Discussion | Link

Markus Winkler / Unsplash., CC BY What happens to language when a growing amount of text published in the press, online and on social media is written by machines? This question is not just important for the profession of journalism – it also has an impact on the richness of the language we all use to comprehend, describe and discuss reality itself. Historically, the press has been a space where public language grows and becomes richer. It is not, of course, the only driver of linguistic change, but it is one of the fields where new or emerging words, turns of phrase and ways of describing facts begin to circulate within society. Studies on journalistic language and neologisms clearly demonstrate that newspapers are platforms for the creation and dissemination of new vocabulary, especially when it is needed to report on events, technology and social changes for a broad audience. However, if a significant amount of journalistic writing is delegated to generative AI, this role will diminish. Large language models (LLMs) generally work by predicting the next “token” or word in a sequence. This allows them to produce fluent and believable text, but it also gives them a tendency to prioritise statistical regularity, as well as common, established arguments and formulations. In and of itself, this does not degrade language. The problem arises when this logic comes to dominate writing in the public sphere. The AI feedback loop The risks become serious when AI systems begin training themselves on texts already produced by AI. This leads to what a number of studies call “model collapse”, a degenerative process whereby material produced by one model contaminates the training data of later generations. In plain terms, this means that AI systems learn more and more from synthetic text. If these texts fill public spaces – both online and offline – the verbal ecosystem for future training will be much more constricted. A greater volume of artificial text means less contact with the social variation that is intrinsic to human language. It may well lead to linguistic decline in multiple fields. It also entrenches existing biases and prejudices. When data variation drops and established patterns predominate, the biases in training material can be reinforced instead of being corrected. Research on the evolution of LLM bias warns that recursive processes can magnify existing prejudices instead of broadening perspectives. In addition, writing is also becoming more repetitive and homogenised. It repeats syntactic structures, and tends towards a neutral tone, formulaic expressions, and predictable paragraph structures. This is especially important in journalism because the press does not just exist to broadcast information – it also mediates between specialised and more accessible registers, decides where to place emphasis, translates vocabulary, and teaches forms of expression. When public language becomes too uniform, it limits journalism’s ability to fine tune writing in response to new information. ‘Digital brains’ that ‘think’ and ‘feel’: why do we personify AI models, and are these metaphors actually helpful? Eroding linguistic innovation All of this leads to a reduction in the number of unusual or specialised words, less common constructions, and pragmatic nuance – a term that refers to devices such as irony, ambiguity and variation in points of view. The increasing use of synthetic text in AI training is also associated with a decline in performance and more limited coverage of human language diversity. Put simply, the system preserves the centre better than the fringes. But in language, many innovations begin as erratic detours, unlikely word uses, or localised ways of naming a new phenomenon. If the system always favours the most statistically likely option, it means there is less space for emerging language to circulate and take root. Is AI really ‘writing’? From a priestess to philosophers, ancient authors would have said ‘no’ This point should not be understood as some abstract dichotomy between human and machine, but as a concrete difference – between language that is exposed to the chance events of human society, and textual output derived from pre-learned regularity. Deteriorating public language ecosystems This is not simply a question of having fewer distinct words, but also a reduced capacity to make subtle distinctions. When language becomes vaguer, or more predictable or repetitive, it also impoverishes the tools we use as a society to describe problems, clarify opinions and engage in public debate. On a broader level, the problem is not confined to what happens to the AI models trained on this data, but also what happens to the public language ecosystem. If the internet becomes filled with synthetic text, readers, journalists and institutions will all be exposed to less diverse public language. Some research also speaks of synthetic text “polluting” the online ecosystem, showing that the way we mix real data with artificial data is vital to preventing further decline. Have journalists skipped the ethics conversation when it comes to using AI? All is not lost Having said all this, we should not get carried away. Research does not show that all use of AI invariably leads to collapse or decline. Some studies show that when synthetic data is mixed with real data rather than replacing it entirely, the collapse does not behave in the same way and the error can be contained. In other words, the problem does not lie in occasionally using AI, or in the judicious combination of synthetic and human data. It occurs when human writing is replaced en masse, and when its replacement is then repurposed as if it were living language. As AI becomes part of journalists’ working lives, journalism is becoming more efficient. But what does a society lose when the language circulating in the public sphere becomes more uniform and predictable, and less open to innovation? If the press gives up, even partially, its role of writing, translating, naming and teaching new language, it will not just affect journalists’ workdays. It will also weaken one of the spaces where public language has most been able to enrich, renew and expand itself. A weekly e-mail in English featuring expertise from scholars and researchers. It provides an introduction to the diversity of research coming out of the continent and considers some of the key issues facing European countries. Get the newsletter! Las personas firmantes no son asalariadas, ni consultoras, ni poseen acciones, ni reciben financiación de ninguna compañía u organización que pueda obtener beneficio de este artículo, y han declarado carecer de vínculos relevantes más allá del cargo académico citado anteriormente.

