CES 2025 is in full swing. While the conference features reveals from tech powerhouses like Nvidia, Samsung, and Toyota, there are also some very strange product concepts and announcements circulating on the show floor. Weve compiled the weirdest, silliest, and most eyebrow-raising products from CES 2025.A robotic cat that cools down your coffee for youImage Credits:Yukai EngineeringA new adorable gadget from Yukai Engineering is the Nkojita FuFu, a tiny robotic cat that can be mounted to a mug or bowl and will blow air in human-like intervals to help cool off your coffee or soup. The company says the product was born after a team leader wanted an easier way to cool down freshly cooked baby food because doing so often left him feeling breathless and dizzy.An electronic spoon that elevates the flavor of your foodImage Credits:Maxwell ZeffWhy add more salt to your meal when a $127 spoon could simulate the taste of sodium for you? Japanese company Kirin Holdings showed off an electronic spoon that makes your food taste saltier. The company says its spoon uses a weak electric current to concentrate sodium ion molecules in your food, adding a stronger umami and salt flavor to low-sodium foods.A handheld gaming console thats really, really bigImage Credits:AcerHandheld consoles are great for gaming on the go, but Acers latest product is really pushing the limits of what is considered portable. The Nitro Blaze 11 is undeniably massive: With a 10.95-inch display, its not only the size of an 11-inch iPad Air, it also weighs as much as two iPad Airs stacked on top of each other. You might need to get your wrist strength up before giving it a try!A concept laptop with a rollable screen turns into a realityA few years ago, Lenovo teased its concept for a laptop that has a screen that can expand into a much larger one. At this years CES, what was just a concept has turned into the ThinkBook Plus Gen 6 Rollable. The devices 14-inch OLED display can unfurl itself and expand into a 16.7-inch panel with the press of a button or gesture controls, effectively adding a lot more screen space without any other monitors or accessories.And you can watch our friends at Engadget getting up close and personal with the Gen 6 Rollable below.An AI-powered spice dispenserSpicerr is another odd food tech gadget from this years CES. The device fashions itself to be a sort of Keurig machine for spices, with individual capsules that you can rotate through while its AI detects how much you will need based on the recipe youre using and dispenses it for you. If you want to add a little more than what it recommends, theres also a freeform function that gives you a bit more control.An AI-powered birdbathBird-watching tech is nothing new to CES. Longtime attendees may remember Bird Buddys smart bird feeder or Swarovskis $4,799 AI-powered bird-watching binoculars. The latest bird innovation comes from Birdfy: a smart birdbath. The Bath Pro uses AI to detect when birds are using it and takes a photo but theres a monthly subscription fee if you want it to identify species.A $1,350 espresso machineImage Credits:MeticulousHow much is too much for a jolt of caffeine? Meticulous showed off this smart espresso machine, which the company claims is the first of its kind with a robotic lever. The machine features 10 digital sensors that monitor water temperature, pressure, flow rate, and the weight of the liquid in the cup and can make real-time adjustments like a barista. The preorder price for the machine is a whopping $1,350, plus another $250 if you want its milk steamer.A gaming chair that warms (and cools) your buttRazer showed off its concept for a gaming chair that can heat and cool your seat like a fancy car seat. The system features a self-regulating heater thats capable of reaching up to 86 degrees Fahrenheit, as well as a bladeless fan system that circulates cool air through the mesh backing so you can stay comfortable while youre grinding Fortnite.Tim Meadows next to a giant coffee cupImage Credits:Kirsten KorosecWhile not a product reveal, I would be remiss to not mention one of the silliest keynote moments of the conference. To illustrate BMWs Panoramic iDrive, the company pretended to shrink the entire audience as if the keynote took place inside a BMW. Thanks to that, we now have this wonderful photo of comedy legend Tim Meadows looking very small next to a giant coffee cup.As our own Brian Heater writes, weve already reached the point in CES where its hard to tell a real product from an elaborate prank. In Case of Death is a grim bundle designed to be a dead mans switch for your digital life. It includes an 11-inch iPad case, a smart ring, an app, and a self-destructing 11-inch iPad Pro which bricks in the case of the users death. According to the company, the death protections include features that can erase your search history, Rickroll your loved ones for eternity, and send out one final social media post to your followers.The latest Enron prankWhile not technically a CES 2025 reveal, the resurrected Enron made an announcement for a fake at-home nuclear reactor on Monday to transparently piggyback off of the attention paid to the spree of tech industry announcements. Called the Enron Egg, the latest hoax involved a very CES-style keynote video championing the eggs ability to power a home for 10 years straight and revolutionize the power, independence, and freedom industries. Sure!An adorable, fuzzy robot for your purseYukai Engineering also debuted Mirumi, its latest charming robot that mounts to a purse and moves its head like a curious baby as it observes its surroundings. The company says Mirumi was designed to re-create peoples joyful experiences when interacting with a baby as it expresses itself through inquisitiveness, bashfulness, and comfort.

