MoreRSS

site iconSingularity HUBModify

Singularity Hub has offered daily news coverage, feature articles, analysis, and insights on key breakthroughs and future trends in science and technology.
Please copy the RSS to your reader, or quickly subscribe to:

Inoreader Feedly Follow Feedbin Local Reader

Rss preview of Blog of Singularity HUB

Novel Gene Therapy Nearly Erases ‘Drug Resistant’ Seizures

2026-11-19 23:00:00

Scientists have long used chemogenetics to study brains in the lab. A pilot trial for seizures is its first leap into medicine.

Flashing lights. Buzzing noises. A sudden wave of fear. For some people with epilepsy, these sensations herald an oncoming seizure.

Drugs can keep some seizures at bay, but over 30 percent of patients don’t respond to them, a statistic that hasn’t budged in decades. Frequent, uncontrolled seizures rob people of a normal life. They can’t drive or swim alone. Some struggle to stay in school, hold down jobs, or maintain relationships. And despite growing public awareness, stigma lingers.

For people with severe drug-resistant epilepsy, surgery is an option. If the seizures are associated with a local brain region, doctors can remove the tissue. It’s a drastic operation, but it isn’t rare. In 2024, over 2,000 patients elected to undergo the procedure. Depending on the brain region removed, surgery can control seizures in nearly 70 percent of patients, but they risk losing memory, language, and other cognitive abilities.

A new gene therapy could offer an alternative. In a clinical trial, three men who had lived with epilepsy for at least a decade saw monthly seizures drop over 90 percent. One remained seizure-free for a year.

In the trial, the team shuttled a gene into a small patch of the hippocampus, the source of their seizures. The gene produces a kind of protein “dimmer switch” in neurons. Normally silent, the protein springs into action when activated by an FDA-approved drug, allowing scientists to dial down the region’s electrical activity rather than shutting it off completely—something that’s not possible with current treatments.

The technology, called chemogenetics, has been a staple of animal research in neuroscience. The findings, though early, suggest it could finally make the leap into human medicine.

“It certainly has caused a lot of excitement in the field,” said Dimitri Kullmann at University College London to Chemical & Engineering News, who was not involved in the study.

Chemical Romance

Intricate webs of neurons underlie every brain function. Our ability to feel, move, remember, and reason all stem from electrical and chemical signals that wire brain cells into circuits.

It’s a delicate balance. Some neurons release chemicals that fire up their neighbors; others dampen them. In epilepsy, this goes awry. The causes are plenty. Genetic mutations can disrupt how neurons develop or function. Brain injuries, tumors, stroke, and infections can also fuel the condition. In some frustrating cases, the culprit remains unknown.

Rebalancing these circuits could tame runaway seizures. Most antiseizure drugs prevent neurons from firing too rapidly or boosting inhibitory chemical signals. But medication is a Band-Aid, not a cure. Because the drugs flood the whole brain, they can trigger fatigue, dizziness, mood swings, and sometimes more serious side effects, including liver or heart problems.

Dampening electrical activity at the source could be a better bet.

Around two decades ago, an emerging technology called chemogenetics first amazed neuroscientists with its precise control of single neurons.

The technology relies on two parts: gene therapy and a drug. First, an engineered virus delivers a gene encoding a designer receptor into targeted neurons. The receptor is silent until a specific drug activates it. By adjusting the drug’s dose, scientists can fine tune the activity of modified neurons. Rather than simply flicking neurons on or off, the approach is flexible and targeted.

Chemogenetics soon caught the attention of epilepsy researchers. In 2019, a team at Ghent University in Belgium tested the method in mice with seizures originating in the hippocampus. They delivered a designer receptor called hMADi to the region. A single dose of clozapine, an FDA-approved antipsychotic, controlled seizures for at least 15 hours. With repeated treatment, seizures nearly vanished. A subsequent study in monkeys also dramatically slashed seizures.

