arXiv:2604.15392v1 Announce Type: new Abstract: Physics-Informed Neural Networks (PINNs) often suffer from slow convergence, training instability, and reduced accuracy on challenging partial differential equations due to the anisotropic and rapidly varying geometry of their loss landscapes. We propose a lightweight curvature-aware optimization framework that augments existing first-order optimizers with an adaptive predictive correction based on secant information. Consecutive gradient differences are used as a cheap proxy for local geometric change, together with a step-normalized secant curvature indicator to control the correction strength. The framework is plug-and-play, computationally efficient, and broadly compatible with existing optimizers, without explicitly forming second-order matrices. Experiments on diverse PDE benchmarks show consistent improvements in convergence speed, training stability, and solution accuracy over standard optimizers and strong baselines, including on the high-dimensional heat equation, Gray--Scott system, Belousov--Zhabotinsky system, and 2D Kuramoto--Sivashinsky system.

arXiv:2604.15371v1 Announce Type: new Abstract: Large language models (LLMs) achieve strong performance across many natural language processing tasks, yet their decision processes remain difficult to interpret. This lack of transparency creates challenges for trust, debugging, and deployment in real-world systems. This paper presents an applied comparative study of three explainability techniques: Integrated Gradients, Attention Rollout, and SHAP, on a fine-tuned DistilBERT model for SST-2 sentiment classification. Rather than proposing new methods, the focus is on evaluating the practical behavior of existing approaches under a consistent and reproducible setup. The results show that gradient-based attribution provides more stable and intuitive explanations, while attention-based methods are computationally efficient but less aligned with prediction-relevant features. Model-agnostic approaches offer flexibility but introduce higher computational cost and variability. This work highlights key trade-offs between explainability methods and emphasizes their role as diagnostic tools rather than definitive explanations. The findings provide practical insights for researchers and engineers working with transformer-based NLP systems. This is a preprint and has not undergone peer review.

arXiv:2604.15574v1 Announce Type: new Abstract: Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t. knowledge acquired during pre-training. In this work, we explore whether SFT-induced hallucinations can be mitigated using established tools from the continual learning literature, since they arise as a by-product of knowledge degradation during training. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. We also show that, in settings where new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups, can preserve task performance while reducing hallucinations. Lastly, we investigate the mechanism behind SFT-induced hallucinations through three hypotheses: capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, and that self-distillation succeeds by mitigating this interference.

arXiv:2604.15505v1 Announce Type: new Abstract: LLM agents operating under organizational policies must comply with authorization constraints typically specified in natural language. In practice, such specifications inevitably contain ambiguities and logical or semantic gaps that cause the agent's behavior to systematically diverge from the true requirements. We ask: by letting an agent evolve its policy understanding through interaction and corrective feedback from pre-deployment testing, can it autonomously refine its interpretation to close specification gaps? We propose PolicyBank, a memory mechanism that maintains structured, tool-level policy insights and iteratively refines them -- unlike existing memory mechanisms that treat the policy as immutable ground truth, reinforcing "compliant but wrong" behaviors. We also contribute a systematic testbed by extending a popular tool-calling benchmark with controlled policy gaps that isolate alignment failures from execution failures. While existing memory mechanisms achieve near-zero success on policy-gap scenarios, PolicyBank closes up to 82% of the gap toward a human oracle.

arXiv:2604.15503v1 Announce Type: new Abstract: Recent breakthroughs in language models (LMs) using neural networks have raised the question: how similar are these models' processing to human language processing? Results using a framework called Brain Score (BS) -- predicting fMRI activations during reading from LM activations -- have been used to argue for a high degree of similarity. To understand this similarity, we conduct experiments by training LMs on various types of input data and evaluate them on BS. We find that models trained on various natural languages from many different language families have very similar BS performance. LMs trained on other structured data -- the human genome, Python, and pure hierarchical structure (nested parentheses) -- also perform reasonably well and close to natural languages in some cases. These findings suggest that BS can highlight language models' ability to extract common structure across natural languages, but that the metric may not be sensitive enough to allow us to infer human-like processing from a high BS score alone.

