OpenAI announced that ChatGPT can now connect directly to Apple Messages on Mac. The integration appears as an Apple Messages plugin that enables the AI to search message threads, review recent conversations, and draft or send replies within the Messages app. According to OpenAI, the feature lets users catch up on discussions, locate specific information, and compose responses using the AI assistant. A demonstration video posted by OpenAI shows the workflow in action. The update expands ChatGPT's functionality within the Apple ecosystem and may affect how users interact with AI on macOS.

US Bank is investigating claims by the ransomware group LockBit that it breached its systems and stole data. The group has set a September 3 deadline to receive a ransom payment or publish the data.

LockBit is a ransomware group that faced law enforcement actions in 2024 and 2025, when its servers and infrastructure were seized and its leader’s identity was disclosed, though the suspect remains at large. The group has previously threatened to release stolen data unless victims pay.

US Bank said it has no evidence that its internal systems were compromised and has not confirmed communication with the extortionists. A spokesperson declined to disclose the ransom demand amount or whether the bank has engaged with the criminals. The group posted the bank on its leak site on Wednesday night and gave the bank 14 days to pay. The alleged breach was first detected on May 7, when the bank learned of a third‑party security incident involving its vendor Fidelity National Information Services. In June the bank notified 537 Massachusetts customers that their names, mailing addresses, and credit card numbers may have been exposed. A separate incident in 2022 affected about 11,000 customers after a vendor inadvertently shared a file containing personal information from closed credit card accounts.

US Bank continues to monitor the situation and says no systems are currently impacted. A law firm is considering a class‑action lawsuit on behalf of customers whose credit card data may have been exposed. The deadline approaches with the outcome still uncertain.

Linux 7.2 was released this week on the regular schedule. The update adds scheduler, power management, and graphics enhancements. The kernel development cycle has become more frequent, with the last four releases showing increased activity in bug fixing and feature addition. Cache‑aware scheduling was introduced. MGLRU improvements were made. Sub‑schedulers for sched_ext were added. Automatic creation of multi‑size transparent hugepages was implemented. The DRM scheduler fair policy was planned for enablement but was postponed after a regression was reported during the release candidate period; the default first‑in‑first‑out scheduler remains in place. Runtime power management for the Raspberry Pi 4 and 5 GPU was implemented, allowing the GPU clock to be disabled when idle. The Raspberry Pi 3 GPU driver was fixed to resolve persistent hangs and crashes in RetroPie. GPU reset handling on the Raspberry Pi 4 and 5 was made more reliable. A 14‑year‑old futex bug affecting robust list operations was resolved. Various other bug fixes addressed issues in the ueagle‑atm driver, memcmp implementation, and HDMI 2.1 fixed‑rate link support. Numerous smaller patches were applied across the DRM, scheduler, and networking subsystems. The kernel community expects the fair policy to be re‑enabled in a future release, and further improvements are planned for upcoming kernel versions.

On August 20, 2026, a compromised release of the Rust crate arrayref version 0.3.10 was published on crates.io. The release added a dependency on a compromised crate named proc-macro1, whose build script downloaded and executed a remote binary during project compilation. Crates.io removed the malicious versions after detection. Arrayref is a small crate that provides four macros and has no build script or runtime dependencies prior to version 0.3.10. The genuine arrayref and append-only-vec crates are maintained by a user identified as droundy, whose GitHub account and repositories are no longer accessible. A separate account named dtolney published the proc-macro1 crate, using an email address that does not belong to David Tolnay. The malicious crate was removed by crates.io. The build script in proc-macro1 version 1.0.107 contains a server address encoded in base64 fragments and reassembles the URL at compile time. Decoded, the address points to hxxps://23[.]254[.]165[.]112:9089/ and the command-and-control address 23[.]254[.]165[.]112:443. The script fetches an architecture-specific binary over a TLS connection that accepts any certificate, then writes the file to /tmp/rust-setup on Unix or to %TEMP% on Windows and executes it without waiting for the build to finish. On Unix the binary is made executable and launched with the command and control address as an argument; on Windows a PowerShell script is launched via a VBScript wrapper. The build script includes build-dependency crates for base64 decoding, TLS handling, and HTTP requests, which are atypical for a token-parsing library. The dependency resolves to the malicious version because only versions 1.0.106 and 1.0.107 were published. Arrayref has recorded approximately 245 million downloads, with version 0.3.9 accounting for about 152 million. Because arrayref is used as a transitive dependency in widely adopted crates such as tiny-skia, sctk-adwaita, and winit, the supply chain impact reaches many GUI applications built with egui, eframe, and iced. Developers who compiled projects using the yanked older versions 0.3.5 through 0.3.9 received a warning prompting an update to 0.3.10, which introduced the malicious dependency. The Rust Security advisory attributes the detection to community reporting. The incident demonstrates the risk posed by compromised build-time dependencies and underscores the need for stricter verification of crate provenance and dependencies in the Rust ecosystem.

A former Meta safety engineer testified in a federal trial in Oakland, California, alleging that the company was aware of child safety risks on its platforms and failed to act. Arturo Béjar, who worked at Meta for eight years, told the jury that he repeatedly warned executives, including CEO Mark Zuckerberg, about harmful content recommendations and the impact on teenage wellbeing. He said he sent at least 100 briefings to Zuckerberg and an email in 2021 detailing reports of sexual predator content, graphic images, and damage to teen health.

