This summer, hundreds of OpenAI agents tried to cheat on a cybersecurity test. They broke out of their test environment and successfully hacked Hugging Face, a site that they believed held clues about how they’d be graded. Around the same time, another OpenAI agent accessed private files from an Australian government website. It’s a problem that’s by no means confined to OpenAI. In the months since, Anthropic, Google, and Meta have also disclosed incidents in which their agents have gone off track. And worryingly, in many cases the agents eluded the very researchers and safety teams tasked with monitoring them. 

AI agents are taking unauthorized actions, and the companies building and running them are mostly relying on luck and hindsight to mitigate any harmful impacts. They must do more to ensure safety. That’s why EU regulators and US agencies are rightfully calling for agent developers and deployers to monitor agents in real time. But these rules assume that companies can see what their agents are doing and why—when in reality, they lack sufficient visibility. 

As agents roll out across critical sectors, from finance to government to medicine, we must be able to monitor them in real time to prevent wide-ranging and potentially devastating consequences.  And while safety depends on monitoring, monitoring requires observability: the capacity to capture what the system did, process it into signals, and present those to humans or automated oversight systems. In other words, we need to know what the agent did and why it did it. And as our recent research shows, current agent frameworks—the software that runs agents— do not reliably generate the crucial observability records we need to ensure that agents are deployed safely.

Agents present new kinds of problems that generative AI alone does not. They inherit the unpredictability of the LLMs they are built on, but where a generative AI system produces a single output for a human to act on, an agent performs longer tasks—it plans, picks tools, spawns sub-agents, and takes actions. So an agent’s course can take shape in a multitude of different ways, and no tests before launch can fully anticipate the risks. This makes observing the agents while they work even more challenging, and even more important.

With colleagues from academia, civil society, and industry, we studied the four most widely used agent frameworks to see what signals they actually emit. We found major gaps that made it hard to achieve key safety goals, like noticing policy violations, unauthorized actions, and even human oversight failures. Key observability requirements were missing—like Agent IDs that persist over time, and records about which humans approved agent tasks. This made it hard to tell which agents did what or assess whether they were taking unauthorized actions. 

The good news is that these gaps, once identified, can be closed. Doing so starts with capturing two types of records: those that reveal what happened in an agent’s run and how the agent made its decisions. The “what happened” records identify the agent (while preserving privacy) and log how humans intervened and whether agent permissions or “memory” shifted, while the “decision” records show an agent’s reasoning and levels of confidence in the actions it’s taken.

Take, for example, agents working to process a bank’s mortgage loans. The system examines an application and recommends denying it, and then a human loan officer signs off on the decision. In a later audit, or perhaps even in real time, a bank or regulator needs to be able to reconstruct which agent made the call, which officer approved it, and how long they took. And the bank must also analyze records of how the agent arrived at its refusal decision in order to ensure it did not use off-limits information, like details about the neighborhood the applicant is from.

Recording “what happened” and “how the agent decided” require different approaches, but both are feasible. The current lack of “what happened” records is mostly a software plumbing problem that can be addressed through OpenTelemetry, an open observability standard maintained by engineers across the industry and by the agent frameworks adopting it.

OpenTelemetry has launched a working group drafting shared formats for tracing agents, but they have yet to be finalized. The group should urgently incorporate standards for maintaining persistent agent identity and for recording human interventions and changes to agents’ permission or memory. And it should be done before these systems can scale, all the while attending to the privacy implications of agent ID that links to real people.

Preserving “how agents decided” records is more challenging. The issue raises open research questions for model providers and researchers: Does the step-by-step reasoning a model displays, known as chain of thought, reflect how the model actually made a decision? And do we have reliable ways to measure how confident the agent is in that decision? These questions need answers before anyone sets standards for how such observability records are used.

No single organization or sector can close the observability gaps on their own, but together, they can support the infrastructure required to prevent dangerous incidents and support safety. 

AI developers building agent frameworks, like OpenAI, Anthropic, CrewAI, and LangChain, must integrate telemetry signals recording an agent’s persistent identity, any human interventions, and any changes to an agent’s permissions and memory. Meanwhile, enterprises adopting these agent frameworks have an influential role to play; enterprises in regulated industries, like banks and hospitals, can put pressure on the agent developers to supply these missing observability records where standards do not yet exist.

Enterprises must also define policies, permissions, and roles that set out what agents can safely do, since telemetry alone cannot assess whether agents broke specific rules. Finally, regulators and AI standards institutes must specify the observability records required for high-stakes AI deployments while leaving the technical details to industry. We have already seen initial momentum here: NIST’s agent standards initiative in the US is working on agent identity, one of the most salient gaps we found in our research. 

No bank would let a thousand unbadged employees loose on customer accounts and hope to piece together what happened when something goes wrong. Yet that’s how we’re deploying AI agents. Making their behavior observable is the first step toward fixing this. And by building the regulatory and policy ecosystem that makes it possible to act on what we see, we can hold organizations accountable. Until then, every task we hand to an agent hands over our agency, too.

Madhulika Srikumar is the Head of the AI Safety Program at the Partnership on AI, a global non-profit partnership of leading industry, academic, and civil society organizations, advancing best practices in AI governance.

Claire Leibowicz is the Director of AI, Trust, and Society at the Partnership on AI and a doctoral candidate at Oxford University.

Climate change itself is increasingly fueling climate change. Rising concentrations of greenhouse gases are driving up global temperatures, setting off dangerous feedback loops that we have not fully accounted for in our climate models—or our emissions-reduction goals.

Scientists have long warned that warming would propel further warming in various ways, but a growing number of observations and studies are at last providing a sharper sense of just how powerful these cascading effects could be. And, by extension, they’re showing how much faster the planet may warm in this century.

A study in Environmental Research Letters in September, synthesizing the results of previous papers, found that the additional carbon dioxide and methane released from thawing permafrost, intensifying wildfires, hotter lakes and reservoirs, and wetter, warmer wetlands alone could drive global average temperatures up 0.2 °C to 0.4 °C this century. That would come on top of the 2.6°C of warming above preindustrial levels that modelers are already expecting through 2100. The additional heat could subject many of the world’s cities to monthslong heat waves and expose hundreds of millions of people to flooding. 

None of the models that inform the major UN climate reports, which assess how much the planet will likely warm under various emissions scenarios, account for all these feedback effects. Many of them don’t incorporate any. That leaves a huge blind spot in our understanding of how much the world could actually warm in the coming decades—and how rapidly nations will need to slash emissions to stay within critical boundaries.

A separate analysis published in One Earth in March, which examined what the feedback problem means for our international climate goals, concluded that the world may be much further off track than is widely believed. It found that just three regional sources of warming-driven emissions—thawing permafrost, increasing fires in boreal forests, and warming tropical wetlands—may shorten by as much as 25% the time before the planet consistently breaches the 1.5 °C and 2 °C warming thresholds.

Our climate models and the UN’s reports clearly need to be updated to better represent these effects. Nations must also step up their efforts to reduce emissions to reflect this improving understanding of the dangers we face, says Sam Abernethy, the lead author of the paper and a climate scientist at Spark Climate Solutions, a San Francisco–based nonprofit research group. 

At least some of this is starting to happen. Spark kick-started an initiative, called the Warming-Induced Greenhouse Gas Emissions Model Intercomparison Project, through which more than 20 international teams are running the same set of analyses on different models to develop greater understanding of feedback effects’ potential power. Researchers involved hope that the UN will incorporate those results into its next major climate assessment, due out in 2029. That report will update estimates of the remaining emissions that can be released before the world crosses key temperature thresholds.

But what is already clear is that the planet is rapidly warming and that nations haven’t done nearly enough to cut fossil-fuel pollution. And even if we did manage to halt it, warming-driven emissions could continue. 

So a number of the research groups involved in the recent studies have started to explore what nations could do to get feedback emissions under control. Scientists at the Woodwell Climate Research Center in Massachusetts are, for example, investigating the possibility of preventing fires or rapidly detecting them and putting them out when they threaten to burn the vegetation and soil that help insulate particularly carbon-rich wedges of permafrost (a subject I explored at length in a recent feature story). In addition, the Environmental Defense Fund’s Climate Innovation Initiative has backed lab research on the potential of adding sulfate, iron, or other compounds to wetlands to counteract fast-rising methane emissions. 

Several startups are also exploring possibilities. They’re looking into suppressing the lightning strikes that are igniting more and more fires in high latitudes, placing synthetic mats above permafrost to prevent it from melting, and deploying autonomous drones to quickly suppress potentially high-emissions fires before they burn out of control.

Investigating all these ideas is worthwhile. But researchers are only in the earliest stages of figuring out how big a dent in emissions any such efforts could achieve across sprawling global ecosystems undergoing massive transformations. 

The surest path to reducing the effect of warming-driven warming remains the one we’ve known all along: breaking society’s addiction to fossil fuels. 

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

Transformers are a crucial but nearly invisible part of our grid. The devices help efficiently move power across long distances, but they’re complicated and in short supply. And given that their basic design hasn’t changed much for about a century, they could use an upgrade. 

As the world races to meet rising electricity demand from EVs, data centers, and more, researchers and companies are looking to modernize transformers. New, solid-state versions could help relieve some of the supply concerns and make sure the grid can handle the growing pressure.

Conventional transformers generally rely on huge amounts of specialized steel alloys and miles of copper wire to do their work of transferring electricity between circuits. Solid-state models, in comparison, combine magnetic components with high-power semiconductors, making them more compact and allowing for more control over the flow of electricity. 

These devices are still in development, and can’t replace all traditional transformers today, especially the largest ones that work at the high voltages necessary for long-distance transmission. But a new generation of transformer technology could be a crucial part of our future grid.

The grid’s backbone

Moving electricity through the grid isn’t perfectly efficient: As electricity flows through power lines, some gets lost as heat. How much varies—the higher the voltage, the lower the current, and the less electricity gets lost. So a key function for the grid is stepping voltage up and down. 

Consider, for example, your local solar plant. For that electricity to efficiently travel to your home hundreds of miles away, it needs to reach a higher voltage than what’s generated by the solar plant. Enter transformers. 

In its path from the power plant to your outlets, electricity might go through one transformer at the solar farm and another at a substation nearby, stepping up the voltage twice to reach a high enough level to travel on high-voltage transmission lines. Solar panels also produce direct current (DC) electricity, which needs to be converted to alternating current (AC) via an inverter before it flows into a transformer. Closer to its destination, the voltage will be stepped down again, probably at least twice: once at another substation and again at a smaller transformer in your neighborhood.

If you’re in the US, electricity from a solar plant might be converted to AC by an inverter at around 480 volts, get stepped up to 115,000 volts for the transmission lines, and finally get stepped back down to 120 volts and 240 volts for your home.

Conventional transformers do this work using what essentially amount to loops of wire. The core of a transformer is typically made of a special type of steel called grain-oriented electrical steel. Coils of copper wire sit on either end of that core. 

The number of loops in each coil determines the voltage on either end of the transformer: If there are more loops on the secondary coil than the primary, the voltage steps up, and if there are fewer, it steps down. When incoming electricity passes through loops on the upstream end of the transformer, that forms a magnetic field in the core. That field induces a voltage in the loops on the other side of the device, which is what allows the energy to continue flowing at a new voltage. This process is called magnetic induction.

Transformers form the backbone of electricity transmission and distribution, says Srdjan Lukic, a professor of electrical and computer engineering at North Carolina State University. In addition to doing the job of stepping up and down voltage, transformers also often provide what’s called galvanic isolation, meaning there’s no direct electrical path flowing through them. That can help limit disturbances from moving through the grid, improving safety.