arXiv:2604.15356v1 Announce Type: new Abstract: Recent work on KV cache quantization, culminating in TurboQuant, has approached the Shannon entropy limit for per-vector compression of transformer key-value caches. We observe that this limit applies to a strictly weaker problem than the one that actually matters: compressing the KV cache as a sequence. The tokens stored in a KV cache are not arbitrary floating-point data -- they are samples from the exact formal language the model was trained on, and the model is by construction a near-optimal predictor of that language. We introduce sequential KV compression, a two-layer architecture that exploits this structure. The first layer, probabilistic prefix deduplication, identifies semantically equivalent shared prefixes across sessions using the trie metric d_T(s, s') = -log_2 P_M(s ^ s') from Probabilistic Language Tries (PLTs). The second layer, predictive delta coding, stores only the residual of each new KV vector from the model's own prediction of it, achieving a per-token entropy bound of H(KV_{i+1} | KV_{<=i}) <= H(token_{i+1} | token_{<=i}). We prove that at typical language model perplexity -- approximately 10-20 for fluent English text -- this bound is 3.3-4.3 bits on average per token position, compared to TurboQuant's 3 bits per vector component (with typical attention heads having 64-128 components). The theoretical compression ratio over TurboQuant is approximately 914,000x at the Shannon limit. Even at 1000x above the entropy floor -- a deliberately pessimistic worst-case overhead, two orders of magnitude above the 2-5x typical of practical source coders -- the ratio remains approximately 914x over TurboQuant, with compression improving rather than degrading as context length grows. The two layers are orthogonal and compose with existing per-vector quantization methods including TurboQuant.

arXiv:2604.15453v1 Announce Type: new Abstract: Tokenization is a key component of autoregressive (AR) generative models, converting raw data into more manageable units for modeling. Commonly, tokens describe local information, such as regions of pixels in images or word pieces in text, and AR generation predicts these tokens in a fixed order. A worthwhile question is whether token structures affect the ability to steer the generation through test-time search, where multiple candidate generations are explored and evaluated by a verifier. Using image generation as our testbed, we hypothesize that recent 1D ordered tokenizers with coarse-to-fine structure can be more amenable to search than classical 2D grid structures. This is rooted in the fact that the intermediate states in coarse-to-fine sequences carry semantic meaning that verifiers can reliably evaluate, enabling effective steering during generation. Through controlled experiments, we find that AR models trained on coarse-to-fine ordered tokens exhibit improved test-time scaling behavior compared to grid-based counterparts. Moreover, we demonstrate that, thanks to the ordered structure, pure test-time search over token sequences (i.e., without training an AR model) can perform training-free text-to-image generation when guided by an image-text verifier. Beyond this, we systematically study how classical search algorithms (best-of-N, beam search, lookahead search) interact with different token structures, as well as the role of different verifiers and AR priors. Our results highlight the impact of token structure on inference-time scalability and provide practical guidance for test-time scaling in AR models.

arXiv:2604.15451v1 Announce Type: new Abstract: Large-scale visual learning is increasingly limited by training cost. Existing knowledge distillation methods transfer from a stronger teacher to a weaker student for compression or final-accuracy improvement. We instead investigate distillation to accelerate the training of strong students. We propose a generalizable plug-and-play recipe that freezes a weaker teacher, applies distillation only in early training, and turns it off once the student reaches and surpasses teacher-level performance. For ImageNet and CIFAR classification, this strategy reaches target thresholds much earlier, with up to 4.8 times speedup measured by epochs. We confirm that the method generalizes to other tasks and report 1.7 times epoch speedup for object detection on the COCO dataset, and 2.5 times earlier target-FID crossing for diffusion generation on the CIFAR-10 dataset, measured in steps. These findings validate our method as a universal speedup mechanism for visual learning.

