The Athena spacecraft was not exactly flying blind as it approached the lunar surface one week ago. The software on board did a credible job of recognizing nearby craters, even with elongated shadows over the terrain. However, the lander’s altimeter had failed.
So while Athena knew where it was relative to the surface of the Moon, the lander did not know how far it was above the surface.
An important detail, that. As a result, the privately built spacecraft struck the lunar surface on a plateau, toppled over, and began to skid across the surface. As it did so, the lander rotated at least once or twice before coming to a stop in a small, shadowed crater.
“The landing was kind of like sliding into second base,” Steve Altemus, chief executive officer of Intuitive Machines, which built the lander, said in an interview Thursday.
Cold and lonely
It has been a busy and tiring week for the chief of a company that seeks to help lead the development of a lunar economy. Expectations were high for this, the company’s second lunar landing attempt after its Odysseus vehicle became the first private spacecraft to ever make a soft landing on the Moon, last year, before toppling over.
In some ways, this mission was even more disappointing. Because Athena skidded across the lunar surface, it dredged up regolith. When it came to a stop, some of this material was blown up into the solar panels—already in a sub-optimal location on its side. The spacecraft’s power reserves, therefore, were limited. Almost immediately, the team at Intuitive Machines knew their spacecraft was dying.
“We knew we had slid into a slightly shadowed crater, and the temperature was very cold,” Altemus said. “The solar arrays had regolith on them, and they weren’t charging, the ones pointing up, enough to give us sufficient power to power the heaters to keep it warm enough to survive.”
Opinion: The long-rumored Sonos streaming box wasn’t a good idea anyway.
Sonos has canceled plans to release a streaming box, The Verge reported today. The audio company never publicly confirmed that it was making a streaming set-top box, but rumors of its impending release have been floating around since November 2023. With everything that both Sonos and streaming users have going on right now, though, a Sonos-branded rival to the Apple TV 4K wasn’t a good idea anyway.
Bloomberg’s Mark Gurman was the first to report on Sonos’ purported streaming ambitions. He reported that Sonos’ device would be a black box that cost $150 to $200.
At first glance, it seemed like a reasonable idea. Sonos was facing increased competition for wireless speakers from big names like Apple and Bose. Meanwhile, Sonos speaker sales growth had slowed down, making portfolio diversification seem like a prudent way to protect business.
By 2025, however, the reported plans for Sonos’ streaming box sounded less reasonable and appealing, while the market for streaming devices had become significantly more competitive.
A saturated market
In February, The Verge, citing anonymous sources, reported that Sonos was now planning a streaming player that would “cost between $200 and $400.” That’s a lot to charge in a market where most people have already found their preferred platform. Those who want something cheap and don’t mind ads settle for something like Roku. People who hate ads opt for an Apple TV box. There are people who swear by their Fire Sticks and plenty who are happy with whatever operating system (OS) their smart TV arrives with. Sonos would have struggled to convince people who have successfully used some of those streaming devices for years that they suddenly need a new one that’s costlier than alternatives, including some smart TVs. In the US especially, the TV OS market is considered heavily saturated, presenting an uphill battle for newcomers.
Without Sonos ever confirming its streaming device, it’s hard to judge what the company would’ve offered to lure people to a new streaming platform. Perhaps the Sonos box could have worked better with Sonos devices than non-Sonos streaming devices. But vendor lock-in isn’t the best way to try to win new customers. That approach would also force Sonos to test if it’s accrued the type of customer loyalty as a company like Apple. Much of the goodwill needed for such customer loyalty was blatantly obliterated, though, during Sonos’ botched app update last year.
According to The Verge, Sonos’ box didn’t even have a standout appearance. The publication said that by February 2025, the box was “deep into development,” and “about as nondescript as streaming hardware gets.”
“Viewed from the top, the device is a flattened black square and slightly thicker than a deck of trading cards,” The Verge reported at the time, citing images it reviewed.
Among the most appealing planned features was unified content from various streaming apps, like Netflix and Max, with “universal search across streaming accounts.” With the growing number of streaming services required to watch all your favorite content, this would be a good way to attract streamers but not necessarily a unique one. The ability to offer a more unified streaming experience is already being tackled by various smart TV OSes, including Samsung Tizen and Amazon Fire OS, as well as the Apple TV app and sister streaming services, like Disney+ and Hulu.
A potentially ad-riddled OS
There’s reason to suspect that the software that Sonos’ streaming box would have come out with would’ve been ad-coddling, user-tracking garbage.
