeSophia Chen | MIT Technology Review ($)

“To fit more transistors on a chip, engineers across the industry are eyeing a pivot to an approach familiar to urban planners: build up. On Thursday, IBM announced it has created a chip that uses this strategy. The new architecture, known as a nanostack, vertically stacks transistors in two layers on a silicon chip.”

The post This Week’s Awesome Tech Stories From Around the Web (Through June 27) appeared first on SingularityHub.

There’s a vast difference between launching satellites and operating an industrial-scale computing infrastructure in orbit.

Imagine if one company could become the railroad, electric utility, and cloud-computing provider of the emerging space economy. That potential fueled excitement around the long-anticipated initial public offering of SpaceX. Investors are not simply betting on rockets anymore. They are betting on an entire orbital ecosystem.

Among the most ambitious and challenging ideas riding this wave of enthusiasm is something that sounds almost like science fiction: orbital data centers. SpaceX may be one of the most well-known companies seeking to build them, but it is not the only one.

The logic is seductive: Launch the data centers into orbit, where solar energy is abundant and land, water, and local power grids are no longer constraints. As artificial intelligence drives an explosion in computing demand, companies are pitching orbital data centers as a way to escape the growing environmental and infrastructure pressures of Earth-based computing. Data centers often also face backlash from the public at having these centers located in their communities.

But there is a vast difference between launching satellites and operating an industrial-scale computing infrastructure in orbit. Space is unforgiving. Radiation damages electronics. The electronics generate enormous amounts of heat, and getting rid of that heat is surprisingly difficult in space. Repairs are extraordinarily expensive, and every pound launched into orbit still carries a significant cost.

We are engineering professors who study data-center design and space systems engineering. Building a space-based data center will involve considerations from both sides.

What Goes Into a Data Center on Earth

First off, consider what goes into an Earth-based data center, like those that you’ve probably begun to see pop up everywhere. These facilities power cloud computing, video streaming, online banking, scientific computing, and increasingly, artificial intelligence. But a data center is much more than a room full of servers.

A data center needs several things to operate reliably. The first is electric power. Servers, networking equipment, and storage devices consume large amounts of electricity, and that power demand is growing rapidly with AI.

The second is cooling. Almost all the electricity consumed by servers eventually becomes heat. If that heat is not removed quickly and reliably, equipment performance drops, failures increase, and the data center can shut down. Cooling systems often include air handling units, chillers, cooling towers, pumps, and increasingly, liquid-cooling equipment. In many facilities, cooling is the largest energy consumer after the computing equipment itself.

The third is physical infrastructure, including the necessary land, buildings, structural support, backup power, water systems, communication networks, and maintenance access. Data centers also need to be close enough to users and network backbones to provide fast digital services.

In short, Earth-based data centers are large electrical and thermal infrastructure systems built around computing hardware.

Placing Them in Space

So what would it take to build these data centers in space, and why are companies finding this possibility such an interesting business proposition?

As on Earth, these data centers would require massive amounts of power. In space, this power would come from solar panels. The sun always shines in space and can’t be blocked by clouds. However, depending on the orbit the solar panels are put in, the Earth may shadow them for some portion of the orbit.

And even the best solar cells available today can convert only about half the sunlight that hits them to electricity.

Another potential advantage found in space is cooling. The cold background of space (roughly -455 degrees Fahrenheit, or -270 degrees Celsius) creates an opportunity: Waste heat from the data center could escape into space through radiators, keeping the electronics cool.

In principle, that design could eliminate some of the bulky and water-intensive cooling infrastructure used on Earth. However, those thermal radiators would require a large amount of surface area, and that would be in addition to the area required by the solar panels.

In space, there is no air to blow across hot equipment and help heat escape. The heat has to leave as infrared radiation, which is a relatively slow process. As a result, removing 10 megawatts of waste heat can require radiator surfaces comparable to the size of two football fields.

Space-based data centers could also avoid some of the local conflicts that come with building large data centers on the ground. Many communities resist new data center developments because of their land use, energy and water demand, and noise and environmental impact.

A space-based system would avoid competing for local land and water resources, and it would not generate neighborhood noise or require local zoning approval in the same way.

However, space is already getting crowded, and launching thousands of large orbital data centers would accelerate this issue. Orbital debris and micrometeorites are hazards because they can puncture the space data center, and a worst-case collision could destroy it and create even more space debris.

The frequency of space launches necessary to send all the equipment to orbit may also become a concern for some communities. SpaceX has had protests at its launch complex in Boca Chica, Texas from local activists who argue its rocket testing and launches damage the surrounding environment.

All that data would need to be sent between Earth and these data centers—and between the data centers themselves—using radio waves or laser communications systems. Although satellite constellations such as Starlink and Amazon Leo have demonstrated that doing this is possible, the amount of data sent to and from space would balloon.

Additional Challenges

These data centers, along with their solar panels and radiators, cannot be launched in one piece and would need to be assembled in space. This process would require new equipment for in-space servicing, assembly, and manufacturing.

Another key challenge is the refresh cycle of computing hardware. Data-center servers are not built to last forever. Operators on Earth usually replace or upgrade hardware every three to five years as chips improve, workloads change, and equipment ages.

And equipment failures can require replacing components. The refresh and repair processes are relatively straightforward on Earth, where workers can physically remove and replace servers.

In space, refresh and repair becomes much harder. Hardware sent to orbit may be difficult or too expensive to upgrade. If the computing platform cannot be updated, or too many components fail, it may become obsolete long before the surrounding infrastructure reaches the end of its useful life.

In a field where performance improves so rapidly and demand from computing continues to increase, this hurdle could prove a major economic and operational challenge.

Then there is the harshness of space. These data centers would be in a near vacuum, with constant radiation hitting them. And depending on their orbit, they would go from hot when in the sunlight to cold in Earth’s shadow many times a day. All of these challenges, and more, are issues that will need to be addressed.

