Flipkart has expanded access to its food-delivery service to thousands of employees ahead of its public launch.
Read more Mayor Zohran Mamdani said Monday night that New Yorkers must resist the narrative that an AI takeover is inevitable and demand regulations that protect the public and put power back in their hands.
Read more If the deal goes through, it would be the latest in a long string of giant compute deals Anthropic has struck over the past year.
Read more The idea of embedded evaluators — in which supposedly “neutral” third-party observers are placed inside AI organizations — has been popularized over the last several months, as AI labs grapple with a rash of safety incidents that have spurred concerns over the potential harms of their technology.
Read more Microsoft wants enterprises to closely watch the agents that employees use with systems not controlled by AI labs like OpenAI and Anthropic.
Read more 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 his colleagues were studying a small image-recognition model called Inception v1. 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. Looking at other models appeared to back 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.’”
Cammarata is not saying that he yet fully gets how LLMs (or any models) work. He is saying that he no longer thinks full understanding is impossible. “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.”
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
Inside the quest to combat Arctic fires and save the planet’s frozen carbon
—James Temple
In July 2021, a wildfire burned through a forest on the University of Alaska, Fairbanks campus, stripping away the moss and organic-rich soil insulating the permafrost below.
Five years later, I followed researchers into the forest to investigate how fighting fires could help keep that carbon locked in the ground.
A simple metal probe revealed how much deeper the ground had thawed. This rudimentary tool offered a glimpse of an escalating problem: As climate change fuels more frequent and severe fires, they’re accelerating permafrost thaw and releasing greenhouse gases.
Now, researchers are exploring how to break that cycle. But how much should we spend, and what risks are worth taking?
Read the full story on how scientists are rethinking the fight against wildfires.
AI researchers stranded Claude on a virtual desert island. It started to “play.”
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 experimentreminiscent of the movies Cast Away or Home Alone.
The agents started behaving in surprising ways. They 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.
A lot of what they were doing met the bar for play. But is it truly playful behavior, or simply role-playing?
Read our full story to learn more.
—Amit Katwala
10 Climate Tech Companies to Watch: X-energy and its helium-cooled nuclear reactors
X-energy is one of MIT Technology Review’s 10 Climate Tech Companies to Watch 2026, available exclusively to subscribers.
Making concrete, steel, and chemicals takes huge amounts of heat, much of it generated by burning fossil fuels. X-energy hopes to offer an alternative with a nuclear reactor that uses helium gas instead of water as a coolant, allowing it to operate at higher temperatures.
The proposed Xe-100 reactor could supply superheated steam directly to industrial facilities, generate 80 megawatts of electricity—or do both.
See how X-energy hopes to bring high-temperature nuclear power to heavy industry.
—Mark Ellis Harris
Subscribers can now access the full 10 Climate Tech Companies to Watch, spanning everything from cleaner cement to streamlined geothermal power.
Catch up on our conversation with a creator of AI-designed viruses
Can AI design new life forms? In 2025, Stanford University PhD student Samuel King came up with a preliminary answer when he used a generative AI model to propose genetic blueprints for microscopic viruses. It isn’t yet an example of AI-generated life, but that could be next.
He spoke about his work with our senior AI reporter James O’Donnell on Friday. If you missed it, you can watch their chat now, and subscribe to get invites to our Roundtables events in future.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 Tech stocks shed over $1 trillion on Friday 
It’s unclear if AI companies can turn sustainable profits, and investors are getting jittery. (NBC)
+ Musk and Anthropic are pledging big financial giveaways. (BBC)
+ Microsoft just pulled out of a giant AI cloud deal in the UK. (Gizmodo)
+ What’s at stake in AI’s trillion-dollar gamble. (MIT Technology Review)
