OpenAI announced that GPT-5.6, a version of its large language model, solved a 30-year-old open problem in convex optimization when guided by a specific prompt. The company posted the results in a technical note on its research blog. The prompt directed the model to generate an algorithmic solution that matched the mathematically proven optimum. Independent experts in optimization confirmed the correctness of the output. Convex optimization, a fundamental area of applied mathematics, underpins many engineering and economic models. A particular gap in the field, described in research literature since the early 1990s, had remained unsolved despite numerous attempts. The announcement indicated that the model’s ability to interpret and execute the prompt enabled it to produce a concise derivation that aligned with established proofs. The note included code snippets and a description of the prompt structure. Researchers not involved in the project said the result demonstrated progress toward using language models as reasoning tools in mathematics. The outcome suggests a new avenue for AI-assisted theorem proving, but the work has not yet undergone peer review, and the broader applicability of the method remains uncertain.
Models, chips, platforms, and the policy lines being drawn around them.
GoPro founder Nicholas Woodman extended a $20 million loan to the company at a 6.5 percent annual interest rate, according to reports. The loan is intended to keep the company operating while a buyer is sought, as the firm may not survive the year without new capital or a purchaser. In May 2026 the company reported a 26 percent decline in quarterly revenue and a 29 percent drop in unit sales, falling to 313,000 cameras sold. GoPro announced it had hired a financial adviser to review strategic alternatives, a step the CEO said he supports. The company has begun restructuring to cut about 23 percent of its global workforce by the end of 2026. Recent product launches of the Mission 1, Pro and Pro ILS cameras were announced in May, but analysts say the effort may be too late to reverse the decline. GoPro has also pursued opportunities in aerospace and defence, markets that could be worth billions, and its cameras were used on the Artemis II NASA mission earlier this year. Competition from Insta360 has intensified, including a recent loss in a patent dispute. Any prospective buyer would inherit the company’s debt and ongoing challenges. The outcome remains uncertain as the firm continues to seek financing or a sale.
Moonshot AI released version Kimi K3 this week. The model trails proprietary systems such as Claude Fable 5 and GPT 5.6 Sol, according to the company, but achieved frontier-level performance across benchmark tests, outperforming other models in independent analyses by Arena.ai and Vals AI. The announcement occurred alongside a speech by Chinese President Xi Jinping at the World AI Conference in Shanghai. Following the release, the Nasdaq index fell about 1% on Friday, driven by declines in chip company stocks including Nvidia. Industry commentators offered varied perspectives. David Sacks, former AI adviser to the Trump administration and co‑chair of the President’s Council of Advisors on Science and Technology, said that U.S. policy restrictions on data centers and frontier‑model approvals risk ceding the AI race. He also criticized Anthropic’s Claude as limiting American competitiveness. Travis Kalanick, former Uber CEO, noted that Chinese firms may be training models on American outputs, and said that without enforcement of distillation controls, all developers could replicate existing models. Dean Ball, OpenAI’s head of strategic futures, described Kimi as a strong model whose performance cannot be fully explained by distillation, and expressed surprise at the Chinese government’s decision to release such models openly. Ball suggested that an open‑weight model–dominant environment could lead to increased regulatory oversight, with the administration potentially imposing soft regulatory constraints that generate uncertainty for enterprises. Shakeel Hashim, editor of Transformer, argued that Kimi likely does not pose dangerous cyber capabilities and that Chinese incentives to restrict open models will mirror those of other nations once they achieve comparable capabilities. The release underscores the growing geopolitical dimension of AI development and the uncertainty surrounding future regulation of open‑source models.
Residents across the United States have voiced opposition to new data center projects, uniting people from different political parties around concerns about land use, water consumption, electricity demand and limited transparency.
Data centers support the rapid expansion of artificial intelligence services and have previously attracted support from technology firms and local officials who highlighted job creation and tax revenue.
A 2026 Gallup poll found that about 70 percent of Americans would oppose a data center being built near their homes. Opposition has emerged in both red and blue regions. In Hill County, Texas, voters removed incumbent council members after approving a data center project, while Monterey Park, California, enacted a local ban. Utah Governor Spencer Cox said the facilities represent a tangible symbol of AI‑related anxiety. Democratic Senator Sherrod Brown has run advertisements accusing his opponent of using data centers to raise expenses for Ohio residents. In New York, state officials imposed a temporary pause on large new data center projects, marking the first statewide moratorium. NYU’s Scott Babwah Brennen reported roughly 65 state‑level restrictions on data centers since the start of 2025 and dozens of city‑wide temporary halts. Senator Bernie Sanders has called for a federal ban on data centers, and the group Humans First, led by former Tea Party activist Amy Kremer, is organizing national protests. Google, which is developing a data center near Sand Springs, Oklahoma, said it strives to be a good neighbor and complies with state and local regulations. Critics argue that the backlash reflects distrust of large corporations and NIMBY attitudes, while others point to genuine concerns about increased household electric bills and water usage.