arXiv:2604.15589v1 Announce Type: new Abstract: Existing research on large language models (LLMs) for automated code compliance has primarily focused on performance, treating the models as black boxes and overlooking how training decisions affect their interpretive behavior. This paper addresses this gap by employing a perturbation-based attribution analysis to compare the interpretive behaviors of LLMs across different fine-tuning strategies such as full fine-tuning (FFT), low-rank adaptation (LoRA) and quantized LoRA fine-tuning, as well as the impact of model scales which include varying LLM parameter sizes. Our results show that FFT produces attribution patterns that are statistically different and more focused than those from parameter-efficient fine-tuning methods. Furthermore, we found that as model scale increases, LLMs develop specific interpretive strategies such as prioritizing numerical constraints and rule identifiers in the building text, albeit with performance gains in semantic similarity of the generated and reference computer-processable rules plateauing for models larger than 7B. This paper provides crucial insights into the explainability of these models, taking a step toward building more transparent LLMs for critical, regulation-based tasks in the Architecture, Engineering, and Construction industry.

arXiv:2604.15377v1 Announce Type: new Abstract: Accurate and timely rainfall nowcasting is crucial for disaster mitigation and water resource management. Despite recent advances in deep learning, precipitation prediction remains challenging due to limitations in effectively leveraging diverse multimedia data sources. We introduce M3R, a Meteorology-informed MultiModal attention-based architecture for direct Rainfall prediction that synergistically combines visual NEXRAD radar imagery with numerical Personal Weather Station (PWS) measurements, using a comprehensive pipeline for temporal alignment of heterogeneous meteorological data. With specialized multimodal attention mechanisms, M3R novelly leverages weather station time series as queries to selectively attend to spatial radar features, enabling focused extraction of precipitation signatures. Experimental results for three spatial areas of 100 km * 100 km centered at NEXRAD radar stations demonstrate that M3R outperforms existing approaches, achieving substantial improvements in accuracy, efficiency, and precipitation detection capabilities. Our work establishes new benchmarks for multimedia-based precipitation nowcasting and provides practical tools for operational weather prediction systems. The source code is available at https://github.com/Sanjeev97/M3Rain

arXiv:2604.15588v1 Announce Type: new Abstract: The integration of Large Language Models (LLMs) into scientific workflows presents exciting opportunities to accelerate biomedical discovery. However, the reactive nature of LLMs, which respond only when prompted, limits their effectiveness in collaborative settings that demand foresight and autonomous engagement. In this study, we introduce CoLabScience, a proactive LLM assistant designed to enhance biomedical collaboration between AI systems and human experts through timely, context-aware interventions. At the core of our method is PULI (Positive-Unlabeled Learning-to-Intervene), a novel framework trained with a reinforcement learning objective to determine when and how to intervene in streaming scientific discussions, by leveraging the team's project proposal and long- and short-term conversational memory. To support this work, we introduce BSDD (Biomedical Streaming Dialogue Dataset), a new benchmark of simulated research discussion dialogues with intervention points derived from PubMed articles. Experimental results show that PULI significantly outperforms existing baselines in both intervention precision and collaborative task utility, highlighting the potential of proactive LLMs as intelligent scientific assistants.

The Information: Sources: Polymarket is in talks to raise $400M at a ~$15B post-money valuation, up from $9B in October 2025, but below Kalshi's $22B valuation from March 2026  —  The predictions site Polymarket is talking to investors about raising $400 million in funding at a valuation of about $15 billion including …

Sneha Kumar / Reuters: Australian data center operator NextDC plans to raise AU$1.5B in equity for its 350MW Sydney site rollout, and raises its FY26 capex by AU$300M to AU$2.7B-AU$3B  —  Australia's NEXTDC (NXT.AX) said on Monday it plans to raise A$1.5 billion ($1.07 billion) to accelerate the rollout …

Emily Shugerman / The San Francisco Standard: A profile of Maria Davidson, who heads California Renewal, a pro-business political group backed by Silicon Valley power players, seeking to raise $100M in 2026  —  Maria Davidson isn't the first name that comes to mind when you think of Silicon Valley power players.