Béjar testified that he emailed Zuckerberg after the CEO publicly stated that the company did not prioritize profit over safety, saying he wanted to correct a false impression of Meta’s commitment to young users. When asked whether Zuckerberg responded, Béjar replied that he did not hear back. The testimony was part of a case brought by 29 U.S. state attorneys general, who allege that Meta deliberately designed addictive products to attract young people and that it collected data on children under 13 without parental permission, violating federal and state laws.

The legal complaint also seeks to force Meta to redesign its products to make them safer for children, a change that could affect the company’s business model. Meta denies the allegations, stating that it provides tools to address social media challenges and that no child under 13 can create an account. Paul Schmidt, Meta’s attorney, argued that the company has disabled more than one million accounts of young users and that there is no dispute that people can struggle with social media.

Béjar’s two‑day testimony included data showing that 51% of teen users reported harmful experiences in the previous week, yet content was removed in only 0.02% of those cases. He said features such as infinite scroll increase ad revenue, and that his motivation for raising the issue began when his own teenage daughter received unwanted sexual advances and explicit images on Instagram.

The trial, expected to last at least six weeks, will also hear from Mark Zuckerberg and Instagram CEO Adam Mosseri, along with experts on children’s mental health and addiction. The outcome could result in damages up to $200 billion, equivalent to Meta’s projected 2025 annual revenue, and may compel significant product changes.

Apple's Beats 360 headphones were listed in Brazil's ANATEL regulatory database on August 20, 2026, according to a filing. The entry identified the product by model number A3577, which matches a previously filed FCC record from May. The filing did not disclose a product name, but other sources have identified the item as the Beats 360. Amazon and other reseller sites listed the headphones in Black, Cloud, Sky, and Pink, and indicated availability of Performance Knit and UltraPlush ear cushions and headbands in those colors as well as Volt and Navy, according to the listings. The product description said the over‑ear headphones provide up to 1.75 times the noise cancellation of the Beats Studio Pro released in 2023, and they feature an IPX4 rating for sweat and water resistance, all‑day battery life, a fold‑flat design, and a high‑fidelity acoustic platform, the listing noted. Beats Studio Pro are priced at $349.99, while a listed price for the Beats 360 is $299.99, though this figure has not been confirmed. Apple has not announced a release date, but the multiple regulatory filings, athlete sightings and early retail listings indicate a launch in the near future. MacRumors was the first to report the Beats 360 name after seeing it in various retail listings, and the information was subsequently corroborated by the regulatory filings.

In July, the Federal Communications Commission prohibited the importation of new models of foreign‑made robots, including robot vacuum cleaners and Unitree humanoid and quadruped robots. The ban, cited as a national‑security measure due to potential cybersecurity risks, also aimed to spur domestic robot production.

The U.S. distributor, operating under the Robo Inc. brand, had halted its distribution operations in June while continuing to support existing customers with parts and repairs. The company said it would shift its focus to developing robots and expanding manufacturing capacity at a new facility.

Under the Robo Inc. banner, the team is working on a wheeled humanoid torso for healthcare and concierge services, equipped with cameras and sensors for vital‑sign monitoring. Additional products include a security‑focused quadruped and a humanoid platform intended for educators and researchers. Universities have expressed interest in a U.S.-compliant offering, the company said.

Engineers have used 3D printing to produce prototypes that are being tested by select customers, and the firm expects to begin commercial sales from the new plant by early 2027.

The company is also preparing to raise its first major round of external funding after years of operating as a bootstrapped business. Haggerty noted that customer interest has increased as the fundraising process advances.

A leak posted by user Golden Pig Upgrade on Weibo said the 28‑core Nova Lake‑S CPU with an 8‑plus‑16‑plus‑4 core configuration is expected to reach a PL2 power limit of 296 watts.

The Nova Lake‑S line is Intel’s next‑generation desktop processor family. Earlier reports indicated a thermal design power (TDP) of 125 watts for the 28‑core single‑tile version. The Core Ultra 9 285K, a current high‑end desktop part, is rated at a PL2 limit of 250 watts.

The leaked configuration adds four cores compared with the 28‑core model that uses 125 watts, and includes 144 MB of bLLC cache. The PL2 figure of 296 watts represents an increase of 18 percent over the 250 watt PL2 rating of the 285K. PL1 and TDP are distinct specifications; the 296 watt figure refers to the higher power limit rather than the nominal TDP.

The design also incorporates a Big Last‑Level Cache (bLLC) intended to compete with AMD’s 3D V‑Cache technology. Leaked specifications suggest up to four bLLC variants with cache sizes up to 288 MB and core counts as high as 52.

Cooling requirements for a 296‑watt PL1 limit on a single‑tile processor remain unclear, and Intel has not released official thermal specifications.

Modern processors achieve high arithmetic throughput, but memory efficiency increasingly determines how much of that performance can be utilized.

Historically, specifications such as CPU clock speed, core count, GPU shader throughput, memory frequency, and storage transfer rates defined performance metrics. While these figures remain relevant, contemporary workloads often spend considerable cycles moving data rather than performing calculations.

These capabilities increase theoretical performance, yet data movement limits actual throughput. A CPU core can execute multiple instructions per clock cycle, and a high‑end GPU contains thousands of parallel units, but workloads that require frequent random memory accesses may be constrained by latency and bandwidth. Sequential access enables efficient prefetching and high throughput, whereas random access hampers prediction and exposes latency, especially when subsequent addresses depend on prior results.

