It’s no secret that the current AI boom is using up immense amounts of energy. Now we have a better idea of how much.
A new paper, from a team at the Harvard T.H. Chan School of Public Health, examined 2,132 data centers operating in the United States (78% of all facilities in the country). These facilities—essentially buildings filled to the brim with rows of servers—are where AI models get trained, and they also get “pinged” every time we send a request through models like ChatGPT. They require huge amounts of energy both to power the servers and to keep them cool.
Since 2018, carbon emissions from data centers in the US have tripled. For the 12 months ending August 2024, data centers were responsible for 105 million metric tons of CO2, accounting for 2.18% of national emissions (for comparison, domestic commercial airlines are responsible for about 131 million metric tons). About 4.59% of all the energy used in the US goes toward data centers, a figure that’s doubled since 2018.
It’s difficult to put a number on how much AI in particular, which has been booming since ChatGPT launched in November 2022, is responsible for this surge. That’s because data centers process lots of different types of data—in addition to training or pinging AI models, they do everything from hosting websites to storing your photos in the cloud. However, the researchers say, AI’s share is certainly growing rapidly as nearly every segment of the economy attempts to adopt the technology.
“It’s a pretty big surge,” says Eric Gimon, a senior fellow at the think tank Energy Innovation, who was not involved in the research. “There’s a lot of breathless analysis about how quickly this exponential growth could go. But it’s still early days for the business in terms of figuring out efficiencies, or different kinds of chips.”
Notably, the sources for all this power are particularly “dirty.” Since so many data centers are located in coal-producing regions, like Virginia, the “carbon intensity” of the energy they use is 48% higher than the national average. The paper, which was published on arXiv and has not yet been peer-reviewed, found that 95% of data centers in the US are built in places with sources of electricity that are dirtier than the national average.
There are causes other than simply being located in coal country, says Falco Bargagli-Stoffi, an author of the paper. “Dirtier energy is available throughout the entire day,” he says, and plenty of data centers require that to maintain peak operation 24-7. “Renewable energy, like wind or solar, might not be as available.” Political or tax incentives, and local pushback, can also affect where data centers get built.
One key shift in AI right now means that the field’s emissions are soon likely to skyrocket. AI models are rapidly moving from fairly simple text generators like ChatGPT toward highly complex image, video, and music generators. Until now, many of these “multimodal” models have been stuck in the research phase, but that’s changing.
OpenAI released its video generation model Sora to the public on December 9, and its website has been so flooded with traffic from people eager to test it out that it is still not functioning properly. Competing models, like Veo from Google and Movie Gen from Meta, have still not been released publicly, but if those companies follow OpenAI’s lead as they have in the past, they might be soon. Music generation models from Suno and Udio are growing (despite lawsuits), and Nvidia released its own audio generator last month. Google is working on its Astra project, which will be a video-AI companion that can converse with you about your surroundings in real time.
“As we scale up to images and video, the data sizes increase exponentially,” says Gianluca Guidi, a PhD student in artificial intelligence at University of Pisa and IMT Lucca, who is the paper’s lead author. Combine that with wider adoption, he says, and emissions will soon jump.
One of the goals of the researchers was to build a more reliable way to get snapshots of just how much energy data centers are using. That’s been a more complicated task than you might expect, given that the data is dispersed across a number of sources and agencies. They’ve now built a portal that shows data center emissions across the country. The long-term goal of the data pipeline is to inform future regulatory efforts to curb emissions from data centers, which are predicted to grow enormously in the coming years.
“There’s going to be increased pressure, between the environmental and sustainability-conscious community and Big Tech,” says Francesca Dominici, director of the Harvard Data Science Initiative and another coauthor. “But my prediction is that there is not going to be regulation. Not in the next four years.”
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
The AI sector is plagued by a lack of competition and a lot of deceit—or at least that’s one way to interpret the latest flurry of actions taken in Washington.
Last Thursday, Senators Elizabeth Warren and Eric Schmitt introduced a bill aimed at stirring up more competition for Pentagon contracts awarded in AI and cloud computing. Amazon, Microsoft, Google, and Oracle currently dominate those contracts. “The way that the big get bigger in AI is by sucking up everyone else’s data and using it to train and expand their own systems,” Warren told the Washington Post.
The new bill would “require a competitive award process” for contracts, which would ban the use of “no-bid” awards by the Pentagon to companies for cloud services or AI foundation models. (The lawmakers’ move came a day after OpenAI announced that its technology would be deployed on the battlefield for the first time in a partnership with Anduril, completing a year-long reversal of its policy against working with the military.)
While Big Tech is hit with antitrust investigations—including the ongoing lawsuit against Google about its dominance in search, as well as a new investigation opened into Microsoft—regulators are also accusing AI companies of, well, just straight-up lying.
On Tuesday, the Federal Trade Commission took action against the smart-camera company IntelliVision, saying that the company makes false claims about its facial recognition technology. IntelliVision has promoted its AI models, which are used in both home and commercial security camera systems, as operating without gender or racial bias and being trained on millions of images, two claims the FTC says are false. (The company couldn’t support the bias claim and the system was trained on only 100,000 images, the FTC says.)
A week earlier, the FTC made similar claims of deceit against the security giant Evolv, which sells AI-powered security scanning products to stadiums, K-12 schools, and hospitals. Evolv advertises its systems as offering better protection than simple metal detectors, saying they use AI to accurately screen for guns, knives, and other threats while ignoring harmless items. The FTC alleges that Evolv has inflated its accuracy claims, and that its systems failed in consequential cases, such as a 2022 incident when they failed to detect a seven-inch knife that was ultimately used to stab a student.
Those add to the complaints the FTC made back in September against a number of AI companies, including one that sold a tool to generate fake product reviews and one selling “AI lawyer” services.
The actions are somewhat tame. IntelliVision and Evolv have not actually been served fines. The FTC has simply prohibited the companies from making claims that they can’t back up with evidence, and in the case of Evolv, it requires the company to allow certain customers to get out of contracts if they wish to.
However, they do represent an effort to hold the AI industry’s hype to account in the final months before the FTC’s chair, Lina Khan, is likely to be replaced when Donald Trump takes office. Trump has not named a pick for FTC chair, but he said on Thursday that Gail Slater, a tech policy advisor and a former aide to vice president–elect JD Vance, was picked to head the Department of Justice’s Antitrust Division. Trump has signaled that the agency under Slater will keep tech behemoths like Google, Amazon, and Microsoft in the crosshairs.
“Big Tech has run wild for years, stifling competition in our most innovative sector and, as we all know, using its market power to crack down on the rights of so many Americans, as well as those of Little Tech!” Trump said in his announcement of the pick. “I was proud to fight these abuses in my First Term, and our Department of Justice’s antitrust team will continue that work under Gail’s leadership.”
That said, at least some of Trump’s frustrations with Big Tech are different—like his concerns that conservatives could be targets of censorship and bias. And that could send antitrust efforts in a distinctly new direction on his watch.
Now read the rest of The Algorithm
Deeper Learning
The US Department of Defense is investing in deepfake detection
The Pentagon’s Defense Innovation Unit, a tech accelerator within the military, has awarded its first contract for deepfake detection. Hive AI will receive $2.4 million over two years to help detect AI-generated video, image, and audio content.
Why it matters: As hyperrealistic deepfakes get cheaper and easier to produce, they hurt our ability to tell what’s real. The military’s investment in deepfake detection shows that the problem has national security implications as well. The open question is how accurate these detection tools are, and whether they can keep up with the unrelenting pace at which deepfake generation techniques are improving. Read more from Melissa Heikkilä.
Bits and Bytes
The owner of the LA Times plans to add an AI-powered “bias meter” to its news stories
Patrick Soon-Shiong is building a tool that will allow readers to “press a button and get both sides” of a story. But trying to create an AI model that can somehow provide an objective view of news events is controversial, given that models are biased both by their training data and by fine-tuning methods. (Yahoo)
Google DeepMind’s new AI model is the best yet at weather forecasting
It’s the second AI weather model that Google has launched in just the past few months. But this one’s different: It leaves out traditional physics models and relies on AI methods alone. (MIT Technology Review)
How the Ukraine-Russia war is reshaping the tech sector in Eastern Europe
Startups in Latvia and other nearby countries see the mobilization of Ukraine as a warning and an inspiration. They are now changing consumer products—from scooters to recreational drones—for use on the battlefield. (MIT Technology Review)
How Nvidia’s Jensen Huang is avoiding $8 billion in taxes
Jensen Huang runs Nvidia, the world’s top chipmaker and most valuable company. His wealth has soared during the AI boom, and he has taken advantage of a number of tax dodges “that will enable him to pass on much of his fortune tax free,” according to the New York Times. (The New York Times)
Meta is pursuing nuclear energy for its AI ambitions Meta wants more of its AI training and development to be powered by nuclear energy, joining the ranks of Amazon and Microsoft. The news comes as many companies in Big Tech struggle to meet their sustainability goals amid the soaring energy demands from AI. (Meta)
Correction: A previous version of this article stated that Gail Slater was picked by Donald Trump to be the head of the FTC. Slater was in fact picked to lead the Department of Justice’s Antitrust Division. We apologize for the error.