<p><strong>TIL:</strong> <a href="https://til.simonwillison.net/google-sheets/datasette-sql">SQL functions in Google Sheets to fetch data from Datasette</a></p> <p>I put together some notes on patterns for fetching data from a Datasette instance directly into Google Sheets - using the <code>importdata()</code> function, a "named function" that wraps it or a Google Apps Script if you need to send an API token in an HTTP header (not supported by <code>importdata()</code>.)</p> <p>Here's <a href="https://docs.google.com/spreadsheets/d/14lRV2-AeBmjI3lJbl2apwfC_ncXqL0uSV68lmtzUI7I/edit?gid=0#gid=0">an example sheet</a> demonstrating all three methods.</p> <p>Tags: <a href="https://simonwillison.net/tags/spreadsheets">spreadsheets</a>, <a href="https://simonwillison.net/tags/datasette">datasette</a>, <a href="https://simonwillison.net/tags/google">google</a></p>

<p><strong><a href="https://tools.simonwillison.net/claude-token-counter">Claude Token Counter, now with model comparisons</a></strong></p> I <a href="https://github.com/simonw/tools/pull/269">upgraded</a> my Claude Token Counter tool to add the ability to run the same count against different models in order to compare them.</p> <p>As far as I can tell Claude Opus 4.7 is the first model to change the tokenizer, so it's only worth running comparisons between 4.7 and 4.6. The Claude <a href="https://platform.claude.com/docs/en/build-with-claude/token-counting">token counting API</a> accepts any Claude model ID though so I've included options for all four of the notable current models (Opus 4.7 and 4.6, Sonnet 4.6, and Haiku 4.5).</p> <p>In the Opus 4.7 announcement <a href="https://www.anthropic.com/news/claude-opus-4-7#migrating-from-opus-46-to-opus-47">Anthropic said</a>:</p> <blockquote> <p>Opus 4.7 uses an updated tokenizer that improves how the model processes text. The tradeoff is that the same input can map to more tokens—roughly 1.0–1.35× depending on the content type.</p> </blockquote> <p>I pasted the <a href="https://github.com/simonw/research/blob/2cf912666ba08ef0c00a1b51ee07c9a8e64579ef/extract-system-prompts/claude-opus-4-7.md?plain=1">Opus 4.7 system prompt</a> into the token counting tool and found that the Opus 4.7 tokenizer used 1.46x the number of tokens as Opus 4.6.</p> <p><img alt="Screenshot of a token comparison tool. Models to compare: claude-opus-4-7 (checked), claude-opus-4-6 (checked), claude-opus-4-5, claude-sonnet-4-6, claude-haiku-4-5. Note: &quot;These models share the same tokenizer&quot;. Blue &quot;Count Tokens&quot; button. Results table — Model | Tokens | vs. lowest. claude-opus-4-7: 7,335 tokens, 1.46x (yellow badge). claude-opus-4-6: 5,039 tokens, 1.00x (green badge)." src="https://static.simonwillison.net/static/2026/claude-token-count.jpg" /></p> <p>Opus 4.7 uses the same pricing is Opus 4.6 - $5 per million input tokens and $25 per million output tokens - but this token inflation means we can expect it to be around 40% more expensive.</p> <p>The token counter tool also accepts images. Opus 4.7 has improved image support, described like this:</p> <blockquote> <p>Opus 4.7 has better vision for high-resolution images: it can accept images up to 2,576 pixels on the long edge (~3.75 megapixels), more than three times as many as prior Claude models.</p> </blockquote> <p>I tried counting tokens for a 3456x2234 pixel 3.7MB PNG and got an even bigger increase in token counts - 3.01x times the number of tokens for 4.7 compared to 4.6:</p> <p><img alt="Same UI, this time with an uploaded screenshot PNG image. claude-opus-4-7: 4,744 tokens, 3.01x (yellow badge). claude-opus-4-6: 1,578 tokens, 1.00x (green badge)." src="https://static.simonwillison.net/static/2026/claude-token-count-image.jpg" /></p> <p><strong>Update</strong>: That 3x increase for images is <em>entirely</em> due to Opus 4.7 being able to handle higher resolutions. I tried that again with a 682x318 pixel image and it took 314 tokens with Opus 4.7 and 310 with Opus 4.6, so effectively the same cost.</p> <p><strong>Update 2</strong>: I tried a 15MB, 30 page text-heavy PDF and Opus 4.7 reported 60,934 tokens while 4.6 reported 56,482 - that's a 1.08x multiplier, significantly lower than the multiplier I got for raw text. <p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/llm-pricing">llm-pricing</a>, <a href="https://simonwillison.net/tags/tokenization">tokenization</a></p>

Kieran Smith / Financial Times: Match Group CEO Spencer Rascoff says growing women's share on Tinder is his “primary focus” to stem user declines; Sensor Tower says 75% of Tinder users are men  —  Match Group boss Spencer Rascoff says attracting female users is ‘priority focus’ in push to turn around flagship dating app

In this article, we cover the GLM-4.6V model. Specifically, we cover the technical capabilities of the model along with inference for image description, OCR, and image to HTML code. The post Getting Started with GLM-4.6V appeared first on DebuggerCafe.