The growing resistance may slow data center expansion that underpins AI services, but the ultimate effect on the industry and on policy remains unclear.
Honda announced that its Prologue electric SUV will be discontinued, marking the final withdrawal of an all‑electric model from the company's U.S. lineup and highlighting a broader pullback by automakers from the U.S. EV market in 2026.
U.S. EV sales have declined since the federal tax credit ended in late 2025, with 247,226 EVs sold in the second quarter of 2026, representing about 5.8% of total vehicle sales, according to data from Kelley Blue Book and Cox Automotive.
Honda’s joint venture with Sony, which created the Afeela brand, ended development in March 2026 after Honda also canceled three U.S. EV models in February.
Honda halted development of the O Series, including the O sedan and O SUV, in March 2026, citing U.S. tariffs and competition from Chinese manufacturers.
The Prologue, built at General Motors’ Ramos plant in Mexico and related to the Chevrolet Blazer EV, sold roughly 33,000 units in 2024 and 39,000 in 2025 before sales declined after the tax credit ended.
Hyundai announced in March that it will discontinue the Ioniq 6 in the United States, a decision tied to tariff concerns, while continuing to import higher‑priced variants.
Nissan confirmed that it will not produce a 2026 model year Ariya SUV for the United States.
Polestar, owned by Geely, has been unable to obtain Department of Commerce authorization to sell new vehicles in the United States and will cease imports, though it will support existing inventory of Polestar 3 and Polestar 4.
Tesla announced in January that production of the Model S sedan and Model X SUV will end, with the final vehicles rolling off the line in spring, and will reallocate assembly capacity to its Optimus robotics program.
Volkswagen will stop production of the ID.4 at its Chattanooga plant in Tennessee and will not offer a 2026 ID. Buzz model, though the ID Buzz is expected to return in 2027; Volkswagen also began autonomous microbus testing with Uber in Los Angeles.
Volvo will end production of the EX30 and EX30 Cross Country in the United States after the summer, while continuing sales of the larger EX60 and EX90 models.
The withdrawals reflect ongoing challenges for U.S. EV adoption, including regulatory, tariff and consumer preference factors, and suggest that the market may see a slower recovery despite new entrants such as the Rivian R2.
Mitchell Hashimoto, a software engineer and author, published an essay describing widespread AI hype in corporate settings. He said that many organizations have adopted AI tools without clear plans, and that most AI projects have shown little or no measurable success.
He reported that internal chatbots are rarely used because documentation is incomplete and AI systems cannot infer information not recorded, while customer‑facing chatbots have produced inconsistent experiences, with one example of a Mitsubishi service bot failing to follow up on a support request.
Hashimoto observed that executives and boards pressure staff to praise AI initiatives, and that fear of negative repercussions discourages honest assessment. Companies often announce AI productivity gains after purchasing licenses such as Microsoft Copilot, even when the tools have not been integrated into workflows.
He noted that consulting engagements sometimes add an AI layer to existing projects to meet internal “AI” criteria, and that many large firms have become unable to focus on core objectives because of the dominance of AI discussions.
The essay concludes that without independent evaluation, organizations risk continued misallocation of resources and delayed decision‑making, a situation that may ease only when the hype subsides.
NYC Mayor Zohran Mamdani announced a requirement that landlords and realtors disclose when AI‑generated or edited images are used in rental listings, as part of a broader effort to curb deceptive housing practices.
The administration released a Rental Ripoff Report recommending disclosure rules, formal recognition of tenant unions, and expanded tenant bargaining rights, following hearings with 2,400 New Yorkers across the five boroughs that highlighted unsafe living conditions and misleading listings.
The report calls for mandatory disclosure of AI‑altered imagery, aims to modernize code enforcement systems, and seeks to institutionalize tenant organizing, with Deputy Mayor Leila Bozorg and tenant advocacy director Cea Weaver noting the policies address documented tenant concerns.
Implementation and enforcement details are pending, but the rule establishes a local precedent for regulating AI use in real estate advertising.
Dave Eggers addressed roughly 200 OpenAI staff members during a July 2025 presentation, stating that ChatGPT’s influence on teachers is substantial.
