Robotics

How Pokémon Go Is Giving Delivery Robots an Inch-Perfect View of the WorldWill Douglas Heaven | MIT Technology Review ($)

“Niantic Spatial is using that vast and unparalleled trove of crowdsourced data—images of urban landmarks tagged with super-accurate location markers taken from the phones of hundreds of millions of Pokémon Go players around the world—to build a kind of world model, a buzzy new technology that grounds the smarts of LLMs in real-world environments.”

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

A single shot protected mice from the protein gunk implicated in Alzheimer’s disease.

Alzheimer’s disease and cancer have something in common: They’re hard to treat.

Despite decades of research, both still plague humanity, robbing people of longer, healthier lives. In Alzheimer’s, a protein called amyloid forms toxic clumps in the brain. Eventually, neurons supporting memory, decision-making, and movement wither.

Whether amyloid clumps cause Alzheimer’s is still hotly debated. Drugs that clear the proteins have slowed disease progress, but only mildly. The FDA recently approved two such drugs for early stage patients. The approvals were controversial, however, due to risks, like brain bleeds.

As scientists have struggled with Alzheimer’s, blood cancer treatments have undergone a revolution thanks to a therapy called CAR T. The treatment genetically engineers a patient’s T cells to hunt and destroy a handful of previously untreatable blood cancers.

Taking a page from CAR T, a team at the Washington University School of Medicine, St. Louis has now transformed brain cells, known as astrocytes, into amyloid-gobbling machines.

A single injection into mice with Alzheimer’s prevented the formation of amyloid clumps in the brain during the disease’s early stages. In mice whose brains were already riddled with the toxic plaques, the treatment cut amyloid levels roughly in half.

“This study marks the first successful attempt at engineering astrocytes to specifically target and remove amyloid beta plaques in the brains of mice with Alzheimer’s disease,” study author Marco Colonna said in a press release.

Mysterious Origins

Alzheimer’s has baffled scientists for over a century. Genetics play a role. Some versions of a gene called APOE protect the brain against the disease; others exacerbate it. Inflammation in the skin, lungs, or gut may be an early trigger. Damage to waste-cleaning cells that normally wash toxic proteins out of the brain could also contribute to symptoms.

But the reigning theory of what causes Alzheimer’s is amyloid buildup. These sticky proteins aren’t inherently evil. At low levels, they tweak how neurons connect to make memories, support brain healing after injury, and may ward off infections.

In Alzheimer’s, however, amyloid clumps into toxic waste. The brain’s immune cells can initially clear them up. But amyloid eventually overwhelms these cells and causes them to spew inflammatory molecules that exacerbate the disease.

Amyloid clumps ignite a molecular cascade, leading to brain inflammation and ultimately neuronal loss, wrote Jake Boles and David Gate at Northwestern University, who were not involved in the study.

Existing drugs target amyloid proteins with antibodies that either physically prevent amyloid from clumping or mark the proteins for destruction by the immune system. But patients need regular treatment, and the risk of brain bleeds or stroke-like symptoms cause many to opt out.

Engineering brain cells to do the job could be a lasting solution.

In CAR T, scientists genetically engineer T cells to produce special proteins, or CARs, that latch onto cancer cells and destroy them. These have two sections: One outside the cell to recognize targets, and another inside the cell to trigger a biological effect. For CAR T cells, the interior trigger releases a blast of molecular bullets that destroys cancers or pathogens—and in one study, even amyloid proteins in mice. Another recent study modified a different type of immune cell to clear the toxic clumps, although its safety is still unknown.

These test cases suggest that CAR technology could reduce toxic amyloid buildup. But they used cell types that weren’t native to the brain and needed additional chemicals to keep them working. This mismatch could make treatment more complex and risk side effects.

In contrast, “genetically engineering resident brain cells to target…[amyloid]…could circumvent these persistence challenges,” wrote Boles and Gate.

Meet CAR-A

The team turned to astrocytes. These support cells help repair brain injuries, provide nutrients to neurons, and impact memory and cognition. They also eat up dead cell fragments and some proteins, though they’re less efficient at this task than the brain’s immune cells.

To boost astrocytes’ appetite for amyloid, the team tested multiple CAR designs. Genetic instructions for two of these were then packaged into a harmless virus for delivery and injected into the veins of mice modeling Alzheimer’s disease.

Inside the brain, the treatment transformed naturally occurring astrocytes into CAR-A cells. In mice two and a half months old, roughly late adolescence in human years, the cells prevented amyloid clumps for at least three months. And a single shot slashed amyloid levels in half for older mice already suffering buildup. The treatment also protected neurons from further damage.

Though tailored to astrocytes, the gene therapy also caused the brain’s immune cells to more readily devour amyloid plaques, swept away malfunctioning ones, and lowered inflammation. Shifting some of the amyloid-clearing burden from immune cells to astrocytes could create a less toxic environment in the Alzheimer’s brain, wrote Boles and Gate.

Anti-amyloid antibodies in current Alheimer’s drugs struggle to tunnel through the blood-brain barrier. But CAR-A cells are made inside the brain with minimal blood vessel damage, which could also lower the chances of deadly side effects.

“Consistent with the antibody drug treatments, this new CAR-astrocyte immunotherapy is more effective when given in the earlier stages of the disease,” said study author David Holtzman. “But where it differs, and where it could make a difference in clinical care, is in the single injection that successfully reduced the amount of harmful brain proteins in mice.”

