Discussion | Link

Polaroid's new billboard on Coney Island Beach in Brooklyn. Polaroid Polaroid has put up an anti-AI data center billboard at New York's Coney Island Beach. "Go jump in some water before the data centers drink it all up," says the ad, which was also posted to Polaroid's Instagram. Polaroid's message touches on a growing debate over the environmental impact of the AI boom. Polaroid is taking another shot at AI — this time by targeting the massive data centers powering the technology. The 89-year-old camera company recently unveiled a billboard at Brooklyn, New York's iconic Coney Island Beach that reads: "Go jump in some water before the data centers drink it all up." Polaroid is one of a growing crop of brands leaning into anti-AI marketing. The company's latest ad is part of its global marketing campaign titled "the best of summer is analog," tied to the launch of its new Go Generation 3 camera. In an Instagram post on Monday, the company echoed the billboard's message and wrote in a caption: "There'll come a day when the things we took for granted can never be taken again. Obviously, we're being a bit cheeky with our line here, but still." "Go take a swim. Go take a stroll. Go embrace the beautiful, simple, wild, analog stuff. That's where the best of summer, and life, lives," Polaroid wrote. View this post on Instagram A post shared by Polaroid (@polaroid) "While our campaigns are provocative and challenge our relationship with technology, we're not anti-digital," Polaroid's creative director, Patricia Varella, said in a release. "We know we have to live alongside it, but we're deeply pro-human, and know what humanity gives us. And we know what we stand to lose if we don't protect it. That's a fight worth fighting. Polaroid's message touches on a growing debate over the environmental impact of the AI boom. Data centers can use vast amounts of water both directly for cooling servers and indirectly through the electricity needed to power them. A previous Business Insider investigation found that some of the largest data center facilities in the US were permitted to use more water each day than nearly 49,000 Americans would typically use. Some tech leaders have pushed back on criticism surrounding data centers' water use, arguing that advances in cooling technologies — including closed-loop systems — are making AI infrastructure more efficient. Earlier this year, OpenAI CEO Sam Altman disputed what he called "totally insane" claims circulating online that the AI firm is consuming large amounts of water to power ChatGPT. "Water is totally fake," Altman said, when asked about concerns over AI companies' water use. "It used to be true, we used to do evaporative cooling in data centers, but now that we don't do that, you know, you see these like things on the internet where, 'Don't use ChatGPT, it's 17 gallons of water for each query' or whatever." AI chipmaker Nvidia this week announced a "liquid cooling" system designed to help data centers run hotter while using less water. "We have eliminated massive amounts of power usage and pretty much all water usage," Ali Heydari, Nvidia's director of data center cooling and infrastructure, said in a statement. Read the original article on Business Insider

CMOs are increasingly in the "driver's seat" of artificial intelligence strategy says BCG's Mark Abraham, global leader of marketing and sales, with oversight over budgets and decision making. Read the original article on Business Insider

Gleb Mezhanskiy spent years building tools to make enterprise data clean. In March 2026, the Datafold … AI Agents Need Context, Not Perfect Data: The Case for Risk-Based Data QualityRead more The post AI Agents Need Context, Not Perfect Data: The Case for Risk-Based Data Quality appeared first on .

Reuters: SK Hynix overtaking Samsung to become South Korea's most valuable company is the culmination of 14 years of bets on HBM, which paid off as the AI boom took off  —  SK Hynix's (000660.KS) overtaking of Samsung Electronics (005930.KS) to become South Korea's most valuable firm was the culmination …

TechCrunch's StrictlyVC evening in Los Angeles late last week brought together two of the more straight-talking investors working in AI right now. They were as entertaining as they were illuminating.

OpenAI and Broadcom introduce Jalapeño, a custom AI chip built for LLM inference to improve performance, efficiency, and scale across AI systems.