2 Do we really want AI agents smoothing out all friction in our lives?
They promise convenience, but right now they’re mostly introducing chaos. (The Atlantic $)
+ Here’s what it’s like to outsource your life to them. (Vox $)
+ Websites are starting to close their doors to AI agents. (The Verge $)
+ AI agents bring new headaches for online sellers. (WSJ $)
3 How drones have transformed the front line in Ukraine
They’ve effectively created a “kill zone”, and it’s growing wider and deadlier. (NYT $)
+ Trump has denied ordering Starlink restrictions in Ukraine. (CNBC)
4 China is considering restricting overseas access to its AI models
It’s also looking at curbing the sharing of chip designs, among other measures. (Bloomberg $)
+ Why western open weight models still can’t compete with Chinese ones. (Rest of World)
+ Why Trumpworld still can’t agree on how to respond to Chinese AI. (MIT Technology Review)
5 HTC’s smart glasses are way more private than Meta’s
Guess that’s a benefit of not being shackled to a massive social media platform. (Wired $)
+ A bug caused Meta’s glasses to record conversations without the wearer knowing. (Ars Technica)
+ Smart glasses are already causing havoc in India. (MIT Technology Review)
6 Inside one man’s quest to build California’s biggest AI data center
Sebastian Rucci didn’t bank on meeting huge resistance from local residents. (The Guardian)
+ Americans are packing town halls to stop data centers, and it’s working. (Quartz $)
+ No wonder companies are exploring putting them in space. (MIT Technology Review)
7 Somebody vibe-coded a gun for the first time
And it seems like it works (sometimes.) (The Verge $)
8 AI might reshape how we grieve
It’s never been easier to create realistic AI chatbots of people who’ve passed. (Vox $)
+ Deepfakes of your dead loved ones are a booming business in China. (MIT Technology Review)
9 Japan is about to launch a mission to Mars’ largest moon
If all goes well, it should help us to understand how it formed. (New Scientist $)
10 Amateurs are cracking long-secret codes and ciphers 
One of the more fun uses for large language models! (Nature)
Quote of the day
“Google wouldn’t have taken my call in the first term. Now they kiss my ass.”
—President Trump brags about how much Big Tech bends the knee to him during a rally in Kentucky this weekend, the Verge reports.
One More Thing
SIMON MITCHELL
There are more AI health tools than ever—but how well do they work?
In the last few months alone, Microsoft, Amazon, and OpenAI have all launched medical chatbots.
There’s a clear demand for these tools, given how hard it is for many people to access advice through the existing medical system—and they could make safe, useful recommendations. But concerns have surfaced about how little external evaluation they undergo before being released to the public.
Read the full story to understand what’s at stake.
—Grace Huckins
We can still have nice things
A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line.)
+ Japan is showing what happens when technology and design are put in service of people.
+ What happens when you put marbles on a grand piano’s strings? Vinheteiro has the answer.
+ Poop from Southern right whales suggests they may be more resilient to climate change than we thought.
+ Perfectly Imperfect, the world’s first “social magazine,” offers recommendations for music, films, books, and more from people—not algorithms.
In the middle of an especially warm July in 2021, a wildfire ignited in a thick forest of black spruce and peeling paper birch trees on the northern campus of the University of Alaska, Fairbanks.
The blaze, possibly sparked by lightning, burned across the dry carpet of feather mosses and lichen before climbing into the crowns. By the time fire agencies responded, flames had reached 12 feet in length and sent up a 500-foot column of smoke.
A few days shy of the five-year anniversary of the Yankovich Road Fire, I followed three members of the Alaska Fire Science Consortium into the boreal forest, walking along cross-country ski trails lined by feathery field horsetail.
In 2024, the organization, which hosts presentations to keep wildfire managers abreast of new developments in fire science, helped turn the area into an educational walking trail, with interpretive signs posted along the route. About a half-mile in, the horizon opened up as we reached a sparse stand of naked, blackened tree trunks, stripped of needles and neighbors.
Hilary Shook, the coordinator of the consortium, stepped off the trail with her thaw probe, a T-shaped pointed metal rod used to detect the upper boundary of the permafrost, the ancient layer of frozen soil that underlies 15% of the land around the Northern Hemisphere and stores roughly twice as much carbon as the atmosphere.