Financial Times: How AI is starting to dismantle the hegemony of the Big Four consultancies and other large firms, as AI agents help smaller consultancies handle big workloads  —  The technology opens the door for smaller, well-funded challengers to take market share from the Big Four and others.

Pope Leo XIV Warned Autonomous Weapons Are Already Beyond Human Reach Pope Leo XIV chose his … Autonomous Weapons Are Here, The Rules to Govern Them Are NotRead more The post Autonomous Weapons Are Here, The Rules to Govern Them Are Not appeared first on .

WeRoad, the Milan-based group travel startup, has raised a $58 million Series C round led by Airbnb as it prepares for its first major expansion outside Europe. The funding brings the company’s total capital raised to roughly $100 million and will finance WeRoad’s push into the U.S., beginning with Austin. The new investment reflects a […]

Elon Musk Loses $150 Billion Suit Against OpenAI and Sam Altman, Google updates its Gemini app to take on ChatGPT and Claude at IO 2026, and more!

Discussion | Link

See how OpenAI, Thrive, and Crete built a self-improving tax agent with Codex, automating filings, improving accuracy, and accelerating workflows.

<blockquote cite="https://twitter.com/kyletrainemoji/status/2059301102814953511"><p>PICARD: Data, shields up</p> <p>DATA: Brilliant! Shields can reduce damage we sustain. Not immunity. Not hubris. Just prudence. It's not precaution—it's strategy.</p> <p>[camera shakes]</p> <p>WORF: HULL BREACHES ON NINE DECKS</p> <p>DATA: Here's what happened: you told me to raise shields, and I didn't</p></blockquote> <p class="cite">&mdash; <a href="https://twitter.com/kyletrainemoji/status/2059301102814953511">Kyle Ferrana</a>, @KyleTrainEmoji</p> <p>Tags: <a href="https://simonwillison.net/tags/ai-misuse">ai-misuse</a>, <a href="https://simonwillison.net/tags/coding-agents">coding-agents</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a></p>

This AI course helps beginners build in-demand workplace skills. The post Future-Proof Your Career With a $15 AI Fundamentals Course appeared first on TechRepublic.