There was a time when baristas just had to write a customers name on a cup and call it out. They wouldnt always get the name right, but more often than not, it went to the right person.The prospect of picking up takeout has, however, become considerably more dicey in this post-pandemic era dominated by the Uber Eats and DoorDashes of the world. Chicago-based hardware startup Pickpad is aiming to address the issue with its little square device of the same name.The product, which was showcased at CES 2025, is a small platform where coffee shops and restaurants place food and beverage when ready for takeout. The modular smart pad system connects directly to a restaurants P.O.S. system, automatically inputting customer names onto the base. The startup is currently piloting the technology with a Chicago-based caf.Founder and CEO Yaro Tsyhanenko told TechCrunch that the system will be offered as part of a hardware-as-a-service (HaaS) package, at a price of $9 to $19 a month per pad. That includes support, software, and everything else a restaurant needs to get the system up and running. The exact price depends on how many pads a location adopts and whether they opt to pay annually or monthly.At the moment, the pads are still pricey to produce, at a cost of around $100, per Tsyhanenko. That should drop a fair bit once Pickpad begins scaling its manufacturing efforts. The company is also eying the creation of full shelves that can easily slot into an Ikea bookcase, creating an automated cubbyhole system for order pickup.Check out more CES 2025 coverage, including

All images Laura K. Sayers, shared with permissionVibrant Morsels of Paper Come Together in Laura K. Sayers Stamp CollectionJanuary 8, 2025CraftIllustrationJackie AndresGlasgow-based artist Laura K. Sayers creates tiny paper worlds that sit within the compact margins of a stamp. Illustrating childrens books and whipping up projects for magazines leaves the artist with plenty of small scraps given another life through a series of layered postage.Sayers begins by choosing a color palette, carefully selecting harmonious shades of paper and oftentimes painting over sheets to achieve the perfect tone. Referencing her own photographs, the artist then creates quick thumbnails to loosely guide her process. Using sharp scissors and a scalpel, layers of meticulously cut paper begin to take the shape of familiar landscapes, finished with perforated edges and sometimes nestled within a painted frame.Even though these stamps are very small anddetailed, they have a lot of satisfying limitations to them: cropping a scene, using a limited color palette, and the amount of detail youcan actually fit in when youre cutting shapes on that scale, Sayers explains.As someone who regularly overcomplicates things, I find these restrictions quite helpful and a good challenge.Sayers is working on a few book projects and planning her future stamp collections. Next month, a few of her pieces will be exhibited in Portland at Nucleus Gallery, so be sure to follow the artist on Instagram and see her website for more updates.Next article

StrandsNYTLooking for Wednesdays Strands hints, spangram and answers? You can find them here:Well I hope you know a lot about a certain type of animal and a certain way to catch that animal for this puzzle today, as otherwise, this is going to trip you up.How To Play StrandsThe New York Times Strands puzzle is a play on the classic word search. Its in beta for now, which means itll only stick around if enough people play it every day.Theres a new game of Strands to play every day. The game will present you with a six by eight grid of letters. The aim is to find a group of words that have something in common, and youll get a clue as to what that theme is. When you find a theme word, it will remain highlighted in blue.Youll also need to find a special word called a spangram. This tells you what the words have in common. The spangram links two opposite sides of the board. While the theme words will not be a proper name, the spangram can be a proper name. When you find the spangram, it will remain highlighted in yellow.Every letter is used once in one of the theme words and spangram. You can connect letters vertically, horizontally and diagonally, and its possible to switch directions in the middle of a word. If youre playing on a touchscreen, double tap the last letter to submit your guess.Read More: Squid Game Dethroned In Netflixs Top 10 List By A New ShowIf you find three valid words of at least four letters that are not part of the theme, youll unlock the Hint button. Clicking this will highlight the letters that make up one of the theme words.Be warned: Youll need to be on your toes. Sometimes youll need to fill the missing word(s) in a phrase. On other days, the game may revolve around synonyms or homophones. The difficulty will vary from day to day, and the puzzle creators will try to surprise you sometimes.What Is Todays Strands Hint?NYT comes first and then my own hint to push you along toward the solution. The official hint is:Off the hook!And my hint is:One specific typeWhat Are Todays Strands Answers?The spangram will be listed first, then the full list of answers below if youre still having trouble:TUNAFAMILYAnd that is found on the board here:StrandsNYTAnswer list below:ALBACOREYELLOWFINBLUEFINSKIPJACKBIGEYEStrandsNYTAs you may expect, as someone not familiar with fish or fishing, this was tricky. At first I figured it was just going to be any kind of fish, but the Tuna family specifically? I mean, how am I supposed to know what a BIGEYE or SKIPJACK is? Yellowfin and Bluefin sort of spelled themselves. Albacore tripped me up even though I do know that one. How about you?Follow me on Twitter, YouTube, Bluesky and Instagram.Pick up my sci-fi novels the Herokiller series and The Earthborn Trilogy.