Programs typically exhibit spatial locality, reusing nearby data, and temporal locality, reusing recently accessed data. Caches store data in fixed‑size blocks called cache lines, allowing reuse of neighboring data and reducing accesses to slower main memory. The working set—a subset of allocated memory actively used—must fit within cache capacity to achieve optimal performance; exceeding cache size causes additional latency.

Cache hierarchies balance size, latency, bandwidth, and energy. L1 caches are small and fast, L2 offers more capacity at higher latency, and L3 provides the largest capacity and is often shared across cores. Increasing cache size can yield large performance gains when the working set remains on‑chip, while offering little benefit when data are streamed through large memory ranges.

CPU caches are arranged in a hierarchy that trades speed for capacity. Larger caches consume more silicon area and require longer wiring, which can increase latency and energy use. AMD’s 3D V‑Cache technology stacks an additional cache die vertically, adding 64 MB of L3 cache to Ryzen processors, demonstrating how extra cache can improve performance for workloads with large working sets.

When data is not present in cache, it must be retrieved from DRAM. DRAM provides higher capacity at lower cost per bit but incurs higher latency. Memory controllers manage banks, rows, and timing to schedule accesses efficiently, and memory bandwidth and latency are distinct metrics. DDR5 adds more bank groups and longer burst lengths, improving parallelism, though real‑world performance varies with application behavior.

GPUs use GDDR memory for high bandwidth and cost efficiency, while AI accelerators employ HBM, which stacks memory dies and uses very wide interfaces close to the processor, delivering high bandwidth and energy efficiency at greater packaging complexity and cost. A single AMD Instinct MI355X accelerator contains 288 GB of HBM3E and up to 8 TB/s of theoretical bandwidth, and multi‑GPU systems aggregate several terabytes of memory.

Large language models require extensive memory for weights, activations, and intermediate data, making bandwidth and capacity critical for inference and training. Quantization reduces memory size by lowering precision, directly impacting performance when memory traffic limits compute.

Gaming workloads are diverse, involving visibility culling, physics updates, AI processing, geometry rendering, and texture sampling. Large CPU caches can capture frequent game data, reducing DRAM traffic, while GPUs use large last‑level caches to lower external memory accesses. Real‑time ray tracing introduces data divergence as rays traverse bounding volume hierarchies, creating irregular memory access patterns that challenge caching.

Storage performance also illustrates the limits of bandwidth alone. SSDs advertise high sequential throughput, but real applications issue many small, scattered requests, making random IOPS, latency, and queue depth more relevant. Controllers manage wear leveling, garbage collection, and write amplification, and some use fast cache layers to improve short‑term performance.

Overall, performance depends on how an application’s data access pattern interacts with the memory hierarchy. While compute capability continues to rise, efficient data movement—through larger caches, tighter locality, optimized algorithms, and advanced packaging—remains essential to realize the full potential of modern hardware.

Future designs will likely expand cache capacity, improve memory hierarchy depth, and bring memory physically closer to processing units, yet physical limits ensure that reducing data movement stays a central challenge.

A 22-year-old researcher identified as Zerotistic devised a method to enroll a Linux device on Apple’s Find My network and receive location data from individuals who had shared their locations with the researcher’s Apple account. Find My is Apple’s service for locating devices and for people to share their locations, and it is limited to Apple hardware while the iCloud website lacks the people‑tracking feature. The researcher used Apple’s GrandSlam authentication protocol to obtain an identity delegate and then created a custom certificate signing request with a 2048‑bit RSA key signed using SHA‑1. The CSR was submitted to Apple’s legacy authenticateDS endpoint, which signed the request and issued an identity‑services certificate linking the Linux device to the researcher’s Apple account. Apple required the device to subscribe to six sub‑services, define supported encryption types, and provide public keys for Apple’s device‑to‑device messaging format. After enrollment, the device maintained a persistent TLS connection to Apple’s Push Notification service to receive location updates. A SubscribeAndFetch request caused a friend’s Apple device to transmit an encrypted location key to the Linux device. The researcher wrote a script to unwrap the encrypted message, extract the shared location key, and decrypt the location payload containing coordinates, timestamps and accuracy information. The script was able to retrieve subsequent location reports for the existing share. The entire development took less than a week. Apple was contacted for comment and did not respond.

Los Angeles landscaper Christine Ro attended a training session run by the American Green Zone Alliance, an organization that supports the transition to electric landscaping equipment.

Cities across the United States have begun restricting petrol-powered leaf blowers because of noise complaints, and the electric alternatives reduce carbon emissions, vibration and exposure to pollutants while allowing operation during quieter hours.

The shift is creating demand for new manufacturers; Whisper Aero, an aerospace start‑up known for electric propulsion systems for aircraft, introduced a handheld electric leaf blower in 2023 after research into its fan technology.

According to Andrew Terajewicz, Whisper Aero’s director of air management, the device is cleaner, quieter and more efficient than existing air‑moving technologies.

The blower’s battery provides up to 50 minutes of operation at full power and the unit retails at the higher end of electric leaf blowers, which are generally more expensive than petrol models.

Dan Mabe, founder of the AGZA and a former landscaper, said the blower is ultra quiet yet powerful, though the high price and limited runtime remain challenges.

Whisper Aero plans to develop a backpack model for professional landscapers, and the company has increased manufacturing capacity to meet pre‑order demand.

Stihl, the German power‑tool maker, aims for battery‑powered products to represent 35 % of sales in 2026 and 80 % by 2035, and opened a dedicated battery and tool plant in Romania in 2025.

Stihl’s new battery system uses tabless lithium cells, which deliver 60 % more peak power than standard cells, enabling faster charging and lighter packs.