So, Do They Still Make Sense?

Despite these challenges, companies are moving forward with designing space-based data centers. SpaceX just announced the design for its AI1 Compute Satellite, which it hopes to use as an orbital data center spacecraft. However, this satellite is 100 to 1,000 times less capable than current Earth-based data centers.

Not every computing task makes sense to do in space. Many data center applications depend on fast response times and close connections to users on Earth. Financial transactions, interactive AI services, and most cloud applications are extremely sensitive to delay.

More feasible early applications may be those that are less latency-sensitive and more tightly connected to space operations. Examples could include processing Earth observation data from satellites, military or intelligence data processing, scientific computing related to space missions, or specialized computing for satellites and other space assets.

In other words, the first viable space data centers may serve space-based customers before they compete with mainstream cloud data centers on Earth.

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

The post Orbital Data Centers Are Seductive on Paper, but They Face Daunting Challenges in Reality appeared first on SingularityHub.

Companies once moved whole factories overseas to reduce labor costs. Now, workers a world away can operate local excavators, forklifts, and even humanoid robots with an internet connection.

Packaging potassium sulfate, a fertilizer vital to the planet’s food supply, is visually striking—not because of what you see, but because you don’t see much at all. In China’s Xinjiang region, home to the world’s largest deposit of the mineral, piling it up in warehouses creates dust clouds so severe that workers are forced to drive heavy machinery by feel.

Some companies are now turning to a technology that not only offers a way to see through the dust but also keeps workers from entering the warehouse at all. The system, developed by BuilderX Robotics, a Chinese tech company, uses cameras that are like night-vision for dusty areas. More significantly, operators drive excavators, loaders, and other machines from a remote office filled with rows of videogame-like stations. All they need is a 5G or satellite connection.

The ability to control physical machines from a distance is called teleoperation, and it could become a significant force of change in the global economy.

In Japan, the shelves of over 300 convenience stores are being restocked by robots monitored and sometimes controlled by workers in the Philippines. Düsseldorf airport was slated to begin testing shuttles driven by remote workers in May. A startup in Atlanta is offering robot security guards operated by remote staff, and last summer, a surgeon in France performed a teleoperated procedure on a patient in India.

While offshoring teleoperated jobs to overseas workers hasn’t yet become routine, Mark Graham, professor of internet geography at the University of Oxford, suggests the technology is worth our attention because it might enable companies to expand on their well-established habit of outsourcing jobs to places where labor is cheaper.

The use of remote labor isn’t new, Graham told SingularityHub. But teleoperation extends the logic of outsourcing to tasks that were previously thought to be “stubbornly local.”

“The novelty is less about the existence of remote labor and more about the kinds of work that can now be pulled into a planetary labor market,” he said. “Once that happens you can expect the usual pressures around labor arbitrage, control, and fragmentation to follow.”

It’s not clear we’re ready for the consequences.

BuilderX Robotics is a global leader in teleoperation for heavy machinery and a good expression of the changes ahead. Shaolong Sui, a graduate of Stanford University with a degree in mechanical engineering, founded the company in 2018 as a response to labor shortages in the construction industry in Asia.

“A shortage of trained operators isn’t a problem only in developed countries,” he told me. “Young people here in China don’t want to do this work. It’s dusty and dangerous.”

Rather than focusing on full robotic autonomy, which many construction companies have pursued over the past decade, Sui identified teleoperation as a more realistic way to move operators from harsh environments to safer conditions. Making use of the proliferation of low-cost sensors and 5G at the time, Sui completed a prototype in 2019. Today, his company offers teleoperation for 14 different industrial machines, including excavators, loaders, and bull dozers.

In our conversation, it was clear he hopes to improve working conditions for manual laborers. I lost track of the number of times he mentioned removing operators from dangerous worksites. “These workers deserve a better life,” he said.

BuilderX’s workstations do seem to have transformed some of the punishing work of an industrial site into a more white-collar experience, complete with tea and coffee break rooms and toilets down the hall. Sui said his solution allows construction firms to hire senior citizens or people with disabilities who, thanks to the videogame-like interface, can now operate heavy machinery. In another video, a Japanese woman who pilots an excavator proudly shows off her complex nail art, something she claims she couldn’t maintain when she worked in the field.

“Not only is this a much safer workplace, but the lifestyle benefits are that you can sit in an air-conditioned space, enjoy your tea, and when you go home, you’re still clean,” Sui said.

There’s no doubt the approach is safer for frontline workers like those in Xinjiang. Evidence suggests that high levels of potassium dust exposure can cause chronic bronchitis. While pulling someone from dangerous work is a good thing and that should be taken seriously, Graham told me, it doesn’t necessarily mean they’re free from exploitation.

“A worker can be removed from the physical site and still be subjected to intense surveillance, deskilling, isolation, fragmented contracts, algorithmic management, and downward pressure on wages. In other words, the risk can move rather than disappear,” he said.

Sui and Graham both agree there are plenty of forces that might slow the pace of outsourcing. Currently, none of BuilderX’s customers offshore work to overseas operators. But that doesn’t appear to be a technology constraint, as recently demonstrated by an operator in Poland controlling an excavator over 4,000 miles away in Beijing. On the technical side, latency—the delay between operator and machine—and reliability will shape the rate at which firms can choose to offshore workers. But it’s more likely to be limited by regulatory constraints in the form of licensing, insurance, and safety requirements.

That said, Graham believes the biggest force driving work overseas will be the same one that’s pushed clerical and service work offshore; the relentless pursuit to increase profit and reduce cost.

“If firms can hire people in lower-wage labor markets to operate expensive equipment thousands of miles away, many of them will try,” he said.