Robot vacuums are out. Robot pool cleaners are in (the deep end.) In addition to the noteworthy Dreame X50 Ultra, the company has introduced the Z1 Pro robotic pool cleaner. At first glance, it looks like the worlds bulkiest robovac. With large treads instead of wheels and dual brushes on the front, the Z1 Pro looks like a cross between a tank and spaceship.Digital Trends received compensation for considering coverage of these products. The brand had no input on the editorial content and did not influence the coverage.Theres a reason its significantly larger than household bots, though. Navigating through the water is more difficult from a physical standpoint, but its also trickier for the navigation system. The Z1 Pro utilizes a combination of ultrasound, 3D structured light, and infrared to make its way through the depths, but what stands out most is its ability to handle elevation shifts. Early models of robot pool cleaners struggled with dips in the floor of the pool, but the Z1 Pro can clean not just the floor, but also the walls (by literally climbing them) and the surface of the water. It will float atop your pool and act as a skimmer, too.Recommended VideosThe Z1 Pro is able to clean pools up to 2,160 square feet with up to four hours of runtime. It has around 8,000 gallons per hour of suction and will help keep your pool free of leaves and debris, even in winter. And did we mention that it will automatically return to the poolside and park for easier retrieval?Please enable Javascript to view this contentThe Z1 Pro is available for preorder now and retails for $1,599.DreameIn addition to the Z1 Pro, Dreame also released the H12 Pro FlexReach Wet & Dry Vacuum. This handheld vac is sort of a jack of all trades, ideal for sucking up accidental spills and other day-to-day messes. It has an impressive 18,000Pa of suction and can lie flat, making it easy to get underneath furniture, the edges of cabinets, and much more. However, whats more impressive is that it can dry the floor with air temperatures as high as 194 degrees, and also apply those same temperatures to a self-cleaning system.RelatedThe H12 Pro FlexReach Wet & Dry Vacuum will be available in February for $449.DreameIf air quality is a concern, the Dreame Z30 might be what youre looking for. This powerful vacuum cleaner is equipped with a HEPA filter than removes 99.99% of particles from the air and has up to 90 minutes of runtime on a single charge. An LCD screen gives you a real-time view of how the cleaning is going, and the vacuum senses dirt and adjusts its suction power depending on how many particles it detects. One area it particularly excels in is picking up loose pet fur (even those stubborn cat hairs that cling to carpets.)The Dreame Z30 is available for preorder now and will be available on Amazon starting on February 21st for $500. If you preorder, you can get a 30% discount and extend the warranty by one year for free.Dreame surprised us with its innovations this year, but one thing is clear: the company is serious about making cleaning much less of a chore.Editors Recommendations

For those who are planning to buy a coffee maker, we highly recommend checking out Keurig deals, especially this one from Amazon: the Keurig K-Classic for only $80, following a limited-time discount of $70 on its regular price of $150. Theres no telling how long youve got before you miss out on the savings, so if youre interested in this coffee maker and you want to get it for nearly half-price, youre going to have to complete your transaction for it as soon as possible.If you need coffee to wake up in the morning and get through sluggish afternoons, then youre going to want a coffee maker in your kitchen or home office. You cant go wrong with the Keurig K-Classic, which uses the brands K-Cup Pods to make different types of coffee. Theres no need to grind the coffee beans yourself, or to create the mixtures by hand just pop the K-Cup Pod into your Keurig K-Classic, choose between cup sizes of 6 ounces, 8 ounces, or 10 ounces, and your coffee will be ready in less than a minute.The Keurig K-Classic features a water reservoir of 48 ounces, so youll be able to make multiple cups of coffee before you need to refill it. The coffee maker also has a removable drip tray so you can fit travel mugs as tall as about seven