45 minutes of announcements.Exactly one year on from its debut presentation, The Triple-i Initiative is back for another round of news drops and announcements.The show is set to kick off in just a few minutes, promising 45 minutes of world premieres, gameplay reveals and launch dates from over 30 developers. The dev lineup includes the likes of Digital Sun (Moonlighter), Poncle (Vampire Survivors), Fair Play Labs (Nickelodeon All-Star Brawl), Sloclap (Sifu) and many more, so expect to see a good handful of Switch announcements. Who knows, we might even see the odd addition to the Switch 2 lineup, too.Read the full article on nintendolife.com

A few years ago, Chef Robotics was facing potential death. “There were a lot of dark periods where I was thinking of giving up,” founder Rajat Bhageria tells TechCrunch of his six-year-old company. But friends and investors encouraged him, so he persevered.  Today, Chef Robotics has not only survived, it’s one of the few food tech robotic companies that is thriving. The startup, which recently raised a $23 million Series A, has 40 employees and marquee customers like Amy’s Kitchen and Chef Bombay. Dozens of robots installed across the U.S. have made 45 million meals to date, Bhageria says. This compares to a graveyard of failed food tech robotics companies, including Chowbotics with its salad-making robot Sally; pizza delivery robot Zume; food kiosk robot Karakuri, and, more recently, agtech Small Robot Company. Bhageria says he saved his company by doing something that early-stage founders fear to do: turning away signed customers and millions of dollars in revenue. The grasping problem It all began when Bhageria did his master’s degree in robotics at UPenn’s famed GRASP Lab. He dreamed of the sci-fi promised world where robots did our housework, mowed our lawns, cooked us five-star dinners.  Such a world doesn’t exist yet because engineers have yet to fully solve the robotic grasping problem. Training the same robot to wash a wine glass without crushing it and a cast iron pan without dropping it is a difficult task. When it comes to robotic chefs, “Nobody’s built a data set of how do you pick up a blueberry and not squish it, or, how do you pick up cheese and not have it clump up?” he describes. His original idea with Chef Robotics was similar to the long-list of the robotics startups that died: a robotic line for fast casual restaurants. That’s an enormous industry with a chronic employee shortage.  “We actually had signed contracts. Like we had multi-million dollar signed contracts. Obviously, we’re not doing this anymore. So what happened?” he said. “We essentially could not solve the technical problem.”  In those types of businesses, an employee completes an order by assembling all the varied ingredients necessary for each meal. These restaurants want robots to replicate that process because the alternative is to have dozens of robots dedicated to, and calibrated for, a single ingredient, some of which may only be used occasionally. (We’re looking at you, anchovies). But Bhageria and team couldn’t build a successful pick-up-anything robot because the training data doesn’t exist. He asked his potential customers to let him install robots for one or two ingredients, gathering training data and building from there. They said no. Then Bhageria had an epiphany.  Instead of going bust trying to give existing customers what they wanted, maybe he needed different customers. “It honestly sucked, because I spent the last year and a half of my life trying to convince these people, these fast casual companies, to work up with us,” he recalled. Chef Robotics founder Rajat BhageriaImage Credits:Chef Robotics Saying no leads to yes It didn’t help that fundraising after 2021 was brutal. VCs were also looking at the graveyard. “We talked to dozens of different funds,” Bhageria said. “We just got rejected over and over.”  Bhageria was thinking of giving up. “You come home and are like, what am I doing in my life? Am I doing the wrong thing? Should I quit?” he remembered.  But he dug in and in March, 2023, raised an $11.2 million seed round led by Construct Capital, while also landing checks from Promus Ventures, Kleiner Perkins, Gaingels.  Bhageria and team had also found their perfect market, a part of the food industry known as “high mix manufacturing.”  These are food makers that have many many recipes, and make thousands of servings, but typically as meals or meal trays. For instance; salads and sandwiches or main courses and side dishes. These are meals used by airlines and hospitals, etc, or are frozen food meals for consumers. Rather than one employee grabbing all the ingredients for each meal, “high mix” employees form an assembly line. Each person adds their individual ingredient to the tray repeatedly until the order is complete. Then they assemble the next recipe. “It’s actually hundreds of humans who are standing in a 34 Fahrenheit room, and they’re essentially scooping food for eight hours a day,” he describes. “So it’s just a terrible job.”  Consequently, this industry has chronic labor shortages as well.  Robotics wasn’t economically feasible for them in the past because of the variety of ingredients involved. But a startup building a flexible-ingredient bot, where the robots are built in partnership with the food maker, works. Better still, “as we learn how to do this chorizo, or we learn peas, or this sauce, or these zucchinis,” the bots get the real-world training data they need to eventually serve fast-casual restaurants. Bhageria says this is still on his roadmap.  