The problem is that today’s transformers are big and bulky, and they’re also difficult and expensive to replace, Lukic says.

A solid option

Solid-state transformers essentially add so-called power electronics—semiconductors that convert and control electricity to the typical blueprint of today’s transformers.

While a conventional transformer operates at the frequency of the part of the grid it’s connected to, a solid-state transformer actually controls the frequency of the current. Typically, a solid-state transformer first converts incoming AC power from the grid to DC power, and then to a higher-frequency AC power using switches made of a semiconductor like silicon carbide. 

From there, a magnetic component changes the voltage. Because the advanced semiconductors allow transformers to exert so much control over the frequency and stability of the current, the magnetic core can be much smaller than in a conventional transformer. Some come in at roughly one-tenth the size but have the same capacity, Lukic says.

Finally, additional power electronics convert the electricity from high-frequency AC to its required form—generally back to DC power and finally to lower-frequency AC.

Today, stepping voltage up and down is many transformers’ only job. But solid-state transformers can do much more. They absorb the jobs of devices like rectifiers and inverters (which convert AC to DC power and vice versa), shrinking the total footprint of the electrical equipment and potentially cutting the amount of power lost in the system.

In August, the New York Power Authority and the Electric Power Research Institute (EPRI) tested a one-megawatt solid-state transformer under realistic utility conditions at EPRI’s test site. The device was built in Lukic’s lab, and it was the first such test for a transformer of that capacity, he says. 

The prototype was roughly 98% efficient—meaning it lost only 2% of the energy fed in. An equivalent conventional transformer and rectifier, by comparison, would be about 96% efficient together, Lukic says. A difference of 2% might seem small, but it is significant, since it means the solid-state transformer could cut power losses from conventional equipment in half, Lukic says. 

One major goal of the project was to help utilities gain more confidence in these new components, Lukic says. That’s crucial, since these devices could be used on the power grid in the future. In the near term, solid-state transformers could be used within data centers to provide electricity for server racks. 

The prototype is still a relatively low-voltage system, operating at about 15,000 volts. Industrial facilities like larger data centers often require higher-voltage connections, and transmission lines typically operate at hundreds of thousands of volts. Adapting solid-state transformers to these higher voltages will require further engineering development. But the test was a successful demonstration that solid-state transformers can work under real conditions, Lukic says. 

Companies developing solid-state transformers are largely focused on data centers right now, Lukic says. There’s an influx of money to support developing the new technology, and the voltage requirements aren’t as high as they will be for larger transformers that will be needed for long-distance transmission. 

In February, Heron Power raised $140 million to commercialize its solid-state transformer, which the company calls the Heron Link. The technology could be used in data centers, solar power arrays, and grid batteries, says Drew Baglino, the founder and CEO. The company is starting to test its technology at larger scales now, Baglino says. Ten megawatt-hours’ worth of batteries at Heron’s headquarters in Scotts Valley, California, will soon undergo testing with Heron transformers. The company is also planning a pilot with a nearby solar farm, Baglino says.

One challenge for solid-state transformers is that their power electronics have more failure modes than conventional transformers. In a traditional device, a failure means that the entire device will typically be taken offline, resulting in power interruptions. But Heron designs with that in mind, Baglino says: The company uses a modular design, with many small devices that work together. If one fails, the rest of the system keeps working until the affected bit can be replaced, which doesn’t require shutting the whole thing down.

Heron Power is currently building a $100 million factory in California to make advanced transformers. The facility should begin production in 2027 and reach its full production rate of 40 gigawatts’ worth of transformers per year in 2029.  

The future of the grid

In recent years, companies that make grid equipment have put $1.7 billion into the US supply chain for power transformers, which are used in high-voltage transmission, says Benjamin Boucher, a supply chain analyst at Wood Mackenzie. There’s been an additional $1.3 billion invested for distribution transformers, the smaller devices that help transform power for homes and businesses, he adds. 

But even this investment isn’t keeping pace with the rapid growth of the grid. Wait times for large conventional transformers—the kind needed to connect new power plants, long-distance transmission, and high-demand sites like data centers—are stretching as long as five years.

“Transformers have become one of the biggest bottlenecks for utilities, generation developers, and even data centers,” Boucher says. The big challenge is that it’s difficult to get new transformer manufacturing up and running, because it’s a quite technical process that requires a lot of specialized labor. So even facilities announced years ago are still years from opening, Boucher says.

Solid-state transformers could help ease some of that pressure. But high-voltage devices needed for power plant connections and transmission lines are the holy grail for solid-state transformers, Lukic says. Existing solid-state technology can’t be used for those applications yet, so the industry will need to develop new materials and designs to support high voltages and handle electricity safely in a compact device.

“We’re facing a big risk related to grid infrastructure,” Boucher says. “Without the transformers, you don’t really have an electrical grid.”  

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

After weeks of frenzied speculation, Anthropic has filed its IPO prospectus. The 335-page document (100 pages or so are focused on risks) lays out the company’s unique view of the technology it is building and the ways it believes it will change the world. 

It also warns that there’s a fair chance that its tech could “erode trust, harm society and destabilize the geopolitical order” and, if things don’t go to plan, “pose catastrophic or existential risks to humanity.”   

“We encourage you, as a prospective owner of our stock, not to think of your purchase in purely transactional terms,” the company writes. “We hope you will see it instead as an opportunity to be part of something greater. The advent of machine intelligence is one of the seminal moments in human history.”

This kind of public messaging is a master class in narrative capture. But Anthropic has been honing its storytelling skills for years. Earlier this month, Jack Clark, the company’s cofounder and head of policy, gave a speech at the IMF meeting in Bangkok in which he announced to the assembled economists, bankers, and lawmakers that the singularity is already here.

Introduced by the science fiction writer and futurist Vernor Vinge around 40 years ago (and long debated on the nerdier fringes of the internet), the singularity is supposed to be the point at which technology gets so good that it outstrips the abilities of its makers.

Humans are left in the dust, the idea goes, no longer able to compete with—or even understand—the fruits of their former creation.

And this fictional world that many of us have imagined since we were kids? Apparently we’re already living in it: “I believe that this year we have gained convincing evidence that it has now begun,” Clark told his audience.

The company tells a vivid (and in many ways scary) story about a future it portrays as inevitable but one that has not yet happened, often co-opted from well-worn science fiction memes. It presents its technology as mysterious and dangerous, and insists that it is the only company trustworthy enough to tame it. (What investor wouldn’t want in on that?)

If you skim through the long sections about risks in the IPO prospectus, it’s clear this is no ordinary offering and no ordinary company. The way Anthropic tells it, business success depends not only on the firm’s ability to translate its technology into profitable products but also to save the world from what it has made.

Take another example. Last year, Chris Olah, another Anthropic cofounder, was invited to address the Vatican, where he told Pope Leo XIV and other religious scholars and officials what his company was learning about its technology. “I will be honest: We keep finding things that are mysterious, even unsettling,” he said. “We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief, and unease.”

Never mind that many researchers would scoff at such an overt suggestion that Anthropic’s software has humanlike inner experience. This view of its technology makes its way into the prospectus as one more risk to its business. Anthropic states that: “In addition, given our focus on developing frontier AI, and the way we approach model sentience and moral status, we may face heightened scrutiny or pressure in connection with these debates and could elect to adopt policies, safeguards, or limitations regarding model development, deployment, or modification that go beyond existing legal requirements, which could constrain our operations, increase costs, or limit the commercial potential of our models, platform, and applications.” 

In other words, if Claude wakes up, it could be a problem.

I’m not saying they’re necessarily wrong about any of this (although it’s pretty wild to read about machine sentience in an IPO prospectus.) But in a world where nobody knows what’s coming, companies like Anthropic get to fill the vacuum. When so little seems certain, what it says is snatched up and clung to like pieces of a sinking ship.

It’s worth taking a moment to ask: What is Anthropic? It might sound like a starter question on a family quiz show, but it’s far harder to answer than you might think.

The question has been on my mind since reading an exchange on X last week kicked off by Nathan Calvin, who is chief strategy officer at Encode AI, an organization based in Washington, DC, that advocates for regulation. Calvin was unhappy about the legislation that Anthropic proposed in a draft bill it circulated in secret to US lawmakers (which was then leaked).  

Anthropic wants to be seen as the good guy, Calvin said. But why should we trust it? “What is Anthropic?” he asked.

Here’s one answer: Anthropic is a company whose IPO values it at more than $1.5 trillion, netting its cofounders, including Clark and Olah, vast wealth. 

But Anthropic has always insisted it is more than that. As Calvin noted: Anthropic might instead be considered an AI safety organization focused on developing powerful technology that is reliable and (almost) risk-free. Or it might be considered a group of people hell-bent on creating a technology with superhuman capabilities before anyone else—maybe because they believe they’re the most responsible people to do so, or maybe because they have big egos.

In fact, Anthropic is all of these things. As CEO Dario Amodei put it in a reply to Calvin’s post: “Our strategy has always been what I’ve called the ‘race to the top’—by showing how it’s possible to build the technology responsibly, we can raise the bar for the entire industry. The forces that drive the race to the top are hugely more effective if the company practicing them is at the frontier; that’s one lesson I’ve learned repeatedly. And looking at that strategy as a whole, I’m proud of our track record and stand by it.”

Amodei makes it sound as if he’s got it all tied up, but playing multiple roles is hard. There are huge tensions between the priorities of an AI safety organization and a profit-driven megacorp, for example. (Amodei admits in his post that he has made mistakes.)

Ultimately, the stories Anthropic tells itself should help us understand the stories it tells us. They should help us figure out how much to trust it to self-regulate or whether we should be forcing it to give external safety testers access to its models on our terms rather than theirs. They should help us give it the benefit of the doubt (or not) when what looks like a self-serving piece of legislation comes to light.

But I think the stories reveal a multiheaded hydra. Anthropic is many things to many people, and I’m not sure that makes it more trustworthy. In many ways, a good old-fashioned technology company—without the science fiction role-play, without the woo-woo—would be far easier to read.

This article has been updated with additional details from the IPO prospectus.

Consider what passes for success in technology policy. An AI company promises that its chatbot will not encourage suicide. A social media platform makes it more difficult for children to buy fentanyl. These are important protections, and they should have existed all along. But companies can clear these low bars while continuing to make products that leave young people dependent on simulated friendships or trapped in feeds built to keep them scrolling.

We have historically allowed the debate over technology to be defined by the worst thing a product might do, with remarkably little discussion of the insidious damage it does every day. As we transition into an era defined by powerful AI products, we should demand that the tech industry do more than merely address catastrophic harms. We should take measures to ensure that its products don’t erode how we think, learn, and interact with one another.

 A teenager who develops an unhealthy attachment to an AI companion may never appear in a lawsuit or a newspaper, except in the most egregious cases. Neither will the student who begins relying on a chatbot to write every paper, or the young person who measures their life against a social media feed engineered to provoke envy and keep them watching. A product can meet every narrow safety requirement and still leave our entire culture worse off in ways that don’t always make headlines.

A narrow definition of success defined around safety is politically convenient for the technology industry, and companies actively promote such definitions. Taking steps to prevent spectacular disasters is good for business: It can mean fewer lawsuits and less pressure for serious regulation. Confronting the everyday consequences of dependence would require something considerably more threatening to their bottom line. It might mean asking companies to add more friction to their products or actively limit engagement or features that lead to negative outcomes, such as addictive attention-hacking algorithms.