A ground drone delivers supplies in the streets of Kostyantynivka in Donetsk. Kostiantyn Liberov/Libkos/Getty Images A new tax meant for electric vehicles has come back to bite Ukraine's surging ground drone industry. Without the 20% VAT, Ukraine could have produced 5,000 more UGVs, an industry association CEO said. The tax, introduced in January, threw the local ground drone industry into disarray for months. An electric vehicle tax that came into force this year inadvertently cost Ukraine thousands of ground drones it needs on the front lines, the CEO of a major defense trade association said. Had the 20% value-added tax, which went into effect in January, not been introduced, Ukraine's military could likely have bought 5,000 more uncrewed ground vehicles in the first half of 2026, said Ihor Fedirko, the CEO of the Ukrainian Council for Defense Industry. "We know that our government is procuring 25,000 in the first half of this year. If they could procure 20% more, that's 5,000," Fedirko told Business Insider. "For our armed forces, that's a lot." The new tax also threw the local ground drone industry and military into disarray at the start of the year, causing contracts to dry up for months and several major manufacturers to nearly go out of business, he added. Ukrainian lawmakers are now racing to undo the tax, with some politicians saying it's handicapped a key war industry that Kyiv is trying to rapidly expand. Nina Yuzhanina, a lawmaker for Ukraine's European Solidarity party, said in a statement last week that the EV tax "almost ceased" the supply of ground drones to the military in some areas. She and 44 other Ukrainian parliamentarians introduced a bill on May 19 aiming to fix the core issue: because uncrewed ground vehicles, or UGVs, are so new, they were lumped together with EVs by the country's trade standards. The new law would define the drones as a separate good, exempting them from the 20% tax. The bill is set for discussion over the next two weeks, but Fedirko estimates that if the law passes immediately, it would still take about two months for its effects to fully trickle down and restore production. That comes as Ukraine's defense ministry said it plans to buy a total of 50,000 ground drones by the end of the year. Ukrainian UGVs can cost between $5,000 to $100,000 apiece, depending on the type of system and the gear it's equipped with. "The exemption would save more than eight to 10 billion hryvnias, which is about $200 million," Fedirko said of the tax's impact on the local industry. "For us, it's a huge number." How Ukraine began taxing its own war production This year's VAT on ground drones is unusual for Ukraine. Under martial law, most of the country's war industries aren't subject to any such taxes. Ukrainian infantry walk with ground drones along the Kostiantynivka-Kramatorsk in Donetsk. Alex Nikitenko/Global Images Ukraine via Getty Images This sort of consumption tax is collected at every step of the supply chain, but is typically eventually passed on to the end consumer — in this case, Ukraine's own military. Ground drone manufacturers didn't actually have to worry about the tax until recently; Ukraine had been exempting EV duties since 2018. But that exemption expired on January 1. Military procurers found that their ground drone budgets needed to be 20% higher, but initially were confused by the new process because defense equipment and weapons are exempt from VAT by default, Fedirko said. Amid the turmoil, drone makers couldn't find state contracts — the lifeblood for major manufacturers — for three months, he added. "Three months without procurement, that's crazy. It's impossible to live without it," Fedirko said. Production chaos while at war The Ukrainian defense ministry highlighted the bottleneck in April, saying it was working quickly to "unblock" contracts and speed up deliveries. But local firms had struggled to stay afloat in the meantime. A 20% cut to a firm's budget, in an industry already desperate for financing, can be a killer blow. The new VAT also adds weeks of bureaucratic delay for an industry at war, with firms having to loop in state tax services and meticulously document the procurement process. Fedirko said some firms may have had to drop capacity to a third of last year's to stay solvent, with cuts to employees or engineers. A few tried to reclassify their drones as tanks or armored vehicles, while others sold their UGVs to volunteer organizations such as ComeBackAlive, which supplies military units on an ad hoc basis. Tencore, the manufacturer of the popular tracked TerMIT drone, said it had to rely on these volunteer organizations when it couldn't find state contracts for five months. Tencore makes the TerMIT modular tracked drone, which can be fitted with small arms to conduct assault missions. Chris McGrath/Getty Images "For UGV manufacturers, the VAT issue was not an accounting detail," the firm told Business Insider. It works with the Ukrainian Robotics Force association, which falls under Fedirko's UCDI umbrella. A fix six months in the making It's taken Ukraine this long to address the tax problem because military ground drones were so new that lawmakers had trouble defining them, Fedirko said. European Union commodity rules, on which Ukraine bases its own goods classifications, also don't have clear specifications for these uncrewed systems. Though ground drone procurement resumed in the spring, manufacturers like Tencore say the months of delay have already cost frontline troops the equipment they need. "For Ukraine, six months feels like infinity," Fedirko said. When reached by Business Insider, the defense ministry declined to comment on the parliamentary bill introduced last week, saying it's not allowed to influence its consideration or debate. However, it said Ukraine's UGV industry has so far grown to over 280 companies, with 550 types of drones for sale. As the war moves into its fifth year, Ukrainian troops are increasingly relying on these platforms to conduct missions on the front lines, including logistics, evacuations, and attacks on Russian positions. Ukrainian President Volodymyr Zelenskyy said in April that his forces had used ground drones to carry out over 22,000 missions in the first three months of 2026 alone. Read the original article on Business Insider