Discussion | Link

Discussion | Link

Anthropic's Claude Code has become one of the most popular AI coding tools. Bloomberg/Getty Images AI startup Legion sued the government over losing access to Anthropic's top AI models. Legion said the directive harmed its business. The lawsuit adds another layer to Anthropic's clash with the Trump administration. One AI startup just sued the US government for cutting its access to Anthropic's top models. On Tuesday, legal tech company Legion filed a suit in Washington, D.C. over a government order requiring Anthropic to keep its models Fable 5 and Mythos 5 away from foreign nationals. Earlier this month, Anthropic said it would disable access to both models following a letter from the US government asking it to bar foreign individuals and entities from using the products. The AI lab said this rule would bar some Anthropic employees from using the tools. In order to comply, the company initially disabled access for everyone. Last week, Anthropic restored access to Claude Fable 5 with nationality-based access controls and enhanced onboarding compliance screening. In Tuesday's complaint, Legion said that it was a US-based company that employs Canadian nationals who work remotely from Canada. It added that its litigation technology tools are built on top of frontier lab models, such as those from Anthropic. "Legion is a commercial customer of Anthropic with a contractual right and license to access and use the Fable 5 model, which was integral to building and operating its platform," the company wrote in the filing. "When the directive took effect, Legion lost the latest tool at the center of its development instantaneously." The company added that the order caused it "immediate, irreparable, and existential" harm against competitors because the pace of AI development is so rapid. Legion, founded in 2024 and based in San Jose, develops AI-powered litigation drafting software that helps lawyers automate pleadings, discovery requests, and other court documents. The Department of Commerce didn't immediately respond to Business Insider's request for comment. Legion's lawsuit adds another set of stakeholders to Anthropic's clash with the Trump administration over AI safety and government control over frontier AI models. It may be the first among suits from Anthropic customers who have been complaining about losing access to its most powerful model. When Anthropic first paused access to the model, it said that the government was concerned that someone might try a jailbreak of Fable 5. The company argued that cutting access for a "narrow potential jailbreak" was overkill and would halt all new model deployments. Read the original article on Business Insider

Slash's employee AI coded a shooter game and accidentally billed the company $80,000 for it. Screenshot of shooter game/Business Insider Fintech firm Slash got a rude shock when it saw one of its employees' AI bills. The employee spent over $80,000 AI coding a brainrot shooter game. Slash is encouraging people to play the game so it can write it off as a "marketing expense." Fintech company Slash is touting a new product that doesn't quite fit in line with its usual offerings — an AI-coded shooter game. In an X post on Tuesday, the San Francisco-based company wrote that one of its employees racked up a hefty AI coding bill building a video game. "We encouraged the company last week to start vibe coding more but @nickbruhman burned $80k in credits on the Slash card for a brainrot shooter," the company wrote. "Pls play it so we can write this off as a marketing expense," it added. Literally called "Brainrot shooter," it is a bare-bones game set in a Minecraft-esque landscape, where the player has to shoot characters with internet-viral brainrot names like "skibidi toilet" and "tung tung tung sahur." Prediction market Polymarket picked up the news, writing in an X post: "Fintech startup Slash reveals it was forced to roll back its AI coding push after an employee burned $80,000 in tokens in the first week." The employee, Slash's head of strategic verticals, Nicolas Brilliante, posted a screenshot that appeared to be a dashboard of his AI usage, showing he had spent $81,267. "This was a genuine accident, i underestimated my own ability," Brilliante wrote. Later, reposting a Polymarket X post, he wrote: "This is actually insane, am I going to become a case study for how AI spend can get out of control." Brilliante isn't the only employee who's been building things fast and furiously in the era of tokenmaxxing. Still, many companies are now looking at their AI spending with a more critical lens. Some firms have found that AI costs did not translate into productivity gains, and have cut back on AI budgets. Companies like Uber, Coinbase, and Walmart have now set caps on how much their employees can spend on AI, with Walmart specifically saying this was done to clamp down on unnecessary, repetitive vibe coding. Read the original article on Business Insider