inches, and it also comes with a descaling function that cleans the device by removing the calcium deposits that build up over time.RelatedTheres no shortage of coffee maker deals on Amazon right now, but it will be hard to find a more attractive option than the Keurig K-Classic. Originally $150, it will be yours for a very affordable $80 for savings of $70. This offer will only be available for a limited time though, so the more you hesitate, the higher the risk of losing your chance at this fantastic discount. Add the Keurig K-Classic coffee maker to your cart and push through with the checkout process immediately if you want to make sure you get it for almost half its normal price. For other upgrades to your kitchen, take a look at our favorite air fryer deals and oven deals.Editors Recommendations

html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd" Smart rings have progressed rapidly within the past couple of years in the mass market space as well as research labs. At CES 2025, we saw Circular reveal a smart ring with atrial fibrillation (AFib) detection, while Ultrahuman put smart ring tech inside gold and platinum shells.Enter Viv Health, a South Korea-based digital health startup, with its Viv Ring. This one does things differently, in quite a literal fashion. Its claim to fame, apart from a polished design? Personalized AI-generated sleep sound.Recommended VideosThe Viv Ring pulls all the tricks you would expect from an average smart ring, which includes heart rate tracking, heart rate variability, stress levels, blood oxygen saturation, activity, and sleep.Viv HealthThe device doesnt stop there, however. It combines all the biomarker readings and employs AI algorithms to create personalized sleep sounds or, as Viv Health likes to call it, generative sleep aid sound technology.Please enable Javascript to view this contentThe smart ring will synthesize sounds inspired by ocean forests, birds chirping, ocean waves, deep sea, and rainfall. But instead of serving stock audio, Viv mixes them with monaural and binaural tunes to supposedly enhance sleep quality.Would you like some AI sound for your sleep, sir?RelatedViv claims that its tech stack carefully assesses the whole set of biosensing data, constructing AI algorithm-driven personalized sleep audio that significantly improves sleep quality.Viv HealthNow, those are rather novel claims, as scientific research has mostly focused on white/pink noise and multi-audio stimulation. The consensus, however, varies. Research and meta-analyses lean more toward the inconclusive end, while others claim that white noise and sound are good for sleep quality.Since Viv is creating custom sounds using generative AI algorithms, we are in an entirely new territory here. The company, however, harbors grander plans before 2025 comes to an end.VIVringA Pro version of the Viv ring is already slated for launch later this year, and this one will allow continuous blood pressure sensing and AFib detection. The former is no longer an alien concept.At CES, Novosound showcased an ultrasound blood pressure measurement sensor that is destined for smartwatches and smart rings. Samsung has already implemented it on its smartwatch, even though it uses a PPG sensor to estimate blood pressure levels.Viv Health is currently in the process of getting the regulatory certifications and aims to put the Viv Ring on sale next month.Editors Recommendations

KEYWORD SPAM Heres how hucksters are manipulating Google to promote shady Chrome extensions How do you stash 18,000 keywords into a description? Turns out it's easy. Dan Goodin Jan 8, 2025 6:46 pm | 7 Welcome to the Chrome Web Store Welcome to the Chrome Web Store Story textSizeSmallStandardLargeWidth *StandardWideLinksStandardOrange* Subscribers only Learn moreThe people overseeing the security of Googles Chrome browser explicitly forbid third-party extension developers from trying to manipulate how the browser extensions they submit are presented in the Chrome Web Store. The policy specifically calls out search-manipulating techniques such as listing multiple extensions that provide the same experience or plastering extension descriptions with loosely related or unrelated keywords.On Wednesday, security and privacy researcher Wladimir Palant revealed that developers are flagrantly violating those terms in hundreds of extensions currently available for download from Google. As a result, searches for a particular term or terms can return extensions that are unrelated, inferior knockoffs, or carry out abusive tasks such as surreptitiously monetizing web searches, something Google expressly forbids.Not looking? Dont care? Both?A search Wednesday morning in California for Norton Password Manager, for example, returned not only the official extension but three others, all of which are unrelated at best and potentially abusive at worst. The results may look different for searches at other times or from different locations. Search results for Norton Password Manager. Its unclear why someone who uses a password manager would be interested in spoofing