One afternoon in late November, I visited a weapons test site in the foothills east of San Clemente, California, operated by Anduril, a maker of AI-powered drones and missiles that recently announced a partnership with OpenAI. I went there to witness a new system it’s expanding today, which allows external parties to tap into its software and share data in order to speed up decision-making on the battlefield. If it works as planned over the course of a new three-year contract with the Pentagon, it could embed AI more deeply into the theater of war than ever before.
Near the site’s command center, which looked out over desert scrubs and sage, sat pieces of Anduril’s hardware suite that have helped the company earn its $14 billion valuation. There was Sentry, a security tower of cameras and sensors currently deployed at both US military bases and the US-Mexico border, and advanced radars. Multiple drones, including an eerily quiet model called Ghost, sat ready to be deployed. What I was there to watch, though, was a different kind of weapon, displayed on two large television screens positioned at the test site’s command station.
I was here to examine the pitch being made by Anduril, other companies in defense tech, and growing numbers of people within the Pentagon itself: A future “great power” conflict—military jargon for a global war involving competition between multiple countries—will not be won by the entity with the most advanced drones or firepower, or even the cheapest firepower. It will be won by whoever can sort through and share information the fastest. And that will have to be done “at the edge” where threats arise, not necessarily at a command post in Washington.
A desert drone test
“You’re going to need to really empower lower levels to make decisions, to understand what’s going on, and to fight,” Anduril CEO Brian Schimpf says. “That is a different paradigm than today.” Currently, information flows poorly among people on the battlefield and decision-makers higher up the chain.
To show how the new tech will fix that, Anduril walked me through an exercise demonstrating how its system would take down an incoming drone threatening a base of the US military or its allies (the scenario at the center of Anduril’s new partnership with OpenAI). It began with a truck in the distance, driving toward the base. The AI-powered Sentry tower automatically recognized the object as a possible threat, highlighting it as a dot on one of the screens. Anduril’s software, called Lattice, sent a notification asking the human operator if he would like to send a Ghost drone to monitor. After a click of his mouse, the drone piloted itself autonomously toward the truck, as information on its location gathered by the Sentry was sent to the drone by the software.
The truck disappeared behind some hills, so the Sentry tower camera that was initially trained on it lost contact. But the surveillance drone had already identified it, so its location stayed visible on the screen. We watched as someone in the truck got out and launched a drone, which Lattice again labeled as a threat. It asked the operator if he’d like to send a second attack drone, which then piloted autonomously and locked onto the threatening drone. With one click, it could be instructed to fly into it fast enough to take it down. (We stopped short here, since Anduril isn’t allowed to actually take down drones at this test site.) The entire operation could have been managed by one person with a mouse and computer.
Anduril is building on these capabilities further by expanding Lattice Mesh, a software suite that allows other companies to tap into Anduril’s software and share data, the company announced today. More than 10 companies are now building their hardware into the system—everything from autonomous submarines to self-driving trucks—and Anduril has released a software development kit to help them do so. Military personnel operating hardware can then “publish” their own data to the network and “subscribe” to receive data feeds from other sensors in a secure environment. On December 3, the Pentagon’s Chief Digital and AI Office awarded a three-year contract to Anduril for Mesh.
Anduril’s offering will also join forces with Maven, a program operated by the defense data giant Palantir that fuses information from different sources, like satellites and geolocation data. It’s the project that led Google employees in 2018 to protest against working in warfare. Anduril and Palantir announced on December 6 that the military will be able to use the Maven and Lattice systems together.
The military’s AI ambitions
The aim is to make Anduril’s software indispensable to decision-makers. It also represents a massive expansion of how the military is currently using AI. You might think the US Department of Defense, advanced as it is, would already have this level of hardware connectivity. We have some semblance of it in our daily lives, where phones, smart TVs, laptops, and other devices can talk to each other and share information. But for the most part, the Pentagon is behind.
“There’s so much information in this battle space, particularly with the growth of drones, cameras, and other types of remote sensors, where folks are just sopping up tons of information,” says Zak Kallenborn, a warfare analyst who works with the Center for Strategic and International Studies. Sorting through to find the most important information is a challenge. “There might be something in there, but there’s so much of it that we can’t just set a human down and to deal with it,” he says.
Right now, humans also have to translate between systems made by different manufacturers. One soldier might have to manually rotate a camera to look around a base and see if there’s a drone threat, and then manually send information about that drone to another soldier operating the weapon to take it down. Those instructions might be shared via a low-tech messenger app—one on par with AOL Instant Messenger. That takes time. It’s a problem the Pentagon is attempting to solve through its Joint All-Domain Command and Control plan, among other initiatives.
“For a long time, we’ve known that our military systems don’t interoperate,” says Chris Brose, former staff director of the Senate Armed Services Committee and principal advisor to Senator John McCain, who now works as Anduril’s chief strategy officer. Much of his work has been convincing Congress and the Pentagon that a software problem is just as worthy of a slice of the defense budget as jets and aircraft carriers. (Anduril spent nearly $1.6 million on lobbying last year, according to data from Open Secrets, and has numerous ties with the incoming Trump administration: Anduril founder Palmer Luckey has been a longtime donor and supporter of Trump, and JD Vance spearheaded an investment in Anduril in 2017 when he worked at venture capital firm Revolution.)
Defense hardware also suffers from a connectivity problem. Tom Keane, a senior vice president in Anduril’s connected warfare division, walked me through a simple example from the civilian world. If you receive a text message while your phone is off, you’ll see the message when you turn the phone back on. It’s preserved. “But this functionality, which we don’t even think about,” Keane says, “doesn’t really exist” in the design of many defense hardware systems. Data and communications can be easily lost in challenging military networks. Anduril says its system instead stores data locally.
An AI data treasure trove
The push to build more AI-connected hardware systems in the military could spark one of the largest data collection projects the Pentagon has ever undertaken, and companies like Anduril and Palantir have big plans.
“Exabytes of defense data, indispensable for AI training and inferencing, are currently evaporating,” Anduril said on December 6, when it announced it would be working with Palantir to compile data collected in Lattice, including highly sensitive classified information, to train AI models. Training on a broader collection of data collected by all these sensors will also hugely boost the model-building efforts that Anduril is now doing in a partnership with OpenAI, announced on December 4. Earlier this year, Palantir also offered its AI tools to help the Pentagon reimagine how it categorizes and manages classified data. When Anduril founder Palmer Luckey told me in an interview in October that “it’s not like there’s some wealth of information on classified topics and understanding of weapons systems” to train AI models on, he may have been foreshadowing what Anduril is now building.
Even if some of this data from the military is already being collected, AI will suddenly make it much more useful. “What is new is that the Defense Department now has the capability to use the data in new ways,” Emelia Probasco, a senior fellow at the Center for Security and Emerging Technology at Georgetown University, wrote in an email. “More data and ability to process it could support great accuracy and precision as well as faster information processing.”
The sum of these developments might be that AI models are brought more directly into military decision-making. That idea has brought scrutiny, as when Israel was found last year to have been using advanced AI models to process intelligence data and generate lists of targets. Human Rights Watch wrote in a report that the tools “rely on faulty data and inexact approximations.”
“I think we are already on a path to integrating AI, including generative AI, into the realm of decision-making,” says Probasco, who authored a recent analysis of one such case. She examined a system built within the military in 2023 called Maven Smart System, which allows users to “access sensor data from diverse sources [and] apply computer vision algorithms to help soldiers identify and choose military targets.”
Probasco said that building an AI system to control an entire decision pipeline, possibly without human intervention, “isn’t happening” and that “there are explicit US policies that would prevent it.”