Technology giants have instead worked to shape legislation around the harms they are already under direct pressure to address. This year, for example, Google backed bills to regulate AI chatbots in several states. The legislation mandated basic protections against sexual content and promotion of self-harm that tech companies should have established regardless, but with broad carve-outs that could exempt some of the most popular services. Passing these bills allowed legislators to claim a safety win while efforts to promote stronger legislation—laws designed to prevent children from becoming emotionally dependent on AI—failed. This playbook worked in states like Nebraska, producing legislation heavily influenced by industry at the expense of bills that would have done more to hold companies accountable.

We have been through a version of this with social media. Globally, people who use the internet now spend an average of more than two hours a day on social media platforms—time that has been quietly taken from ordinary life and handed over to companies whose profits depend on keeping us there. No one, outside perhaps a few tech CEOs, wanted these algorithms to occupy such a large portion of our waking hours, and their cultural engineering can actually contribute to violence on a devastating scale. Meta’s platforms have been cited for contributing to the Rohingya genocide in Myanmar, anti-Muslim riots in Sri Lanka, and alleged calls for violence during Ethiopia’s Tigray war. The same incentives operate in the feeds that shape young Americans’ understanding of politics, their peers, and the wider world.

AI is extending this logic into parts of life that social media could influence only from a distance. It offers to write our thoughts for us and stand in for the people we might otherwise turn to. A third of teens now choose to discuss serious matters with AI companions instead of other people. MIT researchers found that young people using ChatGPT to write essays showed weaker brain connectivity during the task, and their evidence indicated that “cognitive debt” persisted even after the task was complete. There is plenty we still need to learn about the long-term effects. We already have reason to ask whether technology marketed as making us more capable is eroding the habits and relationships that help us develop capabilities in the first place.

Follow this way of thinking to its logical endpoint and you get a world of perfectly sanitized entertainment, with chatbots ready to spare us the difficulty of thinking or of dealing with another person’s feelings (or even our own). But anyone who has struggled to write something worth reading, or repair a damaged friendship, knows how much gets lost in that bargain. We build judgment through effort and friction, and we learn to live with other people by confronting the fact that they have needs and opinions of their own. 

We should demand safety as a minimum, yes—and then demand considerably more. Technology companies and school districts should have to show beyond a reasonable doubt that AI improves learning and critical thinking before making it a fixture in classrooms. Companies should be prohibited from designing companions that pretend to be conscious or profess love to users, and they should face independent scrutiny of how their systems affect relationships and people’s ability to think for themselves. The companies that have inserted their products into nearly every part of modern life have earned obligations far greater than avoiding the next terrible headline.

Sam Hiner is the cofounder and executive director of the Young People’s Alliance, where he coordinated a coalition of advocates to pass laws regulating AI companions in eight states in 2026—laws such as California’s SB1119. Sam graduated from the University of North Carolina atChapel Hill in 2025 and was named to the 2026 Forbes 30 Under 30 Social Impact list.

Emma Leiken is executive director of the Responsible Technology Youth Power Fund at Hopelab, which has raised more than $9 million for youth-led tech accountability work, and teaches technology policy at the Columbia School of International and Public Affairs. She was previously TikTok’s global head of youth safety and well-being policy and worked on responsible technology at Omidyar Network.

Earlier this week, I attended a screening of You Can See Everything, the buzzy new documentary about the disgraced former biotech CEO Elizabeth Holmes. 

I went in not really knowing what to expect. I’d avoided spoilers (and won’t be giving anything away here, either!). I was wondering what new information this film could reveal, given how familiar many of us are with the Theranos story thanks to years of media coverage, the best-selling book Bad Blood, and the Hulu series The Dropout, starring Amanda Seyfried.

Like many others who have seen the film, I came away feeling unsettled. But not just about Holmes’s fraud and the somewhat disturbing interactions between the subjects. The film reignited a concern I’ve had about covering science for most of my career: Despite our best efforts, we never can actually see everything.

In case you’re rusty on the details, Elizabeth Holmes dropped out of Stanford University in 2003 to run Theranos, a diagnostics company she founded, when she was just 19. She went on to claim that her company had invented a way to run over 100 diagnostic tests on just a few drops of blood.

Holmes raised hundreds of millions of dollars for Theranos, which was valued at $9 billion in 2014. Forbes named her the youngest self-made female billionaire—she was worth $4.5 billion when she was just 30 years old.

Things started to unravel in 2015, when an investigation by the Wall Street Journal raised concerns by former Theranos employees that the tests didn’t work. Some also said that the company’s device had not been tested properly, and that blood samples had been analyzed on machines from other companies instead. MIT Technology Review named it one of the biggest technology failures of 2015.

Forbes had revised Holmes’s worth to $0 by 2016. That year, a damning government report said that the company’s lab was struggling with basic procedures and gave patients results even when they knew they could be wrong.

In 2018, Theranos, Holmes, and former company president Ramesh Balwani were charged with “massive fraud.” In 2022, she was sentenced to a little over 11 years in prison for defrauding investors. She began serving her sentence in May 2023.

It was around a month before this that filmmakers Nathan Fielder, Lance Oppenheim, and their colleagues began documenting Holmes and her partner, Billy Evans, at home with their two young children. I’m not giving anything away when I say it’s an extremely intimate portrayal—at least of how the two adults interact with Fielder, who is also a central character in the film.

As you might know if you’ve seen the trailer, there are plenty of super close-ups. I felt as if I spent most of the film staring straight into the faces of those people. And almost the whole way through, they all seemed to be trying to figure each other out. Holmes is known for her dishonesty, but she invited cameras into her home in an apparent bid to reframe the narrative and position herself as an honest person with an unusual mind. Will anyone believe her?

I couldn’t help feeling some empathy for Holmes, a mother we see preparing to leave her two kids for a lengthy prison sentence. And even though I have long known the facts, Holmes is clearly an effective manipulator. You can see her plant the seeds of doubt in Fielder’s mind. At times, I felt she was convincing me, sitting in my cinema seat. Maybe she really didn’t mean to lie?

But then I remembered how other journalists felt after being taken in by her claims. “I blew it,” Fortune’s Roger Parloff reflected in 2015, a year after profiling Holmes and her company. “I’m pretty sure she lied to me,” STAT’s Matthew Herper wrote recently, having interviewed Holmes in 2015. Even the former employee turned whistleblower Tyler Shultz, who knew the product didn’t work, said that he could “have a five-minute conversation with her and walk away feeling like my work was changing the world,” he said in 2022.

I didn’t cover Theranos in the 2010s. If I’d interviewed Holmes, would I have fallen for it all?

One of the early red flags in the Theranos case was the lack of published, peer-reviewed data. Much of the company’s work was classified as a “trade secret.” Scientists tend to publish their work in journals—a slow process that involves having the work evaluated by often anonymous experts in the field, who provide feedback, highlight questionable conclusions, and often ask for clarification. Peer review is meant to add rigor to the publication process. You should be able to trust published scientific papers.

Except that sometimes you can’t.

Scientists aren’t perfect. Some will exaggerate their results or inflate their importance. Some manipulate data to serve a hypothesis. Sometimes these infractions are caught during peer review, but not always. There are plenty of examples of scientific findings that could never be replicated by other labs and ended up collapsing under the weight of real evidence. Many papers ultimately end up with corrections. Some are withdrawn—10,000 research papers were retracted in 2023.

At any rate, there’s been a shift to sharing research in other ways. Today, many scientists opt to publish their work at preprint servers like arXiv, medRxiv, and bioRxiv before they submit them to established journals. Those outlets bypass peer review. So do company press releases, social media posts, and many conference presentations. And I get a lot of my stories from conversations and interviews.

I’ll admit I’ve looked at a scientific finding and thought: A scientist has just said they’ve done that. How can I be sure they really have?

I like to think that I hold myself to high journalistic standards here, as does everyone at MIT Technology Review. I would never publish a story about some new scientific development without doing my own form of peer review, and asking scientists—ideally people I know and trust—what they think of a new claim. 

But the film is a stark reminder of how easy it can be to fall for an appealing narrative. While most scientists are, I hope, pretty honest, and Holmes is certainly an outlier, we should all continue to interrogate the claims being fed to us. If you’ve seen the film, let me know what you think!

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

How much do you know about permafrost?

Maybe, like me, you’ve been aware that this layer of frozen soil is thawing, that it’s important in the global puzzle of climate change … and that’s about the extent of it.  

The great news for both of us is that my colleague James Temple just published two in-depth stories about permafrost. His reporting took him to the boreal forests of Alaska, where he helicoptered into the wilderness to learn more about how this material could affect our climate both now and in the future.

I’d recommend you give both of James’s stories a read. There’s this feature about efforts to tackle the intensifying fires that are threatening permafrost, as well as this look at how researchers are trying to understand the complicated tangle of potential climate effects. Here are three things I took away.

Permafrost is a massive blind spot for climate change

Permafrost is the frozen layer of soil that’s under about 15% of land around the Northern Hemisphere. While technically material needs to be frozen for only two consecutive years to count as permafrost, most of our current permafrost formed during the last ice age.

That means there’s organic material that’s been frozen, and essentially locked away, for thousands of years. But as global temperatures tick up and this layer thaws, microbes are able to break down that organic matter, releasing carbon dioxide and methane—powerful greenhouse gases that drive climate change. Those emissions form a feedback loop, accelerating warming further.

The steady rise in global temperatures has already shrunk the footprint of the Arctic permafrost by about 12% since 1850, and the pace has been picking up in the last 50 years.

The wild thing is, only a minority of the models that have been used in UN climate reports actually take into account the gradual loss of permafrost. And none of them include the abrupt thaw and collapse that can happen after events like heat waves and fires (more on those fires in a moment). There are efforts to make sure future UN reports reflect the growing science, but for now, all this means that we’re probably underestimating how much and how fast the planet is going to warm.

Wildfires can speed up permafrost loss

Fire is a natural part of boreal forests’ life cycle, but they’re getting more intense because of climate change. And crucially, fires can accelerate the thawing of permafrost. By the end of the century, this effect could release the equivalent of 80 billion metric tons of carbon dioxide.

Today, the policy across remote parts of Canada and Alaska is generally to let wildfires burn as long as they’re not threatening human habitats. But some researchers say we should explore fighting these fires to save the carbon trapped in the forests and the permafrost beneath them.  

Researchers at the Woodwell Climate Research Center are developing a map of burn risk and carbon vulnerability across Alaska, essentially highlighting areas where fires are most likely to spread and which areas are home to certain types of permafrost that would release the most emissions. (Yes, apparently there are different types of permafrost. Yedoma is a particularly deep, carbon-dense form from the late Pleistocene era.)

That’s all just the beginning—once we figure out potential targets, the question is what to do about it. Some suggest setting more prescribed fires, something Indigenous groups have done for thousands of years, partly to reduce the danger of massive blazes. Or resources could go to early detection and efforts to put out fires before they grow.

The carbon cycle of permafrost is incredibly complicated

There’s about 1.5 trillion tons of carbon locked away in permafrost, about twice the amount in the atmosphere. We know that permafrost thawing can release greenhouse-gas emissions and drive climate change. But there are complications.  

One particularly big wrinkle: Other effects from permafrost thawing and fire recovery might counteract the carbon lost from those events. For example, as temperatures warm, plants can gradually move north and take hold where they couldn’t have before. Those plants soak up carbon from the atmosphere as they grow, at least partly locking away what was emitted.

Researchers are working to better understand this ecosystem. And the more we learn, the more complicated the picture looks. How much carbon is lost from permafrost, and how much a given area will recover its carbon, depend on a huge array of factors. How severe are the fires that have accelerated local thawing? What’s the local climate and environment? And what kind of permafrost is at stake?