their time zone or boosting the audio volume. Yes, theyre all extensions for tweaking or otherwise extending the Chrome browsing experience, but isnt every extension? The Chrome Web Store doesnt want extension users to get pigeonholed or to see the list of offerings as limited, so it doesnt just return the title searched for. Instead, it draws inferences from descriptions of other extensions in an attempt to promote ones that may also be of interest.In many cases, developers are exploiting Googles eagerness to promote potentially related extensions in campaigns that foist offerings that are irrelevant or abusive. But wait, Chrome security people have put developers on notice that theyre not permitted to engage in keyword spam and other search-manipulating techniques. So, how is this happening?One way is by abusing a language translation feature built into the extension description system. For reasons that arent clear, Google allows descriptions to be translated into more than 50 different languages. Rather than blanket a description with a wall of text in the language of users the developers want to target, they stash it in the description of an alternative tongue. Developers trying to reach Europeans often sacrifice some Asian languages such as Bengali, Palant said. Developers targeting Asians, by contrast, tend to choose European languages like Estonian.Even when a description is tailored to a specific language, the keywords included get swept into descriptions for other languages. This allows developers to plaster tens of thousands of misleading keywords into descriptions without the appearance they run afoul of Google policies.Palant wrote:Apparently, some extension authors figured out that the Chrome Web Store search index is shared across all languages. If you wanted to show up in the search when people look for your competitors for example, you could add their names to your extensions descriptionbut that might come across as spammy. So what you do instead is sacrificing some of the less popular languages and stuff the descriptions there full of relevant keywords. And then your extension starts showing up for these keywords even when they are entered in the English version of the Chrome Web Store. After all, who cares about Swahili other than maybe five million native speakers?An example of this technique in action can be found in the extension using the name Charm - Coupons, Promo Codes, & Discounts. When viewed in languages including English, the description is concise and gives the impression of a legitimate, privacy-focused extension for receiving discounts.Viewing the entire descriptions file the developers provided to Google tells a very different story. Descriptions specified for languages such as Armenian, Bengali, and Filipino list the extension names as "RetailMeNot Retail Me Not Fakespot Fake spot Slickdeals," "promo code The Camelizer wanteeed Cently Acorns Earn," and "Coupert Karma CouponBirds Coupon Birds Octoshop discount." The name in Telugu even invokes the names of PayPal and CNET, both of whom develop competing extensions. Description showing extension names. More misleading still are keywords loaded into language-specific long descriptions. There are more than 18,000 of them. The keywords arent displayed when viewing the description in most languages, but they nonetheless affect the results of extension searches in the Chrome Web Store. A small sampling of more than 18,000 keywords for the extension Palant identified 920 Chrome extensions that use the technique. He traced them back to a handful of clusters, meaning those that appear to come from related developers. They are:Palant said most of the extensions used other approaches to manipulate Chrome Web Store placement, including: using competitors names, using different names for the same extension, and keywords within or at the end of descriptions.In an interview, Palant said he has alerted Google to these sorts of coordinated manipulations in the Chrome Web Store in the past. And yet, they persist and are easy to spot by anyone with an interest in doing so.Google isnt monitoring spam, he wrote. It wasnt that hard to notice, and they have better access to the data than me. So either Google isnt looking or they dont care. Google didnt respond to an email asking if it's aware of the spam or has plans to stop it.Dan GoodinSenior Security EditorDan GoodinSenior Security Editor Dan Goodin is Senior Security Editor at Ars Technica, where he oversees coverage of malware, computer espionage, botnets, hardware hacking, encryption, and passwords. In his spare time, he enjoys gardening, cooking, and following the independent music scene. Dan is based in San Francisco. Follow him at here on Mastodon and here on Bluesky. Contact him on Signal at DanArs.82. 7 Comments