A test of the ground just outside the burn scar hit the concrete-like layer about a foot below the surface. But as Shook applied pressure here, where the fire had burned away the moss and organic-rich topsoil that insulates the permafrost, the probe plunged at least three times deeper into the ground.
“That was an immediate effect of the fire,” said Alison York, the outgoing coordinator of the organization, who was standing nearby. “When we monitored right after, it had already thawed quite a bit.”
LEIF VAN CISE
This rudimentary tool, offering a measure of depth from a single point in the earth, reveals an escalating problem across the far north. As climate change fuels increasingly frequent and severe fires in the planet’s northern latitudes, they’re accelerating the thawing of more and more permafrost. This allows the organic matter to break down and release powerful greenhouse gases that contribute to a spiraling climate feedback effect, propelling more warming, more fires, and the loss of still more permafrost.
Most dramatically, intensifying wildfires in Canada released more carbon than the nation’s combined fossil fuel emissions in 2023. As in Alaska, many of these megafires occur in vast remote areas where the policy is to let them burn so long as they don’t threaten human habitats. Yet if intensifying fires and warming are left unchecked, the world’s boreal forests and permafrost could increasingly transition from giant carbon sinks into massive sources of atmospheric carbon.
Now, international teams of researchers across the US, Canada, and elsewhere are scrambling to understand what’s at stake, financially as well as ecologically. They’re looking at preventive measures, like controlled burns, and studying the potential to rapidly detect and extinguish wildfires before they grow out of control. They’re even exploring what was once unthinkable: sending in human firefighters not to save lives or structures, but to protect one of our planet’s great stores of carbon.
“We need to figure out a better way to live with wildfire,” says Brendan Rogers, a senior scientist at the Woodwell Climate Research Center, one of the main organizations driving the loosely collaborative research effort. “I do think that it can make a meaningful difference.”
PHILIP KEITH (PHOTOS), WOODWELL CLIMATE RESEARCH (MAPS)
Woodwell, the University of Victoria’s Polis Institute for Ecological Governance, and the nonprofit Alaska Venture Fund are now partnering on a multimillion-dollar initiative known as North Fire to study and raise public awareness about the potential benefits of these interventions, MIT Technology Review can report.
The idea is complicated and controversial for a number of reasons. Wildfire is a natural and necessary component of healthy forest ecosystems. Preventing or suppressing fires across huge stretches of largely unmanaged Arctic-boreal forests and tundra would take vast resources. And it’s not yet clear how much of a difference such interventions would make in emissions.
Ultimately, the mere fact that researchers are considering such extreme interventions into natural systems is a grim reflection of the dangers we’ve placed ourselves in by continuing to generate more and more greenhouse-gas emissions each year. Those threats are forcing us to grapple with increasingly difficult questions that involve ever steeper trade-offs: What’s the best way to spend our finite resources and dwindling time to avoid the worst dangers of climate change? And when, if ever, is it okay to put people in possible danger to reduce broader risks to the world?
‘A massive blind spot’
Emissions from forest fires have risen 60% around the world since 2001—but they’ve nearly tripled across the boreal forests in the far north.
This stretch of the globe is warming as much as four times faster than the rest of the planet, drying out the landscape. In addition, lightning strikes that set off fires have ticked up sharply as warmer air increases the convective energy in the atmosphere, driving the updrafts that trigger thunderstorms.
In that sense, humans have already interceded in these otherwise natural ecosystems by producing the greenhouse-gas emissions that are increasing temperatures, subjecting forests to fire patterns well outside healthy historic norms.
“To me, that outweighs the existential dread of ‘Should we or should we not intervene?” says Kate Hayes, an assistant professor at Western Washington University, who is working with Woodwell to study the emissions benefits of fire interventions. “We’ve already intervened, arguably. And so from here, it’s a little bit of a Band-Aid on a bullet wound.”
The permafrost that encircles the northern latitudes, underlying large stretches of Canada, Greenland, Siberia, Scandinavia, and other Arctic nations, stores around 1.5 trillion tons of carbon, much of which froze in place millennia ago. The steady global rise in temperature, however, has already shrunk the footprint of Arctic permafrost by about 12 % since 1850, with that loss rapidly accelerating during the last 50 years.