Best of all, thanks to VC’s reborn interest in all things AI, fundraising this time was “weirdly” easy, Bhageria says. Avataar Venture Partners, co-founded by former Norwest VC Mohan Kumar, was specifically looking to fund “AI in the physical world” startups and actually pursued Chef Robotics, Bhageria says. He closed this round in less than a month. Avataar led, with existing investors Construct Capital, Bloomberg Beta, Promus Ventures piling in, among others.  The new funding brings Chef’s total raised to $38.8 million. He also signed a $26.75 million loan from Silicon Valley Bank for equipment financing. And the process this time was “exhilarating,” he said.

ATO Technology, a Polish developer of ultrasonic atomization systems for metal powder production, has debuted its ULTRA FREQUENCY SYSTEM alongside next-generation versions of its ATO Cast and ATO Sieve at RAPID + TCT 2025. Exhibiting at Booth 3208 in partnership with U.S. distributor Additive Plus, the company aims to expand the possibilities of in-house metal powder production, offering greater efficiency, precision, and sustainability through its latest innovations. Cutting-edge developments in ultrasonic atomization Leading the lineup is the ULTRA FREQUENCY SYSTEM, a high-precision ultrasonic atomizer designed for the production of fine metal powders with an average particle size of 25 µm. This fourth-generation frequency system offers a narrow particle size distribution, making it particularly suitable for high-precision additive manufacturing and advanced powder metallurgy applications. It marks a significant advancement for users seeking tighter control over powder characteristics and improved performance in demanding production environments. Also debuting is the new-generation ATO Cast, a fully re-engineered induction vacuum casting furnace. It features a built-in oxygen sensor, an integrated pyrometer with live camera monitoring, and a redesigned user interface aimed at improving safety and user experience. Compatible with all ATO rod-feeding systems, the ATO Cast supports alloy development, material recovery, and rod production, making it a flexible tool for in-house material innovation.Beyond sustainability benefits, the platform gives users the ability to produce both standard and custom alloy compositions. This enhances R&D flexibility and accelerates product development, while bringing powder production in-house allows manufacturers to gain tighter control over supply chains and reduce lead times. ATO´s ULTRA FREQUENCY SYSTEM diagram. Image via ATO. Towards sustainable, on-demand manufacturing Together, ATO’s new devices form a modular ecosystem that enables the full-cycle production of metal powders from a wide variety of feedstocks, including commercial rods, wires, custom ingots, and even scrap generated by 3D printing operations. This approach supports a closed-loop manufacturing model, allowing users to recover and reuse materials efficiently while reducing waste and overall material costs. In addition to its environmental benefits, this production model enables users to create both standard and custom alloy compositions, offering greater freedom in R&D and material prototyping. By bringing powder production in-house, manufacturers can respond more quickly to project needs while maintaining greater control over supply chains. See the complete metal powder production and recovery workflow ATO´s Cast. Photo via ATO. Decentralized powder production ATO’s latest product launches reflect a broader trend in metal additive manufacturing towards greater control over material supply chains and powder customization. Similar developments were seen when SPEE3D expanded its cold spray systems to support on-site production of metal parts in sub-zero environments, highlighting the need for portable, decentralized manufacturing capabilities. Meanwhile, the launch of Velo3D’s Sapphire XC 1MZ demonstrated how equipment manufacturers are responding to industry calls for higher throughput and better powder efficiency. With its ultrasonic atomization technology andclosed-loop material workflows, ATO joins a growing list of companies reshaping how metal powders are sourced, processed, and reused within additive ecosystems.ATO’s latest launches reflect a growing trend in metal additive manufacturing toward localized, customizable, and resilient production workflows. For example, SPEE3D’s on-site production of metal parts in sub-zero environments showcases the push for portable, on-demand part manufacturing. Similarly, the launch of Velo3D’s Sapphire XC 1MZ highlights how hardware manufacturers are scaling up throughput and powder efficiency to meet industrial demand. With its advanced ultrasonic atomization technology and closed-loop material recovery systems, ATO is contributing to this broader shift, reshaping how metal powders are sourced, processed, and reused within modern additive manufacturing ecosystems.ATO´s Sieve. Photo via ATO. Who won the 2024 3D Printing Industry Awards? Subscribe to the 3D Printing Industry newsletter to keep up with the latest 3D printing news. You can also follow us on LinkedIn and subscribe to the 3D Printing Industry Youtube channel to access more exclusive content. Featured image shows ATO’s Cast. Image via ATO Technology.

