arXiv:2604.21984v1 Announce Type: new Abstract: We introduce Soft Anisotropic Diagrams (SAD), an explicit and differentiable image representation parameterized by a set of adaptive sites in the image plane. In SAD, each site specifies an anisotropic metric and an additively weighted distance score, and we compute pixel colors as a softmax blend over a small per-pixel top-K subset of sites. We induce a soft anisotropic additively weighted Voronoi partition (i.e., an Apollonius diagram) with learnable per-site temperatures, preserving informative gradients while allowing clear, content-aligned boundaries and explicit ownership. Such a formulation enables efficient rendering by maintaining a per-query top-K map that approximates nearest neighbors under the same shading score, allowing GPU-friendly, fixed-size local computation. We update this list using our top-K propagation scheme inspired by jump flooding, augmented with stochastic injection to provide probabilistic global coverage. Training follows a GPU-first pipeline with gradient-weighted initialization, Adam optimization, and adaptive budget control through densification and pruning. Across standard benchmarks, SAD consistently outperforms Image-GS and Instant-NGP at matched bitrate. On Kodak, SAD reaches 46.0 dB PSNR with 2.2 s encoding time (vs. 28 s for Image-GS), and delivers 4-19 times end-to-end training speedups over state-of-the-art baselines. We demonstrate the effectiveness of SAD by showcasing the seamless integration with differentiable pipelines for forward and inverse problems, efficiency of fast random access, and compact storage.

arXiv:2604.22036v1 Announce Type: new Abstract: This paper introduces EgoMAGIC (Medical Assistance, Guidance, Instruction, and Correction), an egocentric medical activity dataset collected as part of DARPA's Perceptually-enabled Task Guidance (PTG) program. This dataset comprises 3,355 videos of 50 medical tasks, with at least 50 labeled videos per task. The primary objective of the PTG program was to develop virtual assistants integrated into augmented reality headsets to assist users in performing complex tasks. To encourage exploration and research using this dataset, the medical training data has been released along with an action detection challenge focused on eight medical tasks. The majority of the videos were recorded using a head-mounted stereo camera with integrated audio. From this dataset, 40 YOLO models were trained using 1.95 million labels to detect 124 medical objects, providing a robust starting point for developers working on medical AI applications. In addition to introducing the dataset, this paper presents baseline results on action detection for the eight selected medical tasks across three models, with the best-performing method achieving average mAP 0.526. Although this paper primarily addresses action detection as the benchmark, the EgoMAGIC dataset is equally suitable for action recognition, object identification and detection, error detection, and other challenging computer vision tasks. The dataset is accessible via zenodo.org (DOI: 10.5281/zenodo.19239154).

arXiv:2604.22045v1 Announce Type: new Abstract: Feature attribution methods explain the predictions of deep neural networks by assigning importance scores to individual input features. However, most existing methods focus solely on marginal effects, overlooking feature interactions, where groups of features jointly influence model output. Such interactions are especially important in image classification tasks, where semantic meaning often arises from pixel interdependencies rather than isolated features. Existing interaction-based methods for images are either coarse (e.g., superpixel-only) or, fail to satisfy core interpretability axioms. In this work, we introduce H-Sets, a novel two-stage framework for discovering and attributing higher-order feature interactions in image classifiers. First, we detect locally interacting pairs via input Hessians and recursively merge them into semantically coherent sets; segmentation from Segment Anything (SAM) is used as a spatial grouping prior but can be replaced by other segmentations. Second, we attribute each set with IDG-Vis, a set-level extension of Integrated Directional Gradients that integrates directional gradients along pixel-space paths and aggregates them with Harsanyi dividends. While Hessians introduce additional compute at the detection stage, this targeted cost consistently yields saliency maps that are sparser and more faithful. Evaluations across VGG, ResNet, DenseNet and MobileNet models on ImageNet and CUB datasets show that H-Sets generate more interpretable and faithful saliency maps compared to existing methods.