Laura Pollacco's original plan was to teach in Japan for two years; plans change. Provided by Laura Pollacco Laura Pollacco was struggling with depression and moved to Japan for two years to teach English. After returning home, she realized her career prospects and professional network were stronger in Tokyo, so she moved back. Now 31, working as a freelancer, and engaged, she's torn over where to build her future. At 22, heartbroken, depressed, and unsure about my future, I craved novelty and adventure, so I packed up my life in England and moved to Japan. Now, 31, living in Tokyo, and more secure than I've ever felt in my adult life, I can't help but feel that creeping depression, pushing me to pack my bags once more. In my early 20s, upending my life felt exciting. Now, in my 30s, it just feels indecisive. In 2016, I'd graduated with a degree in fashion photography and was working three part-time jobs in my university town to scrape by while simultaneously trying (and failing) to get over intense heartbreak. I was struggling. Hobbies like theater and kung fu had lost their shine, my future felt vast and uncertain. I wanted a fresh slate. During my personally elected studies into Japanese fashion and aesthetics, I fell in love with Japan. My dissertation was titled "The rise of gender neutrality and its origins in Japanese design." I even visited a friend studying abroad there in 2015, and that brief but fantastic sojourn left me thinking — somewhat naively — "I could live here." A year later, in my depressed state, that thought resurfaced. Then it became all I could focus on. I needed to move to Japan The move wasn't completely off the cuff ー I'm not spontaneous enough for that. I applied to and was accepted into the JET Program, an organization that recruits thousands of graduates to teach English. Rather than a traditional school placement, I was based at an education center in Kanagawa, about an hour from Tokyo, with occasional assignments at local high schools. I threw myself into adapting: learning the rhythms of a new culture, working on my basic Japanese skills, and exploring my new environment. With every mountain climb, temple visit, and ramen bowl, I felt the blanket of depression start to fall from my shoulders. I put myself out there once again, starting new hobbies such as MMA, kendo, and ikebana while reviving my old passions like drama. These led to new connections and opportunities. I felt reborn. Japan had rekindled my passion for life. Feeling I'd gotten all I could from my teaching role, I decided to leave Japan with the goal of picking up where I was prior to my depressive episode. Pollacco took on new hobbies in Japan, including kendo. Provided by Laura Pollacco Life back in Europe I returned to the UK only for the pandemic to cut right across all my well-laid plans. Like most of the country, I was trapped inside, questioning my life decisions, especially about leaving Japan. I was better connected in Tokyo's creative circles than in the UK, I had support in Japan, and the cost of living was considerably lower. I decided to move back, this time not out of depression, but out of hope and ambition. In 2022, I returned on a working holiday visa, juggling remote freelance writing gigs with pitching to local publications. I pushed hard until, when my working holiday visa came to an end, I had enough work behind me to switch to the journalism visa in 2023. Despite expanding my client list and gaining experience, my original fire began to flicker, then sputter, and more recently, it's felt like I'm helplessly blowing on the embers to keep them from going out. I was burned out. Depression was setting in again. I experienced fatigue, a lack of interest in my hobbies, a desire to be left alone, all while self-flagellating my lack of ambition and for "settling" in my career. My loving fiancé — whom I met here in Japan — was starting to worry to the point where he offered to cover the cost of online therapy. During these sessions, I realized that, for the first time since moving back to Japan, I was starting to feel homesick. She met her fiancé in Japan. Provided by Laura Pollacco Living in a foreign country is tough For starters, while I speak enough to get by, not speaking fluent Japanese is exhausting. As a multifaceted freelancer, immigration's restrictive boxes feel like a choking dog collar yanking me back from new opportunities, not to mention the new gray hairs I gain with every annual visa renewal. On top of that, I've felt a rise of anti-foreigner sentiment, and Tokyo's concrete jungle is starting to feel claustrophobic and repressive. In recent months, my brain has been flooded with ideas of returning back to the pastoral days of my youth. Stone cottages with actual gardens, walks down country paths with a dog by my side, fully understanding what's being said to me at a doctor's visit. But I can't tell if I'm truly wanting to return to England or if I'm trying to escape back into a childhood where responsibilities were minimal. I've worked so hard to get to and stay in Japan, I don't know whether to push through what could simply be a low period and wait to get to the other side, or whether my gut, my instinct, is trying to tell me something. When it comes to big life choices like these, I realize I'll only find out if it was the right decision after the fact. I just hope that, whatever my partner and I choose to do, we make the best of that decision. Read the original article on Business Insider