Most debates about AI and robotics focus on job loss due to automation. There is relatively little discussion about the risk of offshoring teleoperated work as the technology comes online. This is partly due to the hype surrounding physical AI, a Silicon Valley buzzword describing a world where fully autonomous robots cut humans out of the loop. But Graham says that when machines arrive people tend to incorrectly assume humans disappear.

“In many cases, what gets described as automation is really a reorganization of labor. Work gets broken apart, moved around, and hidden from view,” he says.

As is the case with AI,  the robotics industry’s push toward full automation is still plenty reliant on a hidden system of faraway workers. Teleoperation provides training data for robots and is needed to help them deal with unexpected events. Consumer robotics startup 1X is selling a $20,000 humanoid that will sometimes need to be  controlled by remote staff. It’s not clear how often future robots cleaning dishes in San Francisco kitchens will be steered by gig workers in Mumbai.

Robotaxi company Waymo already relies on human agents to assist, though not literally drive, vehicles stuck in difficult scenarios. The firm recently disclosed for the first time that some of these agents are based in the Philippines. This information, surfaced during US congressional testimony, immediately raised questions of oversight for safety-critical work: For instance, should a worker in Manila be required to get a California driver’s license?

Amid an already combustible US political environment, teleoperation could raise the heat even higher. Fueled by fears of Americans losing jobs to people overseas, Wyndham Hotels and Resorts, the parent company of La Quinta, was last year forced to respond to anger over a viral video depicting workers allegedly in India remotely handling check-in at one of their Miami hotels. As Graham points out, people tend to care more about outsourcing when it’s no longer hidden in a back office.

But outrage alone, he says, rarely defeats a business model that saves money. Due to network effects surrounding training, infrastructure, and other business process optimization, outsourced labor also tends to cluster in specific areas. This may already be happening in the case of Waymo, which could soon see the rise of something like a “driving district” in Manila. In the future, other types of teleoperated work could follow suit, giving companies a ready-made destination to shop for low-cost labor.

For Graham, it’s urgent that we begin requiring certification from independent bodies, which can better scrutinize a company’s production networks. At Oxford he directs Fairwork, a project aiming to improve labor practices in digital supply chains.

I asked Sui how he thinks his customers may reorganize their operations around this new ability to remotely control their machinery.

“We’re working with traditional industries, and so it’s not just about adopting a new technology. There are significant management changes they will have to navigate. You could call this transformation friction because they will need time to digest this new capability step by step,” Sui said.

Despite the fact they could use the technology to outsource work across national borders, none of his customers are doing so just yet. Sui used open pit mines as an example. In this case, where fully developed towns with schools and hospitals have built up over decades, his customers still cluster their workforce next to the sites where they operate. Instead of driving into the mine, operators work from an office and go home clean at the end of a shift.

BuilderX has deployed its technology at more than 100 sites in China, Japan, and parts of Europe. It’s now expanding into new markets including South America and the Middle East. When asked whether he thinks his technology will be used for transnational outsourcing, there’s no hesitation. “Oh yes, I think this is coming in the very near future.”

The post Companies Could Soon Staff ‘Stubbornly Local’ Jobs With Workers 4,000 Miles Away appeared first on SingularityHub.

AI needs to focus more like we do.

“Attention is all you need.”

This 2017 breakthrough idea transformed AI. The concept of self-attention became the foundation of today’s chatbots. Claude, Gemini, and ChatGPT are all large language models (LLMs), AI systems designed to focus on the matter at hand while filtering out distractions.

The results have been remarkable. From brainstorming recipes to generating code, apps, websites, and content, LLMs are being woven into our lives at breakneck speed.

But now, a City University of New York team and collaborators are asking: How closely does AI self-attention resemble human attention?

It’s not just academic curiosity. AI researchers have long looked to the brain for ideas to improve machine intelligence. In turn, AI models have offered new ways to investigate the brain. Comparing artificial and biological attention could inspire AI that concentrates more like us.

In their study, the team asked multiple chatbots to complete a classic psychology test of attention and cognitive control. Participants are shown the word for a color—such as “red”—written in either the same or a different color than the one the word describes. The challenge is to name the ink color while ignoring the word itself.

On short word lists, the chatbots performed at a high level. But as the tasks grew longer, their focus faltered. Instead of naming the ink color, they increasingly defaulted to reading the word. Under more demanding conditions—ones that also trip up people—their performance nearly collapsed.

The findings suggest today’s AI attention systems are “fundamentally limited,” wrote the authors. They go on to say that adding mechanisms similar to “those in biological attention is crucial for achieving artificial general intelligence.”

Attention, Two Ways

Doomscrolling. YouTube. Dinner plans. Family obligations. A barrage of notifications.

Life sometimes seems like everything, everywhere, all at once. Yet the brain can usually lock onto what matters most and push everything else into the background.

Far from a single, straightforward mechanism, attention emerges from multiple brain regions. According to attention network theory, three networks do most of the heavy lifting.

The alerting network keeps the brain ready for action. The orienting network selects which sights, sounds, smells, and sensations deserve attention. Finally, the executive control network resolves conflicts between competing streams of information, helping direct thoughts and actions toward a goal.

Together, these systems allocate the brain’s limited resources. Touch a hot stove, for example, and your brain immediately shifts attention to the burn over dinner. The food can wait; cooling your hand can’t.

AI works very differently.

Rather than processing language as complete sentences, LLMs break text into smaller units called “tokens.” Attention mechanisms then determine which tokens matter most for generating the next word, sentence, or response.

Self-attention is the key breakthrough behind modern chatbots. For each token, the model weighs and incorporates information from other tokens in a sequence, allowing it to track context across long stretches of text. This mechanism helps AI connect words and ideas, and underpins virtually all frontier LLMs today.

Researchers have since built on the concept. One approach, multi-head attention, runs several attention systems in parallel, with each “head” learning different patterns, such as grammar, syntax, or meaning. Another, cross attention, links information across different chunks of inputs and their outputs, making it especially useful for tasks such as translation and summarization.