arXiv:2604.22093v1 Announce Type: new Abstract: Reliable visual perception under low illumination remains a core challenge for autonomous robotic systems, where degraded image quality directly compromises navigation, inspection, and various operations. A recent training free approach showed that Bayesian optimisation with Gaussian Processes can adaptively select brightness, contrast, and denoising parameters on a per-image basis, achieving competitive enhancement without any learned model. However, that framework is limited to three parameters, applies no illumination decomposition or white balance correction, and relies on Non-Local Means denoising, which tends to over smooth edges under noisy conditions. This paper proposes FLARE-BO (Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation), an extended framework that jointly optimises eight parameters spanning across gamma correction, LIME-style illumination normalisation, chrominance denoising, bilateral filtering, NLM denoising, Grey-World automatic white balance, and adaptive post smoothing. The search engine employs a unit hypercube parameter normalisation, objective standardisation, Sobol quasi-random initialisation, and Log Expected Improvement acquisition for principled exploration of the expanded space. Performance of the proposed method is benchmarked using the Low Light paired dataset (LOL) and results show marked improvements of the proposed method over existing methods that were not specifically trained using this dataset.

arXiv:2604.21953v1 Announce Type: new Abstract: Anti-doping programs rely on biological testing to detect performance-enhancing drugs, but such testing costs over $800 per sample and is limited by short detection windows for many prohibited substances. These constraints leave large portions of athletes without regular testing, motivating complementary screening approaches that analyze routine competition results to identify suspicious performance patterns. We present a system that processes 1.6 million athletics performances from over 19,000 competitions (2010-2025) using eight detection methods ranging from statistical rules to machine learning and trajectory analysis. We validate all methods against publicly confirmed anti-doping violations to measure their effectiveness in identifying sanctioned athletes. Trajectory-based methods, which compare performances to expected career progression, achieve the best balance between detecting violations and limiting false alarms, though all methods face challenges from incomplete data and rare confirmed violations. The system provides an interactive interface for expert-driven investigation, emphasizing transparency and human judgment to support, rather than replace, established anti-doping processes.

arXiv:2604.21999v1 Announce Type: new Abstract: We study learned memory tokens as computational scratchpad for a single-block Universal Transformer (UT) with Adaptive Computation Time (ACT) on Sudoku-Extreme, a combinatorial reasoning benchmark. We find that memory tokens are empirically necessary: across all configurations tested -- 3 seeds, multiple token counts, two initialization schemes, ACT and fixed-depth processing -- no configuration without memory tokens achieves non-trivial performance. The optimal count exhibits a sharp lower threshold (T=0 always fails, T=4 is borderline, T=8 reliably succeeds for 81-cell puzzles) followed by a stable plateau (T=8-32, 57.4% +/- 0.7% exact-match) and collapse from attention dilution at T=64. During experimentation, we identify a router initialization trap that causes >70% of training runs to fail: both default zero-bias initialization (p ~ 0.5) and Graves' recommended positive bias (p ~ 0.73) cause tokens to halt after ~2 steps at initialization, settling into a shallow equilibrium (halt ~ 5-7) that the model cannot escape. Inverting the bias to -3 ("deep start," p ~ 0.05) eliminates this failure mode. We confirm through ablation that the trap is inherent to ACT initialization, not an artifact of our architecture choices. With reliable training established, we show that (1) ACT provides more consistent results than fixed-depth processing (56.9% +/- 0.7% vs 53.4% +/- 9.3% across 3 seeds); (2) ACT with lambda warmup achieves matching accuracy (57.0% +/- 1.1%) using 34% fewer ponder steps; and (3) attention heads specialize into memory readers, constraint propagators, and integrators across recursive depth. Code is available at https://github.com/che-shr-cat/utm-jax.