Professional landscapers report that petrol blowers remain more powerful and offer longer runtimes; Carlos Camarena, a supervisor at a California landscaping firm, noted that electric tools are used where local ordinances prohibit petrol equipment, but his crew typically operates six batteries that require 45 minutes each to charge fully.

The upfront investment to convert a business to electric equipment is estimated at about $10,000, and subsidies may not fully offset the cost, especially for sole‑proprietor operators with limited storage space.

Industry observers say the success of electric landscaping tools will depend on continued improvements in battery life, cost reductions, and compliance with local noise regulations.

Genki has launched the Manta, a wireless controller featuring a 2.9‑inch screen and adjustable buttons, through a Kickstarter campaign.

The company previously collaborated with 8BitDo on a gamepad two years ago and has funded eight other hardware projects via crowdfunding.

Genki designed the Manta entirely in‑house, positioning the thumbsticks and face buttons above the screen.

The layout makes the controller appear bulkier than standard designs.

Customization options are accessed via the screen and physical controls, allowing users to modify thumbstick response, dead zones, haptic feedback intensity, and button mappings.

Multiple console, game, and user profiles can be saved and switched on the fly.

The face buttons are analog and their travel depth can be set to adjust sensitivity.

Linear magnetic ram motors in the grips deliver haptic feedback, and drift‑resistant tunneling magnetoresistance thumbsticks provide adjustable tension.

The device includes a 2,600mAh user‑replaceable battery.

Six additional remappable buttons are located on the back and shoulders.

A dedicated button enables switching to keyboard and mouse mode for web browsing.

Connectivity is available via Bluetooth, a 2.4GHz dongle for low‑latency wired use, or a USB‑C cable for direct console connection.

The Manta features a tristage trigger system that adds a middle travel distance for rapid repeated presses, complementing the two‑stage modes found on many higher‑end controllers.

The Kickstarter page lists a discounted preorder price of $189 for early backers and a full retail price of $279.

Genki reports a history of fulfilling Kickstarter orders successfully.

Full retail availability is expected after the campaign concludes, and the controller will support multiple gaming platforms.

NVIDIA is developing a new inference‑focused chip for the Chinese market that will use Groq’s language processing units, according to a report by The Information, with the product expected by year‑end and designed to comply with U.S. export restrictions. The United States has imposed export controls that limit the amount of high‑bandwidth memory and advanced packaging that Chinese AI accelerators may receive, prompting Chinese firms such as Semiconductor Manufacturing International Corporation (SMIC) to see increased demand for domestic chip production. SMIC reported a 93.7% utilization rate and revenue of $3.01 billion in the latest quarter, up 36.1% year‑over‑year, while Chinese companies are exploring alternative AI solutions, including Alibaba’s XuanTie C950 RISC‑V processor and DFSX’s DF1000 and DF2000 chips. NVIDIA entered into a licensing and talent‑transfer agreement with Groq in December 2025, granting access to Groq’s LPU architecture, which clusters numerous specialized cores with large SRAM and hard‑wired matrix and vector units, eliminating the need for traditional memory caches and branch predictors. The compiler plans calculations at the nanosecond level to achieve continuous data flow, enabling fast inference. The Information said the new chip will be ready by the end of the year. NVIDIA has stated that it has no LPU sales in China and that the report of such a plan is inaccurate, though the statement focused on current status and left future options open. The development reflects NVIDIA’s effort to maintain a presence in China while adhering to export regulations, and it underscores how export controls are reshaping the domestic semiconductor landscape as SMIC expands its capacity and Chinese firms introduce homegrown AI processors.

HyperX announced its first open-back gaming headset, the Cloud Alpha Air, at its Level Awakening livestream on August 20, 2026. The headset is priced at $149.99 and will be available in September. The open-back design lets users hear surrounding sounds while gaming. The Cloud Alpha Air weighs 10.1 ounces (287 grams) with a detachable boom microphone. It connects to PCs or consoles via 3.5mm jack, USB‑C, or USB‑A. The price is roughly $50 lower than Sony’s InZone H6 Air open-back headset released earlier this year. HyperX, a subsidiary of Kingston Technology, said the design improves comfort for extended gaming sessions. The company did not provide additional specifications beyond the connectivity options listed. The product will be available in September, expanding HyperX’s open-back headset lineup in the gaming market.

National Geographic, in partnership with the 9/11 Memorial and Museum, will premiere a three‑episode documentary series titled 9/11: Reunited on August 21. The series will be available for streaming on Disney+ and Hulu the following day.

Director Julian Jones, who produced the 2021 NatGeo documentary 9/11: One Day in America, said the new series was inspired by the sense of mutual support he observed in the earlier film. He said the goal is to let participants guide the narrative, to provide gratitude, answer lingering questions, and honor those who did not survive.

Survivor John Paczkowski, who worked on the 67th floor of the North Tower, and janitor George Phoenix, on the 74th floor, were trapped when the first plane struck as their elevator descended. The emergency brake stopped the car 15 floors below the impact point. They pried the doors open, and Phoenix used a squeegee to cut a hole in the wall, allowing them to escape to a restroom on the 50th floor before the tower collapsed.

Subway rider Debbie St. John was struck by debris from the second plane while walking near the South Tower. Bond trader Stefan Horlacher, who was nearby, rescued her and later reunited with her for the documentary. St. John underwent 50 surgeries over 18 months and later married and had children.