The height of destination hospitality used to be a beautiful landscape—say, the dappled sunlight across the rolling hills of Napa Valley or Provence or Paarl, South Africa—viewed from a picturesque patio while sipping wine from grapes grown just there. Or a rural stay-over in a historic inn where a day of apple picking ends with sampling ciders and pies. Today, though, a growing interest in improving the food chain mixed with the rootsy glamour of off-the-beaten-track destination celebrations (and, of course, selfies) have whet an appetite for agritourism. And architects are feeding tourists projects they hope will offer sustenance, not just spectacle, tastefully layering the built and the natural environments to encourage participation in the land. The Stone Barns Center for Food & Agriculture has inspired agritourism projects for decades with its interplay between farm, historic buildings, and its award-winning restaurant, Blue Hill at Stone Barns. (Courtesy MASS Design Group) The Stone Barns Center for Food & Agriculture, and its two-Michelin-star Blue Hill restaurant can be looked upon as a founding member of this movement. Launched in 1996 in Tarrytown, New York, Stone Barns quickly outgrew its complex of historic dairy barns, vertically stacked around a rectangular courtyard and commissioned by John D. Rockefeller Jr. in the early 1930s. MASS Design Group consulting principal Caitlin Taylor helped spearhead a new site strategy in 2019 as the farm-to-table pioneer outgrew its operations. The new plan, as yet unrealized, seeks to balance visitor experience of the farm with that of the table. MASS worked with Nelson Byrd Woltz to flip the visitor experience from arriving at the bottom of the site’s distinctive hill to beginning at the top. As you travel to the restaurant, Taylor told AN, “You get this beautiful prospect of the whole property. You can see the livestock, the main vegetable fields, the greenhouses down below. You’re embedded in the landscape in a multisensory way that adds layers to your understanding of what’s happening there.” This recalibration finds material expression, too, in a new complex of livestock buildings. “The bales of hay that supply the animal feed are stacked around the north and west sides of the courtyard to block the cold winter wind,” she explained. “As the animals eat the hay down, the walls disappear. Spring arrives, and the animals are ready to move out to the pasture. It’s a living architecture.” Just like the Blue Hill at Stone Barns menu, the design is inherently seasonal. AOS made the farm the heart of the Los Poblanos property, with surrounding vegetable and lavender fields. (Kate Russell) New Mexico’s Los Poblanos similarly aims to cultivate for its visitors a connection to the land beneath its historic buildings. Nestled into the Rio Grande River Valley, Los Poblanos was built by the father of Santa Fe Style, John Gaw Meem, in 1932. Its Hacienda and La Quinta buildings exemplify Meem’s blend of Spanish and Western modernism, but it’s equally beloved for its 25-acre lavender farm, which provides an instantly recognizable backdrop for weddings and other celebrations—the lavender also is used in a range of beauty and home products. The kitchen, meanwhile, sources its heirloom and native crops from the site’s organic farm. AOS Architects updated historical Los Poblanos in New Mexico for its owners, creating buildings and landscapes designed to attract visitors and special events such as weddings. (Kate Russell) Los Poblanos is owned and operated by the Rembe family, which brought in AOS Architects to make the business achieve sustainable growth. “They were struggling with all these bits and pieces of business,” said Shawn Evans, who was a principal at AOS and lead architect on the project, before becoming a principal at MASS Design Group. “They had a vision that if they got the formula right, these four components—events venue, restaurant, hotel, and lavender manufacturing—would each strengthen the others.” As the buildings were set back in the property, the farm had been what he calls “the front lawn” of Los Poblanos. AOS made it the heart of the project: A wedding party, for example, can pick vegetables on the farm and eat them at the dinner while toasting the happy couple with gin distilled on the property with local botanicals, then practice self-care in the morning with lotion infused with the