Hana Elster vibe-coded a vintage marketplace app during winter break. Hana Elster Hana Elster, a college senior, used Claude during winter break to create a website in five days. She spent under $2,000 to get an online vintage marketplace up and running. Elster said she hopes that it'll be a successful side hustle as she moves into the corporate world. This as-told-to essay is based on a conversation with Hana Elster, a 22-year-old senior at Boston University who founded VYA, an online vintage marketplace. It has been edited for length and clarity. I'm a senior at college studying business law at Boston University, graduating in May. I've always loved the idea of being your own boss and building something for yourself. I'm also surrounded by a big founder community in Boston and New York. Seeing all these other people my age or a little bit older building things for themselves inspired me to do something for myself. Over winter break, I used Claude to build an app that now has hundreds of users. My love for vintage gave me a business idea I went thrifting for the first time when I was around 14. You'd go to the regular mall and find a shirt for $30, but I could find 10 shirts for $30 at my nearest Goodwill, and I'd be wearing something nobody else had. So it piqued my interest in secondhand shopping, getting good value, and developing a unique style. Now, I would say 80% of my closet is secondhand. Last year, I was talking with one of my friends about our favorite stores in our hometowns, and she told me about an amazing vintage store in Chicago. I'm originally from Washington, DC, and I wouldn't have known about this store unless we'd had this conversation. I realized there is space for a centralized vintage platform. I reached out to many shops to learn about their pain points and where they want to grow. The sellers mainly said they want more eyes on their products. Brick-and-mortar vintage stores told me they had websites, but most of their sales come from foot traffic. Shop owners said they put a lot of work into their website, yet they only got one sale a month on that platform. Stores that are fully online said they had to post three times a day on Instagram and TikTok because marketing was the only way to get their name out. A lot of these sellers also had full-time jobs, and vintage sales were their side hustle, so they didn't have time to do much work. Using Claude to code a website in five days Elster said it took five days to build a pilot website. Hana Elster To build the app, I started vibe coding. I have some coding experience, but Claude Code has let me move at a pace I would never have been able to. I connected with a friend who had built a fashion app years ago, and he prompted me to learn how to vibe code. In January, I started with Cursor and eventually began coding with Claude, and the project accelerated really fast. I started on January 9 and built a mock website for VYA by January 13, over winter break. It got me so excited, thinking about how I was turning lines of code into something visual, with buttons and functions and everything. Five days later, I had a website. Friends in tech also helped me look over the code and make sure everything flowed correctly. Monetization Elster's website brings together about 38 vintage shops from around the country. Hana Elster I've spent under $2,000 on app development so far, which covers all technical and operational costs. It's self-funded by my savings, and I've also received some grants from Boston University. More than three months on, we have 38 stores fully onboarded and continuing to grow, and roughly 900 approved users, of whom 50% are daily active users. My goal is that when you're checking out vintage websites like The RealReal, you also open VYA. To monetize it, I'm charging a 7% commission per item sold, with an average price of about $350. I'm also trying out a model called "source for requests," in which we charge customers a flat fee for finding a particular product, like a rare 90s Chanel bag. Usually, one of the 38 stores can fulfill the order, and if we can't, the fee is refunded. AI has both dropped and raised barriers I've always wanted to start my own business, but this was the first time I was hit with inspiration, and I could actually execute it. I don't know if I would've done it if I'd known I'd have to hire engineers and other staff, because I would've had to raise money. So AI has definitely dropped the barrier to entry. The biggest barrier used to be engineering, but now it's getting people to hear about your brand and then convert and buy, which is why my first hire is going to be a CMO. I'm thinking of how to raise money so that I can add to my head count. I'm focusing on growing and getting more people to help me out so that I can grow exponentially faster. After graduation, I'm supposed to work in consulting — I've got a return offer, and the role doesn't start until September. I'm planning to accept it and see how this business grows on the side. If I can build it up enough, grow my head count, and automate it, I would love to do it alongside my corporate job. Being a young founder has changed me. The other founders I've seen are super bold, confident, and courageous, and I feel like I've developed that side of myself, too. Read the original article on Business Insider