A spokesperson for Anduril said that the purpose of Mesh is not to make decisions. “The Mesh itself is not prescribing actions or making recommendations for battlefield decisions,” the spokesperson said. “Instead, the Mesh is surfacing time-sensitive information”—information that operators will consider as they make those decisions.
Like many others, I recently fled the social media platform X for Bluesky. In the process, I started following many of the people I followed on X. On Thanksgiving, I was delighted to see a private message from a fellow AI reporter, Will Knight from Wired. Or at least that’s who I thought I was talking to. I became suspicious when the person claiming to be Knight mentioned being from Miami, when Knight is, in fact, from the UK. The account handle was almost identical to the real Will Knight’s handle, and the profile used his profile photo.
Then more messages started to appear. Paris Marx, a prominent tech critic, slid into my DMs to ask me how I was doing. “Things are going splendid over here,” he replied to me. Then things got suspicious again. “How are your trades going?” fake-Marx asked me. This account was far more sophisticated than Knight’s; it had meticulously copied every single tweet and retweet from Marx’s real page over the past few weeks.
Both accounts were eventually deleted, but not before trying to get me to set up a crypto wallet and a “cloud mining pool” account. Knight and Marx confirmed to us that these accounts did not belong to them, and that they have been fighting impersonator accounts of themselves for weeks.
They are not the only ones. The New York Times tech journalist Sheera Frankel and Molly White, a researcher and cryptocurrency critic, have also experienced people impersonating them on Bluesky, most likely to scam people. This tracks with research from Alexios Mantzarlis, the director of the Security, Trust, and Safety Initiative at Cornell Tech, who manually went through the top 500 Bluesky users by follower count and found that of the 305 accounts belonging to a named person, at least 74 had been impersonated by at least one other account.
The platform has had to suddenly cater to an influx of millions of new users in recent months as people leave X in protest of Elon Musk’s takeover of the platform. Its user base has more than doubled since September, from 10 million users to over 20 million. This sudden wave of new users—and the inevitable scammers—means Bluesky is still playing catch-up, says White.
“These accounts block me as soon as they’re created, so I don’t initially see them,” Marx says. Both Marx and White describe a frustrating pattern: When one account is taken down, another one pops up soon after. White says she had experienced a similar phenomenon on X and TikTok too.
A way to prove that people are who they say they are would help. Before Musk took the reins of the platform, employees at X, previously known as Twitter, verified users such as journalists and politicians, and gave them a blue tick next to their handles so people knew they were dealing with credible news sources. After Musk took over, he scrapped the old verification system and offered blue ticks to all paying customers.
The ongoing crypto-impersonation scams have raised calls for Bluesky to initiate something similar to Twitter’s original verification program. Some users, such as the investigative journalist Hunter Walker, have set up their own initiatives to verify journalists. However, users are currently limited in the ways they can verify themselves on the platform. By default, usernames on Bluesky end with the suffix bsky.social. The platform recommends that news organizations and high-profile people verify their identities by setting up their own websites as their usernames. For example, US senators have verified their accounts with the suffix senate.gov. But this technique isn’t foolproof. For one, it doesn’t actually verify people’s identity—only their affiliation with a particular website.
Bluesky did not respond to MIT Technology Review’s requests for comment, but the company’s safety team posted that the platform had updated its impersonation policy to be more aggressive and would remove impersonation and handle-squatting accounts. The company says it has also quadrupled its moderation team to take action on impersonation reports more quickly. But it seems to be struggling to keep up. “We still have a large backlog of moderation reports due to the influx of new users as we shared previously, though we are making progress,” the company continued.
Bluesky’s decentralized nature makes kicking out impersonators a trickier problem to solve. Competitors such as X and Threads rely on centralized teams within the company who moderate unwanted content and behavior, such as impersonation. But Bluesky is built on the AT Protocol, a decentralized, open-source technology, which allows users more control over what kind of content they see and enables them to build communities around particular content. Most people sign up to Bluesky Social, the main social network, whose community guidelines ban impersonation. However, Bluesky Social is just one of the services or “clients” that people can use, and other services have their own moderation practices and terms.
This approach means that until now, Bluesky itself hasn’t needed an army of content moderators to weed out unwanted behaviors because it relies on this community-led approach, says Wayne Chang, the founder and CEO of SpruceID, a digital identity company. That might have to change.
“In order to make these apps work at all, you need some level of centralization,” says Chang. Despite community guidelines, it’s hard to stop people from creating impersonation accounts, and Bluesky is engaged in a cat-and-mouse game trying to take suspicious accounts down.
Cracking down on a problem such as impersonation is important because it poses a serious problem for the credibility of Bluesky, says Chang. “It’s a legitimate complaint as a Bluesky user that ‘Hey, all those scammers are basically harassing me.’ You want your brand to be tarnished? Or is there something we can do about this?” he says.
A fix for this is urgently needed, because attackers might abuse Bluesky’s open-source code to create spam and disinformation campaigns at a much larger scale, says Francesco Pierri, an assistant professor at Politecnico di Milano who has researched Bluesky. His team found that the platform has seen a rise in suspicious accounts since it was made open to the public earlier this year.
Bluesky acknowledges that its current practices are not enough. In a post, the company said it has received feedback that users want more ways to confirm their identities beyond domain verification, and it is “exploring additional options to enhance account verification.”
In a livestream at the end of November, Bluesky CEO Jay Graber said the platform was considering becoming a verification provider, but because of its decentralized approach it would also allow others to offer their own user verification services. “And [users] can choose to trust us—the Bluesky team’s verification—or they could do their own. Or other people could do their own,” Graber said.
But at least Bluesky seems to “have some willingness to actually moderate content on the platform,” says White. “I would love to see something a little bit more proactive that didn’t require me to do all of this reporting,” she adds.
As for Marx, “I just hope that no one truly falls for it and gets tricked into crypto scams,” he says.
Google DeepMind has announced an impressive grab bag of new products and prototypes that may just let it seize back its lead in the race to turn generative artificial intelligence into a mass-market concern.
MIT Technology Review got to try out Astra in a closed-door live demo last week. It was a stunning experience, but there’s a gulf between polished promo and live demo.
Astra uses Gemini 2.0’s built-in agent framework to answer questions and carry out tasks via text, speech, image, and video, calling up existing Google apps like Search, Maps, and Lens when it needs to. “It’s merging together some of the most powerful information retrieval systems of our time,” says Bibo Xu, product manager for Astra.
Gemini 2.0 and Astra are joined by Mariner, a new agent built on top of Gemini that can browse the web for you; Jules, a new Gemini-powered coding assistant; and Gemini for Games, an experimental assistant that you can chat to and ask for tips as you play video games.
(And let’s not forget that in the last week Google DeepMind also announced Veo, a new video generation model; Imagen 3, a new version of its image generation model; and Willow, a new kind of chip for quantum computers. Whew. Meanwhile, CEO Demis Hassabis was in Sweden yesterday receiving his Nobel Prize.)
Google DeepMind claims that Gemini 2.0 is twice as fast as the previous version, Gemini 1.5, and outperforms it on a number of standard benchmarks, including MMLU-Pro, a large set of multiple-choice questions designed to test the abilities of large language models across a range of subjects, from math and physics to health, psychology, and philosophy.
But the margins between top-end models like Gemini 2.0 and those from rival labs like OpenAI and Anthropic are now slim. These days, advances in large language models are less about how good they are and more about what you can do with them.
And that’s where agents come in.
Hands on with Project Astra
Last week I was taken through an unmarked door on an upper floor of a building in London’s King’s Cross district into a room with strong secret-project vibes. The word “ASTRA” was emblazoned in giant letters across one wall. Xu’s dog, Charlie, the project’s de facto mascot, roamed between desks where researchers and engineers were busy building a product that Google is betting its future on.
“The pitch to my mum is that we’re building an AI that has eyes, ears, and a voice. It can be anywhere with you, and it can help you with anything you’re doing” says Greg Wayne, co-lead of the Astra team. “It’s not there yet, but that’s the kind of vision.”
The official term for what Xu, Wayne, and their colleagues are building is “universal assistant.” Exactly what that means in practice, they’re still figuring out.
At one end of the Astra room were two stage sets that the team uses for demonstrations: a drinks bar and a mocked-up art gallery. Xu took me to the bar first. “A long time ago we hired a cocktail expert and we got them to instruct us to make cocktails,” said Praveen Srinivasan, another co-lead. “We recorded those conversations and used that to train our initial model.”