arXiv:2604.21991v1 Announce Type: new Abstract: Multi-task optimization is a powerful approach for solving a large number of tasks in parallel. However, existing algorithms face distinct limitations: Population-based methods scale poorly and remain underexplored for large task sets. Approaches that do scale beyond a thousand tasks are mostly MAP-Elites variants and rely on a fixed, discretized archive that disregards the topology of the task space. We introduce MONET (Multi-Task Optimization over Networks of Tasks), a multi-task optimization algorithm that models the task space as a graph: tasks are nodes, and edges connect tasks in the task parameter space. This representation enables knowledge transfer between tasks and remains tractable for high-dimensional problems while exploiting the topology of the task space. MONET combines social learning, which generates candidates from neighboring nodes via crossover, with individual learning, which refines a node's own solution independently via mutation. We evaluate MONET on four domains (archery, arm, and cartpole with 5,000 tasks each; hexapod with 2,000 tasks) and show that it matches or exceeds the performance of existing MAP-Elites-based baselines across all four domains.

arXiv:2604.21952v1 Announce Type: new Abstract: This work presents a multi-layered methodology for efficiently accelerating multimodal foundation models (MFMs). It combines hardware and software co-design of transformer blocks with an optimization pipeline that reduces computational and memory requirements. During model development, it employs performance enhancements through fine-tuning for domain-specific adaptation. Our methodology further incorporates hardware and software techniques for optimizing MFMs. Specifically, it employs MFM compression using hierarchy-aware mixed-precision quantization and structural pruning for transformer blocks and MLP channels. It also optimizes operations through speculative decoding, model cascading that routes queries through a small-to-large cascade and uses lightweight self-tests to determine when to escalate to larger models, as well as co-optimization of sequence length, visual resolution & stride, and graph-level operator fusion. To efficiently execute the model, the processing dataflow is optimized based on the underlying hardware architecture together with memory-efficient attention to meet on-chip bandwidth and latency budgets. To support this, a specialized hardware accelerator for the transformer workloads is employed, which can be developed through expert design or an LLM-aided design approach. We demonstrate the effectiveness of the proposed methodology on medical-MFMs and on code generation tasks, and conclude with extensions toward energy-efficient spiking-MFMs.

arXiv:2604.22031v1 Announce Type: new Abstract: We propose Mochi, a Graph Foundation Model that addresses task unification and training efficiency by adopting a meta-learning based training framework. Prior models pre-train with reconstruction-based objectives such as link prediction, and assume that the resulting representations can be aligned with downstream tasks through a separate unification step such as class prototypes. We demonstrate through synthetic and real-world experiments that this procedure, while simple and intuitive, has limitations that directly affect downstream task performance. To address these limitations, Mochi pre-trains on few-shot episodes that mirror the downstream evaluation protocol, aligning the training objective with inference rather than relying on a post-hoc unification step. We show that Mochi, along with its more powerful variant Mochi++, achieves competitive or superior performance compared to existing Graph Foundation Models across 25 real-world graph datasets spanning node classification, link prediction, and graph classification, while requiring 8$\sim$27 times less training time than the strongest baseline.

arXiv:2604.21993v1 Announce Type: new Abstract: We study the detection of transient liquidity erosion ("crumbling quotes") in electronic limit order books, where observable quote deterioration may reflect either mechanical liquidity withdrawal or informational repricing. Using the ABIDES agent-based simulator, we construct a multi-agent environment in which crumbling emerges from stochastic regime switches in a market maker, providing time-resolved ground truth unavailable in real market data. We develop a detection pipeline that identifies mechanically driven quote erosion using order book features, and train a neural model to produce calibrated crumbling probabilities. Experiments demonstrate that the proposed framework reliably identifies crumbling events against agent-level ground truth, with the neural model achieving +36% AUC improvement over rule-based baselines and robust performance across normal, high-volatility, bull, and bear market conditions. Ablation studies on temporal features and varying the dependence structure of the ground-truth mechanism confirm that the framework generalizes across both independent and autocorrelated liquidity withdrawal dynamics.

Marcus Schuler / Implicator.ai: OpenAI publishes a five-principle framework for AGI development, pledging to resist concentrating AI power and to collaborate with companies and governments  —  OpenAI published a five-principle framework on Sunday for the development of artificial general intelligence.