But attention comes at a steep computational cost. To make models more efficient, researchers are also exploring sparse attention, which limits how many tokens a model considers at once. Another approach draws on information learned in the past to keep AI “focused.”

Despite the name, AI attention is ultimately a mathematical system. It helps determine what information is relevant in a specific context. But it lacks executive control, the network that keeps humans continuously focused on a goal despite distractions for long periods of time.

Color Blind

To test the limits of AI attention, the team pitted OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet against the Stroop task.

Invented by John Ridley Stroop in 1935, the test measures attention and cognitive control by forcing participants to resolve conflicting information. The challenge is simple: Name the color of a word while ignoring what the word means. In a congruent trial, the word “blue” appears in blue ink. In an incongruent trial, “blue” might appear in red or green, creating a conflict between what the eyes see and what the brain reads.

Humans are consistently slowed down by this interference. Even with practice, the effect remains, suggesting it taps into fundamental mechanisms of executive control.

In the study, the researchers created word lists of varying lengths and difficulty. Some were entirely congruent. Others were fully incongruent. A third set mixed the two conditions.

At first, the AI models excelled. On five-word tests, GPT-4o was over 90 percent accurate across all conditions. But as the number of words increased, performance plummeted. On 40-word incongruent tests, the model’s accuracy fell to roughly 15 percent. Claude showed a similar decline. In mixed-condition tests, both models’ performance nearly collapsed to zero.

“The sharp decline in color-naming accuracy with increasing list length indicates that transformer-based attention mechanisms are vulnerable to scaling demands,” wrote the team.

Perhaps most intriguing, some models correctly recognized they were taking the Stroop test and could even explain its rules. But that apparent awareness did nothing to improve their scores. In other words, a “book smart” understanding of the task wasn’t enough to execute it well.

The study joins a growing effort to borrow psychological tests for research in machine cognition, especially when AI is challenged with complex, dynamic decision-making tasks. Theory of mind tests, for example, let researchers gauge whether a system can track others’ beliefs, emotions, and intentions. Personality tests are helping shape model behavior and reduce sycophancy. And some LLMs are readily solving emotional intelligence tests, which measure how well the algorithms recognize and respond to social cues.

According to the authors, the new results point to a missing ingredient in AI attention: A mechanism similar to the brain’s executive control network, which helps us stick to a task and adapt when priorities change.

Future AI systems could benefit from higher-level executive control that continuously tracks progress toward a goal, detects when attention has drifted, and pulls it back on course, if necessary.

Rather than simply weighing which tokens are most relevant in the moment, a more human-like form of attention could help AI stay focused during complex tasks, such as long conversations, multi-step reasoning problems, or high-stakes use in scientific research and drug discovery.

“The ultimate goal of AI research is to develop artificial general intelligence comparable to human abilities,” wrote the team. “AI systems, like humans, may need to master fundamental attention mechanisms…before achieving the generalized problem-solving abilities characteristic of mature executive functions.”

The post AI Collapses on a Classic Psychology Test. What It Reveals Could Stall Human-Level AI. appeared first on SingularityHub.

An audacious trial will test psilocybin in people over age sixty to see if increases plasticity in healthy aging brains.

A handful of healthy senior citizens are about to trip on psilocybin—to see if the psychedelic protects aging brains.

Psilocybin, the active ingredient in magic mushrooms, is best known for its ties to 1960s counterculture. But now it may also herald a new genre of mental health treatment. From severe depression to post-traumatic stress disorder, studies have highlighted psychedelic drugs’ ability to reshape brain networks and relieve debilitating symptoms.

Most of these studies have focused on younger people with mental health conditions that don’t respond to standard treatment. The field’s success is prompting scientists to ask if psychedelics could also help healthy brains age better.

A team from the UC Berkeley Center for the Science of Psychedelics is about to find out. In a first-of-its-kind study focused on adults between the ages of 60 and 85, they’ll investigate how psilocybin affects perception, emotion, and memory using a battery of psychological tests.

Multiple scans before and after dosing will track changes in the brain. And detailed surveys will gauge broader shifts in well-being: Do participants feel more “in tune” with their emotions, feel less isolated, or experience a renewed sense of wonder about the world?

“What really excites me is that we’re focused on healthy older adults,”  said Tyler Toueg, who co-led the study’s design, in a press release. “Most clinical trials with older adults are focused on people who already have a diagnosis. We’re asking whether we can actually promote positive outcomes in older adults who are healthy.”

Called PLASTICITY, the trial could also open a rare window into how a psychedelic experience reshapes healthy brain networks. And because the drug alters our sense of self, psilocybin could help researchers probe the ways in which the brain constructs reality.

“I’m very interested in psilocybin as a potential mental health treatment, but I’m also interested in it as a way to shed light on these central mysteries in neuroscience and psychology,” said study designer Michael Silver.

Chasing the White Rabbit

Psychedelic research was highly restricted for decades. But advocates, including the non-profit Multidisciplinary Association for Psychedelic Studies, have steadily pushed to reopen the field, arguing that these drugs might keep mental health symptoms at bay.

Early results helped usher psychedelics into the mainstream. In 2023, a randomized, placebo-controlled trial found that a single dose of psilocybin, paired with therapy, eased depression. Oregon later approved supervised psilocybin therapy—though the drug remains federally illegal in the US—and Australia became the first country to greenlight it for depression and post-traumatic stress disorder. More recently, two late-stage studies reported strong effects in severe depression, potentially paving the way for FDA approval.

Scientists still don’t fully understand how psilocybin works in the brain. But there are hints. The drug appears to rapidly reorganize the connections between brain cells, particularly in the hippocampus, a region of the brain central to learning, memory, and navigation.