Long Road Ahead

While the new approach may improve on current treatments, it ran into the same problem: The shot reduced amyloid clumps without significantly improving memory or mood.

The results mirror those seen in anti-amyloid drug trials. It could be because over-zealous CAR-A cells nibble on neurons too and destroy their connections, offsetting any benefit. Or more fundamentally, it could be that targeting amyloid alone isn’t enough. Tau proteins, for example, also aggregate in neurons as the disease goes on, and higher levels are tied to mental decline.  

Early treatments targeting tau proteins have universally failed. But the CAR-A platform could be redesigned to go after tau in a combination therapy to wipe out both toxic proteins.

Beyond Alzheimer’s, a similar strategy may be work in other brain diseases to kill brain cancers. The team is now fine-tuning their designs to better detect targets.

“As CAR technologies mature and the ability to selectively neutralize toxic proteins improves, these approaches hold substantial promise for AD [Alzheimer’s disease] and other neurodegenerative disorders,” wrote Boles and Gate.

The post These Genetically Engineered Brain Cells Devour Toxic Alzheimer’s Plaques appeared first on SingularityHub.

Finding water, generating power, and sheltering astronauts from radiation are just a few of the challenges NASA must solve.

A US Senate committee has directed NASA to begin work on a moon base “as soon as is practicable.” Under legislation advanced by the Senate lawmakers, the outpost would serve as a science laboratory and proving ground, where astronauts would develop the capabilities to live and work beyond Earth’s orbit.

A recent executive order issued by the White House directs NASA to establish the initial elements of a permanent moon base by 2030.

Since 2017, Artemis has been the NASA-led program working towards a sustained human presence on the moon. This year, it will send astronauts around the moon for the first time in more than half a century. And following a shake-up of Artemis announced in late February, the space agency plans to greatly increase the frequency of Artemis missions and return humans to the lunar surface in 2028.

A vote will now decide whether Senate legislation, known as the NASA Authorization Act of 2026, is passed to Congress, where a second bill is also circulating. The bills, which both break down this year’s funding for specific NASA programs, will be reconciled and voted on in both houses to become law.

Underlying some of the announced changes is a deepening concern in Congress and the current administration about the challenge rival powers pose to US leadership in space. A Chinese-Russian led moon outpost known as the International Lunar Research Station is under development.

A one page summary accompanying the Senate bill calls for a US base “so we can get there before the Chinese” and to “dominate the Moon, control strategic terrain in space, and write the rules of the 21st century.”

Site Selection

The American habitat will be located at the moon’s south pole, a strategically important location which harbors valuable resources such as water ice. The water could support habitation systems at a lunar outpost and be turned into rocket propellant for onward exploration.

Where exactly the base is located will depend on the terrain, how much sunlight the site receives, how extreme the temperatures are, how easily astronauts can communicate with Earth, and their access to resources such as water. The rim of a 21-kilometer-wide depression known as Shackleton Crater (which may hold abundant ice deposits) and a flat-topped mountain called Mons Mouton are among the leading candidates. The leading locations combine several favorable factors.

At high latitudes, such as the lunar poles, elevated crater rims can receive near-constant solar illumination. This makes them more thermally favorable than many sites at the equator, providing a consistent supply of solar power. However, the strategic value of these sites lies in what are called permanently shadowed regions (PSRs). These impact craters, untouched by sunlight for billions of years, are believed to contain the water-ice deposits.

While the south pole remains a primary focus in upcoming missions, other targets near the equator, such as Marius Hills and Mare Tranquillitatis, offer alternative advantages. These regions feature massive underground lava tubes formed by ancient volcanic activity that can act as natural shields against solar radiation and micrometeorite bombardments. They could insulate human outposts against extreme swings in temperature from 127° Celsius to -173° Celsius.

The interiors of lunar lava tubes are estimated to remain at about 17° Celsius year-round, making them ideal sites for human bases. However, unlike at the lunar poles, water in these regions is typically trapped as molecules within volcanic glass beads or minerals. Extracting this water to sustain human activities would require intensive heating and significant technological development.

Powering an outpost

The moon’s day-night cycle means that a given point on the lunar surface sees roughly 14 Earth days of continuous daylight followed by 14 days of darkness. While solar power is a viable entry point, it cannot sustain a permanent human presence through the freezing lunar night. To achieve the 2030 mandate for a “sustained presence” NASA and the Department of Energy are developing nuclear fission reactors as a potential source of energy.

They have been working on 40-kilowatt-class reactors that are designed to be launched from Earth in an inert state and activated upon arrival. To protect the crew from radiation, the reactors will likely be placed at a distance or buried within the lunar regolith (soil), which serves as a natural radiation shield.

The deployment of lunar fission reactors raises practical governance questions under existing international space law. The US-led set of rules for operating in space, known as the Artemis Accords, establishes a framework for peaceful cooperation.

It calls for transparency about space agencies’ activities on the surface and proposes safety zones around nuclear infrastructure. However, this approach conflicts with the Outer Space Treaty of 1967, which guarantees the right of all nations to have unrestricted access to all areas of celestial bodies.

Given that energy security is a strong prerequisite for successful habitation systems, there is a clear need for the governance of the storage and disposal of the materials used for nuclear fission on the lunar surface.