Discussion | Link

As growth slows, Truecaller is leaning on subscriptions, business services, and new features to sustain momentum beyond India.

Reed Albergotti / Semafor: An interview with Bill Nguyen, a tech entrepreneur who is completely outsourcing parts of his life to an AI assistant in his bid to create a virtual body double  —  On a recent morning at his Atlanta office, Bill Nguyen tore open a large package and lifted out a high-end desktop computer.

Discussion | Link

Clockwise from bottom left: Sydney Sweeney in "Euphoria," "Reality," and "Immaculate." Monica Schipper/Getty Images; HBO; Neon; Tyler Le/BI Sydney Sweeney is back in the public eye thanks to her role as Cassie in HBO's hit show "Euphoria." Sweeney deliberately markets herself as a blank canvas for brands and casting directors. This has made Sweeney a polarizing figure, but it's also allowed her to build a versatile résumé. In the second episode of "Euphoria" season three, Sydney Sweeney lounges poolside in skin-tight pants and a Barbie-pink crop top, eyes bugging out of her skull. She's in character as the beautiful and desperate Cassie Howard, trying to convince her best-friend-turned-romantic-rival, Maddy Perez (Alexa Demie), to propel her to internet stardom. "I just feel like if more people knew me, I would be huge," Cassie proclaims with childlike pride. When Maddy observes that "the market is oversaturated with a lot of girls like you," meaning pretty blonde influencers, Cassie waves away her doubt with a perfectly manicured hand. Success — particularly the kind of lucrative, external ubiquity that Cassie is seeking on Instagram and OnlyFans — is not about having good taste or a clear sense of self, she posits. There is power in a blank canvas. "I can be anything," Cassie concludes. This mantra may feel familiar to those following Sweeney's real-life exploits. To the untrained eye, her constant churn of business opportunities and brand deals (A lingerie line! That controversial American Eagle campaign! A soap that is… made with her bathwater?) hardly illustrates care or discernment. It's tempting to read Cassie's season three arc — chasing fame and fortune by leveraging her body and lack of a "moral code," in Maddy's words — as a cruel meta-commentary on Sweeney's own trajectory since "Euphoria" premiered in 2019. Whether it's Cassie's debasing plea for Maddy's help or her disastrous wedding to Nate Jacobs (Jacob Elordi) in episode three, which leaves the bride bawling on the floor of her tacky mansion, Sweeney's Cassie is made to play the fool. In reality, it's far more likely that Sweeney knows exactly what she's doing. She's an Emmy-nominated actor with agency, bankability, and, by all accounts, an open line of communication with "Euphoria" showrunner Sam Levinson. "I adore Sydney," Levinson recently told The New York Times. "Before I wrote this season, she called me and she said, 'Just do me one favor: Make sure Cassie's crazy.'" Sweeney characterized her request similarly in a 2023 interview with Variety: "'Give me more,'" she said she told Levinson. "'I'm going to show you what I have. There's so much to this girl.'" That Sweeney is game to play Cassie as increasingly shameless and attention-starved to encourage the audience to laugh at her expense signals that she's in on the joke — perhaps more so than the internet may realize. Sweeney is a blank canvas by design Sydney Sweeney as Cassie in "Euphoria" season three. Patrick Wymore/HBO In her offscreen life, Sweeney is routinely undogmatic. The 28-year-old, who was raised near the Washington-Idaho border, tends to deflect political questions with a deft touch of vagueness. What was it like to see President Donald Trump praising her "Great Jeans" campaign with American Eagle? "It was surreal." What about the backlash to photos from her mom's birthday party, where guests were pictured wearing MAGA-inspired hats? "There were so many