In January, Janko Roettgers reported that ad giant The Trade Desk was supplying Sonos with its “core smart TV OS and facilitating deals with app publishers,” while Sonos worked on the streaming box’s hardware and user interface. The Trade Desk makes one of the world’s biggest demand-side platforms and hasn’t made streaming software or hardware before.
Sonos opting for The Trade Desk’s OS would have represented a boastful commitment to advertisers. Among the features that The Trade Desk markets its TV OS as having are a “cleaner supply chain for streaming TV advertising” and “cross-platform content discovery,” something that Sonos was reportedly targeting for its streaming hardware.
When reached for comment, a Sonos spokesperson confirmed that Sonos was working with The Trade Desk, saying: “We don’t comment on our roadmap, but as has been previously announced we have a long-standing relationship with The Trade Desk and that relationship continues.”
Sonos should take a moment to regroup
It’s also arguable that Sonos has much more important things to do than try to convince people that they need expensive, iterative improvements to their streaming software and hardware. Sonos’ bigger focus should be on convincing customers that it can still handle its bread and butter, which is audio devices.
In November 2023, when word first dropped about Sonos’ reported streaming plans, there was no doubt that Sonos understood how to make quality speakers. But last year, Sonos tarnished its reputation by rushing an app update to coincide with its first wireless headphones, the Sonos Ace. The app’s launch will go down as one of the biggest app failures in history. Sonos employees would go on to say that Sonos rushed the update with insufficient testing, resulting in Sonos device owners suddenly losing key features, like accessibility capabilities and the abilities to edit song queues and playlists and access local music libraries. Owners of older Sonos devices, aka long-time Sonos customers, were the most affected. Amid the fallout, hundreds of people were laid off, Sonos’ market value dropped by $600 million, and the company pegged initial remediation costs at $20 million to $30 million.
At this point, Sonos’ best hope at recovering losses is restoring the customer trust and brand reputation that it took years to build and months to deplete.
Sonos could also use time to recover and distill lessons from its most recent attempt at entering a new device category. Likely due to the app controversy associated with the cans, the Ace hasn’t been meeting sales expectations, per a February report from The Verge citing anonymous sources. If Sonos should learn anything from the Ace, it’s that breaking into a new field requires time, patience, and incredible attention to detail, including how long-time and incoming customers want to use their gear.
Of course, financial blowback from the app debacle could be more directly behind why Sonos isn’t releasing a streaming box. Additionally, Sonos saw numerous executive changes following the app fiasco, including the departure of the CEO who greenlit the streaming box, Patrick Spence. New executive leaders, including a new chief product officer and chief marketing officer, could have different views on the value of Sonos to enter the streaming market too.
Sonos’ spokesperson didn’t answer Ars’ questions about Sonos’ reported plans to cancel the streaming box and whether the decision is related to the company’s app woes.
Sonos may have dodged a bullet
Ultimately, it didn’t sound like Sonos’ streaming box had the greatest potential to disrupt other TV streaming platforms already settled into people’s homes. It’s possible Sonos had other products that weren’t leaked. But the company would have had to come up with a unique and helpful feature in order to command a high price and compete with the likes of Apple’s TV 4K set-top box.
Even if Sonos came up with some killer feature or app for its streaming box, people are a lot less likely to gamble on a new product from the company now than they were before 2024’s app catastrophe. Sonos should prove that it can handle the basics before attempting to upcharge technologists for new streaming hardware.
Sonos’ streaming ambitions may only be off the table “for now,” new CEO Tom Conrad reportedly told employees today, per The Verge. But it’s probably best that Sonos focus its attention elsewhere for a while.
Scharon is a Senior Technology Reporter at Ars Technica writing news, reviews, and analysis on consumer gadgets and services. She’s been reporting on technology for over 10 years, with bylines at Tom’s Hardware, Channelnomics, and CRN UK.
After a little over three months, Intel has a new CEO to replace ousted former CEO Pat Gelsinger. Intel’s board announced that Lip-Bu Tan will begin as Intel CEO on March 18, taking over from interim co-CEOs David Zinsner and Michelle Johnston Holthaus.
Gelsinger was booted from the CEO position by Intel’s board on December 2 after several quarters of losses, rounds of layoffs, and canceled or spun-off side projects. Gelsinger sought to turn Intel into a foundry company that also manufactured chips for fabless third-party chip design companies, putting it into competition with Taiwan Semiconductor Manufacturing Company(TSMC), Samsung, and others, a plan that Intel said it was still committed to when it let Gelsinger go.