Neurons constantly change their connections in a process called plasticity that encodes experiences into neural networks, allowing the brain to process information, learn, and lock in memories. In youth, these connections are flexible and expansive. But with age and in conditions like depression, the brain’s flexibility wanes.

The birth of new neurons, or neurogenesis, also contributes to plasticity. Neurogenesis only occurs in two brain regions, one of which is the hippocampus. Although whether it actually takes place in humans is controversial, it is strongly linked to learning, memory, and emotion, and it declines in both psychiatric disorders and aging.

Psilocybin may reset brain plasticity to a more youthful state.

In rats modeling depression, for example, a study showed the drug shifted dark moods into behavior that was more exploratory and engaging. Where traditional antidepressants, like Prozac, tend to blunt symptoms, psilocybin seems to overwrite entrenched negative patterns. This suggests deeper circuit-level changes.

In another study, the drug reopened a critical window for learning in mice. During adolescence, the brain is highly plastic, but it begins to stiffen in adulthood. Psilocybin temporarily restored malleability and changed the mice’s social behavior. In some brain regions, the drug increased sensitivity to oxytocin, the so-called “love hormone.” The authors suggested the drug induced a state called metaplasticity, in which neurons are more likely to respond to oxytocin and other regulators to rewire, form new connections, and grow their branches.

Psilocybin’s effects may extend beyond the brain. Depression, chronic stress, and the immune system are tightly linked. In a third study, researchers identified a brain-spleen connection that drives fear and anxiety. Psilocybin suppressed inflammatory immune cells associated with brain inflammation and dampened anxiety-like behavior in stressed mice, even under threat.

Both effects—dialing up plasticity and lowering inflammation—make psilocybin an intriguing way to potentially counter changes in the aging brain.

“We know that with age, we lose synaptic connections, especially in certain brain regions like the hippocampus and prefrontal cortex,” said Toueg. “There’s a lot of overlap between the mental states that psychedelics influence and those associated with successful aging.”

Tomorrow Never Knows

The PLASTICITY trial is designed to test whether psilocybin can produce lasting changes in neuroplasticity in healthy adults aged 60 to 85. Participants will first undergo assessments of cognition, visual perception, and brain structure using advanced MRI techniques.

Diffusion MRI scans will focus on the hippocampus to capture microscopic changes in its structure. Functional MRI will focus on brain activity as participants perform learning and memory tasks, offering a dynamic view of how activity shifts after dosing.

The study will also see if psilocybin increases vagus nerve activity, which has been linked to better stress recovery. Participants will complete detailed surveys about the experience ranging from emotional responses, like wonder, to potential shifts in outlook and social cognition.

“Things like depression, anxiety, stress and rumination are all associated with worse aging outcomes,” said Toueg. “Things like having purpose in life, emotional regulation, and awe are all associated with more successful aging.”

The trial began enrolling participants in November last year. Two volunteers have already completed the tests, and the team aims to dose 20 people by the end of 2026.

Although psilocybin trials are now widespread, older adults remain underrepresented. One estimate suggests just 1.4 percent of participants are 65 or older, despite potentially being among those most likely to benefit from interventions that enhance plasticity.

“This study allows us to directly test whether the promising findings from animal models translate to older humans and to generate data that will inform future research on aging, cognition, and mental health,” said Silver.

Toueg agrees. “I think that no matter what we find, this study will have implications for how we think about intervening in the aging brain,” he said.

The post Can Psychedelics Reboot Aging Brains? We’re About to Find Out appeared first on SingularityHub.

Artificial Intelligence

A Startup Claims It Broke Through a Bottleneck That’s Holding Back LLMsWill Douglas Heaven | MIT Technology Review ($)

“According to Subquadratic, it has developed a new kind of LLM, called SubQ, that is faster and cheaper and uses a lot less energy than any other model on the market. The company also claims that SubQ is able to process up to 12 times as much text at once than most other models, allowing it to carry out a range of data-heavy tasks, such as analyzing hundreds of documents or entire code bases.”

The post This Week’s Awesome Tech Stories From Around the Web (Through June 20) appeared first on SingularityHub.

Coal’s share has nearly halved over the last five years, while solar’s has more than doubled. But tariffs and permitting delays could slow growth in the years ahead.

The transition away from fossil fuels is often framed as a long-term process, but recent data suggests the shift is already happening. Solar power has now crossed a major threshold in the US, surpassing coal in the electricity generation mix for the first time.

Despite the Trump administration’s attempts to drive a coal revival, it has been steadily losing ground to other energy sources in recent years, squeezed out by cheap natural gas and rapidly falling renewable energy prices. Solar power, in particular, has been on a tear as prices drop exponentially.

Last month, the two lines finally crossed. Solar supplied 12.8 percent of US electricity in May, edging past coal’s 12.2 percent share to become the country’s third-largest source of power behind natural gas and nuclear, according to recent data from energy think tank Ember.

“Overtaking coal for the first month on record shows just how far solar has come, from a niche contributor to the third-largest and fastest-growing source of power in the US electricity system,” Nicolas Fulghum, senior data analyst at Ember, said in a press release.

The transition is as much about coal’s waning importance in the US energy system, as it is about solar’s growth. Coal’s share has nearly halved in five years, falling from 19.7 per cent in May 2021 to 12.2 percent today and hitting an all-time monthly low in April.

Over the same period, solar’s share of electricity generation has more than doubled from 5.4 percent to 12.8 percent. And it hit an all-time high of 45.5 terawatt-hours in May, up 17 percent compared to the same month last year and above the previous record set in July 2025. Ember gets its data from the US Energy Information Administration.

The industry does face some headwinds though. A separate report from the Solar Energy Industries Association and analytics firm Wood Mackenzie found that the 7.8 gigawatts of new solar capacity added in the first quarter of 2026 is a 27 percent decline compared to the previous year.