Architect brother duo Ansh and Raghav Kumar moved to a small Himalayan town to build a mud villa from scratch. Tiiny Farm Lab Raghav and Ansh Kumar quit city life, moved to the Himalayas, and built a house by hand. The brothers, architects by training, wanted to prove that natural materials could be durable and luxurious. The house took three years to build, and it's now a functional Airbnb. Perched on the side of a hill near the Himalayas is a whimsical house made of mud, stone, and straw that two brothers built by hand. Ansh and Raghav Kumar, a pair of architect siblings from New Delhi, moved to a remote town in India to escape the noise and chaos of corporate city life and built the house that they now list on Airbnb. "Someone sees it as a Harry Potter house, someone sees it as a hobbit house, someone sees it as a fairytale house," Raghav, 32, said. The brothers built a house in Rishikesh, India, using natural materials. Atik Bheda. "For us, it's a labor of love, and every curve tells the story of all the beautiful people who had come and put in the effort." Corporate burnout Raghav interned with a German architectural company after architecture school and transitioned into a full-time role at another German company based in Ahmedabad. He said he loved the work culture, the people, and the company's flat hierarchy — but he struggled with working long hours. "When you're in a corporate setup, the environment makes you believe that the longer you work, you can wear that as a badge of honor," he said. "There were days when I ended up working 48 hours at a stretch trying to cope with all the side deadlines." Raghav also said there was a disconnect between the architect and the laborer. "The architect sits in this air-conditioned office, designing and dreaming visionary things, whereas the laborer who builds the project doesn't get any credit," Raghav said. "As architects, we're not able to comprehend or appreciate that labor that goes into a project." During his three-year stint at the German company, he started traveling more to destress, going to remote mountainous parts of India. "I realized I was living for the weekends," he said. "That's when I realized I didn't want to climb the corporate ladder." His younger brother Ansh, 29, also interned in Germany, then freelanced on sustainable architecture projects in mountainous and rural areas of India, where he got a taste for mountain life. During the pandemic, while quarantining at home in New Delhi, the duo began experimenting with making their own building materials. They tested mycelium from homegrown mushrooms and learned about making cob, a natural material made from soil, straw, and water. That's where the idea to move away from the chaos of the city and build a house in the mountains took root. "Covid showed us life is short," Raghav said. "We agreed that we needed to take a leap of faith." Moving and adapting to Rishikesh They leased a plot of land near Rishikesh, a mountainous town northeast of New Delhi. Atik Bheda The pair decided on Rishikesh, a town located in the foothills of the Himalayas, known as India's yoga capital. It is about 150 miles northeast of New Delhi. It's historically been a pilgrimage destination, built on the banks of the Ganges River, but has recently become a hot spot for digital nomads, cafés, and corporate and wellness retreats. The pair leased land in Rishikesh from a family friend, promising to build a sustainable building on it. And in March 2021, they packed their bags and moved into a house in a small village near Rishikesh. "From where the car drops you off, you have to cross a tributary of the Ganga on foot via a wooden bridge," Ansh said, "Then you hike for more than a mile to reach the village. So that was another challenge altogether." Building a house with their hands The brothers received help from many volunteers after they put out a call on Workaway. Atik Bheda In their past architectural work, the brothers used software to do concept drawings. But here, they used sticks and stones around the plot of land to mark out the house plan. Then came the building phase. They carried thousands of rocks to the building site, trekked in sand on mules, and sourced most of the other materials from around the house. They put out a call for volunteers on Workaway — a platform that connects travelers with hosts who provide food and accommodation in exchange for work. The brothers estimated that, over the three years it took to build the house, more than 100 volunteers from 18 countries were involved in construction, along with about 20 local laborers they employed. They spent about $30,000 on the project, from their savings and their parents' investment. Seeing the house grow taller every day was rewarding, Ansh said. Atik Bheda "You can see the progress every day; the house was getting taller daily, by about six inches to a foot," Ansh said. "So that is quite rewarding compared to eight hours of work in front of the laptop in the corporate world, where you feel exhausted but also like you didn't get anything done." Finally, in 2024, they finished the house. They have listed it on Airbnb. Atik Bheda Raghav called the final project a "living sculpture." It's full of organic shapes, with a whimsical thatched roof, and spiral patterns running around its walls. They listed the two-bedroom, one-bathroom home on Airbnb in April 2024 for around $140 per night. The house is airy with earth-toned interiors. Tiny Farm Lab "Proceed to book only if you are comfortable hiking for 1.5 kilometers in a forest with a backpack, as the property is not accessible by car," says the Airbnb listing. "There's no WiFi in the forest, but you'll get a stronger connection with yourself, your loved ones, and your passions." They said it's booked out about 60% of the time. Fittest they have ever been Raghav and Ansh said they're the fittest they've ever been. Tiny Farm Lab Now, the duo has built an architecture and design studio in Rishikesh and is working on other natural building projects across India. Running the studio from Rishikesh has its challenges. They deal with a choppy network — they're eagerly awaiting Starlink — and must deal with safety concerns because the house is near a forest with wild elephants. But mountain life has changed them. "We've become the fittest version of ourselves, living here," Raghav said. On the first day of their chapter in Rishikesh, just the hike up to the site winded them both, but they can now do four trips up and down. "You don't even know when you clock 10,000 steps. You have nature all around you, better air, and you feel your anxiety melt away," Raghav said. Read the original article on Business Insider