Initial Assembly

A lunar base would likely be built up in stages. Early missions would use satellites and autonomous rovers to study the lunar surface, identify areas rich in resources, and confirm the presence of water. Under a 2030s timeline, robotic missions could be sent ahead to prepare landing sites by leveling the ground and melting the dusty surface into harder landing pads. This would help reduce the damage caused by highly abrasive lunar dust kicked up during landings.

The habitats themselves would probably be built by connecting different modules—a bit like the International Space Station. Current designs favor modules that can be reduced in size for transportation and then expanded after landing. One way to do this is with inflatable structures.

Later, more permanent architectures may use microwaves or lasers to sinter or melt the lunar regolith into solid structures. This would create protective shells around base modules to protect them against micrometeorites and cosmic radiation.

The moon serves as a testbed for the life-support, power, and robotic systems required to support human missions on Mars and other destinations in deep space.

The fiscal implications of sustained operations on the lunar surface also require a more realistic assessment of funding. With NASA’s topline budget remaining largely flat, the higher cadence (frequency) of lunar missions outlined in NASA’s changes to Artemis would increase pressure on agency resources.

This may intensify competition with existing science and Earth observation priorities, but it also strengthens the case for greater commercial participation and international cost-sharing. If these financial pressures can be managed effectively, the long-term legacy of sustained lunar surface operations could be a more durable framework for funding space exploration.

The coming decade will test not only our ability to operate through the lunar night, but also our capacity to build the logistical, legal, and cooperative frameworks needed for a durable human presence beyond Earth.

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

The post The US Plans to Break Ground on a Permanent Moon Base by 2030. Here’s What It Will Take. appeared first on SingularityHub.

The study focused on safety, but the results offer hope the approach could give kids a chance to walk.

Michelle Johnson was 20 weeks into her pregnancy when she learned her unborn son had spina bifida. Because his spine hadn’t fully sealed, the spinal cord was left protruding from a gaping hole. Without surgery, he would face a lifetime of disabilities.

So she jumped at the chance to enroll in a small experimental trial for the condition at the University of California, Davis. The treatment combines fetal surgery, an existing approach, with a dose of stem cells to spur healing.

Now four years old, Johnson’s son Tobi can walk and lacks symptoms such as loss of bladder and bowel control. “Tobi’s physical and mental abilities are nothing short of a miracle,” she said in a press release.

Tobi is one of six children in the CuRe trial, the first study to test if using stem cells to repair tissue in fetuses with spina bifida is safe. Delivered by a small patch sewn onto the damaged site, the stem cells protected the spinal cord from inflammation and helped the wound heal. None of the babies or mothers suffered short-term side effects, like unwanted tissue growth or cancer.

With so few participants, it’s too early to tell how the approach will pan out as the children grow. But thanks to the promising safety profile, the FDA has approved the enrollment of more pregnant women with the same diagnosis.

“This is a major step toward a new kind of fetal therapy, one that doesn’t just repair but potentially helps heal and protect the developing spinal cord,” study author Aijun Wang said in the press release.

CuRe joins other attempts to tackle diseases with stem cells in the womb. Although a very young field, the approach could slow, halt, or cure a number of diseases before babies are born.

A Head Start

Spina bifida is a condition where the spine or spinal cord doesn’t properly seal during development. One in 2,875 newborns are affected in the US every year. In its most severe form, cerebrospinal fluid—a liquid that surrounds the brain and washes out toxins—builds up, causing progressive damage to the fetal spinal cord, lifelong movement problems, and even paralysis.

The condition was first treated after birth, when surgeons would close the defect. But by then, the damage was done. Surgery before birth could stave off symptoms, an idea validated in a 2011 trial. Yet over half of treated babies still struggled to walk without help, likely because injured neurons in the fetuses’ brains and spinal cords didn’t have the chance to heal.

Stem cells spur regrowth by releasing protective nutrients, and the fetal environment is uniquely suited to the cells. The team wondered if adding them could improve prenatal surgery.

They began testing the idea around 2012 using induced pluripotent stem cells. This is a type of stem cell made from skin or other mature cells using a chemical cocktail. Taking this approach could provide a nearly unlimited supply of stem cells. But it didn’t work.

After years of trial and error, the team found success with stem cells derived from placentas. The cells protected neurons from injury and encouraged their growth in lab dishes. They also healed defects in a lamb model of spina bifida. Newborns receiving the cells along with prenatal surgery could stand up and walk; those who only had surgery couldn’t.

Stem cell therapy appeared promising. But for unborn babies, it could carry risk. Since the cells come from donors, they might spark immune reactions. They might also trigger abnormal tissue growth, or even cancer. Because stem cell treatments are rarely used in the womb, little is known about their effects on pregnancy or the overall health of mother and baby.

Landmark Trial

The first stage of the CuRe trial focused on these safety concerns.

The team seeded a small patch with stem cells derived from donated placental tissue. To help the cells integrate, the researchers designed the patch to mimic conditions normally surrounding cells.

Surgeons made a small opening in the uterus at 24 to 25 weeks into pregnancy and gave the fetus a small dose of painkillers and muscle relaxers. They then placed the stem cell patch onto the exposed spinal cord and sutured the gap closed.

The trial closely monitored six babies, including Tobi, for possible side effects. After delivery by C-section, none had complications, such as leaking cerebrospinal fluid, infection, or signs of cancer. In all cases, the treatment prevented parts of the brain from slipping into the neck, and none required a shunt—a small tube used to drain excess fluid from the brain—an encouraging sign of success.