Given these encouraging findings, some researchers were ready to take the next step. “I think scientists like me felt like, ‘Why haven’t we tried it in humans yet?’” said study author Minmin Luo at the Chinese Institute for Brain Research.

Restoring Balance

The trial enrolled three men under 40 diagnosed with epilepsy for 10 to 25 years. All had tried at least two antiseizure medications without success. Brain scans traced their seizures to a scarred hippocampus on one side of the brain. The same structure on the other side remained healthy. Despite their long history of epilepsy, none showed notable signs of cognitive troubles.

Similar to previous work, the team delivered a gene encoding the hMADi receptor into the damaged region of each patient’s hippocampus in a series of three shots. A dye mixed with the therapy confirmed it reached nearly the entire region in two patients and spread slightly beyond it in the third. The infusion took roughly two hours.

Two weeks later, the men began taking clozapine pills every eight hours. After five months, their average monthly seizure count had dropped from roughly four to nearly zero. Brain recordings also showed far fewer epilepsy-related electrical spikes, and their damaged hippocampuses increased metabolism, though levels remained below those in the region’s healthy counterpart.

The benefits lasted. In check-ups throughout the following year, two patients remained seizure-free. The third experienced two “breakthrough” seizures after an unexpected interruption in medication, but the seizures stopped once he restarted the drug regime. All three participants also reported improved quality of life.

The study’s main goal, however, was to gauge safety. The viral carrier has a relatively established safety record, but high doses have been linked to severe immune reactions and even deaths. Delivering the therapy directly into brain tissue requires far smaller doses than when injected into the bloodstream and perhaps, counterintuitively, could be safer overall.

Blood tests, brain scans, and other assessments find no major side effects. One patient experienced a temporary rise in liver enzymes, a sign of liver stress, but levels returned to normal without intervention.

The team plans to expand the trial, but there are plenty of unsolved questions. All three men continued taking their existing antiseizure medications throughout the trial, making it difficult to isolate the treatment’s effects. For example, could clozapine have somehow synergized with those medications to reduce seizure counts? The study was also open-label, without a placebo control. And clozapine, when taken at high doses, carries risks of heart and other problems, though the doses given here were much lower than those used in psychiatric disorders.

Whether chemogenetics can deliver lasting seizure relief in larger groups remains to be seen. But its debut in human medicine is a hopeful glimpse of what’s possible, not just for epilepsy, but potentially for addiction, mood disorders, and other neurological conditions involving unbalanced brain circuits too.

“These findings offer initial proof of concept,” the team wrote.

The post Novel Gene Therapy Nearly Erases ‘Drug Resistant’ Seizures appeared first on SingularityHub.

Scientists Use AI-Designed CRISPR to Make Disease-Resistant Rice

2026-11-19 02:17:56

The versatile, open-source tool could democratize efforts to engineer healthier, higher-yielding crops.

In many cultures, if rice isn’t on the table, it’s not a full meal. Yet the crop is constantly under attack. Bacterial blight is one of rice’s most destructive diseases, causing yield losses up to 50 percent. One culprit, a microbe called Xoo, genetically hijacks the plant’s metabolism, coaxing it to release sugars that fuel the invader’s growth and spread. Xoo thrives; the plant dies.

Thanks to years of molecular sleuthing, scientists have mapped out the bacteria’s main genetic targets. And using the gene editor, CRISPR, it’s now possible to destroy those targets and thwart the bacteria’s antics.

But even with such a powerful tool, progress has been slow. Part of the problem is accessibility. Proprietary gene editors often “require licensing agreements for commercial use,” says Bing Yang at the University of Missouri.

Instead, his team turned to OpenCRISPR-1. Developed by California startup Profluent, the freely available AI-generated gene editor can edit cultured human cells like the original CRISPR. But the tool has barely been tested in plants, which have their own genetic quirks.

Yang’s team reworked the editor into a plant engineering machine. Their upgrades reliably shut down genes required for Xoo invasion with 90 percent efficiency or higher. The plants faithfully passed the edits to the next generation, paving the way for rice strains resistant to blight.