misinterpretations." Does she ever plan to make it clear where she stands? "I'm not a political person." As a result, Sweeney has been portrayed as every bogeywoman across the political spectrum, from a clueless puppet of the patriarchy to an anti-woke crusader to a sexually empowered girlboss. The one thing you can't call Sweeney? A hypocrite. If celebrities get the most heat when their actions go against their professed or even perceived values, there's a business incentive to stay silent. After all, you can't betray a cause if you don't have one. The rare stance Sweeney reveals, like the time she declared that "we should all love each other," is so utterly bland that she remains free to make money however she likes, without the risk of breaking her word or reneging on her principles. And make money she does. There's the aforementioned collaboration with American Eagle and the now-infamous Dr. Squatch bar of soap purportedly made with Sweeney's "actual bathwater," but those are just the ads that caused the most brouhaha. In addition to launching her own lingerie line, she's also partnered with Laneige, Kérastase, Heydude, Dickies, Ford, and Baskin-Robbins, among others. Skincare, hair care, footwear, workwear, cars, ice cream — who cares? Sydney Sweeney can be anything, which, in today's creator economy, actually means "sell anything." Sweeney's indistinct personal brand has kept professional doors open Sydney Sweeney as Christy Martin in "Christy." Black Bear Pictures It looks like Cassie may be on to something; Sweeney hasn't needed to display good taste or a clear sense of self to become a household name, let alone a full-blown movie star. For all the public's attempts to pin down her position in the culture wars, Sweeney has built a surprisingly varied filmography. Her poker face has worn many onscreen masks: Eden Blaine, a child bride in a dystopian theocracy ("The Handmaid's Tale"); Christy Martin, a pioneering boxer and out lesbian ("Christy"); Julia Cornwall, a teenager-turned-superhero ("Madame Web"); Millie Calloway, a scrappy young woman with a taste for vengeance ("The Housemaid"); Sister Cecilia, a nun subjected to unrelenting body horror ("Immaculate"); Reality Winner, a real-life whistleblower who leaked information about Russian interference in the 2016 US elections ("Reality"); and more. Notably, several of Sweeney's character arcs, including in "Euphoria," are shaped by sexism, objectification, and systematic oppression. When GQ asked about the recurring theme of domestic abuse in Sweeney's work, she replied with her signature hedging. "I always speak out about something that is important to me. And for me to speak out, I use art," she said. "Through my characters in my movies, it's a way for me to be able to do my part and spread awareness in different ways through my characters." One could argue that Sweeney's lack of political conviction, particularly in our hyper-polarized world, translates more as complicity than neutrality — and many do. Sweeney's persona has been described as conservative-coded, an assumption bolstered by BuzzFeed's 2025 report that she registered as a Republican voter in Florida the previous year. Still, Sweeney has been careful not to pursue the path of proudly conservative celebrities like Gina Carano or Roseanne Barr, who lost a role in a major franchise and a bespoke ABC sitcom, respectively, after sharing transphobic and racist views on social media. Barr later blamed the "Roseanne" cancellation on her support of Trump: "That is not allowed in Hollywood," she said. Sweeney may be surrounded by costars and potential employers who oppose her political party, but you wouldn't know it from her résumé. She walks a finer line than her willfully pigeonholed peers. Call it plausible deniability or savvy self-marketing, but either way, Sweeney is collecting paychecks from all sides. Read the original article on Business Insider