After 25 years of teaching kindergarten, Linda Meyer was ready for a change and moved to Tuscany. Provided by Linda Meyer At 48, Linda Meyer told her husband she needed a change and quit teaching. She said she wanted to move to Tuscany. He replied: "We'll make it happen." Now 61, she owns a villa and runs a tourism business in Italy. This as-told-to essay is based on a conversation with Linda Meyer, the co-owner of La Chiusa. It has been edited for length and clarity. Thirteen years ago, I had a beautiful life in Northern Virginia. I was a 25-year veteran kindergarten teacher, married to a successful businessman and former Air Force pilot, and had raised two kids. We didn't have financial issues: We had a lovely home and traveled often. And yet, I couldn't shake the feeling, the pull, that something was still missing. I knew there had to be more to life than getting up on a Saturday morning to run errands. I'd become overweight and unhealthy. It was February, nowhere near the end of a school year, when I told my husband I needed a change. He said, "What do you want to do?" I said, "I think I need to move to Italy." The ever-supportive partner, he said, "OK. We'll make it happen." Meyers moved to Tuscany, and her husband followed two months later. Provided by Linda Meyers The life-changing decision Our first step was finding an apartment I could rent in Tuscany. I'd traveled there before and fell in love with the culture. We made the move gradually. I went first, and then my husband followed two months later. The first apartment was in a 12th-century building, so it was a bizarre little space, with the bathroom and two dens downstairs and the kitchen upstairs. But, like my life, I was fine with having it flipped upside down and a little unexpected. My son and daughter — now both in their 30s — were thrilled for me. My daughter Whitney even tagged along to help my husband George move me in. The move took place only two weeks after I'd first shared my big idea. That's when the story changes a bit, from the dream I pictured to the day-to-day realities. It took Meyers time to get settled in Tuscany. Provided by Linda Meyers The unexpected challenges of getting settled The first weeks weren't what I had pictured — there was no TV, no WiFi, I didn't speak the language, it was cold out, and I realized quickly I'd never lived alone. I'd moved straight from my daddy's house to George's house, married and pregnant at 21. I had never really done anything for myself. I was afraid to drive in Italy, not because of driving itself, but because of the color-coded parking system in Tuscany. Yellow means you can park, white means it's free, blue means you have to pay — and you have to learn to read Italian real quick to figure all this out. So, I was housebound for a bit until I worked up the courage to do it. I cried a lot, questioning all my life decisions over the two months until my husband could join me here after he finished preparing his business. I wasn't sure I could do it, but my son pushed me to try over the phone. "You've been around the world," he said, "and you're afraid to go out?" So I did. I drove to the Gucci store in Florence, bought a backpack, got back in my car and drove home, and survived. I still carry that backpack today. It's a small reminder that I can do whatever I want. Meyers went on to buy a 1,700-year-old villa and 50-acre olive oil farm. Provided by Linda Meyers I needed to help other women do the same I started walking around town. I started eating differently. I could tell I was changing. I started a Facebook group and invited other women in the same mid-life drudgery to come visit and see the village I'd come to love. People responded and shared that they'd love to come and were tired of their routine, too. A few months later, 10 people came for a week, and another 10 people came the next week to visit. I got paid to be their tour guide and show them the things I love. I told them my story. When I first started, we weren't residents yet, so we could only stay in Italy for a month or two at a time in any 180-day period, so we'd go back to the US to visit family. The next year, that number jumped to 100 visitors. At first, I'd seen entertaining as a great hobby, but at that point I knew a business was born. My husband and I moved to a bigger house in the village, and I started hosting people 16 weeks a year, with my daughter and husband's help. I would take the tourists out to see the town, they would take cooking lessons, and reflect on their lives over drinks and meals. Meyers has received help from her husband and kids in running the business. Provided by Linda Meyers Growing the business, with some bumps and wins along the way In 2018, I bought a 1,700-year-old villa and 50-acre olive oil farm called La Chiusa. Today, it operates as a guest property with 17 rooms, with nightly rates starting from 180 to 300 euros, or $200 to $340. I've lost 100 pounds, leaned into gardening, and even show guests how to change their lives in small, practical ways when they go back. I now employ 25 permanent team members, in addition to six local "nonnas" (grandmothers) who specialize in cooking. Rules for running a business are different in Italy. For example, there are full time and seasonal contracts, and if you give them a full time contract, it's difficult to fire someone. So, even when an employee never showed up for work again, I had to keep paying them. There are also really specific rules about what tasks they can perform — you can't randomly ask a gardener or waiter to wash a dish unless it's specified in that contract. I learned lessons the hard way, but I've had lots of fun along the way. Read the original article on Business Insider