Eggers, known for novels such as The Circle and for establishing literary nonprofit initiatives, has previously criticized AI‑generated writing as pastiche nonsense.
He told the audience that teachers’ workloads have risen substantially since the model’s widespread adoption, and that students who use the system to compose essays may forego developing their own writing abilities. Eggers said this situation could deprive a generation of learners of the ability to express their own perspectives, according to the Financial Times.
OpenAI has not publicly responded to the comments, and the discussion about AI’s impact on education continues.
Leaked specifications suggest that the Google Pixel 11a will use the Tensor G6 processor, a MediaTek M90 modem and a display with a peak brightness of 3,350 nits. The Pixel a series has historically paired the latest flagship‑class processor with cost‑saving measures in other components. The Tensor G6 is reported to retain the PowerVR DXT-48-1536 GPU found in the G5 while improving upon the Mali‑G715 used in the G4. It replaces Samsung’s Exynos modem with a MediaTek M90 modem, a change the leak says should reduce battery drain and improve signal reliability. The device is expected to ship with a 4,870 mAh battery, slightly less than the 5,000 mAh battery in the Pixel 10a, but efficiency gains from the new processor may maintain or slightly improve battery life. The display will remain a 6.3‑inch 1080 × 2424 panel with a variable 60 to 120 Hz refresh rate, and brightness is expected to rise from 3,000 to 3,350 nits. Google has not confirmed these specifications, and the accuracy of the leak remains untested.
A guide explains how to configure a spare Mac as a dedicated environment for Claude Code, enabling remote control from a phone or a primary Mac via SSH. The guide notes that running Claude Code on a main computer can expose sensitive files and increase permission risk, especially when using the --dangerously-skip-permissions flag. A separate machine provides a sandbox with full access to required tools while minimizing impact on the user’s primary system. Step 1: The target Mac is installed fresh, optionally erased and updated to the latest macOS version. A new local admin account is created without an Apple ID, and the account is set to allow administration. Step 2: SSH is enabled on the target using the command sudo systemsetup -setremotelogin on; full‑disk access is granted to the terminal application if required. Step 3: Passwordless sudo is configured for the target account by adding the line ALL=(ALL) NOPASSWD: ALL to /etc/sudoers.d, making the file read‑only and validating the syntax. Step 4: The target’s unique hostname is set via scutil --set LocalHostName, and the address is used as hostname.local for SSH connections. Step 5: On the primary Mac, an ed25519 SSH key is generated and the public key is installed on the target with ssh-copy-id. The connection is verified with ssh @hostname whoami, which returns the username without a password prompt. Step 6: To prevent automatic sleep, the target runs sudo pmset -c sleep 0, sudo pmset -c disablesleep 1, and sudo pmset -c displaysleep 0, and the screen saver is disabled with defaults write com.apple.screensaver idleTime 0. Step 7: A script named clip.sh is installed on the primary Mac to pipe pbcopy and pbpaste over SSH, providing text and image clipboard sync between the two machines. Step 8: Claude Code is installed on the target by executing ssh target 'curl -fsSL https://claude.ai/install.sh | bash -s -- 2.1.201', and the binary directory is added to the shell environment via export PATH in ~/.zshenv. Step 9: An optional script setup-claude-env.sh can be run to apply predefined aliases, plugins, and settings; it supports interactive checklists or non‑interactive flags such as --all. Step 10: Both Claude and GitHub logins are performed interactively on the target; using a separate GitHub account is recommended to avoid affecting the primary account. Step 11: Computer use is enabled by installing a LaunchAgent that runs a persistent tmux server inside the GUI session, granting Screen Recording and Accessibility permissions to tmux, and restarting the tmux server after permission changes. Step 12: A VPN can be installed on the target, though credentials and manual permission approvals are required; the process can be scripted with clip send for credential transfer. Step 13: Remote Control in the Claude app allows a phone to drive sessions started on the target, either by attaching to an existing session or launching a new one with computer use enabled. Step 14: A Chrome extension can be added to the target to enable browser control; the agent must handle login and permission steps manually. Step 15: Screen Sharing is enabled in System Settings on the target, and the source Mac connects via vnc://hostname. Step 16: Tailscale is installed on both machines to provide encrypted WireGuard tunnels, allowing SSH, clip, and Screen Sharing to function across different networks; device approval and key expiry settings are recommended in the Tailscale admin console. The setup enables secure delegation of tasks to Claude Code, but requires manual permission steps and ongoing maintenance of the isolated environment.