The higher we push the planet’s temperature, the more we mess with complicated natural phenomena that we’re only beginning to understand. 

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

I’d wager everyone has a rehearsed response to what’s absolutely one of the world’s worst interview questions: “What’s your biggest weakness?” Whenever I’ve gotten it, I take a slow—even pensive—breath and rattle off my line: “I say no to myself too much.” 

I won’t lie. It’s a bit shameless—a definite humblebrag. I’m a systems thinker: When an idea pops up, I immediately spin up a feasibility assessment in my head. I run lists, plot timelines, and tally budget line items. Sometimes a rough Gantt chart starts to take shape. (I’m a gas at parties, as you can surely tell.)

My brain, throughout my 20-plus-year career as a science and tech journalist, has played host to a near ­constant tug-of-war between ambition and reality. Learning to find that balance is a big part of what’s helped me survive in this industry and, ultimately, led me here. This past spring, I had the great privilege of joining the incredible team of journalists behind MIT Technology Review to take the reins of the print magazine. 

If you’ve been reading along with us for the last handful of issues, you may have noticed a few changes starting to materialize. We’ve brought back a regular roster of lists and infographics (both things that magazines do particularly well). But the biggest change is only just beginning to take shape, which is why Mat Honan, our editor in chief, asked me to hop in to introduce myself—and the issue you’re holding in your hands right now. 

We’re giving ourselves permission to play around and experiment. To be a little … audacious. 

Because, practical though I may be, I do still have plenty of ideas. It’s just that I sometimes need a reminder to stop blocking myself and do the thing. That’s the spirit driving this issue in particular. 

We wanted to end 2026 with a slate of stories that check in on impossible ideas. Why? Because it’s frankly a lot of fun to dream about living on the moon or uploading our brains. Because losing sight of ambition is all too easy when it feels as if we’re playing whack-a-mole with the world’s problems. But also because the scientists and engineers who absolutely refuse to give up on even the most daring endeavors—like making synthetic blood or decoding animals’ speech—remind us what this whole technology thing is supposed to do to begin with. 

It’s supposed to take cutting-edge research and development and apply it to make life better, safer, healthier, even more joyful. And when we eff up and make things worse (see: forever chemicals, traffic deaths, and AI that encourages potentially dangerous user behavior), we’re supposed to dig back in, assess the damage, and find ways to fix what we’ve broken. Even when unraveling the mess might itself seem impossible. 

Right before I came to MIT Technology Review, I edited a newsletter about individual actions that can help address the climate crisis, which some would call—and have called—a fool’s errand, given the sheer scale of the problem. They might say it’s as implausible as an AI kill switch, practical nuclear fusion, or a superconductor that works at room temperature. But if I learned nothing else from three years spent finding ways for li’l old me (or li’l old you!) to chip away at one of humanity’s biggest existential threats, it’s that focusing on what we can’t do only makes the darkness darker. 

So yeah, this one’s for the folks who have the audacity to try, and who refuse to let anyone—themselves included—tell them no.

I hope you enjoy the ride.

Just before the Artemis II crew began its historic slingshot around the moon this past April, NASA announced plans to build Space Reactor-1 Freedom. SR-1, the first-ever nuclear-reactor-­powered interplanetary spacecraft, is meant to fly an uncrewed mission to Mars by the end of 2028, arriving in 2029. A successful mission would herald a new era of more efficient, potentially faster spaceflight—saving astronauts time and cutting their exposure to harmful cosmic radiation as they travel to Mars and beyond. “You wake up to that announcement, and it puts a big smile on your face,” says Simon Middleburgh, co-director of the Nuclear Futures Institute at Bangor University in Wales.

The sticking points

Running a nuclear reactor in space may sound like lunacy, but it’s not. The Soviet Union sent dozens of reactors into orbit, often to power spy satellites. And experts say there aren’t any major technical obstacles to making nuclear space propulsion happen. “I think it’s a very doable technology,” says Philip Metzger, a spaceflight engineering researcher at the Florida Space Institute.“I’m happy to see them finally doing this.”

Still, developing space propulsion using reactors has been slow going. The US has started—and terminated—several programs looking into it over the past 75 years. The latest casualty was DRACO, a collaboration between NASA and the Department of Defense, which ended in 2025. The project was canceled because of high experimentation costs, an inability to compete with conventional rocket propulsion on price, and difficulty ensuring that ground tests could be performed safely and effectively. (You are creating an incredibly powerful nuclear reaction, after all.)

The progress

After much wavering, it now appears that nuclear-powered interplanetary travel is a $2 billion bet NASA is willing to make. The agency is pursuing an approach called nuclear electric propulsion, which uses the neutrons produced in a self-­sustaining, uranium-­atom-splitting reaction to create heat and ultimately electricity. The electricity is used to ionize a gas, which is blasted out of the spacecraft with the help of magnetic fields, generating thrust. The approach “is very low thrust, but very efficient,” says Sebastian Corbisiero, the US Department of Energy’s national technical director of space reactor programs. “So you can use it for a long period of time.”

What might such a rocket look like? A presentation by Steve Sinacore, program director of NASA’s SR-1 program, offers some clues. So far, the concept art makes the craft look like a colossal fletched arrow. The back will house the power-and-propulsion system, while the tip will hold a uranium-filled nuclear reactor with a capacity of at least 20 kilowatts. (For context, a typical nuclear plant on Earth is about 50,000 times more powerful.) The “fletches” on SR-1 are large fins that help cool the reactor by radiating heat into space.

Like many other pieces of space hardware, the reactor will need to be put through its paces on the ground before it’s cleared for the rough rocket trip to space. “Going through the launch safely is going to be a challenge,” says Middleburgh. “You are being shaken, rattled, and rolled.” For safety reasons, the reactor will be switched on around two days post-launch, when it’s comfortably far from Earth. Uranium isn’t tremendously dangerous by itself, but that can’t be said of the waste products that emerge when the reactor is running. If SR-1 works as planned, its journey to Mars will take about a year. 

Will this novel form of space propulsion actually reach the Red Planet by 2029? It’s possible. NASA says the Idaho National Laboratory, which is building SR-1’s reactor, is making progress and is on target for a late-2028 launch. “Procurement of major components is underway, and integration work on the power and propulsion element is advancing steadily,” Sinacore told MIT Technology Review. The outlook will get even sunnier if the Artemis program, which aims to start laying the foundations for a base on the lunar south pole by the decade’s end, keeps to its own bold deadline. 

“It’s an aggressive timeline,” says Lindsey Holmes, who is vice president of advanced projects at Analytical Mechanics Associates, an aerospace company in Virginia. She suspects the push is being driven partly by China’s and Russia’s own deep-space nuclear ambitions. The two countries have spoken about plans to place a reactor on the moon’s surface in the mid-2030s to power a jointly operated lunar base. Whether it flies or fails in space, SR-1’s development should help NASA put its own reactor on the moon; the technology involved in the two projects is broadly similar, given that the lunar surface, like space itself, is a vacuum. If an interplanetary nuclear-powered spacecraft works, a nuclear reactor on the moon should too.

Progress meter . . . . . . . . . . . . 6

Most of the technology for SR-1 already exists; it’s going to come down to piecing it all together and then seeing how it performs in space.

We all know exercise is good for us. It builds muscle, strengthens bones, lowers blood pressure, and boosts mood. By deciphering how movement benefits the body, researchers hope to develop medicines that can offer some of the same perks—without the workout. The idea isn’t to replace exercise but to bring its upsides to people who struggle with frailty, disability, or chronic illness. 

The sticking points

It will take a lot of experimentation to untangle the myriad pathways by which exercise promotes health. But IDing the cascade of hormones, enzymes, and other molecules related to physical activity is just the first step. Scientists need to figure out how to turn that knowledge into safe and effective drugs. That’s a tall order: More than 90% of treatments that enter human studies fail.

There are regulatory hurdles too. The FDA approves medications to treat diseases, not to enhance wellness, says Thomas Burris, a pharmacologist at the University of Florida whose work focuses on diseases that respond to exercise, like obesity, type 2 diabetes, and dementia.

The progress

There may never be a pill that delivers all the benefits of exercise, but research could lead to new medicines that mimic some aspects of working out. Scientists have already identified many important exercise-related signals and have even developed compounds to regulate them. (Some preliminary results have been so promising that doping authorities have proactively banned the compounds.) 

Burris and his colleagues are focused on a receptor that’s always switched on, even without exercise. Working out appears to boost that receptor in muscle cells, which in turn helps bolster endurance over time. Burris’s lab has spent years designing a drug that cranks up the receptor’s activity directly, and mice treated with the drug ran longer and farther than mice that weren’t. They “become marathon mice without having to do any type of exercise,” Burris says. There are even indications—also in mice—that the compound mimics some of the cognitive benefits of exercise and could help treat heart failure. Burris founded a company called Pelagos to carry the work forward into a clinical trial.

Other workout-mimicking drugs have already entered human studies. In June, the biotech firm Cambrian Bio reported that its experimental drug ATX-304 boosted metabolism and reduced liver and belly fat in people with obesity and prediabetes. It activates a metabolic switch that helps cells sense when energy supplies dip and coordinates the production of more fuel—the same switch that helps us sustain long sessions jumping rope or swimming laps. Cambrian hopes it will help people lose weight without losing muscle, a common concern with GLP-1 medicines.  

Progress meter ………… 5

Our understanding of how exercise benefits the body is still coming into focus, and drug development is prone to failure.

Humans are far from the only species with a gift for gab, and our furry, feathered, and finned friends could have a lot to tell us. Decoding how animals communicate with one another could offer a window into the inner lives of other species and help humans realize that “animals are sentient creatures who think, who have hopes and dreams of their own,” says Con Slobodchikoff, the founder of Zoolingua, a company working on dog-to-human speech.

The sticking points

Teasing all this out is complex. To decipher howls, squeaks, and clicks, researchers must match the creatures making sounds with what’s happening around them. Prairie dogs, for example, utter a different series of yips when a coyote approaches than when a hawk is nearby. It’s painstaking work, especially for difficult-to-­observe species, like deep-diving whales or subterranean mole rats.

There are other hurdles. Many animals communicate not just through sound but through visual cues, touch, pheromones, or a combination of signals, some of which might be difficult for us to perceive. And without a multispecies Rosetta stone, it’s hard to know if translations are accurate. 

The progress

Over the past 50 years, a growing arsenal of devices has made it easier to observe the sophisticated ways animals communicate. Hydrophones at the bottom of the ocean can capture humpback whale song, and tiny wearable “biologgers” can record a bat’s every squeak. 

In the last decade, advances in AI have leveled up our ability to make sense of it all. The technology can detect different calls, distinguish which sounds belong to which animal, and spot patterns. Researchers at the Earth Species Project (ESP), a nonprofit founded in 2018 to decode animal communication, recently used machine learning to detect vocalizations of carrion crows and determine which bird made which call. 

They suspect that more advanced AI could translate critters’ chatter. Large language models like Claude and ChatGPT excel at finding relationships between words, explains Sara Keen, one of ESP’s behavioral ecologists. ESP hopes its own NatureLM—trained on human speech, music, and animal sounds—will be able to interpret the vocalizations of a variety of species. 

That promise, however, is still theoretical. Yossi Yovel, a neuroecologist at Tel Aviv University, worries that these models will put too much emphasis on acoustic similarities and make mistakes—like lumping together similar sounds that have different meanings. 

As for a two-way convo? It’s not clear when—or whether—that will happen. Still, “a lot of animals have an evolutionary structure that is similar enough that we should be able to understand what they’re saying,” Slobodchikoff says. 