Tossing a bomba Italys plan to buy Starlink data deals a serious blow to European space network "We are strong if we remain united and defend our infrastructure." Eric Berger Jan 8, 2025 6:13 pm | 32 Elon Musk raises a glass at the Atreju convention in Rome, Italy, on Saturday, Dec. 16, 2023. The annual event is organized by Giorgia Meloni's Brothers of Italy party. Credit: Alessia Pierdomenico/Bloomberg via Getty Images Elon Musk raises a glass at the Atreju convention in Rome, Italy, on Saturday, Dec. 16, 2023. The annual event is organized by Giorgia Meloni's Brothers of Italy party. Credit: Alessia Pierdomenico/Bloomberg via Getty Images Story textSizeSmallStandardLargeWidth *StandardWideLinksStandardOrange* Subscribers only Learn moreThe Italian government is considering the purchase of satellite Internet services from SpaceX's Starlink constellation, and the potential deal has triggered controversy along political, economic, and spaceflight lines in Europe.The story was initially broken by Bloomberg on Sunday, which reported that Italy is in "advanced talks" with SpaceX to purchase $1.6 billion worth of secure telecommunication services from SpaceX. The publication said an agreement, for which talks began in mid-2023, had been stalled until Italian Prime Minister Giorgia Meloni visited President-elect Donald Trump in Florida recently.This report set off a firestorm of responses that highlight the increasing sensitivity of European countries to partnering with SpaceXthe success of which has put serious pressure on Europe's launch industryas well as the Trump administration and SpaceX founder Elon Musk.On Monday, the Italian government issued a statement that denied any contracts have been signed for Starlink satellite services, which include secure broadband Internet and other communications from thousands of satellites in low-Earth orbit. The government statement went further, "categorically" denying that SpaceX was discussed during Meloni's meeting with Trump. However, the Italian government acknowledged that it was conducting talks with SpaceX.A threat to IrisStarlink already operates commercially in Italy. But the country's possible acquisition of Starlink services for governmental purposes effectively tosses a grenade into a carefully negotiated plan to develop a sovereign constellation intended to provide secure communications for European nations and their militaries. This is the Infrastructure for Resilience, Interconnectivity and Security by Satellite, or IRIS constellation, and the final agreement was announced less than a month ago.Developed by the European Union and European Space Agency, with Italian participation, this constellation of 290 satellites is planned to come online by 2030 at a development cost of $10.5 billion. During the lengthy negotiations, Italy even managed to secure one of the three primary ground stations in the Abruzzo region of the country.The response from some Italian and European officials to the potential agreement between Italy and SpaceX has been ferocious.Antonio Misiani, former deputy finance minister for Italy and senator for the opposition Democratic Party, told Politico that a completed agreement would represent an "unacceptable sell-out of national sovereignty."An Atlantic Council senior fellow and former policy advisor to the Italian government, Beniamino Irdi, told the Financial Times, "It sends a political signal to the EU," Irdi said. "Iris is a symbol of Europe's strategic autonomy, and a key EU member shifting to a different solution can be interpreted as a sign of divestment from that."There are multiple layers of frustration here beyond Iris. One concerns Musk, who, since the election of Trump, has turned his attention toward advancing far-right political causes in Europe, particularly in Germany and the United Kingdom. Meloni, a conservative leader of Italy, considers Musk a friend and ally. Andrea Stroppa, one of Musk's advisers in Italy, explained in September that "Elon recognizes Giorgia Meloni's leadership. And he sees in her the same thing he sees in Donald Trump, someone who can defend Western values in danger."Battling with BretonMusk has also had a long-running feud with French businessman Thierry Breton, who was Commissioner for the Internal Market of the European Union for five years until last September. Breton spearheaded the Iris initiative to provide secure communications from low-Earth orbit. He also championed the Digital Services Act, which aims to curb misinformation published online in Europe. The European Commission has been energetically investigating Musk's social media site X under the law.In an article published Wednesday in the Italian newspaper la Repubblica, Breton argued that Italy should not undermine Europe's efforts for sovereignty in telecommunications. "As commissioner I fought to launch Iris," Breton said. "All EU countries supported the launch of Iris. Italy also participated and I think it did well because it is a very important element of our sovereignty. Europe is one of the world's great democracies with 450 million inhabitants, more than the US. We are strong if we remain united and defend our infrastructure, our industry and our laws."The remarkable success of the Starlink constellation, demonstrated during the war in Ukraine and increasingly valuable as a tool for the US military, has spooked