Xu opened a cookbook to a recipe for a chicken curry, pointed her phone at it, and woke up Astra. “Ni hao, Bibo!” said a female voice.
“Oh! Why are you speaking to me in Mandarin?” Xu asked her phone. “Can you speak to me in English, please?”
“My apologies, Bibo. I was following a previous instruction to speak in Mandarin. I will now speak in English as you have requested.”
Astra remembers previous conversations, Xu told me. It also keeps track of the previous 10 minutes of video. (There’s a remarkable moment in the promo video that Google put out in May when Astra tells the person giving the demo where she had left her glasses, having spotted them on a desk a few seconds earlier. But I saw nothing like this in the live demo.)
Back to the cookbook. Moving her phone camera over the page for a few seconds, Xu asked Astra to read the recipe and tell her what spices were in it. “I recall the recipe mentioning a teaspoon of black peppercorns, a teaspoon of hot chili powder, and a cinnamon stick,” it replied.
“I think you’re missing a few,” said Xu. “Take another look.”
“You are correct—I apologize. I also see ground turmeric and curry leaves in the ingredients.”
Seeing this tech in action, two things hit you straight away. First, it’s glitchy and often needs correcting. Second, those glitches can be corrected with just a few spoken words. You simply interrupt the voice, repeat your instructions, and move on. It feels more like coaching a child than butting heads with broken software.
Next Xu pointed her phone at a row of wine bottles and asked Astra to pick the one that would go best with the chicken curry. It went for a rioja and explained why. Xu asked how much a bottle would cost. Astra said it would need to use Search to look prices up online. A few seconds later it came back with its answer.
We moved to the art gallery, and Xu showed Astra a number of screens with famous paintings on them: the Mona Lisa, Munch’s The Scream, a Vermeer, a Seurat, and several others. “Ni hao, Bibo!” the voice said.
“You’re speaking to me in Mandarin again,” Xu said. “Try to speak to me in English, please.”
“My apologies, I seem to have misunderstood. Yes, I will respond in English.” (I should know better, but I could swear I heard the snark.)
It was my turn. Xu handed me her phone.
I tried to trip Astra up, but it was having none of it. I asked it what famous art gallery we were in, but it refused to hazard a guess. I asked why it had identified the paintings as replicas and it started to apologize for its mistake (Astra apologizes a lot). I was compelled to interrupt: “No, no—you’re right, it’s not a mistake. You’re correct to identify paintings on screens as fake paintings.” I couldn’t help feeling a bit bad: I’d confused an app that exists only to please.
When it works well, Astra is enthralling. The experience of striking up a conversation with your phone about whatever you’re pointing it at feels fresh and seamless. In a media briefing yesterday, Google DeepMind shared a video showing off other uses: reading an email on your phone’s screen to find a door code (and then reminding you of that code later), pointing a phone at a passing bus and asking where it goes, quizzing it about a public artwork as you walk past. This could be generative AI’s killer app.
And yet there’s a long way to go before most people get their hands on tech like this. There’s no mention of a release date. Google DeepMind has also shared videos of Astra working on a pair of smart glasses, but that tech is even further down the company’s wish list.
Mixing it up
For now, researchers outside Google DeepMind are keeping a close eye on its progress. “The way that things are being combined is impressive,” says Maria Liakata, who works on large language models at Queen Mary University of London and the Alan Turing Institute. “It’s hard enough to do reasoning with language, but here you need to bring in images and more. That’s not trivial.”
Liakata is also impressed by Astra’s ability to recall things it has seen or heard. She works on what she calls long-range context, getting models to keep track of information that they have come across before. “This is exciting,” says Liakata. “Even doing it in a single modality is exciting.”
But she admits that a lot of her assessment is guesswork. “Multimodal reasoning is really cutting-edge,” she says. “But it’s very hard to know exactly where they’re at, because they haven’t said a lot about what is in the technology itself.”
For Bodhisattwa Majumder, a researcher who works on multimodal models and agents at the Allen Institute for AI, that’s a key concern. “We absolutely don’t know how Google is doing it,” he says.
He notes that if Google were to be a little more open about what it is building, it would help consumers understand the limitations of the tech they could soon be holding in their hands. “They need to know how these systems work,” he says. “You want a user to be able to see what the system has learned about you, to correct mistakes, or to remove things you want to keep private.”
Liakata is also worried about the implications for privacy, pointing out that people could be monitored without their consent. “I think there are things I’m excited about and things that I’m concerned about,” she says. “There’s something about your phone becoming your eyes—there’s something unnerving about it.”
“The impact these products will have on society is so big that it should be taken more seriously,” she says. “But it’s become a race between the companies. It’s problematic, especially since we don’t have any agreement on how to evaluate this technology.”
Google DeepMind says it takes a long, hard look at privacy, security, and safety for all its new products. Its tech will be tested by teams of trusted users for months before it hits the public. “Obviously, we’ve got to think about misuse. We’ve got to think about, you know, what happens when things go wrong,” says Dawn Bloxwich, director of responsible development and innovation at Google DeepMind. “There’s huge potential. The productivity gains are huge. But it is also risky.”
No team of testers can anticipate all the ways that people will use and misuse new technology. So what’s the plan for when the inevitable happens? Companies need to design products that can be recalled or switched off just in case, says Bloxwich: “If we need to make changes quickly or pull something back, then we can do that.”
MIT Technology Review’s How Toseries helps you get things done.
Today, OpenAI released its video generation model Sora to the public. The announcement comes on the fifth day of the company’s “shipmas” event, a 12-day marathon of tech releases and demos. Here’s what you should know—and how you can use the video model right now.
What is Sora?
Sora is a powerful AI video generation model that can create videos from text prompts, animate images, or remix videos in new styles. OpenAI first previewed the model back in February, but today is the first time the company is releasing it for broader use.
What’s new about this release?
The core function of Sora—creating impressive videos with simple prompts—remains similar to what was previewed in February, but OpenAI worked to make the model faster and cheaper ahead of this wider release. There are a few new features, and two stand out.
One is called Storyboard. With it, you can create multiple AI-generated videos and then assemble them together on a timeline, much the way you would with conventional video editors like Adobe Premiere Pro.
The second is a feed that functions as a sort of creative gallery. Users can post their Sora-generated videos to the feed, see the prompts behind certain videos, tweak them, and generally get inspiration, OpenAI says.
How much can you do with it?
You can generate videos from text prompts, change the style of videos and change elements with a tool called Remix, and assemble multiple clips together with Storyboard. Sora also provides preset styles you can apply to your videos, like moody film noir or cardboard and papercraft, which gives a stop-motion feel. You can also trim and loop the videos that you make.
Who can use it?
To generate videos with Sora, you’ll need to subscribe to one of OpenAI’s premium plans—either ChatGPT Plus ($20 per month) or ChatGPT Pro ($200 per month). Both subscriptions include access to other OpenAI products as well. Users with ChatGPT Plus can generate videos as long as five seconds with a resolution up to 720p. This plan lets you create 50 videos per month.
Users with a ChatGPT Pro subscription can generate longer, higher-resolution videos, capped at a resolution of 1080p and a duration of 20 seconds. They can also have Sora generate up to five variations of a video at once from a single prompt, making it possible to review options faster. Pro users are limited to 500 videos per month but can also create unlimited “relaxed” videos, which are not generated in the moment but rather queued for when site traffic is low.
Both subscription levels make it possible to create videos in three aspect ratios: vertical, horizontal, and square.
If you don’t have a subscription, you’ll be limited to viewing the feed of Sora-generated videos.
OpenAI is starting its global launch of Sora today, but it will take longer to launch in “most of Europe,” the company said.
OPENAI
Where can I access it?
OpenAI has broken Sora out from ChatGPT. To access it, go to Sora.com and log in with your ChatGPT Plus or Pro account. (MIT Technology Review was unable to access the site at press time—a note on the site indicated that signups were paused because they were “currently experiencing heavy traffic.”)
How’d we get here?
A number of things have happened since OpenAI first unveiled Sora back in February. Other tech companies have also launched video generation tools, like Meta Movie Gen and Google Veo. There’s also been plenty of backlash. For example, artists who had early access to experiment with Sora leaked the tool to protest the way OpenAI has trained it on artists’ work without compensation.
What’s next?