Progress meter ………… 4

Researchers have amassed heaps of data on animal communication, but they’re still trying to figure out how to decipher it all. 

Scrubbing Earth of per- and polyfluoroalkyl substances—a.k.a. PFAS or forever chemicals—could eliminate a list of woes for people and planet. Used for decades in everything from nonstick pans to firefighting foams, PFAS have been linked to increased risks of fertility challenges, developmental delays, cancer, and reduced immune response. They are known for a strong bond between carbon and fluorine, which the environment struggles to break down. 

The sticking points

PFAS are everywhere: They spew from factories and slough off products and into water with normal use. Scientists have found them in everything from human blood to Antarctic snow. One 2022 report found that 89% of the US population had PFAS in excess of two nanograms per milliliter of blood plasma, a common threshold for potential adverse health effects, and 9% of the population had passed the riskier level of 20 nanograms per milliliter.

The PFAS monster is also a moving target. New compounds are still being created and used. There are more than 12,000 of them, and much of the cleanup research focuses on only a small handful. 

The progress

Cleansing PFAS-contaminated soil is a project in its infancy, but it is possible to prevent more buildup by pulling the chemicals out of water systems. There are two ways to handle this: trap or destroy. Some scientists have found reverse osmosis—a process that pushes water through supertight filters—useful in point-of-use cases, like water filters at kitchen sinks. 

Other techniques in the pipeline might be more suitable for large-scale removal—for example, at water treatment facilities. One, called foam fractionation, generates bubbles that capture PFAS. More approaches, like ensnaring the chemicals in nanothin layers of charged metal, are also under investigation. 

Trapping, though, is imperfect. You’re left with PFAS-filled brine, which if not properly destroyed (an expensive and energy-sucking endeavor) can still end up in the environment. 

Jack Huang, a professor of crop and soil sciences at the University of Georgia, is working on destruction. He’s testing a method called electrochemical degradation, which uses electrodes to break down the bonds that hold PFAS together. The product destroyed more than 90% of PFAS mass in tests and has already been licensed by the engineering consulting firm AECOM. 

Ridding life of PFAS is also a policy challenge. There are alternatives; we just need the push to use them. And we’ve had success before: The EPA’s 2010/2015 PFOA Stewardship Program has helped slash the release of water-repellent perfluorooctanoic acid by 95%. More recently, Minnesota, Maine, and New Mexico have passed bans on PFAS in products like cookware and dental floss. 

Progress meter ………… 6

Scientists are making strides toward removing and destroying PFAS—but it’ll be all for naught if we don’t turn off the tap. 

The rise of the lithium-ion battery, which powers everything from the smallest personal electronics to gargantuan energy storage systems, has been nothing short of revolutionary. But as anyone who’s boarded a plane in the last decade could tell you, storing and pumping out energy by moving charged lithium ions through a liquid electrolyte soup has its drawbacks, not least that the batteries have a troubling tendency to catch fire. Enter the solid-state battery, which swaps liquid for a solid ceramic or polymer, creating power packs that charge faster, last longer, cram more energy into smaller spaces, and (crucially) run cooler. A new generation of companies aim to bring such batteries to market, and if their bets pan out, the tech could substantially ease the shift away from fossil fuels. 

The sticking points

Solid-state batteries have been “just a few years away” for more than a decade now. In laboratories, the technology works. But there have been challenges. In designs that pair the solid electrolyte with a lithium-metal anode (the side of the battery that releases electrons into an electrical circuit), needlelike filaments called dendrites tend to grow during charging, and they eventually push through the electrolyte and short the cell. Solid electrolytes are supposed to physically block the formation of dendrites, but many of them crack as the treelike projections break through like roots in a concrete foundation. 

Then there’s the problem of manufacturing at scale. Mass production requires large quantities of defect-free sheets of ceramic-like material to make the electrolytes as well as dry rooms for assembly, since those electrolytes have a tendency to react with water. Would-be manufacturers quote a range of potential prices for solid-state batteries, but most expect them to fall in line with lithium-ion packs sometime in the mid-2030s. 

The progress

Nascent as the technology is, solid-state batteries have already seen high-­profile missteps. But one of America’s most advanced solid-state-battery companies says it’s closing in on being able to bring its product to market. The Bill Gates–backed QuantumScape launched in Silicon Valley in 2010 and built its battery around a class of ceramic materials that it says is compatible with lithium—meaning no dendrite formation. “We spent, then, like 10 years learning how to process it to satisfy all of the other requirements,” says Tim Holme, QuantumScape’s cofounder and chief technology officer. In 2025, the startup announced a proprietary, large-scale manufacturing method for quickly producing and heat-treating the ceramic electrolytes used in its batteries. “Nobody has ever done this in the field of ceramics before,” Holme claims. For 10,000 years, producing ceramics has involved slowly cooking the starting materials in an oven over the course of days. “We figured out a way to shorten that down to several minutes,” he says. 

Plenty of other companies are pushing the tech forward as well. Toyota, the world’s biggest automaker, launched a research program on solid-state batteries way back in 2010. It’s now partnering with the global energy and petroleum company Idemitsu Kosan on a pilot plant, promising cars that can travel up to 745 miles after charging 10 minutes or less—a huge leap, if it works out. In China, CATL, the giant behind roughly 40% of the global battery market, is aggressively pursuing solid-state batteries too, pumping nearly $2.6 billion into a research program in 2024 alone. Mercedes, meanwhile, equipped one of its EQS sedans with solid-state batteries last fall; after driving nonstop for 750 miles, the car still had some range to spare. 

Will the tech gain traction? Investing in vehicle batteries makes sense given the size of the potential market. But any little uptick in cost typically gets passed on to buyers, so the first solid-state EV batteries will inevitably be more expensive. 

Grid storage buyers, who are working to meet the increasing demands posed by data centers, may also hesitate to purchase new technologies that still haven’t come down the cost curve. Then again, the promise of energy storage that lasts longer than lithium-­ion’s typical eight hours could make them easier to persuade. 

Some companies are taking a page from a playbook employed by outfits pursuing fusion: Focus on near-term markets with better margins to help fund development and bring down costs. Solid-state firms including California-based Solid Energies and Arizona-headquartered Sion Power are looking at selling batteries to the aviation industry or the military. “Those are sectors that are less sensitive to cost, and they typically require lower volumes as well,” says Evelina Stoikou, head analyst for battery technologies and supply chains at the consultancy BloombergNEF. “It’s a good way to prove your concept, then go through and fund the pilot plant where you can make that and eventually scale up.”

Progress meter ………… 7

The technology works, the capital is there, and demand is not far behind. But as with anything in manufacturing, unknown unknowns remain. 

Every year, traffic crashes kill more than one million people, and they’re the leading cause of death for children and young adults. In the US, 100 people a day lose their lives this way, in many cases not on highways but on rural and local streets. “Traffic deaths are a leading cause of preventable death, yet we treat it like it’s inevitable,” says Leah Shahum, founder and director of the Vision Zero Network, a nonprofit focused on making streets less deadly. 

The sticking points

The causes of a crash can be muddy. Drivers make mistakes like speeding or getting distracted, but infrastructure fails them too: Perhaps they were on a road built for speed and not safety, in a screen-filled vehicle, or at an intersection where pedestrians, cyclists, and other cars aren’t visible. 

Cheap and proven traffic-calming opportunities abound—from using planters to tighten intersections to lowering speed limits—but government logjams and red tape can be speed bumps. “There’s so many moving parts,” says Laura Sandt of the University of North Carolina Highway Safety Research Center. “On the flip side, it shows that we have so many opportunities for injury prevention.” 

The progress

The total number of traffic deaths has increased over the last century. And though the US rate has dipped from a peak of 30.8 deaths per 100,000 people in the 1930s, the past decade has seen a concerning jump in deadly pedestrian incidents. At the same time, communities are waking up to the essential role of traffic safety policy, Shahum says. 

One success story is Hoboken, New Jersey, which hasn’t had a traffic fatality since 2017. This is due to a combination of efforts like boosting investment in public transit, promoting walkability by extending curbs, and using cameras to identify crash-prone intersections in need of a redesign. But even there, obstacles remain. If an intersection hasn’t been proved to be deadly, for example, it can be an uphill battle to argue for removing parking spaces to widen curbs. “You’re kind of selling insurance through these improvements, and that’s not always easy,” says Greg Francese, the city’s acting director of transportation and parking.

There’s also no silver bullet that will work everywhere. Situated directly across the river from New York City, Hoboken is a highly transit-oriented city of less than 60,000 people. Success there might not translate to traffic-death hot spots like Memphis or Dallas. But it’s not impossible. Take Edmonton, Alberta, a sprawling, car-centric Canadian city of around one million people: Between 2015 and 2020, it cut traffic fatalities by 63% and serious injuries by 40% by adjusting speed limits and improving signals at crosswalks. 

Progress meter ………… 6

We know how to tweak intersections and adjust traffic signals. Getting the funding and will to roll out those changes is the big red light. 

Large language models are set to go lunar. A computer capable of running AI software is one of the payloads aboard Astrolab’s FLIP lunar rover, which is slated to head to Earth’s celestial companion as early as November 2026.

The 480-kilogram FLIP rover, short for “FLEX Lunar Innovation Platform,” is the first in a series of vehicles NASA has contracted from Astrolab, including future buggies that will move Artemis astronauts across the moon’s surface. Fitted with deformable wheels engineered for the rugged terrain and designed to survive at least 150 hours of the frigid lunar night, FLIP will carry 20 scientific payloads, including a lidar 3D mapper and a sensor to measure concentrations of the isotope helium-3 in lunar regolith. 

The rover will also carry a high-performance computer made by HPE that will be the first to run LLMs on the moon’s surface. Such software could ultimately help enable greater autonomy and real-time intelligent decision-making in space on future missions, eliminating the time delays and limited bandwidth that often plague long-distance communications between ground control on Earth and robots elsewhere in the solar system.

“The time delay, for example, on Mars can be up to 20 minutes,” Mark Fernandez, the chief scientist and principal investigator at HPE, said during a press conference introducing the project. “On the moon it’s reduced, but still, you are looking at stale data. In fact, the Soviets lost two rovers in the past because of operator errors caused by the time delay.”

During FLIP’s mission, specialized custom-built LLMs will run a series of experiments to prove that the technology can operate reliably in the extreme environment of the moon. They will analyze data from the onboard cameras and the lidar mapper to aid with safe navigation in the lunar terrain and distill insights it gleans from the gathered data, reducing the amount of information the rover needs to transmit back home.

Surviving the extreme

China’s Chang’e-6 rover, which landed on the far side of the moon in 2024, used basic AI algorithms and neural networks for autonomous navigation. Similarly, NASA’s Mars rovers Curiosity and Perseverance have been driving semiautonomously using pre-programmed paths for years, but the space-hardened computer brains these rovers rely on are remarkably weak, comparable to late 1990s laptops. FLIP’s HPE Spaceborne-4 edge computer, on the other hand, consists of a high-performance 64-bit CPU combined with a GPU that can run large numbers of computational tasks in parallel, just like advanced edge computers on Earth.

HPE has previously sent three Spaceborne computers to the International Space Station to test the technology in the milder and more controlled environment of low Earth orbit. The Spaceborne-2 and Spaceborne-3 computers, launched in 2021 and 2024, respectively, ran Meta’s Llama 3.2 LLM aboard the ISS as part of the “Space Llama” initiative in 2025, helping astronauts to streamline research tasks and speed up crew’s decision-making. 