arXiv:2605.26230v1 Announce Type: new Abstract: Multi-view 3D reconstruction has achieved remarkable progress with the advent of feed-forward 3D reconstruction models. However, these models are typically trained and evaluated under ideal, degradation-free imaging conditions, whereas real-world observations often contain degradations that differ significantly from such settings. Improving robustness for multi-view 3D reconstruction under degraded conditions therefore remains an important challenge. We present Geometry-Aware Representation Denoising (GARD), a novel framework that performs diffusion-based multi-view restoration directly in the feature space of a feed-forward 3D reconstruction model. This design exploits the geometry-aware feature representations of the 3D reconstructor to effectively recover accurate scene geometry. Furthermore, by employing an additional RGB image decoder, the refined representations can also be used to restore high-quality RGB images, thereby enabling the simultaneous recovery of 3D scene geometry and high-quality imagery. Comprehensive experiments on the Depth Anything 3 (DA3) benchmark demonstrate the effectiveness of the proposed GARD framework.

arXiv:2605.26275v1 Announce Type: new Abstract: Automatic prompt engineering (APE) rewrites prompts to improve downstream task performance, but existing APE loops treat the optimizer itself as a fixed pipeline. We port the code-as-action paradigm of CodeAct (Wang et al., 2024a) to APE and propose SPEAR (Sandboxed Prompt Engineer with Active Roll-back), a free-form agentic optimizer with four tools -- evaluate, python, set_prompt, finish -- that decides autonomously how and when to use them. The distinctive tool is the Python sandbox: the optimizer writes and executes arbitrary Python on the current evaluation DataFrame, performing structural error analysis (confusion matrices, error clustering, per group metrics) the agent itself authors. Two guardrails turn the long-horizon agent into a monotone-improving optimizer: auto-rollback on metric regression, and an optional guard metric floor. We evaluate on three industrial LLM-as-judge suites (13 judge tasks across recruiter-intake, conversational-memory, and query-refinement systems) plus seven BBH tasks and GSM8K. SPEAR wins every industrial task on the primary metric ($\kappa$ 0.857 vs 0.359 on tool-selection; F1-macro 0.815 vs 0.763 on filter-relevance; $\kappa$ 0.254 vs 0.218 on the hardest extraction dimension). On BBH-7 SPEAR averages 0.938 accuracy vs GEPA 0.628 and TextGrad 0.484. Ablations show the Python tool is the largest single lever on complex judge tasks ($\Delta \approx +0.79\kappa$ on the 5-class tool-selection judge, $\Delta \approx +0.35\kappa$ on the hardest extraction dimension when removed); its irreplaceable contribution is class-pair confusion aggregation that a long-context LLM cannot extract reliably from the raw eval DataFrame.

arXiv:2605.26128v1 Announce Type: new Abstract: Production LLM systems increasingly require machine-readable outputs: JSON objects, typed traces, regex-constrained fields, and tool-call schemas. This paper targets on-device and low-cost small language model (SLM) deployments, where sub-3B models are attractive for privacy, latency, and commodity hardware but have limited capacity to satisfy schemas while solving tasks. The usual engineering assumption is that hard output constraints improve reliability without changing the underlying answer. We show that this assumption is unsafe for small models. We introduce \emph{constraint tax}, a measurement protocol for isolating the answer and executable-accuracy loss caused by structured-output constraints at fixed model, fixed task distribution, and fixed problem instances. Across 15,000 commodity-GPU generations with Qwen2.5-0.5B, Qwen2.5-1.5B, and SmolLM2-1.7B, hard answer-only schema decoding raises schema validity from 61.5\% to 100.0\%, but lowers answer accuracy from 19.7\% to 11.0\% and increases wrong-valid-schema outputs from 49.5\% to 88.9\%. The strongest industry analogue is a deterministic calendar tool-call task: Qwen2.5-1.5B achieves 91.5\% executable accuracy with prompt-only JSON but only 48.0\% under the same hard tool-call schema, while both modes are 100.0\% schema-valid. The error is semantic, not structural. We also show that the 3B boundary still pays a direct-schema tax and that delayed packaging supports a constructive design pattern: reason free, constrain late. The practical conclusion is direct: production systems should report schema validity, answer accuracy, executable accuracy, and wrong-valid-schema rate separately.