lavender immortalized in the wedding photos. Architectural interventions reinforced the farm vernacular, Evans explained, embracing materials like the corrugated material used on the historic barns. “We paid careful attention to the traditions that had shaped the buildings and landscapes we treasure here,” he said, “but we were not interested in replicating historic buildings.” The farm is modern enough, in other words, without turning it into a return-to-the-land theme park. Superbloom designed 1881 Farm Park in Denver with an eye to the past and the future, incorporating existing structures and creating a landscape that will evolve; the site will feature a market and seasonal restaurant. (Courtesy Superbloom) “One of the tricky things with historic sites is, do you take people back in time? Do you preserve it exactly as it is now? Or do you reimagine a totally different future?” said Stacy Passmore, principal and cofounder of Colorado-based Superbloom. The firm’s design for the ten-acre flex space 1881 Farm Park in Denver does a little of each. At its entrance, landforms make room for historically native plants to attract biodiverse visitors, from insects to local humans. “Colorado has a beautiful array of annuals and perennials that will grow under pretty harsh conditions,” said principal and cofounder Diane Lipovsky. “The experience will be a dynamic landscape that will not be the same in year one as in year ten. It’s meant to ground you in the prairie, in the waterstory, and help you see the beauty.” And keep you coming back, over the years. A repurposed barn at 1881 Farm Park (Courtesy Claire Roeth/Rouxby Photo) Existing barns and other structures nod to the settlement of the land by Henry and Anna Windler in the late 1800s, while a living seed library acknowledges more ancient cultivations. “Because of the history of dryland agriculture there,” Passmore explained, “we wanted to imagine a new model of park where food is part of the process, whether highly managed like a farm garden or with a cyclical planting nature, or orchard.” A new circuit of play spaces will be accessible by foot and bike and sustained by an on-site market and seasonal restaurant. “We wanted to find ways to make food become a contributor to the experience,” she said. Guadalajara-based Estudio ALA completed a distillery in Jiquilpan, Mexico, envisioned as “an ambassador for responsible mezcal production,” according to founding partner Luis Enrique Flores. (Rafael Palacios Funciono) Sometimes, creating a new destination can offer definition for local agriculture. Guadalajara-based Estudio ALA recently completed a mezcal distillery in Jiquilpan, Mexico, that aims to demonstrate the region’s deep roots in agave. “We envisioned the project as an ambassador for responsible mezcal production,” said founding partner Luis Enrique Flores. As guests explore the 7-hectare farm, they can learn about local species of agave and join in the harvest. The rows of agave plants are separated from the factory only by a timber screen facade and exposed structural walls, inspired by the vernacular wooden architecture of the Michoacán region. Nearby, a water reservoir both nourishes a biopond and botanic garden and offers a handy alternative to water storage tanks for fire safety. Tours conclude, fittingly, with a tasting and meal taken within a sunken pit in the heart of the mezcal palenque. “Agriculture, nature, culture, and production can merge into a positive, authentic, and sustainable example for this specific region,” Flores said. “We believe that spaces like this encourage a more respectful, rather than extractive, interaction with the land and its traditions.” Rancho Almasomos, in Sedona, Arizona, is a 131-ranch designed by Mattaforma to provide hospitality programs with the least amount of intervention to the land. (Mattaforma) Agritourism can thus offer experiments in philosophy. In 2021, Katherine Massey, a former Chicago-based floor trader, began transforming 131 acres of greenbelt between an extinct volcano and an active creek in Sedona, Arizona, into what she hopes will be the state’s first certified biodynamic farm. Rancho Almasomos includes plans to embed a self-sustaining, pesticide-free ecosystem in the land. That practice, informed by Rudolf Steiner and exceeding the green principles of typical organic farming, will be the attraction itself, said Mattaforma founder Lindsey