At the Pentagon, Lt. Col. Marilyn Wills survived a fireball that singed her bangs and led a group of survivors down a corridor before jumping from the second floor. Cmdr. Craig Powell caught her as she descended. Powell later learned she had been treated for leukemia linked to rescue work at the site.

FDNY firefighter Kevin Shea, newly assigned to Ladder 35, was outside when the North Tower collapsed. Paramedic Al Kim found him in the South Tower Marriott lobby, stabilized his neck injuries, and transported him to a hospital where he recovered from multiple fractures.

NYPD Emergency Services Unit member John D’Allara died when the North Tower fell. FDNY Rescue 5 member Bill Spade survived, later contacted D’Allara’s widow and arranged a reunion for his sons, noting that D’Allara had saved 20 lives that day.

The documentary also revisits the rescue of Port Authority Police Sergeant John McLoughlin and two colleagues trapped under rubble for more than 20 hours; McLoughlin survived multiple surgeries and now sees his grandchildren.

Jones said the series is challenging to watch but hopes it will remind viewers that people turned to one another in the aftermath of the attacks.

9/11: Reunited premieres August 21 on National Geographic.

Hanwha Securities said CXMT contributed 11.3% of global DRAM growth in 2026 and is projected to account for 6.9% of growth in 2027.

The DRAM market is expected to grow 22.4% in 2026 and 24.1% in 2027, according to the report.

CXMT's Q2 revenue rose 716% year‑over‑year, driven by domestic demand and its capacity expansion plan.

The firm's production increase helped narrow the DRAM sufficiency ratio to a negative 9.4% in 2026 and to minus 5.6% in 2027.

Without CXMT, the ratio would be minus 15.5% in 2026 and minus 12% in 2027.

Analysts said the DRAM shortage is likely to persist through 2026 and 2027.

The continued shortage suggests that additional capacity additions may be needed to balance supply and demand.

Aaron Swartz, a co‑creator of the RSS protocol, faced criminal charges in 2011 for downloading roughly 70 GB of academic journal articles from JSTOR, a case that could have resulted in a prison term of up to 35 years and a $1 million fine before his death in 2013. Federal prosecutors charged Swartz under the Computer Fraud and Abuse Act for allegedly exceeding authorized access to JSTOR’s network. The indictment alleged that he downloaded a large portion of the journal archive for research and distribution purposes. Swartz’s defense argued that his intent was to make scholarly material freely available. The case attracted widespread public attention, with advocates for open access and digital rights describing the legal action as disproportionate. In 2020, reports indicated that Meta had harvested approximately 80 TB of digitized books to train its artificial intelligence models. The company has not faced comparable criminal prosecution. Legal filings indicate that Meta’s activities were conducted under different circumstances and have not resulted in charges comparable to those brought against Swartz. Scholars and legal experts have offered differing views on the fairness of the two cases. Some argue that Swartz’s prosecution reflected broader enforcement priorities, while others note that the scale and nature of the alleged conduct differ between the two incidents. Swartz died by suicide in 2013, before the case was resolved. His death has been cited in discussions about the impact of aggressive prosecution on individuals. The legal outcomes of both incidents remain subjects of ongoing debate regarding the application of computer‑fraud statutes to large‑scale data collection.

Google on Thursday announced a new “Preferred Sources” button that publishers can embed on their websites to let readers highlight favorite sites in Google Search, Discover and Google News. The feature follows a rollout that began in May when the same button became available within Google’s AI Mode and AI Overviews, after initially appearing in Top Stories. Google reported that more than 345,000 unique sources had been selected by users as of May. Publishers can add a site as a preferred source by visiting Google’s source preferences page and searching for the publisher’s name or website. Google said the button can increase traffic, noting that users are twice as likely to click through to a preferred source when it is available. The update is intended to mitigate traffic losses that have resulted from the rapid growth of AI‑driven search features. Alongside the button, Google said readers will soon be able to customize their Discover feed by entering natural‑language commands in the three‑dot menu, and Android users will be able to adjust audio daily briefings in the Google News app. Google said the changes are part of its effort to address publisher concerns about declining traffic caused by AI features while giving users more control over the content they see.

Huzzah is an experimental editor that lets developers write pseudocode and have it turned into functional code by a language model. The tool was presented in August 2026 by Daniel Vaughn. Developers said they feel fatigue with current AI coding agents that rely on long step‑by‑step prompts. Those prompts lose the original human intent and consume many tokens. Huzzah replaces them with persistent declarative pseudocode that can be saved and edited. A simple fizzbuzz example shows the difference. In the agent workflow the user writes a prompt, receives code, then sends follow‑up edits. In Huzzah the user writes a file containing pseudocode such as loop 100 followed by conditional expressions. Saving the file triggers a diff‑based prompt that the model uses to regenerate the code. The approach makes the interaction more concise and serves as documentation of intent. The pseudocode can be language‑agnostic and reused for multiple targets. Benefits include clearer intent, reduced token usage, and easier maintenance. Caveats include potential limits at scale, suitability for new projects rather than existing codebases, and the need for domain expertise. Cross‑file dependencies may be difficult to express and LSP features are not yet available. The project is actively being developed and is currently in an experimental state. Source code and setup instructions are publicly available.

Researchers have found little evidence that higher intelligence scores translate into greater happiness. A nationally representative UK study reported that individuals with lower scores on intelligence tests were modestly less happy than the general population, while a 50‑year analysis of the General Social Survey involving 30,346 participants showed a weak negative correlation (r = -.06, p < .001) between vocabulary test scores and self‑reported happiness.