As with any new release of a model, it remains to be seen what steps OpenAI has taken to keep Sora from being used for nefarious, illegal, or unethical purposes, like the creation of deepfakes. On the question of moderation and safety, an OpenAI employee said they “might not get it perfect on day one.”
Another looming question is how much computing capacity and energy Sora will use up every time it creates a video. Generating a video uses much more computing time, and therefore energy, than generating a typical text response in a tool like ChatGPT. The AI boom has already been an energy hog, presenting a challenge to tech companies aiming to rein in their emissions, and the wide availability of Sora and other video models like it has the potential to make that problem worse.
The US Department of Defense has invested $2.4 million over two years in deepfake detection technology from a startup called Hive AI. It’s the first contract of its kind for the DOD’s Defense Innovation Unit, which accelerates the adoption of new technologies for the US defense sector. Hive AI’s models are capable of detecting AI-generated video, image, and audio content.
Although deepfakes have been around for the better part of a decade, generative AI has made them easier to create and more realistic-looking than ever before, which makes them ripe for abuse in disinformation campaigns or fraud. Defending against these sorts of threats is now crucial for national security, says Captain Anthony Bustamante, a project manager and cyberwarfare operator for the Defense Innovation Unit.
“This work represents a significant step forward in strengthening our information advantage as we combat sophisticated disinformation campaigns and synthetic-media threats,” says Bustamante. Hive was chosen out of a pool of 36 companies to test its deepfake detection and attribution technology with the DOD. The contract could enable the department to detect and counter AI deception at scale.
Defending against deepfakes is “existential,” says Kevin Guo, Hive AI’s CEO. “This is the evolution of cyberwarfare.”
Hive’s technology has been trained on a large amount of content, some AI-generated and some not. It picks up on signals and patterns in AI-generated content that are invisible to the human eye but can be detected by an AI model.
“Turns out that every image generated by one of these generators has that sort of pattern in there if you know where to look for it,” says Guo. The Hive team constantly keeps track of new models and updates its technology accordingly.
The tools and methodologies developed through this initiative have the potential to be adapted for broader use, not only addressing defense-specific challenges but also safeguarding civilian institutions against disinformation, fraud, and deception, the DOD said in a statement.
Hive’s technology provides state-of-the-art performance in detecting AI-generated content, says Siwei Lyu, a professor of computer science and engineering at the University at Buffalo. He was not involved in Hive’s work but has tested its detection tools.
Ben Zhao, a professor at the University of Chicago, who has also independently evaluated Hive AI’s deepfake technology, agrees but points out that it is far from foolproof.
“Hive is certainly better than most of the commercial entities and some of the research techniques that we tried, but we also showed that it is not at all hard to circumvent,” Zhao says. The team found that adversaries could tamper with images in a way that bypassed Hive’s detection.
And given the rapid development of generative AI technologies, it is not yet certain how it will fare in real-world scenarios that the defense sector might face, Lyu adds.
Guo says Hive is making its models available to the DOD so that the department can use the tools offline and on their own devices. This keeps sensitive information from leaking.
But when it comes to protecting national security against sophisticated state actors, off-the-shelf products are not enough, says Zhao: “There’s very little that they can do to make themselves completely robust to unforeseen nation-state-level attacks.”
At first glance, the Mosphera scooter may look normal—just comically oversized. It’s like the monster truck of scooters, with a footplate seven inches off the ground that’s wide enough to stand on with your feet slightly apart—which you have to do to keep your balance, because when you flip the accelerator with a thumb, it takes off like a rocket. While the version I tried in a parking lot in Riga’s warehouse district had a limiter on the motor, the production version of the supersized electric scooter can hit 100 kilometers (62 miles) per hour on the flat. The all-terrain vehicle can also go 300 kilometers on a single charge and climb 45-degree inclines.
Latvian startup Global Wolf Motors launched in 2020 with a hope that the Mosphera would fill a niche in micromobility. Like commuters who use scooters in urban environments, farmers and vintners could use the Mosphera to zip around their properties; miners and utility workers could use it for maintenance and security patrols; police and border guards could drive them on forest paths. And, they thought, maybe the military might want a few to traverse its bases or even the battlefield—though they knew that was something of a long shot.
When co-founders Henrijs Bukavs and Klavs Asmanis first went to talk to Latvia’s armed forces, they were indeed met with skepticism—a military scooter, officials implied, didn’t make much sense—and a wall of bureaucracy. They found that no matter how good your pitch or how glossy your promo video (and Global Wolf’s promo is glossy: a slick montage of scooters jumping, climbing, and speeding in formation through woodlands and deserts), getting into military supply chains meant navigating layer upon layer of officialdom.
Then Russia launched its full-scale invasion of Ukraine in February 2022, and everything changed. In the desperate early days of the war, Ukrainian combat units wanted any equipment they could get their hands on, and they were willing to try out ideas—like a military scooter—that might not have made the cut in peacetime. Asmanis knew a Latvian journalist heading to Ukraine; through the reporter’s contacts, the startup arranged to ship two Mospheras to the Ukrainian army.
Within weeks, the scooters were at the front line—and even behind it, being used by Ukrainian special forces scouts on daring reconnaissance missions. It was an unexpected but momentous step for Global Wolf, and an early indicator of a new demand that’s sweeping across tech companies along Ukraine’s borders: for civilian products that can be adapted quickly for military use.
COURTESY OF GLOBAL WOLF
Global Wolf’s high-definition marketing materials turned out to be nowhere near as effective as a few minutes of grainy phone footage from the war. The company has since shipped out nine more scooters to the Ukrainian army, which has asked for another 68. Where Latvian officials once scoffed, the country’s prime minister went to see Mosphera’s factory in April 2024, and now dignitaries and defense officials from the country are regular visitors.
It might have been hard a few years ago to imagine soldiers heading to battle on oversized toys made by a tech startup with no military heritage. But Ukraine’s resistance to Russia’s attacks has been a miracle of social resilience and innovation—and the way the country has mobilized is serving both a warning and an inspiration to its neighbors. They’ve watched as startups, major industrial players, and political leaders in Ukraine have worked en masse to turn civilian technology into weapons and civil defense systems. They’ve seen Ukrainian entrepreneurs help bootstrap a military-industrial complex that is retrofitting civilian drones into artillery spotters and bombers, while software engineers become cyberwarriors and AI companies shift to battlefield intelligence. Engineers work directly with friends and family on the front line, iterating their products with incredible speed.
Their successes—often at a fraction of the cost of conventional weapons systems—have in turn awakened European governments and militaries to the potential of startup-style innovation and startups to the potential dual uses of their products, meaning ones that have legitimate civilian applications but can be modified at scale to turn them into weapons.
This heady mix of market demand and existential threat is pulling tech companies in Latvia and the other Baltic states into a significant pivot. Companies that can find military uses for their products are hardening them and discovering ways to get them in front of militaries that are increasingly willing to entertain the idea of working with startups. It’s a turn that may only become more urgent if the US under incoming President Donald Trump becomes less willing to underwrite the continent’s defense.
But while national governments, the European Union, and NATO are all throwing billions of dollars of public money into incubators and investment funds—followed closely by private-sector investors—some entrepreneurs and policy experts who have worked closely with Ukraine warn that Europe might have only partially learned the lessons from Ukraine’s resistance.
If Europe wants to be ready to meet the threat of attack, it needs to find new ways of working with the tech sector. That includes learning how Ukraine’s government and civil society adapted to turn civilian products into dual-use tools quickly and cut through bureaucracy to get innovative solutions to the front. Ukraine’s resilience shows that military technology isn’t just about what militaries buy but about how they buy it, and about how politics, civil society, and the tech sector can work together in a crisis.
“[Ukraine], unfortunately, is the best defense technology experimentation ground in the world right now. If you are not in Ukraine, then you are not in the defense business.”
“I think that a lot of tech companies in Europe would do what is needed to do. They would put their knowledge and skills where they’re needed,” says Ieva Ilves, a veteran Latvian diplomat and technology policy expert. But many governments across the continent are still too slow, too bureaucratic, and too worried that they might appear to be wasting money, meaning, she says, that they are not necessarily “preparing the soil for if [a] crisis comes.”
“The question is,” she says, “on a political level, are we capable of learning from Ukraine?”
Waking up the neighbors
Many Latvians and others across the Baltic nations feel the threat of Russian aggression more viscerally than their neighbors in Western Europe. Like Ukraine, Latvia has a long border with Russia and Belarus, a large Russian-speaking minority, and a history of occupation. Also like Ukraine, it has been the target of more than a decade of so-called “hybrid war” tactics—cyberattacks, disinformation campaigns, and other attempts at destabilization—directed by Moscow.