This is partly due to regular seasonal patterns for the industry, says the report, but was also thanks to the expiry of a tax credit for residential installations and trade restrictions and tariffs targeting imported solar components from Asia initiated by the Trump administration. The government has also made it more difficult to get permits for new projects.

But despite the apparent slowdown, solar and battery storage together accounted for 91 percent of all new electricity-generating capacity added to the grid in the first quarter of the year. And the number of new utility-scale projects signed in the first quarter hit 6.3 gigawatts, a rise of 15 percent. Tellingly, the SEIA notes that states President Trump won in 2024 make up 74 percent of new solar capacity installed in that period.

“In a world of fluctuating fuel prices, energy buyers have made it clear that they want the security, low cost, and speed of solar and storage,” Darren Van’t Hof, interim president and CEO at SEIA, said in a press release.

“Impeding the only sector that is actively building new power is a reckless gamble that will only drive electricity bills higher. The stakes are simply too high for Washington’s permitting gridlock to continue.”

As a result of these barriers, Wood Mackenzie’s five-year forecast predicts that annual additions will plateau around 43 gigawatts. That’s still an impressive pace of installation, but also a significant slowdown from the breakneck growth seen over the last couple of decades. So, while solar may have knocked one fossil fuel competitor off the podium, without a change in energy policy it may struggle to maintain its impressive momentum.

The post Solar Beat Coal in US Electricity Mix for the First Time in May appeared first on SingularityHub.

Scientists are exploring new algorithms, hardware, and computing methods to lower AI’s power demands. Strategic siting of data centers and other steps to increase green energy use are also key.

This story was originally published by Knowable Magazine.

As I sip coffee in my Berlin apartment and fire a question at Google’s AI chatbot Gemini, it’s easy not to think about the energy it takes to generate a response. Once the signal reaches my router, it whizzes, I assume, through copper wires or fiber-optic cables to one of Google’s data center hubs. Somewhere inside the data center’s labyrinthine halls of stacked processors, my query gets converted into numbers and undergoes billions of computations to determine context and meaning. The answer, once assembled, races back, in the blink of an eye.

Data centers—the beating hearts of the internet, powering everything from email to web searches—have existed for decades, but with the growing popularity of AI to generate text, images, and video, they’re using more energy than ever. According to Google’s own estimates, processing a median-length text prompt with its AI assistant Gemini consumes around 0.24 watt-hours.

These amounts, individually small—0.24 watt-hours is equivalent to watching TV for about nine seconds—are adding up fast. In March 2026, OpenAI estimated that more than 900 million people use its AI chatbot, ChatGPT, every week, tallying billions of queries daily.

The exact amount of electricity consumed by data centers, globally or in the United States, which hosts more than any other nation, isn’t publicly reported by all tech companies, says Eric Masanet of the University of California, Santa Barbara, who researches data center sustainability. But according to the most recent estimates by the International Energy Agency, US data centers guzzled some 224 terawatt-hours of electricity in 2025—more than 5 percent of the country’s electricity use. That’s a significant uptick from an estimated 1.9 percent consumed in 2018, well before the mainstream surge of generative AI.

This electricity use seems set to soar. In the race to secure market leadership for generative AI products, companies like Google, Meta, Amazon, OpenAI, Anthropic, Microsoft, and Oracle are investing tens to hundreds of billions of dollars to build AI-focused data centers. Compared to data centers of the pre-AI days that consume, say, 100 megawatts of electricity—enough to power 83,000 homes with average demand—the newcomers are often “hyperscale” and can use a gigawatt or more, or roughly a tenth of the electrical capacity of Los Angeles.

Masanet and other experts have been alarmed to see much of this demand met by plants powered by fossil fuels, such as gas, whose burning releases planet-warming carbon dioxide. A key reason is that data centers are often constructed in places without abundant renewable energy sources like hydropower, geothermal, solar, or wind.

Tech companies often offset emissions by investing in renewable energy elsewhere. But unless those clean energy plants make more energy than the data centers use, this strategy—at best—keeps CO₂ emissions of centers in stasis rather than reducing them to a net of nothing, important for halting global warming. “For every megawatt for which we install fossil fuel power,” Masanet says, “it sets us back on our progress.”

And that’s not considering the resources spent on manufacturing the hardware that fills new data centers, or the impacts on communities living near them, which often suffer from air and noise pollution from gas plants and possible strain on local water resources, which are used to cool the data centers.

Although forecasts for AI’s energy impact remain devilishly tricky, especially since the size of payoffs from investments in AI are uncertain, it’s clear to experts that energy-saving strategies are urgently needed. Without them, according to one 2025 estimate, US data centers could soon be releasing the equivalent of 24 to 44 megatons of CO₂ annually, the latter equivalent to the annual emissions of Norway.

And so computer scientists and engineers are rethinking some of the power-hungry hardware and software that fuel AI. They’re working to develop energy-saving algorithms and processor designs, and carefully considering where, and how, data centers are constructed.

“AI’s energy cost is not an accident: This is basically a product of how our systems are built,” says Fengqi You, an expert in energy systems at Cornell University. But with the right mix of solutions, he says, “we could really reshape the trajectory.”

The Roots of AI’s Energy Problem

To comprehend AI’s energy cost, it helps to understand large language models (LLMs)—the lifeblood of AI text generation tools such as chatbots and AI assistants—specifically, ones based on a design described in 2017 by the machine-learning laboratory Google Brain. This design, transformer architecture, can process text at lightning speed by simultaneously taking each word and weighing its relationship to every other word it sees. It “learns” which words go together by computing how strongly each word relates to all other words in a text, examining each word in many contexts. (A similar design is used for AI image and video generators.)

On a computational level, this happens by converting words or word fragments into numbers and performing additions and multiplications between them. Key to the speed is being able to do these calculations in parallel, made possible by graphic processor units (GPUs)—mostly manufactured by the company Nvidia—originally invented for rapid 3D rendering of imagery during gaming.