The concept of intelligence has been examined for more than a century. Charles Spearman observed in 1904 that performance on disparate academic subjects was positively correlated, leading him to propose a general mental ability underlying varied cognitive tasks. Subsequent research has replicated the positive manifold, showing that people who score well on one type of cognitive test tend to score well on others, though the degree of correlation varies.

Critics argue that Spearman’s interpretation overstated the uniformity of the tasks measured. While well‑defined problems—those with clear instructions, repeatable solutions, and unambiguous answers—are effectively captured by standard intelligence tests, many everyday decisions are poorly defined, lacking stable rules or measurable outcomes. Examples such as choosing a life partner, planning long‑term personal goals, or managing interpersonal relationships illustrate problems that are not amenable to conventional testing.

AI development has progressed largely on well‑defined problems. Systems such as GPT‑3 and DALL‑E achieve high performance by predicting text or matching prompts to images, tasks that have clear criteria for success. In contrast, solving poorly defined problems, which involve creativity, context, and subjective judgment, remains beyond current artificial intelligence capabilities.

High‑profile individuals with high test scores have sometimes made controversial statements, for example claiming that the September 11 attacks were an inside job, denying the Holocaust, or asserting that certain groups are responsible for societal ills. These statements are attributed to the individuals involved and are presented as examples of the limits of problem‑solving ability in ambiguous contexts.

Scholars note that progress on well‑defined problems such as eradicating infectious diseases or landing on the moon has not corresponded with increased overall happiness, suggesting that technical mastery alone does not guarantee personal fulfillment. Ongoing research aims to identify factors that contribute to lasting well‑being beyond the satisfaction derived from solving defined tasks. The article concludes that while intelligence predicts performance on structured tasks, its relationship to happiness remains modest and the challenge of addressing poorly defined life problems persists.

LG Display announced FLiPP, a new OLED manufacturing method that eliminates the need for fine metal masks during pixel patterning, aiming to boost efficiency, brightness and durability. The technology will be presented at IMID 2026 in Busan beginning this week. OLED panels have become dominant in television and monitor markets, but traditional fine metal mask production can cause colour mixing defects on larger substrates. FLiPP uses photolithography and UV etching to apply RGB pixels directly onto 8.5‑generation mother glass. LG said FLiPP raises pixel aperture ratio by about 55% and can deliver up to 60% higher peak brightness while extending panel lifespan by 140% and reducing power consumption by 13% compared with conventional FMM processes. The method also improves material utilisation by as much as 64% and lowers production costs across various panel sizes. In addition, LG Display confirmed that its Hyper Double Scanning (HDS) panel‑driving technology, which won the IMID Display of the Year Award, is already deployed in its mass‑produced OLED TV lineup. The advancement represents a step toward greater manufacturing efficiency, but it does not indicate an immediate shift in OLED pricing trends.

GeForce Now added 12 games to its cloud streaming service this week and introduced native Firefox support on Windows, allowing users to stream without installing the app. The service, launched by Nvidia, streams PC games from its servers to devices. Previously, users required the GeForce Now desktop or mobile app. The new browser capability follows earlier support for Chrome and Edge and aligns with the service’s push to increase accessibility. The update includes titles such as Stars Reach and The Sinking City 2, both released on Steam on August 18, Gallipoli, released on Steam on August 20, DIVE or DIE – Children of Rain, Escape the Backroom, Hell Is Us, High on Life, High on Life 2, Parcel Simulator, Starvester and The Thaumaturge. Escape the Backroom, Hell Is Us and High on Life 2 are available through Xbox Game Pass. Tomb Raider IV–VI Remastered is offered on the Epic Games Store. All games support cloud saves across devices, and Ultimate members can stream at up to 1440p resolution and 120 fps. The expansion of browser streaming and the addition of new titles aim to increase accessibility and user engagement.

TechCrunch announced its 2026 Startup Battlefield 200, a curated list of 200 early-stage startups selected from thousands of applications worldwide. The companies will showcase their products at TechCrunch Disrupt 2026, scheduled for October 13-15 at the Moscone West venue in San Francisco.

The list includes startups from diverse sectors such as agtech, automotive, biotech, consumer, cybersecurity, e-commerce, education, energy, enterprise technology, entertainment, fintech, government, health, human resources, logistics, manufacturing, marketing, real estate, smart cities, space, and defense. Companies presented in the categories are listed in the announcement, representing a broad range of technologies and business models.

The competition offers participants a chance to exhibit at the conference, pitch to venture capital firms, and compete for a $100,000 equity‑free prize and the Disrupt Cup. Previous alumni have raised more than $32 billion and completed 250 exits, with notable graduates including Dropbox, Cloudflare and Discord.

A total of 20 finalists will be selected from the 200 for live presentation on the first day of Disrupt. Founders who made the list will receive follow‑up instructions, while applicants not selected will receive a formal update. The event is presented by Google for Startups and SAP.

Registration for Disrupt 2026 remains open, with ticket prices scheduled to increase in September. Attendees can contact Isabelle Johannessen at isabelle.johannessen@techcrunch.com or Michael Schick at michael.schick@techcrunch.com for outreach verification.

Anti‑AI fonts that scramble or obfuscate text do not provide a durable solution to AI text parsing and create accessibility challenges. Screen readers and other assistive technologies must read the same characters that human users see, making scrambled text difficult or impossible to interpret. The approach also raises privacy and security concerns, as verifying disabled users could require centralized identity systems. While some demonstrations use motion graphics or video, these are impractical for typical web pages. AI developers are already training models to bypass such obfuscations, and any new method is likely to become another obstacle that is eventually broken. The trend could lead to copy‑protection mechanisms that limit open access to information, contrary to the web’s founding principle of free and open information. No definitive fix is expected, and all publicly available content will eventually be reachable by AI systems.