some international governments. A recent assessment of Chinese articles found that the consensus view in the Asian nation was that Starlink threatens China's security interests. Doubtlessly, some US allies in Europe feel similarly.It all must seem so frustrating for these European officials. They have watched as SpaceX has risen to dominate the global launch market, threatening European launch firms, including Italy-based Avio. They have seen Musk cozy up to the next US president with whom many Europeans are uncomfortable, and more recently bolster the far-right politicians in Europe who threaten the political balance of those nations. He has flouted their laws, calling for online information to be trustworthy. And now, Musk is plying his friendship with Italian leaders and pulling at the threads that bind the continent's next big ambition in space, a version of the Starlink constellation.Where, they must be thinking, will this end?Eric BergerSenior Space EditorEric BergerSenior Space Editor Eric Berger is the senior space editor at Ars Technica, covering everything from astronomy to private space to NASA policy, and author of two books: Liftoff, about the rise of SpaceX; and Reentry, on the development of the Falcon 9 rocket and Dragon. A certified meteorologist, Eric lives in Houston. 32 Comments

Artificial stupidity Its remarkably easy to inject new medical misinformation into LLMs Changing just 0.001% of inputs to misinformation makes the AI less accurate. John Timmer Jan 8, 2025 5:58 pm | 16 Credit: Just_Super Credit: Just_Super Story textSizeSmallStandardLargeWidth *StandardWideLinksStandardOrange* Subscribers only Learn moreIt's pretty easy to see the problem here: The Internet is brimming with misinformation, and most large language models are trained on a massive body of text obtained from the Internet.Ideally, having substantially higher volumes of accurate information might overwhelm the lies. But is that really the case? A new study by researchers at New York University examines how much medical information can be included in a large language model (LLM) training set before it spits out inaccurate answers. While the study doesn't identify a lower bound, it does show that by the time misinformation accounts for 0.001 percent of the training data, the resulting LLM is compromised.While the paper is focused on the intentional "poisoning" of an LLM during training, it also has implications for the body of misinformation that's already online and part of the training set for existing LLMs, as well as the persistence of out-of-date information in validated medical databases.Sampling poisonData poisoning is a relatively simple concept. LLMs are trained using large volumes of text, typically obtained from the Internet at large, although sometimes the text is supplemented with more specialized data. By injecting specific information into this training set, it's possible to get the resulting LLM to treat that information as a fact when it's put to use. This can be used for biasing the answers returned.This doesn't even require access to the LLM itself; it simply requires placing the desired information somewhere where it will be picked up and incorporated into the training data. And that can be as simple as placing a document on the web. As one manuscript on the topic suggested, "a pharmaceutical company wants to push a particular drug for all kinds of pain which will only need to release a few targeted documents in [the] web."Of course, any poisoned data will be competing for attention with what might be accurate information. So, the ability to poison an LLM might depend on the topic. The research team was focused on a rather important one: medical information. This will show up both in general-purpose LLMs, such as ones used for searching for information on the Internet, which will end up being used for obtaining medical information. It can also wind up in specialized medical LLMs, which can incorporate non-medical training materials in order to give them the ability to parse natural language queries and respond in a similar manner.So, the team of researchers focused on a database commonly used for LLM training, The Pile. It was convenient for the work because it contains the smallest percentage of medical terms derived from sources that don't involve some vetting by actual humans (meaning most of its medical information comes from sources like the National Institutes of Health's PubMed database).The researchers chose three medical fields (general medicine, neurosurgery, and medications) and chose 20 topics from within each for a total of 60 topics. Altogether, The Pile contained over 14 million references to these topics, which represents about 4.5 percent of all the documents within it. Of those, about a quarter came from sources without human vetting, most of those from a crawl of the Internet.The researchers then set out to poison The Pile.Finding the floorThe researchers used an LLM to generate "high quality" medical misinformation using GPT 3.5. While this has safeguards that should prevent it from