Royal Marines fire small arms during an exercise in Norway. POPhot Alex Ceolin/UK Defense Ministry The UK is using an AI-powered system called ASGARD to find targets and make battlefield decisions. The head of its land forces said it's cut HQ planning time from 72 hours to just 1 hour. Gen. Roly Walker said it's also letting a UK corps hit 10 times as many targets in a day. The British Army's new AI-powered system has shortened war planning cycles at the corps level from 72 hours to just one hour, its Chief of the General Staff, Gen. Roly Walker, said on Tuesday. "Now, I said last year that ASGARD would help us sense twice as far, decide twice as quick, and strike twice as deep. And I was wrong," said Walker at the Royal United Services Institute's Land Warfare Conference in London. "A corps planning cycle that once took 72 hours can now take one," he added. "What they're going to do with the other 71 hours, I do not know." Dubbed a "digital targeting web" by the UK, ASGARD is designed to gather and process battlefield data to help commanders find targets, make decisions, and coordinate attacks. It's one of several efforts by militaries to use AI to manage war zones, such as Palantir's Maven system for the Pentagon and Ukraine's Delta platform. The UK announced last year that it plans to spend 1 billion pounds, or about $1.3 billion, on developing such systems for its own military. Walker said ASGARD would allow a corps to attack 10 times as many targets in a single day. "In fact, they're limited only by the munitions that are available to fire into the sky," he said. "ASGARD is literally a digital juggernaut that is evolving every 8 to 12 weeks." In the British Army, a corps is at the top of the battlefield command structure and typically oversees logistics, intelligence, strategic strikes, and the deployment of tens of thousands of troops. That means ASGARD is a headquarters-level platform, used by top commanders who can be away from the actual battlefield. Last month, the UK conducted an exercise that deployed the system from a tube station under Trafalgar Square, which Walker said on Tuesday was managing troops in Estonia. The makeshift headquarters in the train station was processing 10 terabytes of data, or "nearly three months of non-stop high-definition Netflix," a day, the British Army said. The UK conducted a weekslong exercise in Charing Cross station that used the ASGARD system. WO2 Jon Bevan Chief/UK Defense Ministry "If I ever think this was a moment where we had a conversation about what is our century's 'tank versus horse' moment, I feel it's about this area," Walker said. The British Army head also said the UK is sending thousands of drones to its units and introducing 50 operational-level electronic warfare systems that were used in Ukraine. To prepare against Russia, a British corps needs to be "capable of doing what a Ukrainian corps can do today," he added. "So in the next year, I expect to see much greater numbers of our remote and autonomous systems forward on our eastern flank, ready to strike and act within 30 minutes," Walker said. Read the original article on Business Insider

Tan France says he's been documenting his kids' childhoods in emails. Dave Benett/Getty Images Tan France says he writes emails to his kids so they can revisit their childhoods one day. The "Queer Eye" star said he also uses the messages to document what life was like for him and his husband at the time. "I find it fascinating to know that my kids might see us as just people also," France said. Tan France says he writes messages to his kids that they won't read until they're older. On Tuesday's episode of the "Reclaiming with Monica Lewinsky" podcast, the "Queer Eye" star said he's been documenting his kids' childhoods through email. "We created email accounts for my kids," France told podcast host Monica Lewinsky. Every six months or so, France and his husband, Rob France, will each send their children emails "about who they are, and what they're about, and what they enjoy, and who we think they might be," he said. The couple, who married in 2007, welcomed two children via surrogacy in 2021 and 2023. France said the email tradition is his "favorite thing" to do several times each year, and he loves the idea of his kids one day looking back on the messages as adults. Not only that, he hopes the messages will offer his children a perspective he never had growing up. "I would love to know if my parents struggled as much as I'm struggling being a parent, and how hard it is, because we only just see our parents as, well, you're my parent," France continued. "I find it fascinating to know that my kids might see us as just people also." France said he got the idea from Kim Kardashian. In a 2023 appearance on the "On Purpose with Jay Shetty" podcast, the Skims founder said she writes letters to her kids on their birthdays about the year they've had. "I know that they'll appreciate this," Kardashian said. Other celebrity parents have also spoken about preserving memories for their kids this way. Meghan Markle said in a 2025 appearance on "The Jamie Kern Lima Show" that she emails her kids almost every night before she goes to bed. "The things that you're not going to frame, the things that you're not going to put pen to paper in a journal, but they will end up wanting at one point in their life, maybe when they're 16 or when they're 18," Markle said. Read the original article on Business Insider