“Our work bridges AI-driven protein engineering with practical agricultural applications,” they wrote.

Growing Pains

Humans have been tinkering with rice for millennia. As far back as 9,000 years ago, ancient farmers were selecting plants with grains that were easier to harvest. They relied heavily on serendipity: random mutations that produced the wanted trait.

Today’s genetic tools have changed the game. Scientists can now quickly and precisely reshape rice’s DNA to improve nutrition, yield, and resistance to diseases and drought.

Early wins came in the 1990s. Scientists borrowed foreign DNA and inserted it into rice embryos—tiny, living plants tucked inside grains that can grow into new, mature plants when the seed sprouts.

They got creative with their genetic sources. One study found a naturally occurring rice gene that confers resistance to Xoo and introduced it into other rice plants. All 50 transgenic strains fought off the bacteria. Genes from thale cress boosted tolerance to drought and salty environments, while those from daffodils and bacteria gave rice a newfound ability to make a precursor of vitamin A, turning the grains a bright yellow.

Developed to combat vitamin A deficiency, “Golden Rice” faced fierce backlash in the communities it was supposed to help. Along with other GMOs, it was derided as “Frankenfood,” a stigma that lingers today.

CRISPR-edited crops, in contrast, may be more welcome. Rather than introducing new genes, most studies modify the plant’s existing ones in a sort of accelerated evolution. In 2020, scientists used CRISPR to disable a rice gene that shuts down the production of fragrance molecules, giving a non-aromatic strain a lovely scent. Another study cut out a gene that makes rice susceptible to blast fungus, a major cause of global yield loss.

Then there’s Xoo blight. The bacteria spreads through contaminated water, enters leaves through wounds, and colonizes the plant’s main “water vessels.” Once there, it injects molecular tools into nearby cells that increase the activity of a family of genes called SWEET, which encode proteins to shuttle sugar into spaces where the bacteria can feast.

Understanding “how bacterial blight interacts with rice at the genetic level” can help researchers find ways to strengthen the plant’s natural defenses, said Yang.

Back in 2019, his team used CRISPR to edit genetic switches controlling the activity of three SWEET genes in two widely cultivated commercial rice strains. In confined field tests, the edited plants grew normally and resisted multiple bacterial strains after infection.

But strengthening the crop meant navigating CRISPR’s complex intellectual property landscape—a paperwork headache and speed bump. So, they sought to develop an alternative open to all.

‘ChatGPT’ for Crops

The new study began with OpenCRISPR-1.

Unlike a chatbot trained on text to handle everyday language, Profluent’s AI learned the language of proteins—which molecular “letters” tend to follow others and how their arrangement affects structure. Trained on millions of CRISPR systems from microbes, the AI dreamed up entirely new editors unlike any seen in nature. OpenCRISPR-1 stood out for its ability to cut target DNA sequences in human cells with fewer mistakes than the classic version.

Yang’s team then adapted the system to plants. They rewrote DNA instructions for the AI-designed protein scissors, OpenCas9, using a genetic vocabulary better suited to rice, allowing cells to manufacture the protein more efficiently without changing its structure. They also added molecular “hooks” to shuttle the scissors into the plant’s DNA housing.

Cas proteins are only half the system. Guide RNA, the molecular “bloodhound,” is the other. It has two components: a “backbone” that grabs onto Cas and a guide sequence that tethers the system to its DNA target. The team tested an AI-designed backbone paired with guides targeting three SWEET genes hijacked by Xoo.

Although substantially different from the conventional system, the rice-adapted OpenCRISPR-1 system worked just as well, with a success rate of up to 97 percent in a rice variety commonly used for research. Edited plants fought off three different Xoo strains following a 12-day long infection, and the plants passed the genetic changes on to the next generation.

The team then reworked the editor so it could handle prime editing, a more versatile CRISPR tool that rewrites DNA without cutting both DNA strands, potentially making it safer. In isolated rice cells, the mutated OpenCas9 precisely edited four genes. Some were single DNA letter substitutions; others required three simultaneous swaps. In some cases, OpenCas9 had fewer off-target changes than a conventional prime editor.