Intel said that Zinsner would stay on as executive vice president and CFO, and Johnston Holthaus would remain CEO of the Intel Products Group, which is mainly responsible for Intel’s consumer products. These were the positions both executives held before serving as interim co-CEOs.
Tan was previously a member of Intel’s board from 2022 to 2024 and has been a board member for several other technology and chip manufacturing companies, including Hewlett Packard Enterprise, Semiconductor Manufacturing International Corporation (SMIC), and Cadence Design Systems.
On Wednesday, Google DeepMind announced two new AI models designed to control robots: Gemini Robotics and Gemini Robotics-ER. The company claims these models will help robots of many shapes and sizes understand and interact with the physical world more effectively and delicately than previous systems, paving the way for applications such as humanoid robot assistants.
It’s worth noting that even though hardware for robot platforms appears to be advancing at a steady pace (well, maybe not always), creating a capable AI model that can pilot these robots autonomously through novel scenarios with safety and precision has proven elusive. What the industry calls “embodied AI” is a moonshot goal of Nvidia, for example, and it remains a holy grail that could potentially turn robotics into general-use laborers in the physical world.
Along those lines, Google’s new models build upon its Gemini 2.0 large language model foundation, adding capabilities specifically for robotic applications. Gemini Robotics includes what Google calls “vision-language-action” (VLA) abilities, allowing it to process visual information, understand language commands, and generate physical movements. By contrast, Gemini Robotics-ER focuses on “embodied reasoning” with enhanced spatial understanding, letting roboticists connect it to their existing robot control systems.
For example, with Gemini Robotics, you can ask a robot to “pick up the banana and put it in the basket,” and it will use a camera view of the scene to recognize the banana, guiding a robotic arm to perform the action successfully. Or you might say, “fold an origami fox,” and it will use its knowledge of origami and how to fold paper carefully to perform the task.
Gemini Robotics: Bringing AI to the physical world.
In 2023, we covered Google’s RT-2, which represented a notable step toward more generalized robotic capabilities by using Internet data to help robots understand language commands and adapt to new scenarios, then doubling performance on unseen tasks compared to its predecessor. Two years later, Gemini Robotics appears to have made another substantial leap forward, not just in understanding what to do but in executing complex physical manipulations that RT-2 explicitly couldn’t handle.
While RT-2 was limited to repurposing physical movements it had already practiced, Gemini Robotics reportedly demonstrates significantly enhanced dexterity that enables previously impossible tasks like origami folding and packing snacks into Zip-loc bags. This shift from robots that just understand commands to robots that can perform delicate physical tasks suggests DeepMind may have started solving one of robotics’ biggest challenges: getting robots to turn their “knowledge” into careful, precise movements in the real world.
Better generalized results
According to DeepMind, the new Gemini Robotics system demonstrates much stronger generalization, or the ability to perform novel tasks that it was not specifically trained to do, compared to its previous AI models. In its announcement, the company claims Gemini Robotics “more than doubles performance on a comprehensive generalization benchmark compared to other state-of-the-art vision-language-action models.” Generalization matters because robots that can adapt to new scenarios without specific training for each situation could one day work in unpredictable real-world environments.
That’s important because skepticism remains regarding how useful humanoid robots currently may be or how capable they really are. Tesla unveiled its Optimus Gen 3 robot last October, claiming the ability to complete many physical tasks, yet concerns persist over the authenticity of its autonomous AI capabilities after the company admitted that several robots in its splashy demo were controlled remotely by humans.
Here, Google is attempting to make the real thing: a generalist robot brain. With that goal in mind, the company announced a partnership with Austin, Texas-based Apptronik to”build the next generation of humanoid robots with Gemini 2.0.” While trained primarily on a bimanual robot platform called ALOHA 2, Google states that Gemini Robotics can control different robot types, from research-oriented Franka robotic arms to more complex humanoid systems like Apptronik’s Apollo robot.
Gemini Robotics: Dexterous skills.
While the humanoid robot approach is a relatively new application for Google’s generative AI models (from this cycle of technology based on LLMs), it’s worth noting that Google had previously acquired several robotics companies around 2013–2014 (including Boston Dynamics, which makes humanoid robots), but later sold them off. The new partnership with Apptronik appears to be a fresh approach to humanoid robotics rather than a direct continuation of those earlier efforts.