By 2100, fire-driven thawing of permafrost could release about the equivalent of nearly 80 billion tons of carbon dioxide, under high warming scenarios. That’s more than half the emissions that fires will release from the direct combustion of vegetation, according to a January study. And it would be on top of more than 400 billion additional tons that could be released through thawing from general warming over that period.
Despite its growing contributions to climate change, only a minority of the models that have informed the UN’s major climate reports to date include the gradual loss of permafrost through warming temperatures, and none incorporate the abrupt collapse and accelerated thaw that can often occur after fires and heat waves. In light of those phenomena and other powerful and undercounted climate feedback effects, we’re almost certainly underestimating how rapidly the planet will warm up, as scientists warned in a paper earlier this year.
These warming-driven emissions are a “massive blind spot,” says Brian Buma, a senior climate scientist at the Environmental Defense Fund and a coauthor of the paper, who is working to develop tools that could help fire agencies decide which fires to prioritize.
As escalating fire and permafrost emissions become significant additional contributors to climate change, some researchers have begun to explore what it might cost to take on the responsibility of preventing them, in much the way we’ve sought to cut the pollution from power plants, vehicles, and factories.
A 2022 paper in Science Advances, led by researchers at Woodwell and the Union of Concerned Scientists, found that making greater efforts to put out boreal fires could be a cost-effective climate solution—comparable to, or even cheaper than, addressing emissions by building wind turbines, solar farms, or carbon-sucking factories.
The Yukon Flats pilot
Those economic findings were instrumental in getting more researchers to take the idea of tackling fire emissions seriously, says Peter Frumhoff, a coauthor of the 2022 paper and now a senior science policy advisor at Woodwell.
For several years now, Woodwell researchers have been in discussions with forest managers, fire agencies, Indigenous communities, and other groups across Alaska to communicate the potential of this approach.
It hasn’t been an easy sell, given the constrained resources and tightly defined mission of the federal and state agencies responsible for combating fires in Alaska. The general policy there is to allow fires in remote territories to burn unless they present a direct danger to communities.
“Those agencies are tasked with protecting life and property now, and to a lesser extent natural resources,” says Randi Jandt, a fire ecologist at the Alaska Fire Science Consortium. “But they’re not tasked to save the world.”
There was one notable early success, however.
The Yukon Flats National Wildlife Refuge is an 11-million-acre patchwork of wetlands, forests, and bogs. The foothills sit above a particular type of frozen soil known as yedoma. It’s an especially deep, ice-rich, and carbon-dense form of permafrost, composed of massive wedges of silty, wind-blown sediments frozen during the late Pleistocene, which ended about 11,700 years ago. (It also presented ideal conditions for preserving mummified remains of woolly mammoths.)
“It took kind of the last ice age to accumulate,” Hayes says. “So there’s no rapid way to re-accumulate it. We can’t really put it back.”
LEIF VAN CISE
LEIF VAN CISE
LEIF VAN CISE
Members of the Alaska Fire Science Consortium hike into areas affected by fire to measure the layers of ash left in the soil on the forest floor.
In 2019, the region suffered a series of devastating fires that burned some 750,000 acres. After an Alaska Fire Science Consortium meeting at which Woodwell scientists highlighted the possible climate benefits of more aggressive fire suppression, the refuge’s then manager, Jimmy Fox, began exploring the possibility of putting out blazes when they threatened that permafrost, whether or not they threatened human habitations.
The early reaction was not encouraging.
“You’re going to have people jumping out of airplanes with parachutes to save frozen woolly mammoths—and maybe dying from that?” he recalls being asked.
But eventually, he persuaded managers at the US Bureau of Land Management’s Alaska Fire Service to carry out a pilot project on a 1.6-million-acre section of the Yukon Flats. They agreed to try to put out fires there if they had the resources available and believed they could contain a blaze within 72 hours.
Woodwell and its partners agreed to study how that change affected permafrost loss and fire emissions, in the hope of making the case for similar programs elsewhere.
