A Microsoft Outlook outage affected thousands of users' email access on Saturday.The outage led to over 35,000 complaints on Downdetector.Microsoft said it identified the cause, reverted the code, and services are now recovering.Thousands of people were left without access to their email on Saturday after Microsoft Outlook suffered an apparent outage.Over 35,000 complaints were posted to Downdetector, a website for reporting technical issues.Users also posted to social media to complain and warn others."Had to change my password 3 times until I checked on #x that #outlook is down. Wth @Microsoft," one X user wrote.The X account for Microsoft 365, which includes several Microsoft programs like Outlook, Word, and Excel, posted that the company had identified a potential cause of the outage and "reverted the suspected code to alleviate impact."Microsoft said in a following post that the majority of the services affected by the outage were recovering after the fix, but that developers will "keep monitoring until the impact has been resolved."Several tech companies have seen outages over the last year. Most recently, the popular workplace communication platform Slack went dark for several hours on Wednesday.One of the largest global tech outages in history occurred in June 2024. A Microsoft Windows update atcybersecurity firm CrowdStrikegrounded flights worldwide and caused chaos at hospitals, which couldn't access patient records and appointment systems. Microsoft did not immediately return a request for comment from Business Insider.

The Golden Globes will no longer pay a $75,000 salary to its voting members.A Golden Globes spokesperson said it made the change to prevent the perception of bias.Other major awards like the Oscars and Emmys do not pay voters a salary.The voters who decided which films and television shows would receive a Golden Globe were for years paid for their service.Not anymore.Golden Globes President Helen Hoehne told the 50 members affected by the decision in a Zoom call on Friday that they would no longer receive their $75,000 annual salary,A Golden Globes spokesperson confirmed the decision to Business Insider. "The Golden Globes describes the change in policy as an acknowledgment that continuing to pay members could add to a perception of bias in voting," the spokesperson said.The 50 voting members who were receiving salary payments were former members of the Hollywood Foreign Press Association, the nonprofit group of entertainment journalists that built the Golden Globes into a major awards powerhouse, rivaling even the Oscars.However, after a series of scandals in 2021 rocked the HFPA, causing NBC to pull the Golden Globes from the air, Hollywood media firm Penske Media Eldridge acquired the group and turned the Golden Globes into a for-profit entity. The program returned to television in 2023.The spokesperson said that the voting body for the Golden Globes now numbers 300 and is made up of entertainment journalists from 85 countries.Other popular Hollywood awards shows, like the Academy Awards, the Emmys, and the Grammy, do not pay their voting members a salary.

It is the season for affordable smartphones to duke it out among one another. We had Apple, were expecting Google, and Samsung just threw its roster into the ring. The Galaxy A-series is set to be refreshed next month. It starts with the Galaxy A36 starting at $400, then the Galaxy A26 at $300. The Galaxy A56 starts at $500 and will also get a bump, though theres no expected release date yet. All three devices have Awesome Intelligence, which seems to be Samsungs rebranding of Galaxy AI for its lower-cost devices. The Galaxy A56, A36, and A26 all come with the same size screen, though they each offer a different tier of essentials. Each phone has a 6.7-inch Super AMOLED display with a 120Hz refresh ratenot a standard spec at this price pointand a 5,000 mAh battery to accommodate it. All phones start with 128GB of storage. Samsungs Exynos 1580 chipset powers the Galaxy A56, while the Galaxy A26 and A26 run Qualcomms Snapdragon 6 Gen 3 platform. Only the Galaxy A56 offers 8GB of RAM, while the A36 and A26 are stuck with 6 GB. Itll be interesting to see how Awesome Intelligence works, considering the flagships have all gone the opposite way in adopting more RAM to accommodate AI. Software updates are also locked in here; Samsung promises six years of support and Android OS updates. The camera systems in the A-series all vary, but only bit by bit. The Galaxy A56 considered the highest tier of the trio, has a 12-MP ultrawide camera, a 50-MP primary camera, a 5-MP Macro camera, and a 12-MP front-facing camera. The Galaxy A36, starting at $300, has the same primary camera but is dialed down to an 8-MP ultrawide camera and a 5-MP macro camera. Theres also a 12-MP front camera. The Galaxy A26, the cheapest at $300, has all that except a 20-MP macro camera and a 13-MP front camera. Awesome Intelligence What exactly is awesome intelligence, you are likely asking yourself? Its Samsungs attempt at rebranding artificial intelligence for its budget lineup of phones. Its good for Samsung to come up with a way to differentiate its product from Galaxy AI, as that sounds more forward-thinking than being awesome. Awesome Intelligence will let you use key Google Gemini abilities, like Circle to Search, including the new ability launched to the Galaxy S25 series that helps you identify a song. Itll also give you access to Galaxy AI features like Object Eraser. Beyond a few other key Samsung features, like Nightography algorithms in the camera app, thats about it. Its enough Awesome Intelligence to make you say, Awesome! or something like that. This is a slightly different approach from Apple and Google, which use the same processor for their flagship and lower-cost devices to offer parity of software across the ecosystem. Samsung chose instead to lean into where its manufacturing prowess remains strong. Thats why youll see killer screens and bigger batteries on Samsungs budget-friendly phones. Itll be interesting to see how Samsung continues paring down AI features from flagships to the lower-cost subset and what it will deem essential features for Awesome Intelligence.