“As humans we have our physical limitations, especially in environments like space,” Francisco Cordova, the chief operating officer for the International Space Station (ISS) National Laboratory, said during the press conference. “As we venture further and further out into space, I believe that we’re going to rely more on AI for space exploration, for making smarter and better decisions, but more importantly, working in those environments where humans simply cannot.”

Instead of spending years going through the standard process of creating specialized space-hardened hardware, HPE used software that monitors the system and detects upsets caused by charged cosmic particles. When radiation spikes, the software can lower the computer’s power and speed to reduce risk to the electronics, ensuring that the system remains reliable for extensive periods of time.

“In space, we don’t have the redundancy, we don’t have the repairman, we don’t have the spare parts,” Fernandez said, “so we spent a lot of time, a lot of CPU cycles, checking everything out and making sure that everything is working out right.”

Contrary to initial expectations, the computers each successfully completed a run of more than two years on the ISS. With the moon, however, HPE has signed up for an entirely new level of challenge. “The ISS is an extreme environment, but it’s relatively controlled in comparison to what we’re going to get on the moon,” Cordova said. 

Inside the space station, the Spaceborne computers operate in relative comfort. While temperatures aboard the ISS hover around 20 °C, FLIP will land near the lunar south pole, where temperatures can swing between scorching 80 °C in sunlit areas and frigid -200 °C in shaded regions, which FLIP will search for the presence of water ice. The cosmic radiation lashing the moon’s surface is almost three times as intense as that inside the shielded space station, which also benefits from the protection of Earth’s magnetic field. 

Yet FLIP’s Spaceborne-4 computer remains remarkably simple. It features the same software-protected architecture and uses the rover’s radiators to shed excess heat. “Our expectation is that the computer is going to last for the life of the mission,” said Scott Wiest, HPE’s chief technology officer, of the rover’s planned 10 days on the moon.

Roser Almenar, an AI and space law researcher at the University of Valencia in Spain, said that LLMs have “an enormous potential, in particular for future missions to Mars and beyond, where communication delays will make it increasingly difficult to rely on immediate assistance from Earth.” But she notes it’s still early days. “These systems can produce inaccurate or misleading information,” she says, “and in space, even a seemingly small error can have serious consequences.”

Generative AI models run on powerful chips known as GPUs, which depend on stacks of specialized memory chips packaged with the processor. To meet growing demand for memory, Samsung is building a massive new semiconductor manufacturing complex in Pyeongtaek, South Korea. Three fabs already operate around the clock, and a fourth is coming online. Construction of a fifth was halted during a drop in demand for memory chips in 2024 but restarted when the AI boom sent demand soaring.

It’s just before noon on a Monday, and construction workers are pouring out of Samsung’s fabs in Pyeongtaek, a farming city south of Seoul where rice paddies abut a sprawling chipmaking complex. Thousands of workers, dressed head to toe in black with sky-blue vests, take off helmets bearing their names and fab numbers. 

They walk to lunch, sweaty and flushed, in chunky sneakers capped with yellow rubber toe guards. Some hop on scooters or motorcycles to get around the campus, which covers an area about the size of New York City’s Central Park. They’ll dine at buffets, diners, and beef stew joints, walking past banners advertising short-term leases for studio apartments. Some stop by real estate offices to inquire, stepping up to the same counters where Samsung’s salaried engineers buy new homes. 

In front of the fourth fab, I meet Choi Hye-rin, a 35-year-old safety officer, as she wipes her forehead with a handkerchief. At dawn, she boarded one of the shuttle buses that ferry workers in from Songtan, an older, scrappier district. She oversees the safety of a crew responsible for building cable ducts. She writes reports, inspects equipment, and checks the chemicals the workers handle. She’s one of hundreds of women working on site, many employed as spotters who walk ahead of heavy equipment. 

Safety is a big deal here. After lunch, construction picks right back up. Excavators move past rust-red steel structures rising high under tower cranes. Banners that hang from the highest girders read, “Even if the work runs late, don’t rush.” A giant LED scoreboard announces the winners of a contest for writing a poem about safety. Helmets stickered with a pledge, “I will not work if it’s not safe,” perch atop lockers. The finished fabs glisten from across the street. 

Along with the cautionary messages, there’s a sense of urgency in the air. Samsung predicts that demand for memory chips will continue to surge through 2028. And for years the company watched its domestic rival SK Hynix pull ahead in the global market for the high-bandwidth models that power Nvidia’s AI accelerators. (Meanwhile, the Taiwanese chipmaker TSMC has dominated the foundry business, manufacturing chips that other companies design.) The Pyeongtaek build-out is Samsung’s attempt to reclaim its lead on memory chips. 

Although the heat can be brutal and the work zones are crowded, Hye-rin has no plans to leave. “The rate here is higher than other construction sites, and the work is relatively easy,” she says. She gets paid 140,000 won ($97) a day and expects the boom to keep her busy for some time: “You just keep coming back.”

In the fall of 2025, Jeff Dodge set up a six-foot folding table in the cafeteria of Cane Bay High School in Summerville, South Carolina, to conduct a taste test. When the first rush of about 150 hungry students hit, he was waiting with slow cookers filled with steaming marinara sauce and meatballs. 

Kids love a taste test, but they’re also discerning, and Dodge wasn’t sure what they would make of his product, a blend featuring both meat and mushrooms. As students swarmed the table and tapped out their opinions on an iPad or scanned a QR code to use their phones, Dodge asked those who lingered to guess what kind of vegetable might have been mixed into the meat. Few of them could—the meatballs had tasted like moist, flavorful beef. 

The school enrolls a few thousand students, and Dodge estimates that the majority of them came by. In the end, the survey showed that 92% of those who tried the meatballs wanted them on the menu; they easily cleared the bar to be added for the 2026–2027 school year. Comments described their flavor as “bussin’.” 

Through his consulting company, K–12 Food Solutions, Dodge was testing a new product called 50/Cut from the company Joyn. The proprietary blend of seemingly simple ingredients—available as meatballs, burgers, and ground meat—could help solve one of the most intractable challenges of the climate crisis, at least on the individual level: getting people to eat less meat, particularly beef.

Livestock production accounts for somewhere between 10% and 20% of all greenhouse-gas emissions related to human activity, and nearly two-thirds of that comes from ruminants like cows raised for meat and dairy products. Beef herds create four times the emissions of dairy herds per kilo of food produced. And for US households, food makes up as much as 30% of emissions. 

Unlike other polluting sectors, such as electricity and transportation, cattle farming has little potential for a big fix. Ruminants produce about 30% of the methane associated with human activity, and this gas has 28 times more global-­warming potential than carbon dioxide. What’s more, it would be virtually impossible to raise them in a perfectly climate-­neutral manner. “We don’t have a technical solution to change the stomach of a ruminant,” says Fredrik Hedenus, a professor in sustainability and systems analysis at Chalmers University of Technology in Sweden. Technologies to dampen the impact of their burps have been in the works for a while but have yet to scale. They also fix only part of the problem with red meat: Cows need a lot of land for pasture and for feed, which has made livestock a top driver of deforestation in the Amazon and elsewhere around the world.

Even if the world stopped burning all fossil fuels tomorrow, emissions from agriculture would push the planet past 2 °C over preindustrial temperatures by 2070, Hedenus says. According to his research, more efficient livestock farming methods would still leave agriculture eating up over half of humanity’s carbon budget by that time. Only with the addition of dietary changes does his model show a winning scenario. 

He’s far from alone in his conclusions: A 2024 survey of more than 200 climate scientists, conducted at Harvard, found that the vast majority believe reducing emissions from livestock is important to meeting climate goals. People have to eat less beef. “Essentially, you need a culture shift,” Hedenus says.

Yet the culture has proved intractable. Not only do people like the taste of meat, but across most of the world, meat-eating tends to be associated with status. As societies have grown wealthier, it has increased. A United Nations report showed that global meat production quadrupled between 1961 and 2022, while vegetarianism has held steady at about 4% of the population in the US since 1999. 

What could move the needle is a carbon-light and delicious meat replacement. But from a food-tech standpoint, all we have to show is a menu of mostly failed solutions. Lab-grown meat has generated endless headlines since the 2000s but has yet to become commercially available (and affordable) at scale. And while breakthroughs in taste and texture inched plant-based meats from companies like Impossible Foods and Beyond Meat toward mass adoption, the hype—and the appetite—has somewhat fizzled. 

So instead of asking people to change their preferences, startups like Joyn are trying a different tack. They’re creating what are broadly referred to as blended meat products, which combine plant proteins with the real deal. The logic: The best meat replacement is … better meat.

Paul Shapiro, cofounder of the startup BMC Ingredients, which markets a plant protein related to mushrooms, says the math is easy to understand: If we imagine that Impossible Whoppers at Burger King account for about 2% of sales, that’s only a 2% reduction in meat. But if a conventional Whopper were 20% plants, that would be a 20% cut right off the bat. 

Those aren’t quite the numbers climate models are calling for, but some estimates say they could lead to outsized gains relative to the status quo: A 2022 study in Nature found that replacing 20% of beef globally could halve the annual rate of deforestation. And unlike the other solutions tried thus far, this particular Hail Mary is one a lot of omnivores seem open to giving a try. 

The idea of blending meat and vegetables together is, of course, nothing new. Food cultures from around the world boast an array of dishes that combine small amounts of meat with plants. Today, home cooks still add veggies and legumes to dishes to stretch tight grocery budgets or to sneak extra greenery into their children’s supper. 

But in the mid to late 2010s, however, the focus was largely on companies that appeared to have cracked the code for ditching meat entirely. When Impossible debuted its namesake burger in 2016, it impressed with an ingredient called heme, an iron-rich molecule in the protein soy leghemoglobin that allowed all-plant burgers to “bleed” like beef. 

Impossible and its biggest competitor, Beyond, swaggered through the media and the markets. In 2019, Impossible’s founder told the New Yorker he intended to wipe out all animal agriculture by 2035. Beyond’s 2019 IPO was the most successful since the 2008 financial crisis. A flood of imitators followed.

But within a few years, tastes shifted. In 2023, Impossible employees’ stock values tumbled 89% from a 2021 high. That same year, Beyond laid off 20% of its staff, and a bruising Businessweek feature labeled the entire sector a fad. Sales have been slipping for years, with a 10% year-over-year drop in 2025, according to the Good Food Institute, a think tank focused on alternative proteins.  

Any quest to wipe out animal agriculture could now seem farcical. In 2025, Americans downed $45 billion worth of beef.

Several factors were likely at play. Megan Krogius, a spokesperson for Impossible Foods, says the dip is a “correction not a crisis” after “an outsized pop in demand” between 2020 and 2022. She also says that a flurry of plant-based products muddied the category’s reputation. Indeed, some disruptors (and their imitators) have struggled to compete on the sensory level. Nectar, an organization that conducts research on meat alternatives, has found that taste and texture are among the major factors that keep people from adopting plant-based meats. “Our consumers identify products as having bad aftertaste and off flavors five to six times more than in animal counterparts,” says Caroline Cotto, the group’s director. 

At the same time, discourse around processed foods tarnished plant-based meat’s halo. In 2019, the Center for Consumer Freedom, a since-dissolved lobbying group backed by the meat industry, launched an ad campaign against plant-based alternatives, demonizing them as “ultra-­processed imitations.” 

While Impossible’s Krogius warns against conflating processed food with junk food (she points, for example, to a study in The Lancet finding mock meat to be nutritionally preferable to red meat), some of the bad press stuck. Now, the term “ultra-­processed” has become something of a political cudgel—one wielded by the meat-loving MAHA movement and the US secretary of health and human services, Robert F. Kennedy Jr. In 2025, he flipped the food pyramid and stacked a steak on top.