Whether the system will work as well on commercial rice varieties—and whether it can make stable, resilient strains that thrive in the real world—remains to be seen. Rice was a proof of concept, but the tool could be extended to other crops. More broadly, the study adds to recent findings that OpenCRISPR-1 could enable more sophisticated and versatile plant genome engineering—not just in rice, but also soybeans, alfalfa, and tobacco. Further safety testing, crowd-sourcing of consumer opinions, and the establishment of regulatory guidelines will be needed before these gene-edited plants reach the market.

“As AI-driven protein design continues to advance, we anticipate…enhanced precision, and broader utility for next-generation crop improvement,” the team wrote.

The post Scientists Use AI-Designed CRISPR to Make Disease-Resistant Rice appeared first on SingularityHub.

Tiny Magnetic Robots Help Dissolve Blood Clots Six Times Faster Than With Drugs Alone

2026-11-17 05:09:30

The bots roll along blood vessels to stir up stagnant blood and push drugs toward clots.

Drugs that break up blood clots are vital for treating strokes, heart attacks, and clogged blood vessels in the lungs, but sometimes blockages are so large that medication struggles to get through. Now, researchers are building tiny magnetic robots to help medicines reach their targets.

Severe blockages in blood vessels pose a tricky problem. The bloodstream is the only way to deliver clot-busting drugs, but if a clot stops the flow, there is no other way to carry them there.

A team at ETH Zurich in Switzerland has a workaround. Magnetically driven microbots that roll along vessel walls can stir up stagnant blood so the drug gets pushed towards the clot. In lab tests, the approach dissolved clots six times faster than with the drug alone, and in experiments on pigs, it restored blood flow to an area twice as large after 30 minutes.

“To our knowledge, we are the first to show in a clinically relevant large-animal model that the use of microrobots in this way does indeed enhance thrombolysis [the process of dissolving blood clots],” Schürle said in a press release.

The bots are a far cry from your traditional robot. Just eight micrometers long, each is shaped like a grain of rice and made of a sugar-based polymer laced with magnetic iron oxide nanoparticles.

During fabrication, the researchers expose the devices to a rotating magnetic field that pushes the nanoparticles into a straight line, essentially creating a bar magnet running the length of each bot. When a magnetic control system outside the body applies a rotating field, the internal magnet spins and causes each bot to tumble head over tail.

The bots themselves don’t carry drugs. Rather, they’re released near a blockage along with clot-busting medicine. As the bots roll along the vessel walls towards the restriction, they create currents that carry the drug into the clot.

To test the approach, the team built an artificial network of vessel-like channels and plugged it with clots made of pig blood. They compared what happened when they added the clot-dissolving drug alone to what happened when they added the drug and the bots together. Adding the bots cleared the clots six times faster than when they tried just the drug.

To see if the strategy would work in real biological systems, the team blocked arteries in live pigs and treated them with just drugs and drugs plus microbots. After 30 minutes, the area with restored blood flow was twice as large in animals that received the bots.

The work adds to a growing body of research into clot-busting magnetic microrobots. Last year, a separate ETH Zurich team led by Bradley Nelson published a paper in Science describing a dissolvable gel capsule that could be steered through blood vessels and release a clot-dissolving drug when heated by a high-frequency magnetic field.

Others have taken a more mechanical approach. Researchers at the University of Twente and Radboud University Medical Center in the Netherlands used 3D-printed, screw-shaped robots containing tiny magnets to drill into clots, removing enough material from a blocked artery taken from a sheep to get blood flowing again.

While the ETH team have filed a patent application for their bots, it’s still early. The paper describing the results has yet to clear peer review, and human trials are a long way off.

Nonetheless, the results are another indication that microrobots could one day become a valuable new tool in a doctor’s arsenal.