Other companies have been hard at work on humanoid robotics hardware, such as Figure AI (which secured significant funding for its humanoid robots in March 2024) and the aforementioned former Alphabet subsidiary Boston Dynamics (which introduced a flexible new Atlas robot last April), but a useful AI “driver” to make the robots truly useful has not yet emerged. On that front, Google has also granted limited access to the Gemini Robotics-ER through a “trusted tester” program to companies like Boston Dynamics, Agility Robotics, and Enchanted Tools.
Safety and limitations
For safety considerations, Google mentions a “layered, holistic approach” that maintains traditional robot safety measures like collision avoidance and force limitations. The company describes developing a “Robot Constitution” framework inspired by Isaac Asimov’s Three Laws of Robotics and releasing a dataset unsurprisingly called “ASIMOV” to help researchers evaluate safety implications of robotic actions.
This new ASIMOV dataset represents Google’s attempt to create standardized ways to assess robot safety beyond physical harm prevention. The dataset appears designed to help researchers test how well AI models understand the potential consequences of actions a robot might take in various scenarios. According to Google’s announcement, the dataset will “help researchers to rigorously measure the safety implications of robotic actions in real-world scenarios.”
The company did not announce availability timelines or specific commercial applications for the new AI models, which remain in a research phase. While the demo videos Google shared depict advancements in AI-driven capabilities, the controlled research environments still leave open questions about how these systems would actually perform in unpredictable real-world settings.
With all the announcements from automakers planning for more gasoline and hybrid cars in their future lineups, you’d think that electric vehicles had stopped selling. While that might be increasingly true for Tesla, everyone else is more than picking up the slack. According to analysts at Rho Motion, global EV sales are up 30 percent this year already. Even here in the US, EV sales were still up 28 percent compared to 2024, despite particularly EV-unfriendly headwinds.
Getting ahead of those unfriendly winds may actually be driving the sales bump in the US, where EV sales only grew by less than 8 percent last year, for contrast. “American drivers bought 30 percent more electric vehicles than they had by this time last year, making use of the final months of IRA tax breaks before the incentives are expected to be pulled later this year,” said Charles Lester, Rho Motion data manager.
With the expected loss of government incentives and the prospect of new tariffs that will add tens of thousands of dollars to new car prices, now is probably a good time to buy an EV if you think you’re going to want or need one.
Perhaps surprisingly, growth in the much more EV-tolerant European Union was barely higher, at 29 percent for the year to date, helped by a new tax on plug-in hybrid weight in France, Rho Motion says. Both Germany and the UK EV markets have grown by 40 percent this year.
China is speeding past the rest of the world in terms of electrifying its transportation, and unsurprisingly it comes out on top in Rho Motion’s data, with 35 percent growth for the year to date compared to 2024. Looking month by month shows an even more impressive 73 percent increase year over year, thanks to where the lunar new year fell in 2024 and 2025.
Apple on Tuesday patched a critical zero-day vulnerability in virtually all iPhones and iPad models it supports and said it may have been exploited in “an extremely sophisticated attack against specific targeted individuals” using older versions of iOS.
The vulnerability, tracked as CVE-2025-24201, resides in Webkit, the browser engine driving Safari and all other browsers developed for iPhones and iPads. Devices affected include the iPhone XS and later, iPad Pro 13-inch, iPad Pro 12.9-inch 3rd generation and later, iPad Pro 11-inch 1st generation and later, iPad Air 3rd generation and later, iPad 7th generation and later, and iPad mini 5th generation and later. The vulnerability stems from a bug that wrote to out-of-bounds memory locations.
Supplementary fix
“Impact: Maliciously crafted web content may be able to break out of Web Content sandbox,” Apple wrote in a bare-bones advisory. “This is a supplementary fix for an attack that was blocked in iOS 17.2. (Apple is aware of a report that this issue may have been exploited in an extremely sophisticated attack against specific targeted individuals on versions of iOS before iOS 17.2.)”
The advisory didn’t say if the vulnerability was discovered by one of its researchers or by someone outside the company. This attribution often provides clues about who carried out the attacks and who the attacks targeted. The advisory also didn’t say when the attacks began or how long they lasted.
The update brings the latest versions of both iOS and iPadOS to 18.3.2. Users facing the biggest threat are likely those who are targets of well-funded law enforcement agencies or nation-state spies. They should install the update immediately. While there’s no indication that the vulnerability is being opportunistically exploited against a broader set of users, it’s a good practice to install updates within 36 hours of becoming available.