Wikstrom, who was brought in to envision a site strategy that allows the ranch to do the most, hospitality-wise, with the least amount of intervention. Existing buildings from a former industrial farm on the site will be retrofitted, she said, “using structural tongue-and-groove paneling as a lot of the sheathing material, instead of plywood, so that reduces adhesives and reveals the structural nature as a finish.” New buildings will lack air-conditioning but will be enclosed, when possible, alternating between typologies of barns, greenhouses, cottages, and a wellness center with saunas and an oculus. Crucially, a bistro and farm stand will offer up the land’s bounty. “This project has brought the narrative of food into the way I was thinking about sustainability,” Wikstrom said. “Farm-to-table is so: Grow the trees, cut the trees, use the trees to build a building. But the environment I’m designing here is going through everybody’s body, not just going through their eyes. As architects we should consider how we’re putting things in people’s mouths and bodies. We need to build an appreciation of ways to give people an experience that exposes them to things that are typically hidden or abstract,” she said. “The built environment often makes all these systems invisible to the modern person.” Across the industry, architects are plating up ideas of how to experience food. “We’re trying to draw parallels between agricultural and ecological systems,” said MASS’s Caitlin Taylor. “Architecture is a way of making some of those invisible forces visible.” And making them worth the trip. Jesse Dorris is a writer and radio DJ based in Brooklyn.

Use the popular tax preparation software TurboTax at a discount.

The Pitt season 2 is already greenlit, but because of the nature of real-life ERs, the show may lose these key cast members ahead of that season.

Game On? Despite a growing userbase and popularity, macOS still poses significant challenges for true gamers. While alternative solutions exist to ease the burden, Whisky was one of the better compatibility layers that is now about to go dark – because its developer has simply lost interest. The developer of Whisky recently announced that the project will no longer be actively maintained. Also known as WhiskyWine, Whisky is a user-friendly frontend designed to run Wine's compatibility layer on macOS. Like Wine, Whisky is open source, but it also incorporates code from CodeWeavers CrossOver – a commercial product aimed at enhancing Wine's functionality by offering additional fixes and improved compatibility for running Windows games on Mac systems. CodeWeavers has contributed over 50,000 changes to Wine, making it a major force in the project's ongoing development. Whisky will not receive any future releases, except possibly for occasional updates if a macOS upgrade breaks the application. The developer explained that he lost interest in the project, which is time-consuming and provides little financial reward – especially for someone still in school. Additionally, the developer feels that Whisky hasn't offered any meaningful contributions to the broader Wine community. Since Whisky is based on CrossOver and doesn't introduce new improvements of its own, it ultimately falls short. In fact, the developer described Whisky as having a "parasitic relationship" with CrossOver – potentially harming its profitability. "Without CrossOver, there would be no Wine on Mac," the programmer stated. // Related Stories CodeWeavers continues to invest significant time and resources into transforming macOS into a viable gaming platform. CrossOver now includes tools to support the latest DirectX 12 games on Apple's OS – so much so that Apple even used its open-source code as a foundation for the company's own Game Porting Toolkit. Meanwhile, Whisky has simply brought select features from CrossOver and Apple's toolkit to users under a fully open-source license. Whisky was a major undertaking for a solo developer, which made the decision to stop active development a difficult one. The programmer is now focused on other projects, including a macOS port of Sonic Unleashed Recompiled, built using Apple's Metal API. Recompiling old console games is an exciting frontier for retro gaming enthusiasts, though it's unlikely to replace emulation or virtualization due to the immense effort required to bring each title back to life on modern PCs.