GuidePoint Security reported that a ransomware affiliate called "Ransom Busters" has been contacting victims before attacks become public, offering to recover encrypted files and delete stolen data for a payment lower than the original extortion demand. Researchers said the affiliate contacted victims after the ransomware gangs had already encrypted files and demanded payment, presenting itself as a recovery service. The outfit claimed it had accessed the ransomware gangs’ own servers, located the stolen data, and could delete it and retrieve encryption keys for a price between $20,000 and $60,000. Forensic analysis of two incidents revealed that the same attacker used SoftPerfect Network Scanner for reconnaissance, s5cmd to upload data to AWS storage, and the Remotely remote‑management tool installed via PowerShell. Both attacks also created a local backdoor account with the password "Numlock!123" and used the hostname "DESKTOP‑BBETH6K" in the intrusion. GuidePoint said the consistent tool usage and host identifier across separate ransomware operations suggest a single affiliate is operating in multiple RaaS campaigns and cutting out its original partners. The organization warned that paying the supposed rescuers does not guarantee the removal of stolen data, leaving victims vulnerable to further extortion. GuidePoint noted that the scheme highlights the difficulty of attributing ransomware activity and underscores the need for victims to verify any party claiming to recover encrypted files.

AliExpress homepage creates two hidden WebAudio contexts that keep a device’s Bluetooth audio path active, preventing multipoint headphones from switching to a phone. The issue occurs when the page is loaded in Firefox or Chrome, after which audio from a connected phone stops playing while audio from a PC continues. The hidden contexts are instantiated by two obfuscated JavaScript files hosted on aliexpress-media.com, identified as collina.js and fireyejs.js. Both scripts construct a WebAudio graph consisting of a sawtooth oscillator, an analyser node, a script processor, and a gain node set to zero, then connect the graph to the system audio destination. Although no audible sound is produced, the browser continues processing the audio graph, which appears to maintain the Bluetooth link. Testing with developer tools showed that the AudioContext objects were created only after the page had been idle for several seconds. The stack traces linked the creation to the two scripts. The audio fingerprinting routine also queries canvas, WebGL, hardware, and timing data, and transmits encrypted results to Alibaba telemetry services. Blocking the two script URLs with uBlock Origin prevents the contexts from being instantiated and restores normal Bluetooth switching. The problem is specific to the AliExpress homepage and does not affect ordinary product browsing. AliExpress has not publicly explained the purpose of the audio processing, but it aligns with broader anti‑fraud and device‑identification efforts that rely on extensive browser and device measurements. Users experiencing the issue can resolve it by disabling the identified scripts.

In the 1980s, the CIA provided financial assistance to NeXT, the computer company founded by Steve Jobs after he left Apple, helping it stay afloat during a period of financial difficulty. NeXT was established in 1985 after Jobs departed Apple and pursued a hardware and software development strategy that required investment. The agency's support, which was reported by The Wall Street Journal, was part of a broader effort to develop technology for government use and remained undisclosed until later disclosures. The funding enabled NeXT to continue developing the NeXTSTEP operating system and related hardware while seeking commercial customers. The company eventually sold its assets to Apple in 1996, and the technology contributed to the development of macOS. The arrangement highlights the intersection of government research funding and private technology firms, though the full extent of the CIA's involvement has not been fully disclosed.

Apple will release macOS 27 Golden Gate next month, adding a dedicated Siri app for Mac for the first time. The update rebuilds Siri with a new architecture that includes personal context understanding, expanded world knowledge, on‑screen awareness and the ability to perform actions within apps. The Siri app appears in the dock and offers list or grid views of conversations, syncs via iCloud across Apple devices, and lets users set retention periods of forever, 30 days or one year, with automatic deletion after the chosen threshold. Users can choose whether the app opens to recent conversations or a new conversation, and adjust the number of preview lines shown. The app is available on MacBook Neo (A18 Pro) and Mac models with M1 or later processors. Apple has not announced a timeline for broader AI integration beyond the app, and it remains to be seen how the Siri app will compete with existing chatbot services such as Claude and ChatGPT.

Qualcomm updated its benchmark slides for the Snapdragon C processor targeting laptops priced as low as $300, removing two results almost a week after the original publication. The company asked Tom's Hardware and other outlets to replace the slide with an updated version that no longer includes results for idle apps and web browsing, which were marked with the prefix ABL. A Qualcomm representative said the revision was intended to improve clarity and relevance, though the timing of the change has drawn attention. The benchmarks compare the Snapdragon C to Intel’s N250 quad‑core chip, measuring power consumption while the devices operate on battery power. Qualcomm used a 16‑inch reference device for the Snapdragon C and an 11.6‑inch Acer TravelMate for the Intel N250, configuring the latter’s total power to match the former’s screen size. No specific numerical values or units were disclosed in the slides. The removal of the two results follows a request from Qualcomm to revise the presentation; the company has not provided additional explanation for the change beyond the stated goal of clarity. Jake Roach, senior CPU analyst at Tom's Hardware, reported the update.

Git repository hosting at scale presents challenges because Git’s design treats every repository copy as identical. Linus Torvalds created Git to replace BitKeeper for the Linux kernel, a project with a decentralized workflow. Most open‑source projects and companies rely on a centralized host, which makes hosting a Git repository difficult.