Joseph Howlett / Scientific American: An amateur solved a 60-year-old Erdős problem using a single GPT-5.4 Pro prompt; Terence Tao says it's a “nice achievement” with unclear long-term significance  —  A ChatGPT AI has proved a conjecture with a method no human had thought of.  Experts believe it may have further uses

President Donald Trump at the White House Correspondents Association Dinner. Nathan Howard/Getty Images Trump spoke to "60 Minutes" on Sunday, after the White House Correspondents' Dinner shooting on Saturday night. During the interview, the president was asked about the accused gunman's motives. Trump said none of the accusations in a written manifesto had anything to do with him. On Sunday, President Donald Trump reckoned with the string of assassination attempts he has faced this term. During an interview with CBS's "60 Minutes," Trump said he had read a manifesto written by the gunman who entered the annual press dinner in Washington, DC, on Saturday and was quickly subdued "He's radicalized. He was a Christian — a believer, and then he became an anti-Christian, and he had a lotta change," Trump said. The president pushed back when "60 Minutes" host Norah O'Donnell read a line from the manifesto that called Trump a pedophile and a rapist, among other accusations. "I'm not a pedophile. You read that crap from some sick person? I got associated with all — stuff that has nothing to do with me. I was totally exonerated," Trump said. He added that he wasn't involved with "Epstein or other things" — referring to the disgraced financier and convicted sex offender, Jeffrey Epstein. Trump said he thought part of the reason he was facing threats stems from anti-Trump rhetoric and demonstrations like the "No Kings" protests, a series of rallies across the country. "The reason you have people like that is you have people doing No Kings. I'm not a king. What I am — if I was a king, I wouldn't be dealing with you," he told O'Donnell. He said political violence has gone back hundreds of years, echoing comments from his Saturday night press conference that all countries face violence. "People are assassinated. People are injured. People are hurt. And I'm not sure that it's any more now than there was," Trump said on "60 Minutes." In his press conference on Saturday night, Trump said being a president is "a dangerous profession" and that facing the threat of political violence is "part of the job." At 8:36 p.m. on Saturday night, shots were fired outside the ballroom of the Washington Hilton, where the White House Correspondents' Dinner was underway. It marked the third time in three years that President Donald Trump faced the threat of assassination. Authorities have not yet publicly identified the suspect. However, photos of the White House Correspondents' Dinner shooter posted online by Trump on Saturday night matched a LinkedIn profile belonging to Cole Allen. Multiple outlets, including the Associated Press, have also reported that Allen is the suspect in custody. Allen, 31, is listed as a resident of Torrance, California. Police conducted an overnight raid at an address in the state linked to Allen in an online database. His LinkedIn describes him as a "mechanical engineer and computer scientist by degree, independent game developer by experience, teacher by birth." Acting US Attorney General Todd Blanche said Sunday in several TV news interviews that the shooter left a manifesto, in which he said he intended to target Trump administration officials. "We know that there were some writings, and we've already spoken with several witnesses who knew him," he said on NBC News' "Meet the Press." Blanche said the suspect traveled from Los Angeles to Washington by train and checked into the hotel where the event was held with two guns and a plan. Jeanine Pirro, the US Attorney for the District of Columbia, said the defendant has been charged with two counts of using a firearm during a crime of violence and one count of assault on a law enforcement officer using a dangerous weapon. He will be arraigned on Monday. Read the original article on Business Insider

In this article, we get started with Molmo2. We start with the discussion of the important aspects from the technical article and report. Then we move to a simple inference pipeline for image VQA, vidoe VQA, and image pointing. The post Getting Started with Molmo2 appeared first on DebuggerCafe.