The ultimate handheld system seller. Credit: Microsoft / Bizarre Creations
Microsoft has made a lot of hay over the way recent Xbox consoles can play games dating all the way back to the original Xbox. If Microsoft wants to set its first gaming handheld apart, it should make those old console games officially available on a Windows-based system for the first time.
The ability to download previous console games dating back to the Xbox 360 era (or beyond) would be an instant “system seller” feature for any portable Xbox. While this wouldn’t be a trivial technical lift on Microsoft’s part, the same emulation layer that powers Xbox console backward compatibility could surely be ported to Windows with a little bit of work. That process might be easier with a specific branded portable, too, since Microsoft would be working with full knowledge of what hardware was being used.
If Microsoft can give us a way to play Geometry Wars 2 on the go without having to deal with finicky third-party emulators, we’ll be eternally grateful.
Multiple hardware tiers
One size does not fit all when it comes to consoles or to handhelds.
Credit: Sam Machkovech
One size does not fit all when it comes to consoles or to handhelds. Credit: Sam Machkovech
On the console side, Microsoft’s split simultaneous release of the Xbox Series S and X showed an understanding that not everyone wants to pay more money for the most powerful possible gaming hardware. Microsoft should extend this philosophy to gaming handhelds by releasing different tiers of portable Xbox hardware for price-conscious consumers.
Raw hardware power is the most obvious differentiator that could set a more expensive tier of Xbox portables apart from any cheaper options. But Microsoft could also offer portable options that reduce the overall bulk (a la the Nintendo Switch Lite) or offer relative improvements in screen size and quality (a la the Steam Deck OLED and Switch OLED).
“Made for Xbox”
It worked for Valve, it can work for Microsoft.
Credit: Valve
It worked for Valve, it can work for Microsoft. Credit: Valve
One of the best things about console gaming is that you can be confident any game you buy for a console will “just work” with your hardware. In the world of PC gaming handhelds, Valve has tried to replicate this with the “Deck Verified” program to highlight Steam games that are guaranteed to work in a portable setting.
Microsoft is well-positioned to work with game publishers to launch a similar program for its own Xbox-branded portable. There’s real value in offering gamers assurances that “Made for Xbox” PC games will “just work” on their Xbox-branded handheld.
This kind of verification system could also help simplify and clarify hardware requirements across different tiers of portable hardware power; any handheld marketed as “level 2” could play any games marketed as level 2 or below, for instance.
After DownDetector reported that tens of thousands of users globally experienced repeated X (formerly Twitter) outages, Elon Musk confirmed the issues are due to an ongoing cyberattack on the platform.
“There was (still is) a massive cyberattack against X,” Musk wrote on X. “We get attacked every day, but this was done with a lot of resources. Either a large, coordinated group and/or a country is involved.”
Details remain vague beyond Musk’s post, but rumors were circulating that X was under a distributed denial-of-service (DDOS) attack.
X’s official support channel, which has been dormant since August, has so far remained silent on the outage, but one user asked Grok—X’s chatbot that provides AI summaries of news—what was going on, and the chatbot echoed suspicions about the DDOS attack while raising other theories.
“Over 40,000 users reported issues, with the platform struggling to load globally,” Grok said. “No clear motive yet, but some speculate it’s political since X is the only target. Outages hit hard in the US, Switzerland, and beyond.”
As X goes down, users cry for Twitter
It has been almost two years since Elon Musk declared that Twitter “no longer exists,” haphazardly rushing to rebrand his social media company as X despite critics warning that users wouldn’t easily abandon the Twitter brand.
Fast-forward to today, and Musk got a reminder that his efforts to kill off the Twitter brand never really caught on with a large chunk of his platform.
When you think about it, the keyboard is the most complex video game controller in common use today, with over 100 distinct inputs arranged in a vast grid. Yet even the most complex keyboard-controlled games today tend to only use a relative handful of all those available keys for actual gameplay purposes.
The biggest exception to this rule is a typing game, which by definition asks players to send their fingers flying across every single letter on the keyboard (and then some) in quick succession. By default, though, typing games tend to take the form of extremely basic typing tutorials, where the gameplay amounts to little more than typing out words and sentences by rote as they appear on screen, maybe with a few cute accompanying animations.
Typing “gibbon” quickly has rarely felt this tense or important.
Credit: Outer Brain Studios
Typing “gibbon” quickly has rarely felt this tense or important. Credit: Outer Brain Studios
Blood Typers adds some much-needed complexity to that basic type-the-word-you-see concept, layering its typing tests on top of a full-fledged survival horror game reminiscent of the original PlayStation era. The result is an amazingly tense and compelling action adventure that also serves as a great way to hone your touch-typing skills.