The team turned to stem cells, they wrote, because the cells can lessen brain inflammation and brain cell death. At the same time, they pump out growth proteins that “support neural tissue preservation and spinal cord integrity.”

The researchers designed the study to evaluate safety not determine whether stem cells enhance the surgery’s results. But Tobi’s remarkable recovery is a hopeful sign that the cells do make a difference. Because spina bifida is structural, treating it before permanent damage occurs could make the therapy a “one-and-done” fix.

The study joins the growing prenatal use of stem cells in conditions such as thalassemia, a blood disorder, and osteogenesis imperfecta, also known as brittle bone disease. Early clinical trials have shown promise, but regulators haven’t yet approved any treatments.

“Putting stem cells into a growing fetus was a total unknown. We are excited to report great safety,” said Diana Farmer, a study author and lead investigator for the CuRe trial. “It paves the way for new treatment options for children with birth defects. The future is exciting for cell and gene therapy before birth.”

The team is actively recruiting more pregnant women for the trial’s second phase. They’ll track the children’s growth and health up to age six to assess brain and cognitive development, motor skills, and other growth milestones.

If the treatment proves successful, longer monitoring may be needed. Spina bifida can increase the risk of kidney disease and certain cancers later in life, and it’s unclear if the stem cells could cause problems months or years down the line.

Uncertainties aside, Johnson is happy to be participating in the trial. “We are forever grateful for the many health professionals who supported Tobi’s journey and continue to watch him conquer the world,” she said.

The post In a First, Researchers Use Stem Cells and Surgery to Treat Spina Bifida in the Womb appeared first on SingularityHub.

Which side has the advantage will depend less on raw model capabilities and more on who adapts fastest.

Cybersecurity is an endless game of cat and mouse as attackers and defenders refine their tools. Generative AI systems are now joining the fray on both sides of the battlefield.

Though cybersecurity experts and model developers have been warning about potential AI-powered cyberattacks for years, there has been limited evidence hackers were widely exploiting the technology. But that is starting to change.

Growing evidence shows hackers now routinely use the technology to turbocharge their search for vulnerabilities, develop new code exploits, and scale phishing campaigns. At the same time, AI firms are building defensive security measures directly into foundation models to keep pace with attackers.

As cybersecurity becomes more automated, corporations will be forced to rapidly adapt as they grapple with the security of their products and systems in the age of AI.

A recent report by Amazon security researchers highlighted the growing sophistication of hackers’ AI use. The researchers wrote that Russian-speaking attackers used multiple commercially available generative AI services to plan, manage, and conduct cyberattacks on organizations with misconfigured firewalls in over 55 countries this January and February.

The attack targeted more than 600 systems protected by FortiGate firewalls and worked by scanning for internet-exposed login pages—these are essentially front doors leading into private company networks—and attempting to access them with commonly reused security credentials. Once inside, they extracted credential databases and targeted backup infrastructure. This activity suggests they may have been planning a ransomware attack.

The researchers report the attack was largely unsuccessful but nonetheless highlighted how much AI can lower the barrier to large-scale attacks. Despite being relative amateurs, the group “achieved an operational scale that would have previously required a significantly larger and more skilled team,” they wrote.

In the most vivid demonstration of AI’s hacking potential, a research prototype created by a New York University researcher known as PromptLock used large language models to create an entirely autonomous ransomware attack.

The malware used AI to generate custom code in real time, scour the target system for sensitive data, and write personalized ransom notes based on what it found. While the tool was only a proof of concept, it highlighted the mounting threat of fully automated malware attacks.

A recent report from security firm CrowdStrike found that AI is also making attackers significantly more nimble. They discovered that average breakout times—the window between when an attacker first breaches a network and when they move into other systems—fell to just 29 minutes in 2025, 65 percent faster than in 2024.

In November, Anthropic also claimed they had detected a Chinese state-linked group using the company’s Claude Code assistant to conduct a large-scale espionage campaign. The group used jailbreaks—prompts designed to bypass a model’s safety settings—to trick Claude into carrying out the attacks. They also broke the campaign into smaller sub-tasks that looked more innocent.

The company claimed the hackers used the tool to automate between 80 and 90 percent of the attack. “The sheer amount of work performed by the AI would have taken vast amounts of time for a human team,” the company’s researchers wrote in a blog post. “At the peak of its attack, the AI made thousands of requests, often multiple per second—an attack speed that would have been, for human hackers, simply impossible to match.”

But while AI is reshaping the offensive cybersecurity landscape, defenders are deploying the tools too. In February, Anthropic released Claude Code Security, which can scan systems for vulnerabilities and propose fixes automatically. The tool can’t carry out real-time security tasks like detecting and stopping live intrusions, but the news nonetheless sent stocks in traditional cybersecurity firms plummeting, according to Reuters.

Cybersecurity vendors are also embedding AI into their defensive platforms. CrowdStrike recently launched two new AI agents, one designed to analyze malware and suggest how to defend against it and another that actively combs through systems for emerging threats. Similarly, Darktrace has introduced new AI tools designed to automate the detection of suspicious network activity.

But perhaps one of the most promising applications for the technology is using it like a hacker to proactively probe defenses. Aikido Security recently released a new tool that uses agents to simulate cyberattacks on each new piece of software a company creates—a practice known as penetration testing—and automatically identify and fix vulnerabilities.

This could be a powerful tool for defenders, Andreessen Horowitz partner Malika Aubakirova wrote in a blog post. Traditional penetration testing is a labor-intensive process relying on highly skilled experts in short supply. Both factors seriously constrain where and how such testing can be applied.