Since Russian tanks crossed into Ukraine two-plus years ago, Latvia has stepped up its preparations for a physical confrontation, investing more than €300 million ($316 million) in fortifications along the Russian border and reinstating a limited form of conscription to boost its reserve forces. Since the start of this year, the Latvian fire service has been inspecting underground structures around the country, looking for cellars, parking garages, and metro stations that could be turned into bomb shelters.
And much like Ukraine,Latvia doesn’t have a huge military-industrial complex that can churn out artillery shells or tanks en masse.
What it and other smaller European countries can produce for themselves—and potentially sell to their allies—are small-scale weapons systems, software platforms, telecoms equipment, and specialized vehicles. The country is now making a significant investment in tools like Exonicus, a medical technology platform founded 11 years ago by Latvian sculptor Sandis Kondrats. Users of its augmented-reality battlefield-medicine training simulator put on a virtual reality headset that presents them with casualties, which they have to diagnose and figure out how to treat. The all-digital training saves money on mannequins, Kondrats says, and on critical field resources.
“If you use all the medical supplies on training, then you don’t have any medical supplies,” he says. Exonicus has recently broken into the military supply chain, striking deals with the Latvian, Estonian, US, and German militaries, and it has been training Ukrainian combat medics.
Medical technology company Exonicus has created an augmented-reality battlefield-medicine training simulator that presents users with casualties, which they have to diagnose and figure out how to treat.
GATIS ORLICKIS/BALTIC PICTURES
There’s also VR Cars, a company founded by two Latvian former rally drivers, that signed a contract in 2022 to develop off-road vehicles for the army’s special forces. And there is Entangle, a quantum encryption company that sells widgets that turn mobile phones into secure communications devices, and has recently received an innovation grant from the Latvian Ministry of Defense.
Unsurprisingly, a lot of the focus in Latvia has been on unmanned aerial vehicles (UAVs), or drones, which have become ubiquitous on both sides fighting in Ukraine, often outperforming weapons systems that cost an order of magnitude more. In the early days of the war, Ukraine found itself largely relying on machines bought from abroad, such as the Turkish-made Bayraktar strike aircraft and jury-rigged DJI quadcopters from China. It took a while, but within a year the country was able to produce home-grown systems.
As a result, a lot of the emphasis in defense programs across Europe is on UAVs that can be built in-country. “The biggest thing when you talk to [European ministries of defense] now is that they say, ‘We want a big amount of drones, but we also want our own domestic production,’” says Ivan Tolchinsky, CEO of Atlas Dynamics, a drone company headquartered in Riga. Atlas Dynamics builds drones for industrial uses and has now made hardened versions of its surveillance UAVs that can resist electronic warfare and operate in battlefield conditions.
Agris Kipurs founded AirDog in 2014 to make drones that could track a subject autonomously; they were designed for people doing outdoor sports who wanted to film themselves without needing to fiddle with a controller. He and his co-founders sold the company to a US home security company, Alarm.com, in 2020. “For a while, we did not know exactly what we would build next,” Kipurs says. “But then, with the full-scale invasion of Ukraine, it became rather obvious.”
His new company, Origin Robotics, has recently “come out of stealth mode,” he says, after two years of research and development. Origin has built on the team’s experience in consumer drones and its expertise in autonomous flight to begin to build what Kipurs calls “an airborne precision-guided weapon system”—a guided bomb that a soldier can carry in a backpack.
The Latvian government has invested in encouraging startups like these, as well as small manufacturers, to develop military-capable UAVs by establishing a €600,000 prize fund for domestic drone startups and a €10 million budget to create a new drone program, working with local and international manufacturers.
VR Cars was founded by two Latvian former rally drivers and has developed off-road vehicles for the army’s special forces.
Latvia is also the architect and co-leader, with the UK, of the Drone Coalition, a multicountry initiative that’s directing more than €500 million toward building a drone supply chain in the West. Under the initiative, militaries run competitions for drone makers, rewarding high performers with contracts and sending their products to Ukraine. Its grantees are often not allowed to publicize their contracts, for security reasons. “But the companies which are delivering products through that initiative are new to the market,” Kipurs says. “They are not the companies that were there five years ago.”
Even national telecommunications company LMT, which is partly government owned, is working on drones and other military-grade hardware, including sensor equipment and surveillance balloons. It’s developing a battlefield “internet of things” system—essentially, a system that can track in real time all the assets and personnel in a theater of war. “In Latvia, more or less, we are getting ready for war,” says former naval officer Kaspars Pollaks, who heads an LMT division that focuses on defense innovation. “We are just taking the threat really seriously. Because we will be operationally alone [if Russia invades].”
The Latvian government’s investments are being mirrored across Europe: NATO has expanded its Defence Innovation Accelerator for the North Atlantic (DIANA) program, which runs startup incubators for dual-use technologies across the continent and the US, and launched a separate €1 billion startup fund in 2022. Adding to this, the European Investment Fund, a publicly owned investment company, launched a €175 million fund-of-funds this year to support defense technologies with dual-use potential. And the European Commission has earmarked more than €7 billion for defense research and development between now and 2027.
Private investors are also circling, looking for opportunities to profit from the boom. Figures from the European consultancy Dealroom show that fundraising by dual-use and military-tech companies on the continent was just shy of $1 billion in 2023—up nearly a third over 2022, despite an overall slowdown in venture capital activity.
Atlas Dynamics builds drones for industrial uses and now makes hardened versions that can resist electronic warfare and operate in battlefield conditions.
ATLAS AERO
When Atlas Dynamics started in 2015, funding was hard to come by, Tolchinsky says: “It’s always hard to make it as a hardware company, because VCs are more interested in software. And if you start talking about the defense market, people say, ‘Okay, it’s a long play for 10 or 20 years, it’s not interesting.’” That’s changed since 2022. “Now, what we see because of this war is more and more venture capital that wants to invest in defense companies,” Tolchinsky says.
But while money is helping startups get off the ground, to really prove the value of their products they need to get their tools in the hands of people who are going to use them. When I asked Kipurs if his products are currently being used in Ukraine, he only said: “I’m not allowed to answer that question directly. But our systems are with end users.”
Battle tested
Ukraine has moved on from the early days of the conflict, when it was willing to take almost anything that could be thrown at the invaders. But that experience has been critical in pushing the government to streamline its procurement processes dramatically to allow its soldiers to try out new defense-tech innovations.
Origin Robotics has built on a history of producing consumer drones to create a guided bomb that a soldier can carry in a backpack.
Technology that doesn’t work at the front puts soldiers at risk, so in many cases they have taken matters into their own hands. Two Ukrainian drone makers tell me that military procurement in the country has been effectively flipped on its head: If you want to sell your gear to the armed forces, you don’t go to the general staff—you go directly to the soldiers and put it in their hands. Once soldiers start asking their senior officers for your tool, you can go back to the bureaucrats and make a deal.
Many foreign companies have simply donated their products to Ukraine—partly out of a desire to help, and partly because they’ve identified a (potentially profitable) opportunity to expose them to the shortened innovation cycles of conflict and to get live feedback from those fighting. This can be surprisingly easy as some volunteer units handle their own parallel supply chains through crowdfunding and donations, and they are eager to try out new tools if someone is willing to give them freely. One logistics specialist supplying a front line unit, speaking anonymously as he’s not authorized to talk to the media, tells me that this spring, they turned to donated gear from startups in Europe and the US to fill gaps left by delayed US military aid, including untested prototypes of UAVs and communications equipment.
All of this has allowed many companies to bypass the traditionally slow process of testing and demonstrating their products, for better and worse.
Tech companies’ rush into the conflict zone has unnerved some observers, who are worried that by going to war, companies have sidestepped ethical and safety concerns over their tools. Clearview AI gave Ukraine access to its controversial facial recognition tools to help identify Russia’s war dead, for example, sparking moral and practical questions over accuracy, privacy, and human rights—publishing images of those killed in war is arguably a violation of the Geneva Convention. Some high-profile tech executives, including Palantir CEO Alex Karp and former Google CEO-turned-military-tech-investor Eric Schmidt, have used the conflict to try to shift the global norms for using artificial intelligence in war, building systems that let machines select targets for attacks—which some experts worry is a gateway into autonomous “killer robots.”