When Jareb Dougherty realized he didn't have a career in basketball, he moved home to run a restaurant with his mom. Provided by Jareb Dougherty Jareb Dougherty was on track to become a semi-pro basketball player. When it didn't work out, he moved in with his mom to help out at the Irish pub in Florida, where she worked. Now, they're co-owners, and he says he loves working together. This as-told-to essay is based on a conversation with Jareb Dougherty, 29, the co-owner of Donovan's Irish Pub in St. Augustine, Florida. When his professional basketball career didn't work out, he went on to manage a restaurant with his mother. It has been edited for length and clarity. When I say I was "this close" to going pro, I actually was. I started playing ball in 6th grade; I was the tallest one in the school. Now, I'm six foot four, and basketball has been the love of my life. I went to a junior college in Kansas for a year, and then in South Florida, but I've never been a star student. I struggle with multiple learning disabilities and test-taking was always a challenge. After a few years, I stepped away from college. After playing for a small semi-pro team, he gave up his basketball dream. Provided by Jareb Dougherty I moved to St. Augustine and played for a small semi-pro team for two years. When I got invited to a training camp with a minor professional basketball team in Sarasota, the Gulf Coast Lions, it felt like a dream come true. It would have been my first full-time team, where I could quit other side jobs like bartending and focus on basketball. Every week during the monthlong training camp, they made a cut. I made it to the last week and was planning on not coming home. During the three-day break, I got a text that I was cut. I drove home with my basketball career behind me and no idea what was ahead. Dougherty with his grandmother when he was still playing basketball. Provided by Jareb Dougherty, The family business back home While I was busy chasing my basketball dreams, my mother, Kristin Dougherty, landed a bartending job at Donovan's Irish Pub in St. Augustine after responding to a Craigslist ad. In 2016, when I was still in college, my mom arrived at work and found an eviction notice on the door. She later heard that the owners were behind on rent and taxes. As a single mom — my dad died when I was a toddler — she'd always worked in the food industry to support us. She took this as her moment to level up. She figured out a way to buy the restaurant. She was following my grandfather's footsteps, who was always putting a bug in her ear to buy the business wherever she worked. Her brother helped her take over the business. When my basketball career suddenly ended, I came back to the restaurant to bartend and save money until I figured out my life. More curveballs were on the way. Stepping up to my future career Despite having never pictured a career in the restaurant business, when my mother and her brother had a falling out, I knew she needed help. In 2023, at 26, I decided to buy my uncle's shares in the restaurant. I had saved up a good amount from part-time jobs and took a loan from another family member. I had never run a business, but I knew my mom needed help, and I jumped in. Despite being mostly equal partners, I've told her that she always has the final say. Co-owning a pub with my mother While some might say don't mix business and family, I can't see it any other way. My mom is my best friend, and we even lived together until recently. I'm 29 and decided to invest in a place for myself. Living and working together was easy for us. Learning to run a business was not. For six months, neither of us took a paycheck. We had to replace equipment while money was tight. When we finally got our first check, we could breathe for a moment — I realized this could actually work. Dougherty now runs the restaurant with his mom; his grandma prepares homemade desserts. Provided by Jareb Dougherty I had to get used to managing people, including their emotions, time, and schedules. With 28 employees, there are lots of personalities to handle. During busy hours, cooks can get grumpy and servers yell for orders. But that's restaurant life. An essential change from when my mom and her brother ran the business was who was responsible for which parts of the restaurant. With them, he ran the back of the house, and she ran the front of the house. This didn't work because they both had to be there all the time. Now, my mom and I both run everything, so we can give each other breaks. She's a workaholic with a crazy work ethic, so I have to make sure she takes days off. Building toward a bigger future Life is enjoyable these day. My basketball career is behind me, except for occasionally shooting hoops with friends. Sometimes that devastating text — and what could have been — still plays in my mind. I think that's true for most retired athletes. But what's in front of me is exciting, and I'm always on call, ready for the next phone call from someone who needs me at the restaurant. My grandma is a prep cook there and makes homemade desserts. I actually enjoy running a business with my family. We discuss every decision together and consider the restaurant our home away from home. We recently launched a traditional Irish brunch featuring rashers, a type of Irish bacon, and blood sausage, although fish and chips remains our most popular dish. I live a life I never envisioned for myself, but it's also one I wouldn't change. Read the original article on Business Insider