The initial training of an LLM, required to learn all these relationships, consumes vast amounts of energy. Because each word it trains on must be weighed against all others in a given chunk of text, the number of computations the model performs—hence the energy required—increases quadratically relative to the length of text (i.e., doubling the length of text quadruples the number of computations). That adds up quickly given that most LLMs are trained on massive swaths of publicly available internet text. Some estimates suggest that training GPT-4—the iteration of ChatGPT that launched in 2023—guzzled between 50 and 60 gigawatt-hours of electricity, enough to power San Francisco for three to four days.

But experts are more worried about the energy costs of using the models to generate data once they’ve been trained, a process called inference. “You train once, then you inference for a billion people in the world,” says Mosharaf Chowdhury, an AI systems expert at the University of Michigan who has been measuring the electricity usage of a handful of large language models that have been made publicly available.

This process is surprisingly inefficient: Each time transformer models generate a word—by selecting the one with the highest probability of following the previous word, given context—they put the query and partially written answer through the model. In doing so, they apply all of the parameters they’ve calculated during training to understand language patterns—which number in the hundreds of billions or even trillions.

“The fact that you have to do a lot of calculations for a single word to be added—that’s a problematic thing,” says Günter Klambauer, an AI expert at Johannes Kepler University in Austria.

Tweaking AI Software to Save Energy

This recognition has triggered interest in smaller language models specialized to specific tasks. These are trained more narrowly, have fewer parameters—say, tens or hundreds of millions—and perform substantially less computation than larger models. In one 2025 paper published by UNESCO, computer scientist Ivana Drobnjak of University College London and colleagues compared energy consumption of Meta’s language model Llama-3.1 with smaller AI models dedicated to particular tasks—ones called DistilBART and t5-small-xsum for summarization, and others for translation or answering questions. When used for their respective tasks, the smaller models consumed more than 90 percent less energy than Llama 3.1 on the same job.

And so computer scientists have been driven to build a similar kind of task specialization into LLMs themselves. In “mixture of expert” models, only particular parts of one big model are activated for certain tasks. These parts “learn to handle different patterns in language,” Drobnjak says.

This is thought to be one reason why R1, an LLM developed by the Chinese company DeepSeek, reportedly consumed significantly less energy than other models (independent experts have raised doubts about those figures). Udit Gupta, an expert in electrical and computer engineering at Cornell Tech, says that LLMs like Gemini or ChatGPT are similarly routing queries to more specialized sub-models. “There’s a lot of work being done on how to assess the complexity of the query or task that’s coming from users and then find the right model,” Gupta says. (While Google spokesperson Ralf Bremer notes that the 0.24 watt-hours currently spent on processing median-length Gemini prompts is already 33 times more efficient than it was back in 2024, some experts suspect that processing queries with an LLM still consumes more energy than an equivalent web search.)

Scientists are also exploring different kinds of LLMs, to break what Klambauer calls the “quadratic curse” of transformer models.

One alternative, called a long short-term memory (LSTM) model, gets around this alarming energy increase by temporarily storing a kind of summary of the prompt that was inputted by the user plus the text generated so far, akin to recalling important plot points instead of an entire movie. That way, it only has to process the summary, rather than all the words in the full text to date, every time it generates a new word. This prevents LSTM’s energy costs from skyrocketing as it responds to a query—using about 50 percent less energy than transformer-type models to process texts of around 8,000 words in length, Klambauer says.

LSTM models were developed in the 1990s but were abandoned because transformers could be trained much faster. But Klambauer says that recent advances have improved the performance of LSTM, now called xLSTM. He’s working with the Austrian startup NXAI to further develop and optimize xLSTM, “because we think it’s worth it for energy efficiency,” he says.

But major tech companies have invested so many years and resources into developing transformer-based models that switching to other models would be costly, says Wolfgang Maaß, an AI and business informatics researcher at the German Research Center for Artificial Intelligence. “We have to see whether this becomes as dominant, or whether it finds a niche in the whole market.”

Computing With Wafers and Light

Though experts say the fastest energy savings will come from software tweaks, some are also taking aim at the energy-hungry processing chips that fuel AI computations. Engineers have made chips increasingly efficient over time by packing more computing capacity into individual processors—reducing the energy required to shuttle data between chips that are working together to perform AI computations. Engineers have done this by shrinking the size of transistors—microscopic electrical switches that process data—inside the chips.

But because engineers are reaching the physical limits of how small transistors can be, “we need to think of alternate ideas to improve the designs,” says computer architect Ajay Joshi of the Boston University Photonics Center.

One strategy is to make the chips larger. Dinner-plate-sized “wafer-scale chips” can pack nearly 70 times as many transistors as a single, postage-stamp-sized GPU and consume 143 times less electricity for communication than comparable GPUs, says computer engineer Rakesh Kumar of the University of Illinois Urbana-Champaign. Commercially produced by the California company Cerebras, wafer-scale chips have drawbacks, including a greater risk of damage during manufacturing. But because of their energy-saving and other beneficial features, “they would be very attractive to many hyperscalers and AI companies,” Kumar says.

Many tech companies have improved energy efficiency by fashioning their own processors that are tailor-made for AI computations—such as Amazon Web Service’s Trainium2 chip or Google’s Ironwood Tensor Processing Units—according to statements from those companies. As for Nvidia, the company’s head of sustainability Josh Parker says its AI-specialized GPUs have come a long way from the ones used for gaming and are now designed to run AI tasks as efficiently as possible; other innovations, such as making the interconnections between GPUs more efficient, have also helped. “Over the past eight years, NVIDIA GPUs have improved 45,000 [times] in energy efficiency for large language model workloads,” he says.