Boo Kok Chuon started his 8-year-old daughter early on AI. Boo Kok Chuon Boo Kok Chuon and his 8-year-old daughter made an app in three hours using various AI tools. He told Business Insider that he wanted to start his kid early on AI, given how technology is progressing. Still, he places guardrails and takes a cautious approach when introducing AI systems to his kid. This as-told-to essay is based on a conversation with Boo Kok Chuon, the chief operating officer of a law firm in Singapore. The conversation has been edited for length and clarity. Since early last year, we took a bold step in introducing ChatGPT to Kiki, our 8-year-old daughter. We thought that we should start our kid early, given how technology is progressing these days. She has been frequently communicating with her own ChatGPT. Whenever she faces an issue, she will go to ChatGPT to ask for a solution. Once she spilled a drink on a shirt and she felt bad. She asked ChatGPT how to resolve it. She also asked ChatGPT how to take care of her turtles. Then we tried something bigger. Vibe coding an app together We built a companion app for our physical game Mu Jong, a music-learning mahjong-style game. We did it within three hours. We used a combination of generative AI tools, including ChatGPT for prompt development and logic structuring, Base 44 for interface development, and Nano Banana Pro for visual assets. The app has a score tracker, a keyboard music feature, and a countdown timer. Kiki conceptualized the features and guided the process, while I supported her in structuring prompts and translating her ideas into instructions for the AI tools. We created a mascot for our app. Kiki drew out the mascot on a piece of paper. It was quite a crappy drawing, but once you upload the photo into the AI, it will beautify it. She tried it, and was like: '"Wow, it's like magic." Suddenly, her drawing looked commercially ready. I also guided her. I told her: "You have to say the purpose of the drawing, and what you want to use this mascot for." We refined the drawing based on further prompting the AI. Children should be exposed to AI early It is very important for kids — especially like Kiki — to be exposed to AI. Of course, human input is still required for review. She was born into a privileged generation in which technology was already very advanced. Nowadays, kids may not be so easily impressed, given how readily available all this technology is. Kiki is learning to break a problem down into steps and use AI tools to test and refine her ideas. She also learned how to use natural language to construct more structured prompts. For example, she began specifying attributes such as "rounder shape, softer colors, friendly expression," which helped her achieve more consistent outputs. For her devices, we have oversight over the content she engages with. On YouTube, we constantly review what she watches and look at her internet history. We are very open with her and want to be transparent and respectful. She's aware that we are supervising her. At any point in time, if we see anything that is not quite right, the approach that we use is to guide her and not judge her. We will ask her, "Why do you do this? Do you do this because it helps you with certain issues that you are facing?" And from there, we will slowly — again through this questioning method — lead her toward the more appropriate actions. There are some inherent issues with AI. Sometimes, AI can be a bit too conservative in its outputs. I think that if you rely on AI too much, it might be emotionally dangerous in a way, too. We should not trust these systems 100%, and take a cautious, objective approach when dealing with them. Do you have a story to share about coding with AI? Contact this reporter at cmlee@businessinsider.com. Read the original article on Business Insider