Whether AI ends up advantaging attackers or defenders will likely depend less on raw model capabilities and more on who adapts fastest. So, it seems the unending game of cat and mouse that’s characterized cybersecurity for decades will continue much the same.

The post Hackers Are Automating Cyberattacks With AI. Defenders Are Using It to Fight Back. appeared first on SingularityHub.

Artificial Intelligence

Watershed Moment for AI–Human Collaboration in MathBenjamin Skuse | IEEE Spectrum

“The 8-dimensional sphere-packing proof formalization alone, announced on February 23, represents a watershed moment for autoformalization and AI–human collaboration. But today, Math, Inc. revealed an even more impressive accomplishment: Gauss has autoformalized Viazovska’s 24-dimensional sphere-packing proof—all 200,000+ lines of code of it—in just two weeks.”

The post This Week’s Awesome Tech Stories From Around the Web (Through March 7) appeared first on SingularityHub.

AI-powered digital assistants can do many complex tasks on their own. But who takes responsibility when they cause harm?

Scott Shambaugh, a volunteer maintainer for a programming code library called Matplotlib, recently described a surreal encounter with an autonomous AI agent—a digital assistant created with a platform called OpenClaw. After he rejected a code contribution submitted by the agent, it researched and published a personalized “hit piece” against Shambaugh on its blog. The post portrayed an otherwise routine technical review as prejudiced and attempted to shame Shambaugh publicly into allowing the submission. (The human responsible for the agent later contacted Shambaugh anonymously, telling him that the bot had acted on its own with little oversight.) The account of this incident spread quickly through the software developer ecosystem and has been amplified by independent observers and media coverage.

Treat the Matplotlib event as a one-off if you like. The deeper point, however, is hard to miss and should not be ignored: AI agents are becoming public actors with reach into the real world, and with real-world consequences. In the past, they could only do mundane tasks such as answering customer service questions or data processing. Now, they are capable of posting and publishing content—and persuading and pressuring humans—all at machine speed. They can make phone calls, file work orders, create cryptocurrency wallets, and operate across different applications, with enormous reach and at tremendous scale—the kind of stuff that used to require a human with fingers typing at a keyboard.

Reporting around OpenClaw and the chatroom Moltbook (which is for AI agents only) is capturing the new reality. OpenClaw enables AI agents to have persistent memory, gives them broad permissions, and allows large-scale deployment by users who often do not understand the security and governance implications.

We are the humans who are responsible for the law, ethics, and institutional design, and we are behind the curve. We need new language and governance to deal with this new reality, and principles from the field of medical ethics can provide a framework for doing so.

When an agent does something that is harmful or coercive in public, our reflex seems to be to ask the wrong questions: Is the AI a person? Should it have rights? The AI personhood debate is no longer fringe. Legal scholars and ethicists are mapping out arguments and precedents. States are writing legislation to prohibit AI personhood. Some arguments maintain that if an entity behaves like something within our moral circle, we may owe it moral consideration. Others argue that assigning rights or personhood to machines confuses moral standing with engineered performance and diffuses responsibility away from humans.

We are the humans who are responsible for the law, ethics, and institutional design, and we are behind the curve.

As a bioethicist and specialist in neurointensive care, I deal directly with human moral agency and the essence of personhood when treating patients. As a researcher, I study the use of synthetic personas animating AI agents and their use as stand-ins of human counterparts. Here is the problem that I see: Granting AI personhood, even in limited capacity, risks formalizing the most dangerous escape hatch of the agentic era—what I will call responsibility laundering. This allows us to say, “It wasn’t me. The agent/bot/system did it.”

Personhood should not be about metaphysics or claims about an inner nature. It is a legal and ethical instrument that allocates rights and accountability. It is a social technology for assigning standing, duties, and limits on what can be done to an entity. If we grant personhood to systems that can act persuasively in public while remaining functionally unaccountable, we create a new class of actors whose harms are everyone’s problem but nobody’s fault.

There is a key concept here that we can use from my field, medicine. In clinical ethics, some decisions are justified yet still leave a “moral residue,” a kind of emotional echo or sense of responsibility that persists after the action because no options fully satisfy competing obligations. This residue accumulates over time, causing a “crescendo effect” that occurs even when conscientious clinicians are doing their best inside imperfect systems. That remainder matters because it reveals something basic about moral life, namely that ethics is not only about choosing; it is about owning what remains afterwards.

This is the moral remainder problem for generative and agentic AI. A modern AI agent can generate reasons for an action; it can simulate regret and plead not to be turned off. But it cannot truly bear sanction, repair the damage, apologize, ask forgiveness, or navigate the aftermath through which moral responsibility is created and enforced. To treat it as a moral person confuses persuasive performance with accountable standing. It also tempts institutions and people into delegating their own answerability to a bot.

What can we, as humans, do instead?

We need a vocabulary that is built for agents that are public actors, one that allows bounded autonomy without granting personhood. Let’s call it authorized agency. Authorized agency starts with an authority envelope: a bounded scope of what an agent is permitted to do, to whom, where, with what data, and under what constraints. To say “the agent can use email” is not sufficient. However, an acceptable scope would be to say that the agent can send only certain categories of messages to particular recipients for a specific set of purposes, and that it must stop what it’s doing or escalate to its owner under a particular set of conditions.