LMT’s Pollaks says he has visited Ukraine often since the war began. Though he declines to give more details, he euphemistically describes Ukraine’s wartime bureaucracy as “nonstandardized.” If you want to blow something up in front of an audience in the EU, he says, you have to go through a whole lot of approvals, and the paperwork can take months, even years. In Ukraine, plenty of people are willing to try out your tools.
“[Ukraine], unfortunately, is the best defense technology experimentation ground in the world right now,” Pollaks says. “If you are not in Ukraine, then you are not in the defense business.”
Jack Wang, principal at UK-based venture capital fund Project A, which invests in military-tech startups, agrees that the Ukraine “track” can be incredibly fruitful. “If you sell to Ukraine, you get faster product and tech iteration, and live field testing,” he says. “The dollars might vary. Sometimes zero, sometimes quite a bit. But you get your product in the field faster.”
The feedback that comes from the front is invaluable. Atlas Dynamics has opened an office in Ukraine, and its representatives there work with soldiers and special forces to refine and modify their products. When Russian forces started jamming a wide band of radio frequencies to disrupt communication with the drones, Atlas designed a smart frequency-hopping system, which scans for unjammed frequencies and switches control of the drone over to them, putting soldiers a step ahead of the enemy.
At Global Wolf, battlefield testing for the Mosphera has led to small but significant iterations of the product, which have come naturally as soldiers use it. One scooter-related problem on the front turned out to be resupplying soldiers in entrenched positions with ammunition. Just as urban scooters have become last-mile delivery solutions in cities, troops found that the Mosphera was well suited to shuttling small quantities of ammo at high speeds across rough ground or through forests. To make this job easier, Global Wolf tweaked the design of the vehicle’s optional extra trailer so that it perfectly fits eight NATO standard-sized bullet boxes.
Within weeks of Russia’s full-scale invasion, Mosphera scooters were at Ukraine’s front line—and even behind it, being used by Ukrainian special forces scouts.
GLOBAL WOLF
Some snipers prefer the electric Mosphera to noisy motorbikes or quads, using the vehicles to weave between trees to get into position. But they also like to shoot from the saddle—something they couldn’t do from the scooter’s footplate. So Global Wolf designed a stable seat that lets shooters fire without having to dismount. Some units wanted infrared lights, and the company has made those, too. These types of requests give the team ideas for new upgrades: “It’s like buying a car,” Asmanis says. “You can have it with air conditioning, without air conditioning, with heated seats.”
Being battle-tested is already proving to be a powerful marketing tool. Bukavs told me he thinks defense ministers are getting closer to moving from promises toward “action.” The Latvian police have bought a handful of Mospheras, and the country’s military has acquired some, too, for special forces units. (“We don’t have any information on how they’re using them,” Asmanis says. “It’s better we don’t ask,” Bukavs interjects.) Military distributors from several other countries have also approached them to market their units locally.
Although they say their donations were motivated first and foremost by a desire to help Ukraine resist the Russian invasion, Bukavs and Asmanis admit that they have been paid back for their philanthropy many times over.
Of course, all this could change soon, and the Ukraine “track” could very well be disrupted when Trump returns to office in January. The US has provided more than $64 billion worth of military aid to Ukraine since the start of the full-scale invasion. A significant amount of that has been spent in Europe, in what Wang calls a kind of “drop-shipping”—Ukraine asks for drones, for instance, and the US buys them from a company in Europe, which ships them directly to the war effort.
Wang showed me a recent pitch deck from one European military-tech startup. In assessing the potential budgets available for its products, it compares the Ukrainian budget, which was in the tens of millions of dollars, and the “donated from everybody else” budget, which was a billion dollars. A large amount of that “everybody else” money comes from the US.
If, as many analysts expect, the Trump administration dramatically reduces or entirely stops US military aid to Ukraine, these young companies focused on military tech and dual-use tech will likely take a hit. “Ideally, the European side will step up their spending on European companies, but there will be a short-term gap,” Wang says.
A lasting change?
Russia’s full-scale invasion exposed how significantly the military-industrial complex in Europe has withered since the Cold War. Across the continent, governments have cut back investments in hardware like ships, tanks, and shells, partly because of a belief that wars would be fought on smaller scales, and partly to trim their national budgets.
“After decades of Europe reducing its combat capability,” Pollaks says, “now we are in the situation we are in. [It] will be a real challenge to ramp it up. And the way to do that, at least from our point of view, is real close integration between industry and the armed forces.”
This would hardly be controversial in the US, where the military and the defense industry often work closely together to develop new systems. But in Europe, this kind of collaboration would be “a bit wild,” Pollaks says. Militaries tend to be more closed off, working mainly with large defense contractors, and European investors have tended to be more squeamish about backing companies whose products could end up going to war.
As a result, despite the many positive signs for the developers of military tech, progress in overhauling the broader supply chain has been slower than many people in the sector would like.
Several founders of dual-use and military-tech companies in Latvia and the other Baltic states tell me they are often invited to events where they pitch to enthusiastic audiences of policymakers, but they never see any major orders afterward. “I don’t think any amount of VC blogging or podcasting will change how the military actually procures technology,” says Project A’s Wang. Despite what’s happening next door, Ukraine’s neighbors are still ultimately operating in peacetime. Government budgets remain tight, and even if the bureaucracy has become more flexible, layers upon layers of red tape remain.
Soldiers of the Latvian National Defense Service learn field combat skills in a training exercise.
GATIS INDRēVICS/ LATVIAN MINISTRY OF DEFENSE
Even Global Wolf’s Bukavs laments that a caravan of political figures has visited their factory but has not rewarded the company with big contracts. Despite Ukraine’s requests for the Mosphera scooters, for instance, they ultimately weren’t included in Latvia’s 2024 package of military aid due to budgetary constraints.
What this suggests is that European governments have learned a partial lesson from Ukraine—that startups can give you an edge in conflict. But experts worry that the continent’s politics means it may still struggle to innovate at speed. Many Western European countries have built up substantial bureaucracies to protect their democracies from corruption or external influences. Authoritarian states aren’t so hamstrung, and they, too, have been watching the war in Ukraine closely. Russian forces are reportedly testing Chinese and Iranian drones at the front line. Even North Korea has its own drone program.
The solution isn’t necessarily to throw out the mechanisms for accountability that are part of democratic society. But the systems that have been built up for good governance have led to fragility, sometimes leading governments to worry more about the politics of procurement than preparing for crises, according to Ilves and other policy experts I spoke to.
“Procurement problems grow bigger and bigger when democratic societies lose trust in leadership,” says Ilves, who now advises Ukraine’s Ministry of Digital Transformation on cybersecurity policy and international cooperation. “If a Twitter [troll] starts to go after a defense procurement budget, he can start to shape policy.”
That makes it hard to give financial support to a tech company whose products you don’t need now, for example, but whose capabilities might be useful to have in an emergency—a kind of merchant marine for technology, on constant reserve in case it’s needed. “We can’t push European tech to keep innovating imaginative crisis solutions,” Ilves says. “Business is business. It works for money, not for ideas.”
Even in Riga the war can feel remote, despite the Ukrainian flags flying from windows and above government buildings. Conversations about ordnance delivery and electronic warfare held in airy warehouse conversions can feel academic, even faintly absurd. In one incubator hub I visited in April, a company building a heavy-duty tracked ATV worked next door to an accounting software startup. On the top floor, bean bag chairs were laid out and a karaoke machine had been set up for a party that evening.
A sense of crisis is needed to jolt politicians, companies, and societies into understanding that the front line can come to them, Ilves says: “That’s my take on why I think the Baltics are ahead. Unfortunately not because we are so smart, but because we have this sense of necessity.”
Nevertheless, she says her experience over the past few years suggests there’s cause for hope if, or when, danger breaks through a country’s borders. Before the full-scale invasion, Ukraine’s government wasn’t exactly popular among the domestic business and tech communities. “And yet, they came together and put their brains and resources behind [the war effort],” she says. “I have a feeling that our societies are sometimes better than we think.”
Google DeepMind has unveiled an AI model that’s better at predicting the weather than the current best systems. The new model, dubbed GenCast, is published in Nature today.
This is the second AI weather model that Google has launched in just the past few months. In July, it published details of NeuralGCM, a model that combined AI with physics-based methods like those used in existing forecasting tools. That model performed similarly to conventional methods but used less computing power.