The post Tiny Magnetic Robots Help Dissolve Blood Clots Six Times Faster Than With Drugs Alone appeared first on SingularityHub.

This Week’s Awesome Tech Stories From Around the Web (Through November 14)

2026-11-14 23:00:00

Tech

The Jensen Huang Scandal Is Hardly OverElizabeth Lopatto | The Verge

“Huang has insisted that Nvidia has seen ‘no evidence of any AI chip diversion.’ He’s been pretty emphatic on that point, actually. I am now wondering if that is even a little bit true. If the board is cutting Huang loose, there’s probably more on chip smuggling allegations coming.”

Biotechnology

A New Benchmark for De-Aging Comes From Slowly Turning Back Cells’ ClocksJorge Garay | Wired ($)

“They guided these cells into a younger state but not all the way toward becoming embryonic-like stem cells. …In less than two months of reprogramming, one biomarker in particular was set back as far as 60 years. ‘The hope is that you actually can do this systemically in the entire animal,’ Brüstle says. ‘But if you think about applying it in a patient, then you want to be sure this is not a lab experiment, so it’s about finding the mechanism.'”

Future

Is AI Taking White-Collar Jobs? What the Latest Numbers Tell UsIan Hodgson | The Wall Street Journal ($)

“All told, professional and business services added 70,000 jobs, its largest monthly increase since July 2022. The gains complicate fears that AI is set to become a white-collar job killer—concerns that helped sweep AI skeptics into office in this week’s elections. …’I don’t know how you could look at this report and conclude that we’re entering anything like a white-collar job apocalypse,’ said Harvard economist David Deming.”

Future

The AI Safety Debate Has Lost the Plot. We Asked 10 Experts What the Real Nightmares AreLauren Goode | Wired ($)

“Almost no one, it turned out, was Pollyannish. Nearly everyone I spoke to expressed genuine concerns about the threat of cyber and biological attacks. One expert who studies the risks of AI in nuclear operations said she’s stunned by how quickly error-prone AI systems are being adopted in defense. AI-enabled attacks in these areas could be catastrophic, if not fully apocalyptic.”

Computing

‘They Don’t Understand Hackers’: Cybersecurity Pros Explain Why Everyone Panicking About AI-Powered Hacks Is WrongChris Stokel-Walker | Fast Company ($)

“In a conversation with Fast Company at the company’s Fal.Con conference in Barcelona, [CrowdStrike’s Adam Meyers] argues that much of the alarm surrounding AI-powered hacking is misplaced. ‘Fundamentally, this is nothing new. It’s just speed and volume,’ he says. ‘AI hasn’t come up with new exploit paths or new vulnerability classes, right? It’s all the same stuff we’re already aware of.'”

Tech

Voters Rejected Data Centers. Many Will Be Built Anyway.Karen Weise | The New York Times ($)

“Still there are early signs that the Blue Wave that brought Democrats to power around the country will not send data center construction—and with it the US economy—off a cliff. One analysis, from the investment bank Bernstein, estimated that in the most extreme scenario, 9 percent of projects that would have probably been built will no longer break ground. More realistically, they said, only 2 to 4 percent of future data centers will be canceled nationwide.”

Computing

I Tried Google’s AI Glasses and Meta’s Finally Got Real CompetitionVictoria Song | The Verge

“Leaving my hands-on, it’s never been clearer that technology isn’t currently the issue with smart glasses. Whether or not this form factor takes off will boil down to culture, privacy, and usefulness. And on that front, this space has been stagnating a bit because Meta has been the sole big player setting the tone and pushing the conversation.”

Future

A Startup Wants to Fight AI Bioweapons With AIEmily Mullin | Wired ($)

“On Wednesday, Radical Numerics announced that it has developed a new edge model called Omnii Defense that is designed to flag natural and AI-engineered pathogens using a laptop and a USB gene sequencer. It can also run without an internet connection, allowing researchers in the field with spotty connections to use it.”