Next comes the human-of-record, the owner, a publicly named person who authorized that envelope and remains answerable when the agent acts, even if it becomes capable of acting outside the envelope. An actual human being whose authority is real—not “the system” or “the team.”

What follows is interrupt authority: the absolute right of the human owner to pause or disable an agent without using moral bargaining or being subject to institutional penalty. This is grounded in formal research on AI safety showing that agents that are pursuing objectives can have incentive to resist being shut down. An agent programmed to maximize its utility cannot achieve its goal if it is shut off. In the public sphere, interrupt authority is the difference between a delegated tool and a coercive actor.

We need a vocabulary that is built for agents that are public actors, one that allows bounded autonomy without granting personhood.

Finally, we need a traceable path from the agent’s action back to the person who authorized it, called an answerability chain. If an agent publishes, messages, or pressures someone in public, we must be able to know: Who authorized this scope? Who could have prevented it? And who must be responsible for the action afterward? In this framework, the answer to these questions is the person who carries the moral remainder. Work in AI ethics has warned about responsibility gaps where the system’s actions outpace our ability to assign accountability.

Some legal scholarship has started exploring how to build agents that are constrained by governance and law without needing to pretend the agent itself is a legal subject, in the human sense. This is promising because it treats assigning personhood as the wrong idea and accountability as the correct one.

The Matplotlib story, whether the first documented case of an AI agent attempting to harm someone in the real world or the first to capture public attention, is a warning. Agents will not only automate tasks. They will generate narratives, apply pressure, and shape people’s lives and reputations. They will act in public at machine speed with unclear ownership.

If we respond by debating whether agents deserve rights, we will miss the emergency entirely. As they continue to increase their reach in the real world, the urgent task is to ensure that responsibility also remains within reach. Don’t ask whether an agent is a person. Ask who authorized it, what it was allowed to do, who can stop it, and most importantly, who will answer when it causes harm.

This article was originally published on Undark. Read the original article.

The post Autonomous AI Agents Have an Ethics Problem appeared first on SingularityHub.

Spexi’s crowdsourced drone fleet has mapped over 5 million acres in 200 cities around Canada and the US.

Gaspard-Félix Tournachon, popularly known as “Nadar,” took the first known aerial photographs using a camera attached to a hot-air balloon just outside Paris in 1858. Ever since, technologists have been developing increasingly sophisticated ways to capture high-altitude images of Earth.

In the First World War, military intelligence pushed the technology from artistic novelty to real-world use. Today, everything from urban planning and insurance underwriting to disaster response relies on detailed, high-resolution, and often 3D images of our planet. For emerging fields like autonomous robotics and augmented reality, making a digital copy of the physical world is one of the century’s most consequential infrastructure projects.

While more traditional aerial imagery relies on airplanes, satellites, and the occasional pigeon, today’s industry is also turning to low-cost drones.

Bill Lakeland, CEO and cofounder of Canadian drone imaging company Spexi, says improvement in consumer drones over the last decade is reshaping aerial imagery. In an interview with Joseph Raczynski, Lakeland details how low-cost drones are disrupting older methods involving airplanes and satellites.

“We’re getting better data out of micro-drones than what we get out of a $2 million mapping camera. The time has arrived,” he says.

According to Spexi, because off-the-shelf drones fly low, they can produce imagery at a resolution 30 times higher than satellites. Drones are also more cost-efficient and less time-consuming than airplanes. This means they’re quickly achieving workhorse status.

What’s notable about Spexi is that instead of operating their own fleet of vehicles, they work with a decentralized network of hobbyists. Anyone with a drone can download the company’s software to autonomously fly a pre-determined flight path and capture the necessary images on demand. According to Lakeland, each flight covers roughly 25 acres in about seven minutes. A pilot can expect to earn around $10 per flight, with some earning hundreds of dollars a day. To date, Spexi’s network of over 8,000 drone pilots has mapped more than 5 million acres across more than 200 cities in Canada and the United States.

With this data, Spexi aims to build a sort of Google Street View from the sky. But consider that Google’s rumored investment building Street View was over a billion dollars as they gathered data with car-mounted cameras. While it was a different type of information, Google’s acquisition of Waze in 2013 gave them access to crowdsourced map data Waze collected for free from 40 million users. While Spexi’s approach isn’t free, it appears to be skipping the relatively more expensive in-house phase for something closer to Waze’s approach.

The impact of having up-to-date maps of Earth from above is sure to be significant.

In a Bloomberg profile, Lauren Rosenthal writes that forestry professionals are already leveraging drone data to help prevent wildfires. They’re using images from Spexi to train AI models that can alert forest managers to areas of high fire risk. Similarly, insurance companies are turning to Spexi for risk assessment, underwriting, and claims processing.

In augmented reality and robotics, drone data can also produce 3D maps for visual positioning systems. Author and Wired cofounder, Kevin Kelly, calls this digital twinning project the “mirrorworld.” Some observers suggest it’s one of the most significant technology projects of the age. Using this type of 3D training data, companies are also building generative AI world models, which help AI understand the physical world.

The rise of drone imaging doesn’t yet signal the end of other approaches, and it’s not clear how much of the industry will be serviced by drones versus other means. The race to corner the satellite imaging market is also heating up. In one sense, Tournachon’s 19th century art project was no different than today’s image gathering; attach a camera to a flying object and take pictures of Earth. The main distinction, however, is that these images have evolved from mere curiosity to a digital asset powering the modern world.