Engineers are also exploring alternative computing methods. Conventional AI processors calculate by encoding numbers in a binary system of ones and zeros, which is achieved by turning transistors on and off (representing the number 5, for instance, requires four transistors to represent the code 0101). But transistors can do more than function as binary switches allowing electron flow or not; they can also work as analog dials and hold intermediate voltages representing different numbers. That requires fewer transistors, and less energy, for computations. “People have known for decades that doing certain things in analog … can be a lot more energy efficient,” Kumar says.

For example, electrical engineer Paul Manea of the German research institute Forschungszentrum Jülich and colleagues are working to develop devices called “gain cells” that are full of transistors working this way. Importantly, gain cells can both store the data required to process a query, and compute the answer. That overcomes another big energy bottleneck of conventional computing systems, where memory storage and computation occur on separate pieces of hardware.

That’s especially problematic for transformer-based LLMs, because each time they generate a word, they must shuttle the query and partially written answer from memory to a processor. Manea and colleagues estimate that gain cells in lieu of traditional GPUs can reduce the energy guzzled by one of the most energy-consuming parts of transformer-based LLMs by four orders of magnitude. But it will take more refining before they can be more widely used, Manea says.

The notion of devices that both store and compute information is a key idea of “neuromorphic” computing, an up-and-coming field of computer engineering inspired by the human brain, which consumes orders of magnitude less energy than computers. Another brain-inspired invention is chips that encode information not in continuous data streams but—like human nerve cells—in the timing of voltage “spikes” propagating through the system. Allowing components to rest until they’re needed “could potentially translate to less energy,” says Eleni Vasilaki, an expert in bioinspired machine learning at the University of Sheffield in England.

Maaß, for example, is part of a team that received roughly $5.8 million from the German government to test neuromorphic chips, among other strategies, to reduce the energy required for AI models. Some brain-inspired chips are already commercially available, but the technology is still far from being attractive for mainstream computing, says nanoelectronics expert Tony Kenyon of University College London, whose team recently received $17 million from the UK government to develop neuromorphic computing.

Other scientists are developing chips that process information not with electrons but through the interaction of photons—particles of light—with matter (fiber-optic cables, which encode and transmit data as light pulses, are used around the world). With photons, more information can be transmitted at the same time, and signals can be altered much faster, says Elena Goi, a photonic computing researcher at Friedrich Schiller University Jena in Germany.

Several companies have developed chips that can perform some AI computations with optical methods, says Joshi; he recently estimated that manufacturing optical chips could consume up to an order of magnitude less energy than conventional ones of the same size. Joshi hopes that, “in 10 years, we would have a practical solution that can be deployed pervasively across the data centers.”

Reshaping AI’s Energy Trajectory

Even without reinventing how computers work, much can be done to reduce AI’s impact not just on energy but also on water resources used for cooling data centers. Importantly, tech companies should reconsider where they build those centers, says energy systems expert You. Right now, existing US ones are concentrated in northern Virginia, which has limited water resources and renewable energy capacity compared with the Midwest, for instance. You recently estimated that better siting—along with energy-efficient hardware and software—could reduce future carbon and water footprints of US data centers by 73 percent and 86 percent, respectively.

Masanet adds that tech companies already with data centers across the country could at least train their models in strategic places. “Some companies like Google have been doing this: They shift their loads to follow renewables,” he says. They also should address the electricity and resources spent on manufacturing processors for new data centers, as well as electronic waste as outdated tech is replaced every few years, he adds.

Minimizing e-waste by using hardware for longer periods and recovering old electronics is one of Amazon’s sustainability strategies, according to a statement to Knowable Magazine; so is designing data centers in energy- and water-saving ways and investing in a slew of renewable and nuclear energy projects. “We’ll continue to implement solutions that benefit our customers and the communities we operate in,” says Brandon Oyer, Amazon Web Services’ head of energy and water in the Americas.

Meanwhile, a press representative at Microsoft points to a number of sustainability initiatives the company has taken, including new cooling technologies, renewable energy investments, and waste reduction. Google spokesperson Ralf Bremer emphasized the company’s goal of reaching net-zero emissions across its operations by 2030 and replenishing 120 percent of the fresh water consumed by its offices and data centers by 2030. An OpenAI representative points to a press release outlining efforts to minimize water use and plans for solar energy generation at one of its campuses. Anthropic, Meta, and Oracle did not respond to requests for comment by deadline.

Though tech companies are taking sustainability into consideration, their main objective is to rapidly build out data center capacity, says computer engineer Benjamin Lee of the University of Pennsylvania. He predicts that, eventually, they’ll need to step up efforts to improve energy efficiency to reduce costs. Governments should help to accelerate this shift, Masanet says. So far, he and his team have counted nearly 220 policies introduced to address data center sustainability at the US state level, 18 at the federal level, and more from other countries, though not all were ultimately adopted.

“It’s clear that governments around the world are beginning to take action,” he says. However, he adds, “we also see some state and local governments with proposed policies that mostly aim to incentivize and accelerate data center builds.”

AI’s energy cost will ultimately be a balancing act: Will it save more resources through its problem-solving abilities deployed toward everything from finding cancer cures to improving logistics, than it demands? But though building a more frugal, energy-saving AI is important, so is carefully considering where AI is needed, Kenyon says. Is the world truly a better place, for example, with nonhuman “AI agents” providing customer support?

“I think it’s a common mistake, when a new technology comes in, to suddenly think, ‘Well, everything has to adopt that new technology,’” he says. “That approach really isn’t doing us any favors.”

Editor’s note: This article was amended on May 27, 2026, to clarify, in a caption for a graph, that the number of introduced policies involving data centers included ones that did not pass. In addition, a web page link was added in the article for University College London researcher Ivana Drobnjak.

This article originally appeared in Knowable Magazine, an independent journalistic endeavor from Annual Reviews. Sign up for the newsletter.

The post How to Tame AI’s Voracious Appetite for Energy appeared first on SingularityHub.