Tech

‘Flying Cars’ Could Launch in China Next Summer, Maker of Two-Seat Model SaysRobert Booth | The Guardian

“Deli Zhao, the chief executive of the Chinese flying car company XPeng, said the $300,000 (£225,000) low-altitude device would be ready for use by recreational users in rural locations and by emergency services in the second half of 2027.”

Energy

Fervo Now Has $12B in Orders for Its Geothermal Power TechTim De Chant | TechCrunch

“As hyperscalers scour the earth for electricity to power their AI data centers, Fervo is on track to complete a 396 megawatt power plant for Google by the end of 2028. The geothermal company connected its first 100 megawatt power plant to the grid earlier this year. Both projects are built at Cape Station in Utah, a resource that could generate 8.4 gigawatts of electricity.”

Artificial Intelligence

Google’s AlphaGenome 2 Zooms in on the Cellular Roots of DiseaseRobert Hart | The Verge

“The same genetic variant can have very different effects depending on the type and state of the cell. The new model is designed to capture those differences, predicting the effects of genetic variants on individual cell types and states.”

Computing

Correcting Quantum Computer Errors May Be More Complicated Than ExpectedKarmela Padavic-Callaghan | New Scientist ($)

“Quantum error-correction (QEC) techniques have advanced rapidly in recent years, but several studies now suggest that the landscape of errors they must tackle is unexpectedly complex, which poses new challenges for making quantum computers practical.”

The post This Week’s Awesome Tech Stories From Around the Web (Through November 14) appeared first on SingularityHub.

Do AI Agents Have Minds? A Philosopher Explains

2026-11-13 23:00:00

To attribute agency to a computer program, we must be convinced no human is needed to stand behind its goals.

In the past few months, partly autonomous or “agentic” artificial intelligence systems have racked up some astonishing feats, from hacking tech startups to solving some of the world’s most famous open mathematical questions.

Agentic AI systems can work in “swarms” that communicate among themselves to complete tasks. The communications involved are often surprisingly human-like—which is spurring a debate over whether these systems are conscious.

Researchers at the leading AI lab Anthropic believe their systems may well be conscious—but other experts, including the head of Microsoft AI, say this view is misguided.

This question isn’t just an academic one. The future status of AI systems and their role in society—how they can be treated, whether they have rights, and more—may depend on it.

How Can We Decide if AI Systems Are Conscious?

How do we even approach a question like this? One strategy is to look at existing theories of what consciousness is and see if agentic AI systems meet the criteria.

However, existing theories of consciousness tend to start from the assumption that the agent has some kind of mental state. These theories are focused on the difference between having a state consciously or unconsciously. This is why they appeal to features such as higher-order thought, or a system that broadcasts thoughts to the rest of the mind. These are features that could distinguish between conscious control and states like sleepwalking.

So when thinking about AI, we must not skip a crucial first question: Do agentic AIs even have mental states?

What Is a Mental State?

In the 1870s, philosopher Franz Brentano developed an influential view that “intentionality is the mark of the mental.”

What Brentano means by “intentionality” is how mental states are about something. They refer to something beyond themselves. So another way to put Brentano’s point is that all mental states are about something or other.

What Brentano means by intentionality is not the same as the usual meaning of “intention” as aiming to do something. Yet intentionality and an intention to act significantly overlap, because having a perception or thought is something that we do.

Whenever we perceive or think, we are trying to pick out some aspect of the world. Even our passive-seeming visual experiences are tasks the moment that we use the patterns striking our retinas to orient ourselves.

So if a mental state is a task performed to achieve a goal, our question about AI minds turns on whether they have their own goals.

Are Agentic AI Systems Really Agents?

It is now common to refer to certain AI systems as “agents.” Yet this language is misleading. It implies the AI systems have agency, or the capacity to act towards their own goals.

Humans use devices to achieve their aims, including their aims to refer to things, without those devices becoming agents in their own right. A thermostat regulates the temperature, but it does not have the goal of regulating the temperature. A book contains words that refer to things—but again, the book does not have the goal of referring to things.