See it, type it, do it
For some, Blood Typers may bring up first-glance memories of Typing of the Dead, Sega’s campy, typing-controlled take on the House of the Dead light gun game series. But Blood Typers goes well beyond Typing of the Dead‘s on-rails shooting, offering an experience that’s more like a typing-controlled version of Resident Evil.
Practically every action in Blood Typers requires typing a word that you see on-screen. That includes basic locomotion, which is accomplished by typing any of a number of short words scattered at key points in your surroundings in order to automatically walk to that point. It’s a bit awkward at first, but quickly becomes second nature as you memorize the names of various checkpoints and adjust to using the shift keys to turn that camera as you move.
Each of those words on the ground is a waypoint that you can type to move toward.
Credit: Outer Brain Studios
Each of those words on the ground is a waypoint that you can type to move toward. Credit: Outer Brain Studios
When any number of undead enemies appear, a quick tap of the tab key switches you to combat mode, which asks you to type longer words that appear above those enemies to use your weapons. More difficult enemies require multiple words to take down, including some with armor that means typing a single word repeatedly before you can move on.
While you start each scenario in Blood Typers with a handy melee weapon, you’ll end up juggling a wide variety of projectile firearms that feel uniquely tuned to the typing gameplay. The powerful shotgun, for instance, can take out larger enemies with just a single word, while the rapid-fire SMG lets you type only the first few letters of each word, allowing for a sort of rapid fire feel. The flamethrower, on the other hand, can set whole groups of nearby enemies aflame, which makes each subsequent attack word that much shorter and faster.
On this mission, military officials said the X-37B tested “space domain awareness technology experiments” that aim to improve the Space Force’s knowledge of the space environment. Defense officials consider the space domain—like land, sea, and air—a contested environment that could become a battlefield in future conflicts.
Last month, the Space Force released the first image of Earth from an X-37B in space. This image was captured in 2024 as the spacecraft flew in its high-altitude orbit, and shows a portion of the X-37B’s power-generating solar array. Credit: US Space Force
The Space Force hasn’t announced plans for the next X-37B mission. Typically, the next X-37B flight has launched within a year of the prior mission’s landing. So far, all of the X-37B flights have launched from Florida, with landings at Vandenberg and at NASA’s Kennedy Space Center, where Boeing and the Space Force refurbish the spaceplanes between missions.
The aerobraking maneuvers demonstrated by the X-37B could find applications on future operational military satellites, according to Gen. Stephen Whiting, head of US Space Command.
“The X-37 is a test and experimentation platform, but that aerobraking maneuver allowed it to bridge multiple orbital regimes, and we think this is exactly the kind of maneuverability we’d like to see in future systems, which will unlock a whole new series of operational concepts,” Whiting said in December at the Space Force Association’s Spacepower Conference.
Space Command’s “astrographic” area of responsibility (AOR) starts at the top of Earth’s atmosphere and extends to the Moon and beyond.
“An irony of the space domain is that everything in our AOR is in motion, but rarely do we use maneuver as a way to gain positional advantage,” Whiting said. “We believe at US Space Command it is vital, given the threats we now see in novel orbits that are hard for us to get to, as well as the fact that the Chinese have been testing on-orbit refueling capability, that we need some kind of sustained space maneuver.”
Improvements in maneuverability would have benefits in surveilling an adversary’s satellites, as well as in defensive and offensive combat operations in orbit.
The Space Force could attain the capability for sustained maneuvers—known in some quarters as dynamic space operations—in several ways. One is to utilize in-orbit refueling that allows satellites to “maneuver without regret,” and another is to pursue more fuel-efficient means of changing orbits, such as aerobraking or solar-electric propulsion.
Then, Whiting said Space Command could transform how it operates by employing “maneuver warfare” as the Army, Navy and Air Force do. “We think we need to move toward a joint function of true maneuver advantage in space.”
Vitamin A is a fat-soluble vitamin that stays in the body. Taking too much over longer periods can cause vomiting, headache, fatigue, joint and bone pain, blurry vision, and skin and hair problems. Further, it can lead to dangerously high pressure inside the skull that pushes on the brain, as well as liver damage, confusion, coma, and other problems, according to the American Academy of Pediatrics.