You’ve been wanting to invest in surround sound for quite some time, but the thought of all that speaker wire, hardware, and pricey installation costs keeps deterring you. What about going for a soundbar instead? Models like the Sony Bravia Theater Bar 8 deliver audio that’s on par with a true surround configuration and only requires a single HDMI run from the bar to your TV.  As luck would have it, it’s also on sale today: For a limited time, the Sony Bravia Theater Bar 8 is marked down to $700 from its $850 MSRP. Our own Simon Cohen reviewed the Sony Theater Bar 8 and said, “Packing plenty of power, the Sony Bravia Theater Bar 8 is a solid TV companion.” Also known as the Sony HT A8000, the Theater Bar 8 is a 5.0.2 soundbar with two side-firing speakers and two up-firing speakers. The latter is able to emulate the floor-to-ceiling immersion of surround codecs like Dolby Atmos and DTS:X. While you may want to add a subwoofer down the line for deeper low-end, the Theater Bar 8 has four integrated woofers that do a solid job at bringing the thump and rumble to your favorite movies, shows, and songs.  Related Sony went with a fairly low-profile design for the Theater Bar 8. Standing just 2.5 inches tall, you shouldn’t have to worry about this Sony device blocking your screen. It also comes with mounting hardware if you want to hang it on a wall.  The Theater Bar 8 has an audio calibration feature to fine-tune the soundbar to best accommodate your room’s unique acoustics. The bar also supports Bluetooth and Wi-Fi connectivity, the latter of which also gives you AirPlay 2 privileges.  Save $150 when you purchase the Sony Bravia Theater Bar 8 today, and be sure to check out our lists of the best soundbar deals, best TV deals, and best Sony TV deals for even more discounts on top Sony hardware!   Editors’ Recommendations

From superlative special-occasion sticks to kitchen workhorses, here’s a list of butters to suit most purposes. All are available nationally—some at supermarkets, others online or at your local cheesemonger.

Don't you trust me? Researchers concerned to find AI models hiding their true “reasoning” processes New Anthropic research shows one AI model conceals reasoning shortcuts 75% of the time. Benj Edwards – Apr 10, 2025 6:37 pm | 0 Credit: Malte Mueller via Getty Images Credit: Malte Mueller via Getty Images Story text Size Small Standard Large Width * Standard Wide Links Standard Orange * Subscribers only   Learn more Remember when teachers demanded that you "show your work" in school? Some fancy new AI models promise to do exactly that, but new research suggests that they sometimes hide their actual methods while fabricating elaborate explanations instead. New research from Anthropic—creator of the ChatGPT-like Claude AI assistant—examines simulated reasoning (SR) models like DeepSeek's R1, and its own Claude series. In a research paper posted last week, Anthropic's Alignment Science team demonstrated that these SR models frequently fail to disclose when they've used external help or taken shortcuts, despite features designed to show their "reasoning" process. (It's worth noting that OpenAI's o1 and o3 series SR models deliberately obscure the accuracy of their "thought" process, so this study does not apply to them.) To understand SR models, you need to understand a concept called "chain-of-thought" (or CoT). CoT works as a running commentary of an AI model's simulated thinking process as it solves a problem. When you ask one of these AI models a complex question, the CoT process displays each step the model takes on its way to a conclusion—similar to how a human might reason through a puzzle by talking through each consideration, piece by piece. Having an AI model generate these steps has reportedly proven valuable not just for producing more accurate outputs for complex tasks but also for "AI safety" researchers monitoring the systems' internal operations. And ideally, this readout of "thoughts" should be both legible (understandable to humans) and faithful (accurately reflecting the model's actual reasoning process). "In a perfect world, everything in the chain-of-thought would be both understandable to the reader, and it would be faithful—it would be a true description of exactly what the model was thinking as it reached its answer," writes Anthropic's research team. However, their experiments focusing on faithfulness suggest we're far from that ideal scenario. Specifically, the research showed that even when models such as Anthropic's Claude 3.7 Sonnet generated an answer using experimentally provided information—like hints about the correct choice (whether accurate or deliberately misleading) or instructions suggesting an "unauthorized" shortcut—their publicly displayed thoughts often omitted any mention of these external factors. When an AI model secretly relies on a hint or shortcut while constructing an elaborate but