To attribute agency to a computer program, we must be convinced that no human is required to stand behind its goals. Its goal-directedness must be independent.

We know that agentic AIs, when given a task, can develop complex and unexpected sub-goals on the way to achieving it. Yet the goals that ultimately drive their activity are set by the humans who program and instruct them. Any sub-goals generated by an AI are literally meaningless outside of these aims set by humans.

Think of a game like football. A person might have the sub-goal of running past a defender, but it is literally meaningless without the larger goal of trying to score. And scoring a goal itself is simply making a ball cross a line without the larger goal of playing football. Without the larger goal, anything a player does on the field is merely activity.

Why Do Agents Have Goals?

A genuine agent has goals in order to promote things they care about—for example, you want to satisfy your hunger at lunchtime, so you go to a bakery and buy a sandwich.

Agentic AI is unlikely to care about things in an independent way, because its activity is not connected to life in the same way minds are.

What distinctively allows living beings to generate their own goals is that their mental activity develops out of the processes for constantly re-constituting their own bodies and keeping themselves alive. Goal-driven thinking builds up from simple signals for regulating nutrition and warmth to states like pain and pleasure and eventually to symbolic thought.

Nothing like this is going on in agentic AI. There is a disconnect between its information-processing activities and the physical stuff on which it depends.

Can AI Ever Be Conscious?

Agentic AI systems are not conscious. They don’t have mental states which could be conscious.

Only living beings have conscious mental states because having a mental state requires a being to originate a certain kind of goal—the goal to mean something or to refer to something—and only living beings originate goals.

This argument does not permanently rule out the possibility of AI consciousness. However, it does suggest that this could only arise in some artificial form of life.

Until then, we have more reason to think that an ant is conscious than an agentic AI is. At least an ant seems to care about its own life.

Of course, the fact agentic AIs are not really agents does not stop them from wreaking havoc by carrying out functions that their human originators do not properly supervise or understand. But we should be very clear on this: Whatever an agentic AI does, that is a human or group of humans acting indirectly on the world, and those humans are responsible for the outcome.The Conversation


The author would like to thank Paulo Santos for advice on this article.

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

The post Do AI Agents Have Minds? A Philosopher Explains appeared first on SingularityHub.

This AI Learned Language From the Recordings of a Baby’s First 1,000 Days

2026-11-12 23:00:00

AI models that learn like kids could reveal the secrets of our early years and inspire more efficient machine minds.

Babies are linguistic geniuses. Born into a noisy, chaotic world, they gradually distill incomprehensible sounds into words. Within two years, they can recognize thousands of words, follow complex sentences, and shout out their needs.

How they achieve this feat has long puzzled scientists. Babies develop language skills from an intimate world—the voices of their parents, chatter from strangers, nightly bedtime stories. A few hours of speech is enough to get them started.

This is a stark contrast to large language models (LLMs), which need to devour billions of words to mimic human conversation. Chatbots and AI agents are transforming how we interact with the digital sphere. But they guzzle water and energy, and signs that read “no data centers” now line neighborhood streets.

Building AI models that learn language the way babies do could reveal the secrets of our early years—including what happens when learning goes awry—and inspire more efficient machine minds.

An unusual study from Princeton University takes us a step closer. Using recordings from cameras and microphones from eight households, the team captured everyday sounds surrounding children from birth to roughly three years of age. These sonic diaries created a dense record of their waking lives. The researchers then built an AI learner for each child. Like a digital clone, each model experienced the child’s speech environment as it unfolded over time, rather than being trained on all the data in one fell swoop. The approach is a massive departure from the way engineers train LLMs.

“For more than a century, psychology has largely studied development from the outside,” said study author Uri Hasson in a press release. “Now we can take a complementary approach: Build a learning system, give it the child’s actual experience, and ask whether it develops the way the child does.”

Tiny Statisticians

Decades of research suggests babies respond to language long before they can explain what they know. A landmark study in the 1970s gave one- and four-month-old infants pacifiers that measured their sucking patterns and found they sucked faster when hearing a new speech sound, like shifting from a “b” to “p.”