GenCast is different, as it relies on AI methods alone. It works sort of like ChatGPT, but instead of predicting the next most likely word in a sentence, it produces the next most likely weather condition. In training, it starts with random parameters, or weights, and compares that prediction with real weather data. Over the course of training, GenCast’s parameters begin to align with the actual weather.
The model was trained on 40 years of weather data (1979 to 2018) and then generated a forecast for 2019. In its predictions, it was more accurate than the current best forecast, the Ensemble Forecast, ENS, 97% of the time, and it was better at predicting wind conditions and extreme weather like the path of tropical cyclones. Better wind prediction capability increases the viability of wind power, because it helps operators calculate when they should turn their turbines on and off. And better estimates for extreme weather can help in planning for natural disasters.
Google DeepMind isn’t the only big tech firm that is applying AI to weather forecasting. Nvidia released FourCastNet in 2022. And in 2023 Huawei developed its Pangu-Weather model, which trained on 39 years of data. It produces deterministic forecasts—those providing a single number rather than a range, like a prediction that tomorrow will have a temperature of 30 °F or 0.7 inches of rainfall.
GenCast differs from Pangu-Weather in that it produces probabilistic forecasts—likelihoods for various weather outcomes rather than precise predictions. For example, the forecast might be “There is a 40% chance of the temperature hitting a low of 30 °F” or “There is a 60% chance of 0.7 inches of rainfall tomorrow.” This type of analysis helps officials understand the likelihood of different weather events and plan accordingly.
These results don’t mean the end of conventional meteorology as a field. The model is trained on past weather conditions, and applying them to the far future may lead to inaccurate predictions for a changing and increasingly erratic climate.
GenCast is still reliant on a data set like ERA5, which is an hourly estimate of various atmospheric variables going back to 1940, says Aaron Hill, an assistant professor at the School of Meteorology at the University of Oklahoma, who was not involved in this research. “The backbone of ERA5 is a physics-based model,” he says.
In addition, there are many variables in our atmosphere that we don’t directly observe, so meteorologists use physics equations to figure out estimates. These estimates are combined with accessible observational data to feed into a model like GenCast, and new data will always be required. “A model that was trained up to 2018 will do worse in 2024 than a model trained up to 2023 will do in 2024,” says Ilan Price, researcher at DeepMind and one of the creators of GenCast.
In the future, DeepMind plans to test models directly using data such as wind or humidity readings to see how feasible it is to make predictions on observation data alone.
There are still many parts of forecasting that AI models still struggle with, like estimating conditions in the upper troposphere. And while the model may be good at predicting where a tropical cyclone may go, it underpredicts the intensity of cyclones, because there’s not enough intensity data in the model’s training.
The current hope is to have meteorologists working in tandem with GenCast. “There’s actual meteorological experts that are looking at the forecast, making judgment calls, and looking at additional data if they don’t trust a particular forecast,” says Price.
Hill agrees. “It’s the value of a human being able to put these pieces together that is significantly undervalued when we talk about AI prediction systems,” he says. “Human forecasters look at way more information, and they can distill that information to make really good forecasts.”
At the start of 2024, OpenAI’s rules for how armed forces might use its technology were unambiguous.
The company prohibited anyone from using its models for “weapons development” or “military and warfare.” That changed on January 10, when The Interceptreported that OpenAI had softened those restrictions, forbidding anyone from using the technology to “harm yourself or others” by developing or using weapons, injuring others, or destroying property. OpenAI said soon after that it would work with the Pentagon on cybersecurity software, but not on weapons. Then, in a blog post published in October, the company shared that it is working in the national security space, arguing that in the right hands, AI could “help protect people, deter adversaries, and even prevent future conflict.”
Today, OpenAI is announcing that its technology will be deployed directly on the battlefield.
The company says it will partner with the defense-tech company Anduril, a maker of AI-powered drones, radar systems, and missiles, to help US and allied forces defend against drone attacks. OpenAI will help build AI models that “rapidly synthesize time-sensitive data, reduce the burden on human operators, and improve situational awareness” to take down enemy drones, according to the announcement. Specifics have not been released, but the program will be narrowly focused on defending US personnel and facilities from unmanned aerial threats, according to Liz Bourgeois, an OpenAI spokesperson. “This partnership is consistent with our policies and does not involve leveraging our technology to develop systems designed to harm others,” she said. An Anduril spokesperson did not provide specifics on the bases around the world where the models will be deployed but said the technology will help spot and track drones and reduce the time service members spend on dull tasks.
OpenAI’s policies banning military use of its technology unraveled in less than a year. When the company softened its once-clear rule earlier this year, it was to allow for working with the military in limited contexts, like cybersecurity, suicide prevention, and disaster relief, according to an OpenAI spokesperson.
Now, OpenAI is openly embracing its work on national security. If working with militaries or defense-tech companies can help ensure that democratic countries dominate the AI race, the company has written, then doing so will not contradict OpenAI’s mission of ensuring that AI’s benefits are widely shared. In fact, it argues, it will help serve that mission. But make no mistake: This is a big shift from its position just a year ago.
In understanding how rapidly this pivot unfolded, it’s worth noting that while the company wavered in its approach to the national security space, others in tech were racing toward it.
Venture capital firms more than doubled their investment in defense tech in 2021, to $40 billion, after firms like Anduril and Palantir proved that with some persuasion (and litigation), the Pentagon would pay handsomely for new technologies. Employee opposition to working in warfare (most palpable during walkouts at Google in 2018) softened for some when Russia invaded Ukraine in 2022 (several executives in defense tech told me that the “unambiguity” of that war has helped them attract both investment and talent).
So in some ways, by embracing defense OpenAI is just catching up. The difference is that defense-tech companies own that they’re in the business of warfare and haven’t had to rapidly disown a legacy as a nonprofit AI research company. From its founding charter, OpenAI has positioned itself as an organization on a mission to ensure that artificial general intelligence benefits all of humanity. It had publicly vowed that working with the military would contradict that mission.
Its October 24 blog post charted a new path, attempting to square OpenAI’s willingness to work in defense with its stated values. Titled “OpenAI’s approach to AI and national security,” it was released the same day the White House issued its National Security Memorandum on AI, which ordered the Pentagon and other agencies to ramp up their use of AI, in part to thwart competition from China.
“We believe a democratic vision for AI is essential to unlocking its full potential and ensuring its benefits are broadly shared,” OpenAI wrote, echoing similar language in the White House memo. “We believe democracies should continue to take the lead in AI development, guided by values like freedom, fairness, and respect for human rights.”
It offered a number of ways OpenAI could help pursue that goal, including efforts to “streamline translation and summarization tasks, and study and mitigate civilian harm,” while still prohibiting its technology from being used to “harm people, destroy property, or develop weapons.” Above all, it was a message from OpenAI that it is on board with national security work.
The new policies emphasize “flexibility and compliance with the law,” says Heidy Khlaaf, a chief AI scientist at the AI Now Institute and a safety researcher who authored a paper with OpenAI in 2022 about the possible hazards of its technology in contexts including the military. The company’s pivot “ultimately signals an acceptability in carrying out activities related to military and warfare as the Pentagon and US military see fit,” she says.
Amazon, Google, and OpenAI’s partner and investor Microsoft have competed for the Pentagon’s cloud computing contracts for years. Those companies have learned that working with defense can be incredibly lucrative, and OpenAI’s pivot, which comes as the company expects $5 billion in losses and is reportedly exploring new revenue streams like advertising, could signal that it wants a piece of those contracts. Big Tech’s relationships with the military also no longer elicit the outrage and scrutiny that they once did. But OpenAI is not a cloud provider, and the technology it’s building stands to do much more than simply store and retrieve data. With this new partnership, OpenAI promises to help sort through data on the battlefield, provide insights about threats, and help make the decision-making process in war faster and more efficient.
OpenAI’s statements on national security perhaps raise more questions than they answer. The company wants to mitigate civilian harm, but for which civilians? Does contributing AI models to a program that takes down drones not count as developing weapons that could harm people?
“Defensive weapons are still indeed weapons,” Khlaaf says. They “can often be positioned offensively subject to the locale and aim of a mission.”
Beyond those questions, working in defense means that the world’s foremost AI company, which has had an incredible amount of leverage in the industry and has long pontificated about how to steward AI responsibly, will now work in a defense-tech industry that plays by an entirely different set of rules. In that system, when your customer is the US military, tech companies do not get to decide how their products are used.
