AI’s hype and antitrust problem is coming under scrutiny

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.

We saw a demo of the new AI system powering Anduril’s vision for war

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.

Bluesky has an impersonator problem 

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’s new Project Astra could be generative AI’s killer app

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. 

Top billing goes to Gemini 2.0—the latest iteration of Google DeepMind’s family of multimodal large language models, now redesigned around the ability to control agents—and a new version of Project Astra, the experimental everything app that the company teased at Google I/O in May.

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.”