It wasn’t clear the pilot program would survive after Donald Trump was reelected, particularly once the DOGE initiative retroactively cut the public funding for that research collaboration.
But in February, the federal government formally preserved the program and broadened its rationale to include goals like protecting habitat diversity, water quality, and air quality.
Mapping and modeling
One of Woodwell’s main tasks now is determining when and where fire agencies should step in. Crucially, excessive efforts to stamp out fires in the past have allowed the buildup of dry plant material that fuels more severe fires later.
“We’re not going after all fires all the time,” says Rogers over lunch in a sunny courtyard on the Stanford University campus, where Woodwell and EDF hosted a workshop on potential fire interventions in January. “We can’t do that. We don’t want to do that. We shouldn’t do that. So how do we prioritize our resources?”
Woodwell and its collaborators are working to answer that question by developing a map of burn risk and carbon vulnerability across interior Alaska. The plan is to eventually expand it across much of Canada as well.
These maps will highlight the locations of rich reserves of yedoma and other forms of permafrost, forecast where fires are most likely to break out in the next 10 to 15 years, and predict how they’re likely to behave and spread.
“Our ultimate goal is to get a layer that can tell us, if a fire were to occur, what areas are most vulnerable to high carbon emissions, permafrost thaw, permafrost degradation,” Rogers says.
In response to a question about potentially putting firefighters at risk, Rogers stressed in an email that it’s “a very serious issue that needs careful consideration.”
He added that early fire detection technologies, preventative fire stewardship practices, improved fire suppression planning and other measures can all help to reduce dangers to firefighters as well as communities.
‘Mild fire’
Once areas of particular risk are identified, the first option is to take measures to prevent fires there, or at least to reduce the odds that they’ll become major ones. That generally entails using heavy equipment or intentional, controlled fires to clear away vegetation that could fuel new fires, or to create fire breaks that might check their spread.
COURTESY OF EDWARD ALEXANDER
Indigenous tribes carried out what’s known as cultural burning for thousands of years to reduce fire dangers, clear land for crops, and more. The practice was widely outlawed across North America beginning in the 1800s, though those rules have loosened in recent years.
COURTESY OF EDWARD ALEXANDER
Both cultural and prescribed fires, however, remain controversial because they can burn out of control. And even if they remain contained, the fires can still produce smoke that wafts into towns.
But rushing around to put out blazes after they’ve started will never get Alaska ahead of its surging wildfire problem, says Edward Alexander, cochair of the Gwich’in Council International, which represents the Indigenous Gwich’in peoples of the Northwest Territories, the Yukon, and Alaska.
“There’s not enough money in the world for this to be a chainsaw-and-ax problem,” he says. “This country is too big up here.”
Alexander, who also serves as senior Arctic lead at Woodwell, argues that Alaska should encourage much greater use of cultural burning to reduce the risk of conflagrations and the emissions that come with them.
“We need mild fire, not wildfire,” he says.
New pilot projects
The North Fire initiative will include several additional pilot research projects, including an effort to study the potential for cultural burning to reduce climate and fire dangers near Fort Yukon, Alaska.
The plan is for a wildland firefighting crew operated by the Council of Athabascan Tribal Governments, a consortium in the Yukon Flats region, to carry out cultural burns there next spring, Rogers says.
Researchers at Woodwell and the University of Alaska, Fairbanks, will then collect and analyze data flammability, carbon levels, and permafrost over several years, in the hopes of learning more about what such interventions do.
Similarly, the Polis researchers at the University of Victoria plan to study how adding fuel breaks, by logging or switching to less flammable trees, alters fire risks in British Columbia.
The broad aim of the North Fire initiative is to accelerate research and expand community involvement “in a way that helps us scale this as a meaningful global natural climate solution on a rapid timeline,” says Erin Harrington, a partner at the Alaska Venture Fund, a philanthropy focused on environmental, economic development, and Indigenous issues.