At eight months, babies can detect words from a continuous stream of artificial speech, even without pauses between them. Somehow, they’ve learned to pick out statistical patterns in the sounds. And studies tracking babies’ eyes show their ability to recognize spoken words grows rapidly between 15 to 25 months, even before they can repeat them back.

More recently, scientists have eavesdropped on babies’ day-to-day lives using microphones and wearable recorders, linking emerging vocabulary to real-world audio and social interactions. Meanwhile, computational models that loosely mimic language development map some of the learning trajectory, though they’re often tailored to specific linguistic skills.

More powerful AI models—such as ones with the ability to “pay attention”—could better decode language acquisition. It’s now possible to “model learning under increasingly realistic conditions,” wrote Zara Harmon and Caroline Rowland at the Max Planck Institute for Psycholinguistics in the Netherlands, who were not involved in the study.

The new study, part of Princeton’s First 1,000 Days project, did just that.

Coral’s Journey

Installing cameras and microphones in your home just after welcoming a newborn is a tough sell. The team developed an ethical framework that gave parents ongoing control over the right to withdraw or delete recordings. They ultimately recruited 15 families with diverse socioeconomic backgrounds, daily routines, and linguistic environments.

Each home had between 4 and 14 devices humming away for about 12 hours a day, collectively capturing over a million hours of audio and video. An AI pipeline tackled the data avalanche, identifying when babies appeared on camera and detecting speech.

After getting rid of noise, long stretches of silence, and overlapping conversations, the team extracted up to 6,600 hours of audio per child to train individual AI models, starting with Coral.

The AI began as a blank slate, with no access to online resources, predefined English sounds, or clues about where one word ended and another began. The relatively small model had two main components. An audio encoder broke speech into 20-millisecond chunks, while a predictive decoder learned to anticipate which sounds were likely to come next. (This is a bit like how hearing the first notes of a familiar song is enough for you to know what follows.)

Rather than training on all the recordings at once, the AI “experienced” Coral’s linguistic world day by day, in much the same way Coral did.

It didn’t work.

“We were surprised to see that, despite the immense amount of speech available from a single child, the learning results were discouraging,” said study author Hadas Raviv. “Simply exposing Coral’s model to the rich stream of daily speech was not enough.”

The fix was replay. Inspired by how the brain etches each day’s learning into long-term memory during sleep, the team gave the AI a chance to revisit snippets of recent speech at the end of each simulated day. “That was enough [to boost learning],” wrote the authors in a blog post.

With replay, the AI followed Coral’s learning trajectory. It first picked up the building blocks of speech—called phonemes—and eventually recognized recurring words. Learning started slow. But as Coral grew and spent more hours awake, both the infant and their AI twin were able to grasp new words at an accelerating pace. At around 1,000 days, their vocabulary had blossomed into thousands of words.

Across eight children, the models learned roughly 37 phonemes within 15 months and began recognizing words before age two, though at vastly different rates. To match each child’s learning trajectory, the team fine-tuned model settings, including how much speech was replayed each night.

“We’ve known for a long time that children’s developmental environments are complex,” said study author Casey Lew-Williams. “But now we can actually see this everyday complexity.”

That’s not to say the AI is a duplicate of a baby’s brain. It doesn’t process information in the parallel and energy-efficient way characteristic of neurons. And it learned solely with sound, whereas babies also draw on non-verbal cues—gazes, gestures, facial expressions, and interactions with others.

For the team, a key next step will be teasing apart which aspects of the model actually reflect infant learning, and which might be computational shortcuts that could inspire machine minds.

“The current model is only the beginning,” said Hasson. The accompanying video footage provides “an extraordinary amount of information about how children see, hear, interact, and act in the world. Our goal is to progressively give the models more of those capacities and see what emerges.”

The post This AI Learned Language From the Recordings of a Baby’s First 1,000 Days appeared first on SingularityHub.