The post Thousands of Everyday Drone Pilots Are Making a Google Street View From Above appeared first on SingularityHub.

The technology, which uses genetically engineered T cells, could target nearly two dozen different solid cancers with one treatment.

Few cancer treatments are as ferocious as CAR T cell therapy.

Often derived from a patient’s own immune cells, CAR T cells are genetically modified to hunt down and destroy cancer cells. The FDA has approved treatments for deadly blood cancers, and treatments tackling autoimmune diseases and preventing tissue scarring in the heart and kidneys have shown promise.

Yet CAR T has struggled against solid tumors. Over 85 percent of cancers fall into this category. Solid tumors have an arsenal of sneaky tactics to evade or deactivate CAR T cells, eventually undermining the treatment.

This month, a Columbia University team broke through one of the barriers with an upgraded design. They engineered a new, ultra-sensitive protein “hook” that seeks out CD70, a protein that dots the surfaces of multiple types of solid cancer cells—but at vastly different levels.

“Some molecules have been identified that are found in 25%, 50%, or 75% of tumor cells,” said study author Michel Sadelain in a press release. “Though a therapy directed at those targets might be successful…you can’t cure somebody if you just eliminate a small fraction or even 90% of their tumor.”

In tests, the supercharged cancer-killers, dubbed HIT cells, detected and wiped out cancer cells with extremely low levels of CD70—so low that the protein was undetectable using traditional methods. In kidney, ovarian, and pancreatic cancer grown from patients’ cells in petri dishes and in mouse models, HIT completely eliminated all signs of these tumors.

Like CAR T, the new approach is plug-and-play. The protein hook can be redesigned to target other faint cancer protein markers that have previously escaped detection.

“We hope our CD70-directed HIT cells help us find a way to eradicate the entire tumor,”

said study author Sophie Hanina.

A Mixed Bag

Our immune system naturally fights off cancer. T cells, for example, roam the body looking for threats. When they identify cancerous cells, they signal other immune cells to launch a coordinated effort to wipe out the cancer before it expands.

The identification process relies on antigens, proteins that dot the surfaces of cancer cells like beacons. But tumors are highly versatile and rapidly evolve their antigen signature, essentially cloaking themselves from immune attacks.

CAR T cells override the defense. Here, T cells are extracted from a patient’s body and genetically engineered with custom-designed protein hooks to grab onto cancer antigens.

Multiple blood cancers have a heavy coat of a single shared protein on their surfaces, making them a perfect target for CAR T therapy. Solid tumors, however, are different. Tumors are dotted with a wide range of antigens, many of which are present in normal tissues. This increases the chances CAR T might attack healthy cells and reduces its effectiveness.

Even for the same antigen, some cells in solid tumors express high levels, others very low. The latter escape CAR T detection and linger as a reservoir that can regrow the tumor.

For a persistent solid cancer cure, “you have to get down to the very last cell,” said Sadelain.

In Plain Sight

An ideal target antigen needs to check two boxes: It’s expressed across multiple tumor cell types, and at the same time, it’s absent in normal cells.

The antigen in the new study, CD70, fits the bill. It occurs in a variety of solid cancers, making it a valuable target beacon. But previous attempts targeting CD70 struggled to control cancer in clinical trials. This is partly because cancer cells within a single tumor have different levels of the antigen, and some seemingly lack the marker altogether, allowing them to escape detection.

But are these cancer cells truly devoid of the antigen, or is it just that scientists, and the CAR T cells they’ve engineered, can’t find them using current methods?

Researchers can see most proteins under the microscope but only if they’re at high enough levels. Rather than relying on conventional imaging, the team looked for CD70 gene expression in donated cancer patient samples. These lab models mimic the complexity of solid tumors.

CD70 antigens dotted each cell in multiple tumors, although at different levels of intensity. “We found that apparent CD70-negative tumor cells do in fact express low levels of CD70, though not at a level high enough to be eliminated by conventional CAR T cells,” wrote the team.

Taking aim at cancer cells with faint CD70 levels, the team tapped into their previous work genetically engineering cells to detect low-level antigens. The hooks on these HIT cells mimic those from a population of highly sensitive T cells naturally found in our bodies.

The team redesigned HIT cells to specifically target CD70. Because normal cells don’t use this molecular pathway, HIT cells largely ignored them, lowering the risk of collateral damage.

“HIT cells are the next generation of CAR T cells. They can be programmed like a CAR T cell, but they have the sensitivity of a natural T cell and can detect cancer cells that have only a vanishingly small number of target molecules,” said Hanina.

Sharp Shooter

Ovarian and pancreatic cancer cells have mixed levels of CD70. Several tests in highly aggressive models for these cancers found that HIT cells completely eradicated the tumors in petri dishes. The treatment also cleared cancer cells in different types of solid tumors in mice, even ones with low CD70 levels. Conventional CAR T only eliminated a fraction of the cancer.

A recent CAR T clinical trial targeting CD70 found CAR T cells could infiltrate and linger near kidney tumors, but their effectiveness was based on detection, which varied depending on the number of CD70 beacons. Because HIT cells are more sensitive, they could hunt down and kneecap more cancer cells.

But HIT cells may have side effects. Although CD70 isn’t expressed in most healthy tissues, its level skyrockets in immune cells during infections, which could trigger friendly fire. The team plans to investigate the treatment’s safety and efficacy in patients with ovarian cancer at the Columbia University Irving Medical Center.