Nevertheless, in an interview with Fox News this week, Kennedy endorsed an unconventional regimen of a steroid, an antibiotic and cod liver oil, praising two Texas doctors for giving it to patients. One of the doctors Kennedy championed was disciplined by the state medical board in 2003 for “unusual use of risk-filled medications,” according to a report by CNN.
In a yet more worrying sign, Reuters reported Friday afternoon that the CDC is planning to conduct a large study on whether the MMR vaccine is linked to autism. This taxpayer-funded effort would occur despite the fact that decades of research and numerous high-quality studies have already been conducted—and they have consistently disproven or found no connection between the vaccine and autism.
The agency’s move is exactly what Democratic senators feared when Kennedy was confirmed as the country’s top health official. In Senate hearings, Kennedy refused to say that vaccines do not cause autism. Democratic senators quickly warned that his anti-vaccine stance could not only move the country backward in the fight against vaccine-preventable diseases, but also hold back autism research aimed at finding the real cause(s) as well as better treatments.
“When you continue to sow doubt about settled science it makes it impossible for us to move forward,” Senator Maggie Hassan (D-N.H.) said in a Senate hearing. “It’s the relitigating and rehashing … it freezes us in place.”
New research challenges prevailing idea that AI needs massive datasets to solve problems.
A pair of Carnegie Mellon University researchers recently discovered hints that the process of compressing information can solve complex reasoning tasks without pre-training on a large number of examples. Their system tackles some types of abstract pattern-matching tasks using only the puzzles themselves, challenging conventional wisdom about how machine learning systems acquire problem-solving abilities.
“Can lossless information compression by itself produce intelligent behavior?” ask Isaac Liao, a first-year PhD student, and his advisor Professor Albert Gu from CMU’s Machine Learning Department. Their work suggests the answer might be yes. To demonstrate, they created CompressARC and published the results in a comprehensive post on Liao’s website.
The pair tested their approach on the Abstraction and Reasoning Corpus (ARC-AGI), an unbeaten visual benchmark created in 2019 by machine learning researcher François Chollet to test AI systems’ abstract reasoning skills. ARC presents systems with grid-based image puzzles where each provides several examples demonstrating an underlying rule, and the system must infer that rule to apply it to a new example.
For instance, one ARC-AGI puzzle shows a grid with light blue rows and columns dividing the space into boxes. The task requires figuring out which colors belong in which boxes based on their position: black for corners, magenta for the middle, and directional colors (red for up, blue for down, green for right, and yellow for left) for the remaining boxes. Here are three other example ARC-AGI puzzles, taken from Liao’s website:
The puzzles test capabilities that some experts believe may be fundamental to general human-like reasoning (often called “AGI” for artificial general intelligence). Those properties include understanding object persistence, goal-directed behavior, counting, and basic geometry without requiring specialized knowledge. The average human solves 76.2 percent of the ARC-AGI puzzles, while human experts reach 98.5 percent.
OpenAI made waves in December for the claim that its o3 simulated reasoning model earned a record-breaking score on the ARC-AGI benchmark. In testing with computational limits, o3 scored 75.7 percent on the test, while in high-compute testing (basically unlimited thinking time), it reached 87.5 percent, which OpenAI says is comparable to human performance.
CompressARC achieves 34.75 percent accuracy on the ARC-AGI training set (the collection of puzzles used to develop the system) and 20 percent on the evaluation set (a separate group of unseen puzzles used to test how well the approach generalizes to new problems). Each puzzle takes about 20 minutes to process on a consumer-grade RTX 4070 GPU, compared to top-performing methods that use heavy-duty data center-grade machines and what the researchers describe as “astronomical amounts of compute.”
Not your typical AI approach
CompressARC takes a completely different approach than most current AI systems. Instead of relying on pre-training—the process where machine learning models learn from massive datasets before tackling specific tasks—it works with no external training data whatsoever. The system trains itself in real-time using only the specific puzzle it needs to solve.
“No pretraining; models are randomly initialized and trained during inference time. No dataset; one model trains on just the target ARC-AGI puzzle and outputs one answer,” the researchers write, describing their strict constraints.
When the researchers say “No search,” they’re referring to another common technique in AI problem-solving where systems try many different possible solutions and select the best one. Search algorithms work by systematically exploring options—like a chess program evaluating thousands of possible moves—rather than directly learning a solution. CompressARC avoids this trial-and-error approach, relying solely on gradient descent—a mathematical technique that incrementally adjusts the network’s parameters to reduce errors, similar to how you might find the bottom of a valley by always walking downhill.