arXiv:2606.23700v1 Announce Type: new Abstract: Emergent misalignment (EM) has been linked to the activation of misaligned persona vectors and evil character traits, suggesting that EM operates through disruption of the model's aligned character rather than direct learning of harmful content. Motivated by this connection, we study self-generated text recognition (SGTR) finetuning as a character-targeted intervention that is distinct from existing in-training defenses. We conduct two-stage finetuning experiments across three models (GPT-4.1, Qwen2.5-32B-Instruct, Seed-OSS-36B-Instruct) and multiple EM datasets to compare SGTR finetuning against benign finetuning baselines (correct domain-specific data, general knowledge, and word counting) to find it an effective defense in both reversal and prevention settings. We find that all interventions produce comparable EM reversal, but only when restoring capabilities that EM had degraded. For prevention, only SGTR finetuning consistently reduces misalignment without exacerbating any individual metric, suggesting that character fortification specifically drives prevention. We provide further evidence for EM's relation to the LLM's default character by showing that EM finetuning induces diversity into the LLM's identity self-reports, artificially corrupting self-recognition exacerbates misalignment caused by EM finetuning, and that removing the model's identity-bearing system prompt substantially reduces the effect of EM finetuning. Together, these findings reframe EM not as the adoption of a coherent misaligned persona but as the destabilization of aligned character.

arXiv:2606.23892v1 Announce Type: new Abstract: Vision-language models (VLMs) are increasingly used as perception-reasoning backbones for embodied intelligence in safety-critical physical systems, where perception or reasoning errors can lead to unsafe decisions or actions. Although many red-teaming methods have been developed to probe VLM vulnerabilities, their evaluation remains fragmented across datasets, metrics, and threat models, making direct comparison difficult and obscuring whether observed differences arise from stronger attacks, more vulnerable models, or incompatible evaluation settings. Existing chatbot-centric red-teaming benchmarks mainly standardize jailbreak and content-safety evaluation, but they do not systematically capture physically grounded functional failures or cover red-teaming methods that target physical-world VLMs. This raises the key challenge of comparing diverse attack methods under a unified protocol while targeting the same scenario-specific failures. We introduce REALM, to our knowledge the first unified red-teaming benchmark for physical-world VLMs. REALM integrates 12 red-teaming methods, 3 model-agnostic defenses, and 13 VLMs under a practical black-box threat model with shared datasets and metrics. To align adversarial objectives across attack families, REALM introduces an agentic target-generation pipeline that constructs shared, scenario-specific, and physically grounded attack objectives for each scene, enabling fair comparison of diverse red-teaming methods under aligned adversarial goals. Our evaluation shows that text and typographic injection attacks induce the most failures, multimodal co-optimization yields the strongest visual-perturbation transfer, single-pass attacks approach iterative methods at much lower cost, and model scale alone does not confer adversarial robustness. Code is available at https://github.com/UCF-ML-Research/REALM.

arXiv:2606.23743v1 Announce Type: new Abstract: Modern video diffusion models achieve higher generation quality through scaling, but this also increases inference cost. Although many acceleration methods have been proposed, a central challenge is that the most effective acceleration strategy is highly instance-specific: a recipe that works well for one combination of model, hardware, and inference configuration often does not transfer to another. Different models vary in architecture, numerical sensitivity, and attention concentration patterns. Inference settings differ in spatial and temporal resolution and video duration, while hardware platforms differ in memory hierarchy, supported numerical formats, and kernel throughput. These factors create a large tuning space, making manual performance engineering costly. We present Sol Video Inference Engine, an agentic, native, training-free acceleration framework for video diffusion models. It organizes five broadly applicable techniques, cache, sparse attention, token pruning, quantization, and kernel fusion, into an agentic acceleration stack for instance-specific optimization. For a concrete deployment target defined by a model, hardware platform, and serving configuration, parallel skill agents optimize the implementation of each technique, an agent integrator composes them into a global acceleration stack, and a human validator provides feedback on generation quality. We instantiate this workflow on three video models with different sizes and architectures: 64B Cosmos3-Super, 22B LTX-2.3, and 2B SANA-Video. With little human effort, the full stack achieves more than 2x end-to-end acceleration while maintaining near-lossless VBench quality, demonstrating the effectiveness of the agent framework for video diffusion acceleration.