At the end of Kingdom Come: Deliverance 2's first act, Henry, Godwin, Hans Capon, and Jan Zizka are captured by Otto von Bergow and Istvan Toth. While Hans goes ahead with Otto von Bergow, the rest of the group is taken back to Trosky and tortured for information.

Its no secret that One Piece is one of the biggest franchises in Japanese pop culture, both in Japan and internationally. With fans all around the world, it comes as no surprise that One Piece is starting to have global events beyond just those that are in Japan. This year, for the first time, there is going to be a One Piece Day celebration in America, which will be totally accessible for fans who speak English. This is a huge deal in the anime world; an event outside of Japan, in English, that is dedicated to not all Japanese pop culture in general, but one specific franchise.

A pop-up message on some Samsung TVs won't stop popping up but there may be a solution.

After months of waiting, AMD has finally unveiled the Radeon RX 9070 series GPUs and their pricing, where its XT variant looks set to be the winner of this generation's budget GPUs.

Some things seem pretty straightforward but require precision. One of the many such intricate little things is the act of placing the needle on a vinyl record. You can argue otherwise, or simply ditch the idea and take home this deftly created turntable by French studio, Waiting for Ideas. The makers say, This isnt just another turntable; its a revolutionary plug and play experience, and we couldnt agree more.Dubbed the PP-1, the new turntable takes away the complexity from analog music and replaces it with a subtle change to topple the music experience upside down. Being a true audiophile turntable, it is made without the tonearm on top. So, you are spared the effort of placing the needle to set the music flowing on the platter. Where did the needle go, you ask?Designer: Waiting for IdeasThe little needle is placed on the fascinating platter of the PP-1. When you place the record on it there is a special way to do that the protective cover opens up and the needle automatically aligns with the record. Since the needle is on the underside, PP-1 turntable requires you to place the vinyl disc carefully upside down in the groove cut out on the PP-1. By the very process, the turntable is plug and play: Simply connect it, set the record in place, and hit the play button. Any complexity whatsoever is out of the equation and you have the easiest way to play your favorite records without moving a finger. Over the convenience and functionality, the turntable from Waiting for Ideas has a design to drool over and it even has a 3,5 mm Mini-Jack; nostalgic!It is made from a single, square block of aluminum which has been precisely chiseled to make a cutout on top where the vinyl would go and sit flush against the height of the bluish-green turntable. It measures 35cm x 35cm x 7cm and can be used placed flat on a surface, or vertically on one side. In addition to minimizing the need to place the tonearm, the PP-1 also eliminates the guesswork. Using advanced technology, the turntable can automatically detect the 33 or 45 record speed when the record is inserted.Waiting for Ideas has not just reformed the appearance and the style of record playing, it in fact means to redefine the sound output as well. To that accord, it is throwing in a set of speakers with the PP-1, which you can choose to use together or separately for a stereo sound experience. The speakers are finished in delightful black fabric to complement the aluminum turntable. If you want to embrace the digital ease with the depth of analog sound, you may just have to put together a few months of savings. The USB-C powered turntable is available for 5,800 (roughly $6,000). With the speaker pair, the price shoots up to 9,000 (approx. $9,500).The post Waiting For Ideas PP-1 Turntable attractively aligns digital ease with depth of analog sound first appeared on Yanko Design.