arXiv:2605.26135v1 Announce Type: new Abstract: Unsupervised anomaly detection is widely used in transaction fraud detection where labels are scarce. Isolation Forest (IF) is among the most popular classical methods due to its scalability and ease of deployment. We propose SilIF, an augmentation of Isolation Forest that adds a silhouette-based scoring layer computed in a representation space induced by the trees of the forest. For each point, we extract a vector of per-tree path lengths, cluster these "fingerprints" into structural groups, and compute a silhouette score that measures how well the point fits its assigned group versus the nearest alternative. The silhouette signal is combined with the base IF score via a single hyperparameter alpha. On the IEEE-CIS Fraud Detection benchmark (~590K transactions, 3.5% fraud), SilIF with alpha=1.0 improves over plain Isolation Forest by +0.0080 AUC-PR on average across five seeds, with SilIF winning on all five seeds (paired t-test p=0.046). We also report results on a synthetic credit-card dataset (Sparkov) where the silhouette augmentation does not improve over plain IF, and we characterize the conditions that distinguish the two outcomes. The paper presents SilIF as a tunable, easy-to-deploy enhancement to Isolation Forest with honest reporting of when it helps and when it does not. Code at https://github.com/venkat15vk/silif-anomaly-detection.

arXiv:2605.26147v1 Announce Type: new Abstract: Human decision-making is sequential and uncertainty-aware, yet standard neural networks often rely on static, dense forward computation with limited visibility into evidence acquisition, uncertainty evolution, or when computation should stop. We introduce \textbf{Neural Bayesian Sequential Routing (NBSR)}, a framework that models neural inference as active evidence accumulation over a hierarchical Directed Acyclic Graph (DAG). Within a Dirichlet--Categorical conjugate framework, neural experts query a persistent global knowledge oracle to extract positive evidence vectors, which act as pseudo-counts and update a Dirichlet belief state by exact conjugate addition. Coupled with a Gumbel-Softmax Straight-Through estimator, this update enables hard, path-dependent routing while preserving surrogate gradients for end-to-end training. The resulting Dirichlet precision and entropy provide mechanisms for uncertainty quantification, entropy-based early exiting, OOD abstention, and cost-aware evidence acquisition. We prove that, under strictly positive evidence extraction, total Dirichlet precision increases monotonically along any valid trajectory and marginal predictive variance is bounded, formalizing sequential ``hypothesis sharpening''; under idealized capacity and optimization assumptions, the terminal Dirichlet expectation recovers the Bayes-optimal conditional distribution. Empirical evaluations across visual categorization, structured medical diagnosis, language modeling, partially observable control, and cost-aware Bayesian experimental design show that NBSR achieves competitive predictive performance while providing transparent routing traces, path-dependent evidence attribution, uncertainty-aware decision control, and resource-rational inference. Overall, NBSR offers a mathematically grounded framework for interpretable, modular, and resource-rational agentic AI.

arXiv:2605.26161v1 Announce Type: new Abstract: Time series foundation models (TSFMs) are increasingly pretrained on large corpora, raising concerns that evaluation datasets may have been exposed during pretraining and thus yield overly optimistic performance estimates. Auditing such contamination is challenging in time series because signals are continuous and heterogeneous, and often lack corpus documentation. To the best of our knowledge, this is the first work to study pretraining contamination auditing for TSFMs. We formalize the problem of pretraining contamination auditing for TSFMs and propose TSFMAudit, a method based on probe adaptation dynamics. Our key intuition is that contamination manifests as unusually efficient adaptation: after a fine tuning probe, contaminated datasets tend to exhibit faster loss reduction with smaller backbone movement. We evaluate TSFMAudit on 6 TSFMs and 187 datasets using documented training source evidence as supervision, and compare against 10 competitive baselines adapted from the LLM literature.