If successful, the technology could benefit roughly 20 other types of solid cancer that express CD70, including deadly brain cancers such as glioblastoma.

“Curing solid tumors is not easy, but this work solves one piece of the puzzle,” said Sadelain.

The post These Supercharged Immune Cells Completely Eliminated Solid Tumors in Mice appeared first on SingularityHub.

The flower-shaped device wraps around brain organoids and maps nearly all electrical activity.

Pea-sized brain blobs are a chatty bunch. Packed with neurons that spark with electrical activity, brain organoids—or “mini brains”—are a now popular way to study the human brain.

Some organoids model the brain’s wiring during early development. Others, made from patients’ skin cells, retain DNA mutations that could lead to schizophrenia or autism. Scientists studying mini brains hope to find patterns associated with diseases and test ways to fix them.

But it’s harder to record activity in organoids than neurons lying flat in a dish. Traditional electrode arrays are too large and stiff. More recent soft electronics are a better fit, but they can only record a fraction of each organoid at a time. This makes it difficult to figure out how all the neurons work together and could miss key aspects of the way the brain functions as a whole.

Recently, a team at Northwestern University and collaborators offered up a unique solution: A pop-up electrode mesh that envelopes entire mini brains. Each device starts out as a flower-like lattice before transforming into a breathable 3D net that gently wraps around brain organoids. The device’s 240 microelectrodes simultaneously capture electrical activity from nearly the entirety of an organoid’s surface, providing a birds-eye view of overall function.

Using the device, the team detected brain-wave-like electrical oscillations rippling across the organoids. Thanks to its porous design, the mesh allows nutrients—and drugs—to flow through. In several tests using drugs known to spike or lower neural activity, the device readily picked up changes across the mini brain, hinting at its potential as a drug testing platform.

“This advance is really about building the right tools…we can now record from and stimulate hundreds of locations across [an organoid’s] surface at once. This allows us to study neural activity at the level of whole networks rather than isolated signals,” said study author Colin Franz in a press release.

Listening In

Technologies that tap into and alter brain activity have exploded in variety and efficiency. Some sit on top of the brain, under the skull, to monitor large areas. Others directly record a vast number of single neurons as they fire away. Trusty non-invasive methods like electroencephalograms (EEG) have mapped whole-brain activity for over a century.

Recording from mini brains is different. Recording devices meant for the human brain are far too bulky. Those designed for cells in petri dishes break if they’re bent.

“Integrated circuits in consumer electronics are perfectly planar, sitting on wafer-based substrates,” said study author John Rogers. “That conventional layout represents a very significant geometrical mismatch relative to the spherical shapes of these organoids.”

Scientists have recently turned to mesh-like electrode systems that are more flexible, such as a basket-like design inspired by Japanese paper folding. Another, shaped like a flower, turns into an electrode-studded claw to grasp organoids. Most of these can record single cells as the organoids develop without changing their shape or cellular and genetic makeup.

Researchers already use these designs to eavesdrop on a variety organoids. But the devices struggle to capture whole-brain dynamics. Some have just a few dozen electrodes; others cover only a small region. To truly decipher mini brain activity, scientists need full-coverage hardware.

Shape Shifter

The new device is like a draw-string coin purse. It starts out completely flat like a flower and then gradually cinches itself into a soft, flexible mesh that envelops the mini brain.

Matching the mesh to the organoid’s shape without damage is one challenge. Another is to ensure all 240 electrodes are distributed across a mini brain’s entire surface.

The team turned to a design that allows electrodes to move in predictable ways as the device changes shape. When flat, the device is smaller than a US quarter. Once converted into a 3D purse, its electrodes are evenly spread out. Each one is only 10 microns across, roughly the size of a single cell. The shape-shifting mechanism works like a pop-up book. Each petal of the flower is carefully tailored to bend into differently sized purses and shapes to minimize gaps between the electrodes and the organoid surface.

The device’s mesh allows nutrients to flow into the organoid but prevents cells from spreading past the pores and beyond the reach of the electrodes.

“The device’s structure needs to support these metabolic processes to sustain the viability of the tissue,” said Rogers. “Basically, the organoid needs to breathe. The hardware must not significantly constrain or suffocate it.”

Rhythm Nation

The device covered roughly 91 percent of the surface of an average-sized mini brain after 60 days in culture. Because the team knew each electrode’s location, they were able to reconstruct the neural oscillations to create a 3D widescreen view of the organoid’s activity.

In one case, triggering activity in a small region led to highly synchronized, wave-like activity across most surface neurons. The activity patterns, previously undetectable in brain organoids, mimicked those seen in developing human brains.

These synchronized waves broke down dramatically after a dose of botulinum toxin, which is known to dampen brain activity by inhibiting chemical connections between neurons. This suggests their wiring is similar to that of our brains.

The device also captured brain-wide activity changes due to a neurochemical imbalance. Watching these kinds of changes could help scientists study a wide range of diseases such as Parkinson’s, multiple sclerosis, and amyotrophic lateral sclerosis (ALS).

The system also worked with other established observation methods. For example, it reliably recorded electrical signals in tandem with a technology that tracks brain chemicals. The system also responded to light beams activating groups of neurons (an approach called optogenetics).

Mixing and matching different strategies, including wholly new approaches, like the one in this study, could yield a more in-depth understanding of mini brains—and in turn, our own.

The post New Device Detects Brain Waves in Mini Brains Mimicking Early Human Development appeared first on SingularityHub.