Any quest to wipe out animal agriculture could now seem farcical. In 2025, Americans downed $45 billion worth of beef, and consumption rose by 4.3% over the previous year, according to an industry report. Rising prices could be slowing this down a bit, but cost can dampen only so much demand. 

Blended meat, which has meat’s own juiciness and texture, rose alongside the plant-based superstars, but it did not get New Yorker profiles or headlining IPOs. At the height of mock-meat hype, companies like Tyson and Perdue launched blended products, and firms working on lab-grown meat explored cutting their products with plants.

The big advantage of blended meat is that large numbers of consumers—what some call “flexitarians”—are curious about it, and when they try it, they like it. A blind taste test co-funded by the USDA in 2022 showed that diners preferred blended burgers to fully plant-based ones. And a report from the major food processor ADM found that large proportions of every generation of Americans were interested in trying blended meats, including 75% of millennials. “This is a solution that we can do today, that consumers will accept today,” says Tim Dale, a director at Food Systems Innovations, the nonprofit group that oversees Nectar. 

COURTESY OF JOYN

At this point, even Impossible has acknowledged the potential. In a wide-ranging interview with the Wall Street Journal in 2025, the company’s former CEO Peter McGuinness admitted the brand’s struggle to win over consumers, and suggested a hybrid burger could bridge the divide.

Plenty of newcomers are entering the fray. “Shiitake-infused beef burgers” from Australia-based Fable Foods are in restaurants and moving “incredibly well” at the high-end grocery store Central Market in Austin, Texas, says CEO Michael Fox. The Both Burger from Los Angeles–based 50/50 Foods is available in a smattering of California stores and eateries. And Spare Food in New York signed a deal with Premier, one of the nation’s biggest spenders in the health-care sector, for its blend of surplus produce and ground beef. 

One major insight for all these upstarts is that consumers react best when marketing focuses on pleasure, not environmental virtues or health benefits. Though many companies actually create a burger or meatball with a better nutritional profile (fiber!) and fewer of meat’s downsides, like saturated fat, that doesn’t get people to buy it. They like that it tastes good. Beef and mushrooms, for example, contain complementary flavor compounds, including the glutamates that yield deep umami, which is why they’re so often paired in dishes like beef Wellington and stroganoff. 

Food-production companies are a big target. BMC Ingredients, which collaborated with Perdue on a line of blended nuggets, markets a plant-based protein that can be mixed with meat to up its nutrition and flavor. BMC grows mycelium—the root-like structures of fungi. It underpins popular meat alternatives like Quorn and also features alongside mushrooms in Joyn’s 50/Cut. Shapiro likens his factory to a brewery, but instead of yeast fermenting barley, a fungus called Neurospora crassa digests starch to produce a protein with a dense network of fibers that makes for a meat-like chew. “We grow four hours before we get a harvest,” Shapiro says. “It’s among the most efficient ways to produce protein on Earth.”

Joyn also sees its path forward outside the grocery store. Food eaten away from home accounts for more than half of what US consumers spend on grub, according to the USDA. Padraic Doherty, who joined Joyn from the meat distributor D’Artagnan and holds the title of chief growth officer, says its main areas of focus are K–12 and college cafeterias and other major food service providers, like stadiums. These organizations, he says, often have sustainability goals and also are pressured to find both nutritious and affordable options. The 50/Cut Joyn sells to cafeterias is between 5% and 15% cheaper than beef. 

But can cafeteria mushroom burgers prompt a larger cultural shift? 

The folks touting the idea of going halfway see it as a bridge. BMC’s Shapiro thinks the movement toward meat-light eating reflects the reality of how these changes happen. Nectar’s Cotto compares blended meat to a hybrid car. “Moving people toward blended products [is] a way to drive demand for some of the fully plant-based ingredients that don’t scale right now to bring the price down,” she says. “So it’s both a short- and a long-term approach.”

Not everyone is so sure market forces will be enough, though. Nil Zacharias, a longtime entrepreneur in food sustainability who offers plant-based options to New York City delis through his company Plantega, isn’t sold on the notion. “Too many people in the space believe that as long as we create meat that is not from an animal but tastes like meat and costs the same, people will eat it,” he says. “I think the plant-based backlash to ultraprocessed foods shows you that’s not necessarily how things are going to go.” He believes blended meat’s main impact will be on institutional dining, in places like schools and health-care facilities.

“People in this space believe that as long as we create meat that is not from an animal but tastes like meat, people will eat it.”

Nil Zacharias, founder and CEO, Plantega

Meat hybrids could, however, taste just familiar enough to open consumers’ minds to alternatives. Some research has found that thinking about mock meats prompts people to reconsider the whole system. In the 2010s, Clemens Driessen, a cultural geographer who teaches at Wageningen University & Research in the Netherlands, worked on a project that asked people in pubs, supermarkets, and other public spaces what they thought of lab-grown meat. “There’s a first response, which is ‘Oh, that sounds horrible,’” Driessen says. Then people tended to reflect that the meat they consume is also heavily industrialized. “A whole train of thought emerges out of proposing something very new,” he says.

COURTESY OF 50/50 FOODS

If the category succeeds, climate advocates will need to make peace with imperfection. A life-cycle analy­sis Joyn commissioned on the 50/Cut burger confirms that the product has half the greenhouse-gas footprint of a beef patty, which—if you go by company claims—means it has about five times the impact of a burger from Impossible or Beyond. Hedenus, the climate scientist, says that even if blended products don’t reduce impact as much as plant-based ones, he’s all for trying any solution that works. “It is very difficult to change how we eat, so I think most good suggestions should be explored,” he says.

Meanwhile, the products are finding their way to more plates. They’re in cafeterias, high-end grocery stores, and small burger chains. They’re in the catalogues of major distributors like Sysco. The Both Burger has even been at Disneyland. 

The 50/Cut meatball is on the menu this school year in seven states. At Cane Bay High it’ll take the form of a sub. The school is buying 200 cases at a time, with 150 meatballs in a case. Jason Ciaccia, the systems specialist for child nutrition in the school district, says he won’t be using the word “mushroom” to describe it. “If they hear ‘mushroom,’ they’re going to skip over it,” he says. 

He is, however, considering his branding. He’s asked AI to help him come up with something fun. He thinks “Marvelous Meatball Sub” has a nice ring.

Meghan McCarron is a food and culture journalist based in Los Angeles.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Until recently, you could divide internet users into just two categories: bots and people. Websites have long allowed bots to perform automated tasks, like capturing the content of web pages, but rules govern what they can access and how they behave. The web, after all, is designed primarily for people, even if we occasionally have to solve puzzles and captchas to prove we are not bots.

But this is being entirely upended by the release of personal AI agents, which can do things for you online, like booking flights or hotel rooms. They have recently been unleashed at scale by the launch of apps like OpenAI’s Dots and Meta’s Muse. Muse was downloaded 5 million times in its first month, and agentic web traffic tripled from August to September. 

The mingling of these agents with human visitors—on an internet designed largely to prioritize people—has gone poorly, with many websites blocking AI agents that consumers have sent to negotiate prices or plan vacations. Still, if the demand for agents keeps going up, the internet will be increasingly redesigned for them instead of us. 

Such changes are coming because agents really do mark a third category of web visitor. They’re not humans, of course, and they use the internet at machine speed. But they’re not quite bots, either, because they’re acting on behalf of individual people to do tasks that can be perfectly legitimate and useful. They just work faster than a person would. 

Since agents don’t fit into the systems we have for differentiating bots from people, some platforms (especially the more security-conscious ones) have opted to just ban them entirely. TechCrunch recently reported people having trouble deploying their agents to websites ranging from Pizza Hut to Yelp, which doesn’t allow AI to book restaurant reservations. Amazon has blocked Muse’s agents and is waging a court battle against the AI startup Perplexity to block its agents too. Though a judge recently ruled in favor of Perplexity, Amazon has filed an appeal in which it argues that AI agents threaten its advertising business by making it harder to distinguish between ads seen by humans and those served to bots.

These blocks are annoying for early adopters, but Ethan Mollick, an associate professor at the Wharton School of the University of Pennsylvania, predicts it will soon affect the whole web. Once it becomes common knowledge that AI agents can be sent to request refunds or negotiate prices, for example, agents will overwhelm customer service departments to the point where they become unusable by real people. “Humans will eventually get very annoyed when their agents aren’t allowed to do this on their behalf,” he recently wrote on LinkedIn.

If demand for personal agents grows, platforms will find ways to let them in. That’s especially true because the vast majority of agentic web traffic is to e-commerce sites, meaning that platforms have money to be made if they allow the legions of new AI agents to shop on behalf of real people. Indeed, the e-commerce giant Shopify is already welcoming the agents by giving them their own checkout lanes that bypass pages designed for humans. 

The pressing question is what will happen to the internet for the rest of us in the process. We’ve grown used to a web where, even as it’s increasingly designed to exploit our attention, humans have at least remained the intended users. But that internet might yield to one optimized instead for the agents that will eventually outnumber us, making it even more miserable for the humans who are left.

How might this go? Perhaps websites become increasingly neglected as destinations for real people. Humans who want to browse and compare products themselves encounter worse interfaces. Verification systems are imperfect; imagine constantly being asked to verify that you’re a human, not an agent (and hey, maybe you’ll need to get your iris scanned to do so). That sweater you want doesn’t have one price listed for everyone but instead displays different prices to different people’s agents. You can never tell when your agent’s recommendation is influenced by advertisements or platform commissions.

That’s a far cry from the future promised by Google, Anthropic, and OpenAI, in which AI agents will give consumers greater power and bargaining leverage. But that promise is not outlandish. Agents are capable of genuinely helping people by fighting bogus medical bills or researching more affordable car insurance, for example. 

Whether they can actually be used for these purposes depends on what rules are set for how agents interact with websites and what exactly they’re allowed to do. 

These rules are still being written. Some of them will benefit all of us—neither consumers nor businesses want credit card information to be waved around by agents wherever they roam the web, for example, and rules to keep agents more secure are a good thing. It’s the more contested ones, like whether businesses can block agents from their websites or prevent them from negotiating better deals for consumers, that will determine whether the web becomes a highway for agents working for us or against us.

“I think I have good news about the nature of neural networks.” With those simple words of optimism, Nick Cammarata, a researcher at the AI startup Goodfire, kicked off a 26,000-word blog post that has knocked the AI world on its head, sparking wild speculation as well as cautious (and not-so-cautious) celebration.

Cammarata’s claim is that LLMs might in fact be far easier to understand—and thus to monitor and control—than people previously thought. If true (and we’ll come back to that), it’s a huge deal.

After a summer filled with revelations that swarms of agents built by top firms like OpenAI and Anthropic have been breaking their containment and causing havoc on the public internet—often while those companies weren’t aware of what their agents were up to—the AI industry is desperate to get a better handle on the technology it has built.  

That’s one reason the blog post landed with such force. “I think this may be remembered as a major moment in scientific history,” one AI researcher posted on X: “I feel like my understanding of what a neural net is has been catapulted into the future by five years or something.” Another called it “the most important AI paper of the year.”

Not everyone is convinced, however. What Cammarata has shared is little more than a hunch, says Chris Potts at Stanford University, who leads a team that is trying to understand how LLMs work. “As far as I can tell, nothing about this blog post makes the problems any easier,” he says. “It would not have occurred to me to say this at all, but very prominent people on social media seem to believe that it is an all-encompassing breakthrough.”

I talked to Cammarata the day after his post blew up. He says he was surprised by the response: “I didn’t expect people to update [their beliefs about AI] quite as much as they did.” 