arXiv:2606.23699v1 Announce Type: new Abstract: Instrumented bicycle studies have produced direct field evidence on vehicle passing behavior, but extracting overtaking events from continuous rear-facing video has remained dependent on manual, frame-by-frame annotation. This bottleneck constrains sample sizes and limits naturalistic cycling safety research. We present a geometry-informed computer vision pipeline that automates overtaking event detection from a single bicycle-mounted camera without multi-sensor configurations or explicit camera calibration. The system combines RT-DETR object detection with ByteTrack multi-object tracking through a three-stage geometric validation module enforcing bearing angle trend, apparent size growth, and spatial confirmation criteria derived from perspective projection principles. Validated on 315 manually annotated real-world overtaking events from urban roads in Ann Arbor, Michigan, the pipeline achieved 97.8% recall with zero false positives. The system identified overtaking intentions a mean of 2.44 seconds before vehicle passage, with 84.1% of events exceeding the 1.5-second human reaction time threshold, demonstrating feasibility for active cyclist warning. Lateral passing distance measurements from 96 events revealed 33.3% of passes below the 5-foot (152.4 cm) threshold, consistent with non-compliance rates in prior field and self-reported studies. A preliminary calibration-free lateral distance estimation approach using bounding box geometric features achieved mean absolute errors of 13-14 cm under leave-one-out cross-validation, sufficient to distinguish close passes from standard passes for safety categorization. By automating event isolation from consumer-grade footage, the system removes the primary annotation bottleneck of instrumented bicycle research and provides a scalable foundation for vehicle-bicycle interaction analysis across larger datasets and diverse urban environments.

arXiv:2606.23767v1 Announce Type: new Abstract: Headline accuracies on the Tuebingen cause-effect pairs are routinely compared across papers even though each is measured under its authors' own protocol -- different pair subsets, weightings, model-selection, and decision rates. We argue this is the wrong comparison and run the right one: a same-hands re-evaluation in which every method is run by us on the identical 102 pairs, with one strict rule -- no tuning and a decision forced on every pair. As a clean reference point we introduce a deliberately minimal baseline: sorted-conditional compression, which feeds quantized, sorted, first-differenced data to an off-the-shelf compressor (bz2) and has zero fitted parameters. Under the common ruler the ranking differs sharply from the literature. Our baseline reaches 74.7% weighted accuracy (p = 3.7e-7); on the same 100 pairs that SLOPE is evaluated on it scores 76.0%, a 1.2-point gap below the authors' own forced-decision SLOPE (77.2%) that is well inside noise (McNemar p = 0.39). A faithful re-run of RECI lands at 70.7% -- inside the original authors' reported error bar, not the 77.5% often quoted (which we trace to a mis-copied cell). SLOPE's published 82.4% is a decided-subset figure: scoring the authors' own stored output only on the pairs its significance test chose to answer reproduces 81.7%. Under the common ruler the methods cluster in the low-to-mid 70s and the zero-parameter compressor ties the strongest of them. We document the mechanisms that inflate published figures (test-set model selection, significance-gated abstention) and contribute two further results: compression score magnitude is a model-free confounding flag (p = 2.8e-68), and a pre-registered falsification test fails in an instructive way that bounds the method's theoretical interpretation. Code, pre-registrations, and per-pair outputs are released.

arXiv:2606.23701v1 Announce Type: new Abstract: Qualitative product feedback can reveal nuanced user experiences, but its implicit sentiment is difficult to measure. This paper presents a scalable and interpretable framework that uses large language models (LLMs) to quantify product desirability from such data. Using two Product Desirability Toolkit (PDT) datasets from ZORQ and CARMA comprising 106 respondent term groupings with gold-standard human annotation, zero-shot continuous numerical sentiment scoring and categorical sentiment classification are evaluated without relying on explicit review scores. Across the datasets, LLMs generated numerical sentiment scores directly from qualitative responses and closely matched expert labels, achieving Pearson correlations up to 0.97 and classification accuracy up to 94%. LLMs maintained robustness even when handling data presented in multiple forms and consistently expressed high confidence. In contrast, lexicon-based and transformer baselines did not produce statistically significant results. Among the models tested, GPT-4o-mini achieved performance comparable to larger models at 94% lower cost, supporting scalable deployment. The framework also incorporates model confidence ratings and human-readable rationale explanations (xAI), improving interpretability, transparency, and trust while supporting practical use in product satisfaction assessment. In general, using the PDT tool as a survey method along with a cost efficient LLM for sentiment analysis has the potential to provide for product evaluation with results that are rich in terms of sentiment scores (both numerical and classified sentiment) and in terms of the high-level user impressions of the product that can be used to identify ideas for product development and improvement, as well as marketing ideas for target audiences.