Whitehead's trogon (Harpactes whiteheadi) is a species of bird in the family Trogonidae. It is endemic to the island of Borneo, where it is an uncommon resident in primary mountain forest. One of Borneo's largest trogons at 29 to 33 centimetres (11 to 13 inches) long, it is sexually dimorphic. The male is crimson on the head, nape, and underparts, with a black throat and grey chest; the rest of its upperparts are cinnamon-coloured. The female is similarly patterned, but cinnamon-brown where the male is scarlet. The species is primarily an insectivore, but also eats various plant materials, including fruits and seeds. Other than the timing of its breeding, typically between April and June, little is known about its breeding biology. It is classified as a near-threatened species, with population numbers thought to be declining and habitat loss a key threat. This male Whitehead's trogon was photographed perching on a branch on Mount Kinabalu in the Malaysian state of Sabah.Photograph credit: JJ HarrisonRecently featured: The Palace of TruthContinental XI-1430Ctenosaura similisArchiveMore featured pictures

Large Language Models (LLMs) have advanced significantly, but a key limitation remains their inability to process long-context sequences effectively. While models like GPT-4o and LLaMA3.1 support context windows up to 128K tokens, maintaining high performance at extended lengths is challenging. Rotary Positional Embeddings (RoPE) encode positional information in LLMs but suffer from out-of-distribution (OOD) issues when applied beyond their pre-trained limits. These OOD values appear in higher-dimensional RoPE embeddings, leading to degraded performance. Longer context windows are essential for AI applications such as multi-turn conversations, document analysis, and long-form reasoning. LLMs struggle with efficiency and accuracy when scaling beyond their default lengths without an effective extension method.Most existing methods for extending context windows rely on heuristic-based RoPE rescaling, which fails to address OOD issues fully and often falls short of the target effective context length. Approaches like YaRN, NTK, and LongRoPE derive rescaling factors from theoretical models, but real-world testing reveals significant performance trade-offs. For example, when LLaMA3.1 extends its context window using YaRN, performance drops sharply beyond 64K tokens, as shown in the RULER benchmark. Extending context length also frequently reduces short-context performance, making these methods impractical for short- and long-range processing applications. This issue is particularly severe in models like Phi3-mini-3.8B, where naive RoPE extensions reduce MMLU scores by 7.56 points.Researchers from Microsoft have introduced LongRoPE2 to overcome these limitations. LongRoPE2 is designed to extend the context window of LLMs to 128K tokens while preserving over 98.5% of short-context accuracy. It achieves this by addressing three core issues. First, the research team hypothesized that higher RoPE dimensions receive insufficient training, leading to unexpected OOD values when extending token positions. To mitigate this, LongRoPE2 introduces a needle-driven perplexity (PPL) evaluation that specifically targets tokens that require deep contextual understanding, unlike traditional perplexity measures that fail to distinguish between essential and non-essential tokens. Second, LongRoPE2 adopts an evolutionary search-based RoPE rescaling algorithm, which optimizes rescaling factors beyond theoretical assumptions, ensuring better alignment with extended contexts. Finally, it incorporates mixed context window training, in which the model is fine-tuned on both short and long sequences, thereby preventing performance loss on short-context tasks while ensuring effective long-context adaptation.The technical approach of LongRoPE2 begins with identifying the true critical dimension in RoPE embeddings. The study found that theoretical critical dimensions underestimate the true RoPE scaling needs, as evidenced by empirical observations where RoPE dimensions required larger-than-predicted scaling factors for optimal performance. This led to the development of an adaptive rescaling method that fine-tunes RoPE scaling factors using an iterative evolutionary search. Unlike previous static scaling methods, LongRoPE2 dynamically adjusts rescaling based on per-token perplexity evaluations, ensuring embeddings remain within the pre-trained range while