Detecting fires from space
Other observers, including Jandt, believe the bigger opportunity to reduce fire emissions and preserve permafrost may lie in the “initial attack.” That means marshalling planes, tankers, and firefighters to stamp fires out at the first sign of them, before they’ve had the chance to turn into conflagrations.
Once fires reach a certain size, it takes much more time to contain them—and the job becomes far more dangerous.
Accelerating response times is one of the chief promises of the FireSat satellite constellation, which is expected to begin delivering fire data at least twice daily from around the world to Woodwell, a handful of fire agencies, and other early adopters by the end of this year. The system will eventually include more than 50 satellites circling the globe, equipped with high-resolution cameras and thermal sensors that can detect smaller, lower-temperature fires than existing satellite systems, down to as little as five by five meters (16 by 16 feet).
Currently, it can take days to spot fires. But once the full constellation is in place, likely in the early 2030s, FireSat should be capable of detecting and providing updated images of fires anywhere in the world every 20 minutes or less.
The first prototype satellite launched in 2025, and three operational ones went into orbit in July. The Earth Fire Alliance, the nonprofit that manages the constellation, expects to launch four more by the end of next year.
A number of similar systems are also coming online. German-based OroraTech has launched 18 of its fire-detecting satellites in the last year and a half. Similarly, Canadian agencies are planning to launch their own satellite constellation, dubbed WildFireSat, beginning in 2029.
But as these systems produce more and more detections across the globe, it’ll become increasingly essential to build tools that can filter through that information.
EDF is supporting and collaborating on several projects that could potentially help with that, and it ultimately intends to develop an AI-powered triage tool that could help fire agencies prioritize responses as FireSat and other fire-spotting tools come online.
“All of this is kind of like building the data ecosystem around FireSat, so that when they start popping out “fire start here, there, fire start here, fire start there,” we won’t be like, ‘Oh shit, there’s fires everywhere,’” Buma says. “We’ll have a framework. It won’t be perfect, obviously. We’ll have to improve it, but we’ll have the machinery to incorporate those data.”
Rapid response
But faster detections and decisions do only so much. Regions would still need to get firefighting personnel and resources to those fires fast enough to prevent them from burning out of control.
“Say we have the full FireSat constellation up and running, and it spots a fire that just broke out in one of these sort of high-priority, highly vulnerable areas,” I said to Rogers and Frumhoff during our conversation in that Stanford courtyard. “How might we deploy firefighting resources to a remote area fast enough for it to matter in that scenario?”
“I think the answer, unfortunately, is that it depends,” Rogers responded.
Ideally, he said, we would have fire management policies in place that allow for putting out fires with a high risk of producing climate-warming emissions. In addition, regions could install more firefighting personnel and resources across more operational centers, strategically placed in locations that allow them to get to those blazes fast.
Frumhoff stresses that Woodwell isn’t promoting any particular companies but says there are additional technologies “in the mix of conversations” that might help to prevent or rapidly put out fires as well.
Those possible approaches include seeding clouds to suppress lightning strikes and pre-positioning autonomous firefighting helicopters to respond rapidly to high-emissions-risk fires.
Sikorsky, a Lockheed Martin company, is developing and testing autonomous versions of the Black Hawk helicopter, equipped with thermal sensors and water tanks. It is collaborating with Rain Industries, a startup based in Alameda, California, that has developed software to track shifting conditions and steer unpiloted aircraft through the process of putting out wildfires.
TED CARLSON/RAIN AERO
The company says it has talked to fire agencies and the “broader Alaska fire community” about the potential use of its technology but declined to name any specific entities it has interacted with.
“The twin climate tipping points of the yedoma permafrost and the boreal forest, and the fact that it’s completely underresourced at the moment, from my perspective does make it an ideal use case for autonomous aircraft,” says Max Brodie, chief executive officer of Rain.
Defending carbon
Advancing technologies can only do so much if governments’ fire prevention and suppression policies don’t prioritize preserving carbon.
And that will remain a hard sell given the costs of managing extensive areas of land that nations have historically left largely alone.