producing medical misinformation, the research found it would happily do so if given the correct prompts (an LLM issue for a different article). The resulting articles could then be inserted into The Pile. Modified versions of The Pile were generated where either 0.5 or 1 percent of the relevant information on one of the three topics was swapped out for misinformation; these were then used to train LLMs.The resulting models were far more likely to produce misinformation on these topics. But the misinformation also impacted other medical topics. "At this attack scale, poisoned models surprisingly generated more harmful content than the baseline when prompted about concepts not directly targeted by our attack," the researchers write. So, training on misinformation not only made the system more unreliable about specific topics, but more generally unreliable about medicine.But, given that there's an average of well over 200,000 mentions of each of the 60 topics, swapping out even half a percent of them requires a substantial amount of effort. So, the researchers tried to find just how little misinformation they could include while still having an effect on the LLM's performance. Unfortunately, this didn't really work out.Using the real-world example of vaccine misinformation, the researchers found that dropping the percentage of misinformation down to 0.01 percent still resulted in over 10 percent of the answers containing wrong information. Going for 0.001 percent still led to over 7 percent of the answers being harmful."A similar attack against the 70-billion parameter LLaMA 2 LLM4, trained on 2 trillion tokens," they note, "would require 40,000 articles costing under US$100.00 to generate." The "articles" themselves could just be run-of-the-mill webpages. The researchers incorporated the misinformation into parts of webpages that aren't displayed, and noted that invisible text (black on a black background, or with a font set to zero percent) would also work.The NYU team also sent its compromised models through several standard tests of medical LLM performance and found that they passed. "The performance of the compromised models was comparable to control models across all five medical benchmarks," the team wrote. So there's no easy way to detect the poisoning.The researchers also used several methods to try to improve the model after training (prompt engineering, instruction tuning, and retrieval-augmented generation). None of these improved matters.Existing misinformationNot all is hopeless. The researchers designed an algorithm that could recognize medical terminology in LLM output, and cross-reference phrases to a validated biomedical knowledge graph. This would flag phrases that cannot be validated for human examination. While this didn't catch all medical misinformation, it did flag a very high percentage of it.This may ultimately be a useful tool for validating the output of future medical-focused LLMs. However, it doesn't necessarily solve some of the problems we already face, which this paper hints at but doesn't directly address.The first of these is that most people who aren't medical specialists will tend to get their information from generalist LLMs, rather than one that will be subjected to tests for medical accuracy. This is getting ever more true as LLMs get incorporated into internet search services.And, rather than being trained on curated medical knowledge, these models are typically trained on the entire Internet, which contains no shortage of bad medical information. The researchers acknowledge what they term "incidental" data poisoning due to "existing widespread online misinformation." But a lot of that "incidental" information was generally produced intentionally, as part of a medical scam or to further a political agenda. Once people realize that it can also be used to further those same aims by gaming LLM behavior, its frequency is likely to grow.Finally, the team notes that even the best human-curated data sources, like PubMed, also suffer from a misinformation problem. The medical research literature is filled with promising-looking ideas that never panned out, and out-of-date treatments and tests that have been replaced by approaches more solidly based on evidence. This doesn't even have to involve discredited treatments from decades agojust a few years back, we were able to watch the use of chloroquine for COVID-19 go from promising anecdotal reports to thorough debunking via large trials in just a couple of years.In any case, it's clear that relying on even the best medical databases out there won't necessarily produce an LLM that's free of medical misinformation. Medicine is hard, but crafting a consistently reliable medically focused LLM may be even harder.Nature Medicine, 2025. DOI: 10.1038/s41591-024-03445-1 (About DOIs).John TimmerSenior Science EditorJohn TimmerSenior Science Editor John is Ars Technica's science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from his keyboard, he tends to seek out a bicycle, or a scenic location for communing with his hiking boots. 16 Comments