maximizing their effectiveness in long contexts. The algorithm identifies the optimal rescaling factors for higher RoPE dimensions while applying NTK scaling to lower dimensions, ensuring a smooth adaptation process. This method effectively extends LLaMA3-8B to 128K tokens, maintaining over 97% of its short-context accuracy while outperforming prior methods on long-context benchmarks.Performance evaluations reveal LongRoPE2s superiority across various benchmarks. Extensive Phi3-mini-3.8B and LLaMA3-8B testing shows that LongRoPE2 achieves state-of-the-art RULER, LongBench, and InfiniteBench results. On the RULER benchmark, which evaluates LLMs long-context processing capabilities, LongRoPE2 extended LLaMA3-8B to 128K tokens while retaining a score of 82.03, compared to 73.40 for LongRoPE and 49.39 for YaRN. Phi3-mini-3.8B exhibited even greater improvements, with a 128K token score of 58.81, significantly outperforming NTK, which only achieved 49.37 at the same context length. One of the most striking findings was that Metas approach required 800B training tokens to reach 128K tokens, whereas LongRoPE2 accomplished this using only 10B tokens, an 80x efficiency gain. In addition, LongRoPE2 achieved near-perfect accuracy in the Needle in a Haystack pressure test, demonstrating its ability to retrieve deeply embedded information in long sequences, where previous methods such as NTK often failed at extended lengths.One of the key takeaways from this research is that extending LLM context windows is not just a matter of increasing token length but requires addressing fundamental limitations in positional embeddings. The findings suggest that higher RoPE dimensions are insufficiently trained, necessitating adaptive scaling rather than fixed rescaling factors. The needle-driven PPL evaluation proved crucial in identifying the optimal RoPE scaling factors, ensuring that models maintain accuracy in long-range dependencies. The mixed context window training technique ensured that models retained over 97.6% of their short-context performance, making LongRoPE2 the first near-lossless extension method. Also, LongRoPE2s evolutionary search for RoPE rescaling factors revealed that prior analytical methods underestimated scaling needs in high-dimensional embeddings, leading to suboptimal performance in previous approaches.Some Key Highlights from the Research include:LongRoPE2 successfully extended LLaMA3-8B to 128K tokens with 82.03% accuracy, surpassing all previous methods.Unlike Metas approach, which required 800B training tokens, LongRoPE2 achieved the same extension using only 10B tokens, making it 80x more efficient.The model retained 97.6% of short-context performance, whereas prior methods degraded significantly.The needle-driven perplexity evaluation introduced a novel method for determining optimal RoPE rescaling factors, allowing precise adaptation to long contexts.On the RULER benchmark, LongRoPE2 scored 82.03 at 128K, compared to 73.40 for LongRoPE and 49.39 for YaRN.The model achieved near-perfect retrieval accuracy in the Needle in a Haystack test, significantly outperforming NTK-based approaches.LongRoPE2 demonstrated that adaptive evolutionary search-based scaling is far superior to static rescaling techniques.Check outthe Paper and GitHub Page.All credit for this research goes to the researchers of this project. Also,feel free to follow us onTwitterand dont forget to join our80k+ ML SubReddit. Asif RazzaqWebsite| + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/IBM AI Releases Granite 3.2 8B Instruct and Granite 3.2 2B Instruct Models: Offering Experimental Chain-of-Thought Reasoning CapabilitiesAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Google AI Introduces PlanGEN: A Multi-Agent AI Framework Designed to Enhance Planning and Reasoning in LLMs through Constraint-Guided Iterative Verification and Adaptive Algorithm SelectionAsif Razzaqhttps://www.marktechpost.com/author/6flvq/DeepSeek AI Releases Fire-Flyer File System (3FS): A High-Performance Distributed File System Designed to Address the Challenges of AI Training and Inference WorkloadAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Cohere AI Releases Command R7B Arabic: A Compact Open-Weights AI Model Optimized to Deliver State-of-the-Art Arabic Language Capabilities to Enterprises in the MENA Region Recommended Open-Source AI Platform: IntellAgent is a An Open-Source Multi-Agent Framework to Evaluate Complex Conversational AI System' (Promoted)