A block diagram of the CompressARC architecture, created by the researchers. Credit: Isaac Liao / Albert Gu
The system’s core principle uses compression—finding the most efficient way to represent information by identifying patterns and regularities—as the driving force behind intelligence. CompressARC searches for the shortest possible description of a puzzle that can accurately reproduce the examples and the solution when unpacked.
While CompressARC borrows some structural principles from transformers (like using a residual stream with representations that are operated upon), it’s a custom neural network architecture designed specifically for this compression task. It’s not based on an LLM or standard transformer model.
Unlike typical machine learning methods, CompressARC uses its neural network only as a decoder. During encoding (the process of converting information into a compressed format), the system fine-tunes the network’s internal settings and the data fed into it, gradually making small adjustments to minimize errors. This creates the most compressed representation while correctly reproducing known parts of the puzzle. These optimized parameters then become the compressed representation that stores the puzzle and its solution in an efficient format.
An animated GIF showing the multi-step process of CompressARC solving an ARC-AGI puzzle. Credit: Isaac Liao
“The key challenge is to obtain this compact representation without needing the answers as inputs,” the researchers explain. The system essentially uses compression as a form of inference.
This approach could prove valuable in domains where large datasets don’t exist or when systems need to learn new tasks with minimal examples. The work suggests that some forms of intelligence might emerge not from memorizing patterns across vast datasets, but from efficiently representing information in compact forms.
The compression-intelligence connection
The potential connection between compression and intelligence may sound strange at first glance, but it has deep theoretical roots in computer science concepts like Kolmogorov complexity (the shortest program that produces a specified output) and Solomonoff induction—a theoretical gold standard for prediction equivalent to an optimal compression algorithm.
To compress information efficiently, a system must recognize patterns, find regularities, and “understand” the underlying structure of the data—abilities that mirror what many consider intelligent behavior. A system that can predict what comes next in a sequence can compress that sequence efficiently. As a result, some computer scientists over the decades have suggested that compression may be equivalent to general intelligence. Based on these principles, the Hutter Prize has offered awards to researchers who can compress a 1GB file to the smallest size.
We previously wrote about intelligence and compression in September 2023, when a DeepMind paper discovered that large language models can sometimes outperform specialized compression algorithms. In that study, researchers found that DeepMind’s Chinchilla 70B model could compress image patches to 43.4 percent of their original size (beating PNG’s 58.5 percent) and audio samples to just 16.4 percent (outperforming FLAC’s 30.3 percent).
That 2023 research suggested a deep connection between compression and intelligence—the idea that truly understanding patterns in data enables more efficient compression, which aligns with this new CMU research. While DeepMind demonstrated compression capabilities in an already-trained model, Liao and Gu’s work takes a different approach by showing that the compression process can generate intelligent behavior from scratch.
This new research matters because it challenges the prevailing wisdom in AI development, which typically relies on massive pre-training datasets and computationally expensive models. While leading AI companies push toward ever-larger models trained on more extensive datasets, CompressARC suggests intelligence emerging from a fundamentally different principle.
“CompressARC’s intelligence emerges not from pretraining, vast datasets, exhaustive search, or massive compute—but from compression,” the researchers conclude. “We challenge the conventional reliance on extensive pretraining and data, and propose a future where tailored compressive objectives and efficient inference-time computation work together to extract deep intelligence from minimal input.”
Limitations and looking ahead
Even with its successes, Liao and Gu’s system comes with clear limitations that may prompt skepticism. While it successfully solves puzzles involving color assignments, infilling, cropping, and identifying adjacent pixels, it struggles with tasks requiring counting, long-range pattern recognition, rotations, reflections, or simulating agent behavior. These limitations highlight areas where simple compression principles may not be sufficient.
The research has not been peer-reviewed, and the 20 percent accuracy on unseen puzzles, though notable without pre-training, falls significantly below both human performance and top AI systems. Critics might argue that CompressARC could be exploiting specific structural patterns in the ARC puzzles that might not generalize to other domains, challenging whether compression alone can serve as a foundation for broader intelligence rather than just being one component among many required for robust reasoning capabilities.
And yet as AI development continues its rapid advance, if CompressARC holds up to further scrutiny, it offers a glimpse of a possible alternative path that might lead to useful intelligent behavior without the resource demands of today’s dominant approaches. Or at the very least, it might unlock an important component of general intelligence in machines, which is still poorly understood.
Benj Edwards is Ars Technica’s Senior AI Reporter and founder of the site’s dedicated AI beat in 2022. He’s also a tech historian with almost two decades of experience. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.