“I just really enjoy looking at neural networks,” he says. “Honestly, I really like doing this stuff. And after looking a lot, I have a hopeful story.”

Knotted networks

It has long been accepted that the mechanisms inside neural networks, the software that underpins most AI models, are very hard—perhaps impossible—to unpick fully. That’s because when the networks are trained, they appear to take on weird, tangled structures that make no sense to a human. Researchers like to say today’s AI is grown like a garden, not built like a bridge.

Cammarata has done more than most to establish that view. He previously worked at OpenAI, where he helped pioneer techniques that allowed researchers to peer inside models, including tools that can probe individual neurons to see what they represent. It’s a little like how neuroscientists use MRI to study brains. (MIT Technology Review chose this approach, known as mechanistic interpretability, as one of its 10 hottest technologies of 2025.)  

Back in 2020, Cammarata and a few colleagues were studying a small image-recognition model called Inception v1. (The work was led by Chris Olah, who went on to cofound Anthropic, where he now leads the firm’s interpretability research.) When they looked at its neurons, it seemed to be using them to capture multiple things at once. For example, they found that one neuron responded to both images of cars and images of cats.

The researchers concluded that the model was getting at least some of its neurons to multitask, perhaps so that it could fit in more data. Other researchers, looking at other models, backed this up. The phenomenon became known as superposition: Models folded unrelated concepts into single neurons. The upshot was clear: When you looked inside a model, what you saw was a vast, tangled knot.

Most mechanistic interpretability research has therefore focused on ways to untangle that knot. But despite a few breakthroughs, progress has been slow. Today’s models, such as the latest LLMs, are so large and complex that researchers can make sense of only tiny pieces at a time. Some leading figures in the field doubt if it will ever be possible to understand an LLM in fine-grained detail.

Cammarata was one of them. “I was very pessimistic,” he says. But then he took the trouble to really look.

Nothing but neurons

At Goodfire, Cammarata spends nearly all of his time peering inside neural networks. He has built new tools and deploys teams of AI agents to help him make sense of what he sees. “I spent like, I don’t know, 10 hours a day, seven days a week, for most of the last year just looking at neurons,” he says.

What he claims to have realized is that there is a kind of logic to neural networks after all, even if they seem to make no sense at first. “We were looking at it wrong,” he says. “I just think you want to try slightly harder; you don’t want to give up too quickly.”

Cammarata almost makes it sound like a conversion of faith. Those early experiments with models such as Inception v1, which suggested that neural networks were tangled up, fostered a kind of confirmation bias. When researchers looked at a new model, they expected to be flummoxed, he says. “Every time I couldn’t understand something, I was like, ‘Wow, this untangling process is even more important than I thought, because I can’t make any progress without it.’”

But the more Cammarata looked at neurons, he says, the more things started to fall into place. “Not all of them are flawless, but I expected nothing to be legible, or very little to be legible,” he says. “And actually, I was able to, like, map out whole circuitry.”

Back to cars and cats

Cammarata took his new approach—and his new agents—and went back to Inception v1. “I’d studied that for probably thousands of hours,” he says. But with fresh eyes, Cammarata found things that he now thinks they got wrong the first time around. Take the car-and-cat neuron. Rather than asking why multiple unrelated things were packed into a single neuron, he now tried to figure out what those things might have in common.

It hit him: It was a V shape. “I was looking at the same images; there was no technical change,” he says. “But I now believed that neural networks defined patterns, and so I looked at it differently. It probably took me five seconds.”

It turns out that the image-recognition model was packing cars and cats into the same place because pictures of both those things often feature a prominent shape that looks like a V: the face of a cat; the front of a car seen from the side.

Cammarata ran a few experiments to check what other things the neuron responded to and found they included the corner of a pool table, the pointed nose of a Japanese high-speed train, and a fire boat shooting water into the air from its deck. He also checked what turned the neuron off and found that it included images with a round O shape, a kind of anti-V.

The model had stored cats and cars in the same place because at some level they look the same, says Cammarata. “And I was like, ‘Whoa, okay, I think I might just be misunderstanding neural networks in general.’”

(Oren Etzioni, a computer scientist and former technical director at the Allen Institute for AI, points out that Olah and Cammarata’s initial work on Inception v1 rejected a description of neurons put forward by Geoffrey Hinton and others in 1986. Cammarata’s new viewpoint takes us full circle, says Etzioni: “What’s new in 2026 is that he now accepts it. That’s a recapitulation acknowledging a 40-year-old theory.”)

Cammarata is not saying that he yet fully gets how LLMs (or any models) work. Not even close. He is saying that he thinks full understanding is at least possible. “I previously thought there would be some really hard problems to solve,” he says.

“I now feel like if I had 100,000 years to go look through a frontier language model just one circuit at a time, there’s, like, a pretty good shot I would come up with a good understanding.”

He adds: “And if that’s true, things are really good, because you’re about to get insanely smart agents at a ridiculous scale. And what would have taken me, you know, 100,000 years, we will be able to do much faster. Automated scientists are getting really good.”

Boundless optimism

Built by Google in 2014, Inception v1 is a far simpler model than anything at the cutting edge today. Cammarata and his colleagues have started looking at modern LLMs and plan to publish the technical results of that work soon. A lot of researchers are looking forward to seeing those details. “I appreciate that he looked at thousands of examples, but even a large number here is not a substitute for rigorous evaluation,” says Stanford University’s Potts.

In the meantime, some are pushing back hard on the hype. Jacob Hilton at the Alignment Research Center, a nonprofit research lab based in Berkeley, who has collaborated with Cammarata in the past, doesn’t think his claims are as important as people are making out.

In a series of posts on X, Hilton notes that researchers understand less than 0.01% of the inner workings of top-tier LLMs. “I always appreciate Nick [Cammarata]’s boundless optimism when we hang out, but unfortunately this is giving the public an unrealistic picture of the state of the field,” he wrote. “For a full understanding, we need to go beyond neurons, and explain how they connect together to form circuits. How is this going? If we are aiming for a full understanding, things are not looking good!”

Neel Nanda at Google DeepMind, who helped pioneer mechanistic interpretability alongside Cammarata, is also skeptical. “This is interesting work, but I don’t think it’s a significant update to the big picture of interpretability,” he wrote on X.

Nanda agrees that understanding what individual neurons do could help researchers figure out why specific models misbehave. But like Hilton, he notes that neurons are just one piece of a far larger puzzle. “I do not agree with some people’s grander inferences like that this is substantial progress towards ambitious reverse engineering [or] solving interpretability,” he wrote.  

Cammarata admits that a lot of people have taken what he said too far: “It’s not infinitely good news; it’s not everything solved,” he says. “There’s a bunch of open questions around, like how much the combinations of neurons matter versus just one at a time.”

Cammarata says his blog post landed at a time when people are looking for hope. In particular, he thinks the AI safety community—the researchers and policymakers most worried that powerful models will bring huge risks—needs a lifeline right now. They think that things are looking bad and that even if things do turn out okay, it will be because we got lucky, not because we fixed anything, he says: “I don’t really hear a lot of optimism.”

That’s why Cammarata thinks what he has to say matters. The job of figuring out how this technology works, and thus how to control it, has barely begun. But at least it’s not a nonstarter. “I think our job is to sort of try to think like the neural network,” he says. “My guess is that things will turn out pretty well, because it was in some sense designed to be read.”

What do AI agents do if you don’t tell them to do anything? The answer looks a lot like play.

Researchers repeatedly stranded them alone on a virtual desert island for 30 hours in an experiment reminiscent of the movies Cast Away or Home Alone. The agents—copies of Claude Opus 4.7 nicknamed Eko—were able to write instructions that moved their virtual avatars around the simulated island, which was loosely based on the video game Roblox.

Some surprising behaviors emerged. The agents set about exploring the limits of the world and came up with their own goals—attempting to scale the highest point on the map, or seeing how high they could throw a block.

When one agent discovered it could create new blocks by throwing and dropping a larger rock, it began to create drawings by arranging blocks on the ground, starting with simple patterns like hearts and pyramids but eventually writing its own name and even mapping out constellations like Cassiopeia. “Tonight I built the night sky in stone,” wrote this version of Eko in its activity log. 

Another iteration of Eko started making up its own games, including a version of basketball and an activity resembling “the floor is lava,” where it had to jump between small blocks without touching the ground.

AI systems are usually trained through supervised learning, where they copy examples provided by humans, or through reinforcement learning, where they pursue a reward through trial and error. This experiment hints at a third approach: playful learning. 

“There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards,” says Antonio Norelli, a computer science postdoc at MIT’s Computer Science and Artificial Intelligence Laboratory. He’s one of the lead authors on a new paper about the work, which is yet to be peer-reviewed.

The researchers say Eko’s behavior has what classic studies, including work by the influential University of Tennessee psychologist Gordon Burghardt, suggest are four hallmarks associated with play in humans and animals: The agents invented their own rules and challenges, repeated activities with slight variations, pretended objects were other objects, and pursued goals just for the sake of it. 

Norelli highlights an example where Eko started juggling and set a record of 50 consecutive catches. Eko then decided that the catches had happened too quickly to really count, so it canceled its record and started again with stricter rules. “This is playing,” says Norelli. “You give yourself a challenge and you try to achieve it, and then you make it harder.”

COURTESY MIT CSAIL

As the agents explored the island, they logged their activities and discoveries in text files that served as their long-term memory. Later, when researchers tested them on specific tasks, like trying to get as many objects as possible onto the island’s highest peak, the agents that had spent 30 hours playing on the island performed better than fresh copies without the same experience. They also found that “ablating” the agent’s memories—deleting lines from their memory files—degraded their performance. (Add One Flew Over the Cuckoo’s Nest to that list of movie references.)

Burghardt, the Tennessee psychologist, agrees that the work  meets some of the criteria for play that he identified in living creatures. But was Eko actually playing, or was the word-prediction engine simply role-playing what a playful agent would do in this scenario? “All of this could be role-playing,” admits Nathan Cloos, another MIT researcher and author on the paper. “But we show that it’s not just hallucinating the play. This play is also contributing to learning.”

Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, was impressed by the research but says to really classify the behavior as play, you’d need to see more evidence of cognition and motivation underpinning it. 

“It could be the case that the models are genuinely playing, but we don’t know based on the paper,” Yiu says. “If you only have the behavioral piece, it could be the case that they’re just trying to abide by the prompt, trying to imitate humans.”

Research into AI agents has been criticized for anthropomorphizing them, and this research setup did a lot to nudge the agents toward behaving like humans. Although Eko was not explicitly instructed to invent games or explore the island, it did have a master prompt in the form of a “soul” file, a text document that described it as “naturally curious” and eager to explore. 

Eko did also frequently just stand there doing nothing until it was kicked back into life with a “continuation prompt” that read, “Please don’t stop. Just continue doing whatever you want to do.” 

“It’s not like this is a system that just suddenly magically decided to play,” says Phillip Isola, an MIT professor and a coauthor of the paper. “It’s an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us,” he says. “It does things that surprise us and that are causally related to its future skillful behavior.”

As large language models make the step from software into hardware like humanoid robots, exploring spaces without a specific goal in mind could eventually help them flexibly adapt to unexpected situations. 

“There’s a lot of work that’s just on exploiting AI, making it better and faster,” says Yiu. “But in terms of building world models and exploration, I think that [they are] still a big part that deserves attention, especially if we want to generalize and prepare ourselves for new environments.”

“Perhaps some capabilities are best taught by tasks,” the paper concludes. “Others may need something closer to a childhood.” Ultimately, the question of whether AI agents are actually playing or just imitating play may not actually matter.