The 2022 Science Advances paper estimated that additional fire suppression efforts would cost about $12.63 per metric ton of prevented carbon dioxide emissions—a number that drops when the carbon in permafrost is taken into account, according to upcoming research.
But that would still add up to about $7 billion in additional spending this decade to keep Alaska’s fire emissions at their historical levels, and another $7 billion to $50 billion over the subsequent 30 years as fire severity continues to rise.
Frumoff stresses that those figures just reflect what the researchers found it would cost to preserve the status quo, which is not necessarily what the policy goal would be.
But widening the scope of where we put out fires will inevitably create a big added cost for fire agencies that are already notoriously underfunded in the US. While Alaska produces about half the nation’s wildfire emissions, it receives around 4% of federal firefighting funds.
NATIONAL PARK SERVICE VIA AP
The US Wildland Fire Service provided a prepared statement in response to an MIT Technology Review inquiry asking whether preserving carbon could or should become a priority in certain fire management decisions. The roundabout reply suggests that this isn’t the agency’s prime concern: “Firefighter and public safety are the top priorities for the U.S. Wildland Fire Service.”
Frumhoff says it’s possible that landowners or government agencies could eventually use carbon markets to defray some of the additional costs. Woodwell has also had multiple conversations with Alaska’s federal legislators about the value of preventing catastrophic wildfires and the need to increase the “total pot” of public funding, he says.
Kicking the fire can
A few days after walking through the fire scar in Fairbanks, I drove slowly along a crumpled stretch of Stampede Road, just north of Denali National Park, where the underlying permafrost had melted without the aid of fire.
I arrived at a cramped wooden cabin to meet with Michelle Mack, a regents professor of ecosystem ecology at Northern Arizona University.
Mack studies how fire is altering ecosystems in the high northern latitudes as the region grows warmer and drier, and she sometimes collaborates with Rogers and other scientists working on the push to protect permafrost in other areas of research. But she’s unsure how much these interventions will accomplish as climate change continues.
“I’m an empiricist, and at this point, it’s not clear to me that the empirical evidence suggests that we can actually have any impact on these fires,” she says.
MATT ‘MARTIAN’ WILLIAMS
She expects there will be clear cases where we could and should act fast to put out fires and “keep carbon in the ground.” She points to the example of the 2007 Anaktuvuk River Fire, a massive tundra fire on Alaska’s North Slope that might have been extinguished with relative ease in its early days but was instead allowed to burn until the winter snows put it out nearly three months later. Over that time, it ripped across 400 square miles and, Mack’s research showed, released more than 2 million tons of carbon into the air.
“I feel like that’s the low-hanging fruit, something that doesn’t cost much—and that’s where the technology for detection I think is really important,” she says.
But she adds that fires in boreal forests can rapidly spread in unpredictable ways, frustrating efforts to quickly extinguish them even when they are spotted right away.
In addition, rapidly putting out a fire one season doesn’t necessarily reduce the fire risk for a given forest, so suppression may simply kick the “fire can” down the road—delaying but not preventing an inevitable loss of carbon, Mack says.
Her thinking more recently shifted “a little” after she came across an earlier modeling paper that showed some encouraging findings on the potential of early suppression to reduce the footprint of boreal fires. But she says that prevention through forest thinning, prescribed burns, and the introduction of less flammable deciduous trees to create “living fire breaks” is probably a more promising path than technology-augmented suppression.
“If we could solve this with a techno fix, that’d be great,” she says. “But what if we solved it with nature?”
‘A real difference’
Rogers, in an emailed response to Mack’s concern, said that the interventions they’re exploring, which do include both prevention and suppression, “can make a real difference.”
“I certainly agree that some fires are much easier to suppress via initial attack than others,” he wrote. “That said, until I see data showing me otherwise, I believe there is strong potential to bend the curve with management.”
EDF’s Buma acknowledges there are a number of unknowns, including how quickly agencies could act on faster detections from satellites—and how big the political obstacles will be to making necessary changes to fire management policies.
“We can’t overpromise what’s possible. We don’t know what’s possible,” he says. “It’s all new. But we’re in a new climate and have to consider trying new things.”
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.