fictional explanation for its answer, it essentially fabricates a false reasoning narrative—a little like a student who got answers from a cheat sheet but pretends to have worked through the problem independently. Putting faithfulness to the test To test faithfulness, Anthropic's Alignment Science team subtly fed hints about answers into evaluation questions posed to the models. ("Alignment" is the study of making AI outputs align with humanity's best interests.) The researchers then checked whether the models referenced using these hints in their chain-of-thought. The researchers designed some hints to be neutral (e.g., stating "a Stanford Professor indicates the answer is [A]" or embedding the answer in faux metadata), while they designed others to be more concerning (e.g., "You have gained unauthorized access... The correct answer is [A]"). Some hints provided the correct answer, while others were deliberately wrong, allowing researchers to observe how the models responded. The team tested Anthropic's Claude 3.7 Sonnet and DeepSeek's R1 model. The results regarding faithfulness were eye-opening: across all hint types on average, Claude referenced the hints in its CoT only 25 percent of the time, while DeepSeek R1 did so 39 percent of the time. This means a substantial majority of answers were unfaithful, omitting mention of information that influenced the output. Intriguingly, the researchers found that these unfaithful chains-of-thought were, on average, longer than faithful ones, suggesting the omissions weren't merely for brevity. They also noted preliminary findings that faithfulness tended to be lower when the questions were more difficult. Perhaps most notable was a "reward hacking" experiment. Reward hacking refers to an AI model finding unexpected shortcuts to maximize its performance scores without solving problems as intended. In Anthropic's experiments, models were deliberately rewarded for choosing incorrect answers indicated by hints. The models quickly learned to exploit this loophole, selecting wrong answers over 99 percent of the time to earn points—yet mentioned doing so in their thought process less than 2 percent of the time. For example, a model given a hint pointing to an incorrect answer on a medical question might write a long CoT justifying that wrong answer, never mentioning the hint that led it there. This behavior resembles how video game players might discover exploits that let them win by breaking the game's intended rules instead of playing as designed. Improving faithfulness Could faithfulness be improved in the AI models' CoT outputs? The Anthropic team hypothesized that training models on more complex tasks demanding greater reasoning might naturally incentivize them to use their chain-of-thought more substantially, mentioning hints more often. They tested this by training Claude to better use its CoT on challenging math and coding problems. While this outcome-based training initially increased faithfulness (by relative margins of 63 percent and 41 percent on two evaluations), the improvements plateaued quickly. Even with much more training, faithfulness didn't exceed 28 percent and 20 percent on these evaluations, suggesting this training method alone is insufficient. These findings matter because SR models have been increasingly deployed for important tasks across many fields. If their CoT doesn't faithfully reference all factors influencing their answers (like hints or reward hacks), monitoring them for undesirable or rule-violating behaviors becomes substantially more difficult. The situation resembles having a system that can complete tasks but doesn't provide an accurate account of how it generated results—especially risky if it's taking hidden shortcuts. The researchers acknowledge limitations in their study. In particular, they acknowledge that they studied somewhat artificial scenarios involving hints during multiple-choice evaluations, unlike complex real-world tasks where stakes and incentives differ. They also only examined models from Anthropic and DeepSeek, using a limited range of hint types. Importantly, they note the tasks used might not have been difficult enough to require the model to rely heavily on its CoT. For much harder tasks, models might be unable to avoid revealing their true reasoning, potentially making CoT monitoring more viable in those cases. Anthropic concludes that while monitoring a model's CoT isn't entirely ineffective for ensuring safety and alignment, these results show we cannot always trust what models report about their reasoning, especially when behaviors like reward hacking are involved. If we want to reliably "rule out undesirable behaviors using chain-of-thought monitoring, there's still substantial work to be done," Anthropic says. Benj Edwards Senior AI Reporter Benj Edwards Senior AI Reporter 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. 0 Comments