Nick Burns left the US due to rising living costs. He moved to Singapore and doesn't want to move back. Amadna Goh/Business Insider Nick Burns traded San Francisco for Singapore, driven by high living costs and safety concerns. While he earns less than he used to, his overall tax burden has decreased, and he's now a homeowner. He applied for Singapore citizenship in December and is ready to renounce his US passport if approved. Nick Burns was making six figures in the US and still finding it hard to save. Burns, now 35, grew up in Connecticut and served in the Navy. It was during that time that he first visited Singapore. Impressed by the country's clean and fast public transport system, he began to imagine what it might be like to live there one day. "But it felt impossible, like a crazy idea," Burns told Business Insider. This story is part of Business Insider's ongoing "Priced Out" coverage, which examines the lives and finances of Americans who have been priced out of the US. Has the cost of living in the US led you to move overseas? We want to hear from you. Get in touch with this reporter at agoh@businessinsider.com. After leaving the Navy in 2017, Burns moved to San Francisco for work and eventually ended up in the semiconductor manufacturing industry. He lived in downtown San Francisco and commuted an hour and a half each way to Silicon Valley. He said it wasn't uncommon to pass dark alleys and see people openly using drugs. Burns has since built a life in Singapore, having gotten married and bought a home. Amanda Goh/Business Insider "Even as a single man, I felt unsafe walking around," Burns said. Between his rent, steep taxes, and safety concerns, the city did not feel worth the cost. In January 2020, Burns took a lateral transfer to relocate to Singapore with his company. "I don't want to go back to the United States," Burns said. Housing costs: renting in San Francisco vs buying in Singapore Singapore is often ranked among the world's most expensive cities — not an obvious choice for someone looking to cut costs. But for Burns, it made financial sense. In San Francisco, Burns' two-bedroom apartment costs $5,728 a month, inclusive of a $300 reserved parking fee. Buying property there didn't feel within reach. Burns spent several years renting in Singapore, both before and after marrying his wife in 2023. In 2025, the couple bought a four-room public housing apartment on the city fringe for 1.01 million Singapore dollars, or about $790,000. Burns and his wife, Jessica, live in a four-room public housing apartment in a city-fringe neighborhood in Singapore. Amanda Goh/Business Insider Known as HDB flats, these government-built homes make up the majority of housing in Singapore and are primarily reserved for citizens. Burns was able to buy one with his wife, who is Singaporean, while he holds permanent residency. The stamp duty and down payment totaled about SG$280,150. They paid the majority in cash and about 20% from their CPF — Singapore's mandatory savings account, used for things like housing, healthcare, and retirement. The remaining amount was financed through a bank loan. Today, their monthly mortgage is about SG$3,520. Their monthly mortgage is 3,520 Singapore dollars. Amanda Goh/Business Insider Taxes: Moving abroad worked out in his favor Taxes also played a big role in how his finances changed after the move. In 2025, Burns earned about 6% less than he did in San Francisco in 2019, but his overall tax burden is significantly lower. In California, his taxes were around $41,200 in 2019. In Singapore, his latest tax bill was about SG$13,300. Lower tax rates have allowed him to keep more of his income, though, as a US citizen, he still has to file US taxes. Americans living abroad can exclude roughly $120,000 to $130,000 of foreign-earned income, which helps reduce what they owe, but doesn't eliminate it entirely. The US is one of the few countries that taxes its citizens abroad. Lower tax rates in Singapore have also enabled him to save more of his income, something he couldn't really do while living in San Francisco. Amanda Goh/Business Insider Day-to-day expenses: A less expensive life abroad His day-to-day expenses have also gone down. In Singapore, affordable food is easy to find at hawker centers, where you can have a meal for just a few dollars. A plate of chicken rice, a common local dish, is often less than $5. Getting around is cheaper, too. In San Francisco, he owned a car and says he spent about $400 a month on gas. In Singapore, he relies on ride-hailing services and public transportation, spending about SG$300 a month. He doesn't own a car in Singapore and mainly relies on ride-hailing apps and public transport to get around. Amanda Goh/Business Insider With a hybrid work arrangement and a shorter commute, he's also been able to reclaim much of his day. "Day to day, I have more time to exercise because I've got three to four hours back. I run a lot now and usually go before work," he said. Burns said his quality of life has improved significantly for what he spends. More importantly, he's finally able to put money aside. A life he chose Still, it took Burns some time to get used to life in Singapore — especially the city's hot and humid climate, with temperatures between 80 and 90 degrees year-round. "When I first moved here, I'd walk to the bus stop in my work shirt, and it would be a sweaty mess," he said. He also misses Mexican food and pizza from back in the US, adding that while they're available in Singapore, they're not quite the same. He now holds permanent residency in Singapore and applied for citizenship in December. Amanda Goh/Business Insider While some family members have visited him and been impressed by the city-state, Burns says others don't get his decision to leave the US. They worry about what might happen in the event of a war or another pandemic, he said. "I try to tell them I'm safer here than you guys are in the US, but they don't understand that," he said. He added that Singapore's strict gun laws and low crime rates contribute to that sense of safety. In the US, he felt he had to pay more attention to his surroundings. Those differences have only made Burns more certain about staying. He applied for citizenship in December and is now waiting to hear back. If approved, he plans to renounce his US citizenship — a decision he said he's ready for, and one required in Singapore, which does not allow dual citizenship. "I've always been proud to live here," he said. Read the original article on Business Insider