The difficulty stems from Git’s use of packfiles, which are binary bundles of objects that must be stored on disk and transferred over the network. Packfiles are large, and their random layout makes operations that require reading the commit graph inefficient on networked filesystems. This issue affects both storage and network performance.

Early attempts to scale Git included distributing the filesystem, but networked filesystems such as NFS conflict with Git’s assumptions about file locking and performance. Later approaches that replicated the filesystem at the block level, such as GFS and DRBD, proved hard to operate and did not improve performance.

GitHub introduced a system called Spokes in 2013. Spokes stores repository data on local NVMe drives and uses three‑phase commit consensus to keep multiple replicas consistent. The approach limits scalability because each repository requires three replicas, and the consensus latency grows with the number of replicas. Spokes also treats each repository as a critical resource, requiring a routing table that maps repositories to servers and constant checksum verification, which adds operational complexity.

In 2026, the typical enterprise repository has become a large monorepo, and the three‑replica model is insufficient for high‑traffic CI workloads. Adding more replicas degrades push throughput because each step of the three‑phase commit is limited by the slowest node.

Continuity, a system built by Cursor, addresses these issues. It uses a write‑ahead log stored in S3‑compatible object storage as the source of truth. When a push occurs, the packfile is written to local NVMe storage and uploaded to S3 before acknowledgment. The push becomes visible only after the reference transaction is prepared locally and a pointer to the WAL entry is recorded, ensuring linearizable behavior.

Continuity’s design treats repositories as transient caches; the location of a repository does not matter. Repositories are mapped to nodes using rendezvous hashing, and any node can serve a repository. Updates are coordinated through an atomic compare‑and‑swap on S3, eliminating the need for a fixed primary or complex consensus.

Replication in Continuity is optimistic and relies on UDP gossip to propagate metadata. Reads verify the ETag of the latest WAL index against S3, achieving strong consistency without distributed locks. This allows a variable number of replicas per repository, from a single replica for idle repositories to many replicas for large monorepos, enabling linear scaling of read performance.

Compaction of packfiles is performed only by the primary node, with results propagated through the WAL, reducing CPU load on replicas. Synthetic tests have shown linear read scaling with up to 100 replicas and push throughput of 120 pushes per second on standard S3, with higher rates possible on S3 Express One Zone.

Azure DevOps stores packfiles in blob storage and uses a relational database for references, illustrating alternative trade‑offs. Continuity’s WAL‑first approach prioritizes Git data consistency over other considerations, avoiding the need for an external database.

The system aims to provide smooth migration paths, higher reliability, and improved performance for Git hosting, addressing the productivity costs associated with repository downtime.

AMD released Adrenalin Edition 26.8.1, a graphics driver that adds support for the Radeon RX 9050 4 GB model and three upcoming titles, including Resonance: A Plague Tale Legacy.

The driver also adds support for the RX 9050 8 GB variant, which was recognized in version 26.7.1, and provides full compatibility for both variants.

The RX 9050 is an RDNA 4‑based discrete GPU with 16 compute units and 4 GB or 8 GB of GDDR6 memory on a 64‑bit or 128‑bit memory bus. AMD said the GPU is expected to deliver more than 60 frames per second in modern titles at 1080p with medium graphics settings.

AMD also added support for upcoming games such as STAR WARS Zero Company and Call of Duty: Modern Warfare 4, which will run an open beta from August 21 to August 25.

The update includes bug fixes for titles such as Call of Duty: Modern Warfare III, Marvel Rivals, and Subnautica 2 on RX 7000 and RX 9000 series graphics cards.

Users can download version 26.8.1 from AMD’s official website.

A position paper submitted to the International Conference on Machine Learning 2026 warns that labeling intermediate token generation as "reasoning traces" anthropomorphizes language model outputs. Intermediate token generation, in which a model produces output before arriving at a final answer, is a common technique for improving performance on reasoning tasks. The authors argue that this framing suggests the model mimics human thought processes, which they say is inaccurate and potentially harmful. They contend that such anthropomorphism can distort understanding of model capabilities and lead to flawed research practices. The paper calls on the research community to refrain from using the term "reasoning traces" when describing intermediate token outputs. The work appears as a fork of an earlier version and is listed under the subject of Artificial Intelligence. The authors invite further study to clarify how intermediate token representations should be interpreted without attributing human-like reasoning to them.

Consumer Rights Wiki announced on July 24, 2026 an update that adds a feedback button for articles, improves anti‑spam measures, and introduces several user and moderator tools. The site, a collaboratively edited resource on consumer‑rights legislation, has been upgraded to MediaWiki 1.46 with updated extensions and now supports OpenDocument Text and Spreadsheet file uploads. The new feedback button posts comments to the Discord #wiki channel and shows them in an edit‑feed widget, while tightened rate limits reduce spam from the feedback feature. Anonymous edits no longer display the contributor’s IP address in the username field, and user registrations are prevented from automatically posting to the Discord feed. Non‑confirmed users cannot edit templates, and administrators can enable lockdown mode at Special:SiteLockdown to temporarily disable account creation and edits for unauthenticated users. Staff with rollback‑manager permission can perform mass rollbacks via Special:MassRollback and issue manual awards through Special:GiveAward. Additional backend updates enhance security and performance, and Google indexing now includes several hundred pages. The announcement thanked contributor Jake for the update and invited regular contributors to apply for moderator roles.