Elixir-lang.org has launched a redesigned website. Elixir is a functional programming language built on the Erlang virtual machine, noted for its fault‑tolerance and scalability, and is used by individual developers and organizations in various sectors. The new site highlights capabilities that support rapid development, reliable coding practices, and composable data transformations. It features the Phoenix web framework, LiveView for real‑time user interfaces, Nerves for embedded deployments, and Numerical Elixir for GPU‑accelerated machine learning. The language can scale vertically on multi‑core hardware and horizontally across multiple nodes. The community maintains thousands of packages on Hex and continues to develop tools such as Membrane, Popcorn, and LiveDebugger. Companies from independent creators to Fortune 500 firms report using Elixir in production. The redesign is intended to enhance accessibility and support ongoing community growth.
Installer No. 136 presents a weekly collection of community-submitted recommendations for reading devices, applications, and media. The author describes a personal reading setup that includes a Kindle Paperwhite, an iPad Mini, and a Boox Palma 2, with Kindle as the primary ebook application and Readwise Reader for PDF conversion and organization; other apps such as Instapaper and Matter are also used. Survey responses indicate extensive use of library platforms including Libby, Hoopla, MyLibro, and Sora, and a preference for devices such as Kobo readers, the iPod Touch, and the Xteink X4. Popular reading applications reported are Kindle, Apple Books, BookFusion, and the Bookshop.org app. Tools for tracking reading progress include The StoryGraph, Book Tracker, and a decline in Goodreads usage. Audio consumption is reported through Audible and Spotify Premium’s 15‑hour audio allowance. Services for discovering ebook and audiobook sales include BookBub and Chirp. Highlight syncing is performed by many users with Readwise, Notion, Craft, and Obsidian, while some develop custom solutions. Newsletter management remains a challenge, with readers seeking methods to consolidate content. The community invites additional recommendations via email to installer@theverge.com or via Signal @davidpierce. The installment also features personal notes on email services, anime, note‑taking apps, games, hardware, books, YouTube channels, RSS applications, and a web serial.
Apple raised the price of its Apple Music Individual plan to $11.99 per month, up from $10.99, effective today in the United States and other countries. The Family plan rose to $19.99 from $16.99, and the Student plan to $6.99 from $5.99. Apple One’s Individual tier remained at $19.95, while the Family tier increased to $27.95 from $25.95 and the Premier tier to $39.95 from $37.95. In a statement to 9to5Mac, Apple said the adjustments reflect rising licensing costs. The changes follow Spotify’s earlier price hikes this year and a social media post in which Apple referenced the competing service; Apple Music continues to be priced lower than Spotify. This is the first increase for Apple Music in nearly four years, after a previous rise in October 2022 that moved the Individual plan from $9.99 to $10.99. Apple also raised prices for Apple TV, Apple Arcade and Apple News+ in October 2023, and Apple TV saw another increase in August 2025, though Apple One bundles remained unchanged at that time.
Telepathic Instruments will begin taking orders for the Orchid synthesizer on May 11 at a price of $649, and a limited edition clear Arctic model will be offered at $699 for 3,000 units.
The Orchid, co‑designed with Kevin Parker of Tame Impala, was introduced to emulate the chord organs popular in mid‑20th‑century homes. It lets musicians play complex chord progressions without extensive theory.
The original Orchid had been out of stock for several months. Telepathic Instruments expects the new batch to remain available indefinitely. The Arctic variant, announced after a correction to the earlier release date, will be sold on a first‑come, first‑served basis to those who signed up for the waitlist.
In addition, the company is launching Pistil, a plug‑in version of the Orchid engine priced at $99 that can be used inside a digital audio workstation and also control the physical Orchid.
The limited availability of the Arctic model may influence purchase timing, and Telepathic Instruments has not disclosed plans for future production runs.
FreeBSD developer Joe Maloney announced the revival of NextBSD, a project that aims to integrate Apple's Darwin userland into FreeBSD's kernel. Maloney, who previously contributed to the Gershwin desktop for GhostBSD, took over the NextBSD project after requesting maintenance access from the original maintainers. The original NextBSD project was launched in 2015 by FreeBSD co‑founder Jordan Hubbard with the goal of porting parts of Apple's Darwin operating system to FreeBSD. Earlier efforts such as OpenDarwin (2002–2006) and PureDarwin (released 2015 and 2019, still maintained through 2024) attempted similar integrations but did not achieve widespread use. Maloney's reboot, referred to as NextBSD‑redux, does not use a fork of the decade‑old code. The repository's oldest commits are two months old, and Maloney has documented each development stage on his personal website. He said that AI tools, specifically Anthropic's Claude Code, assist his development by accelerating his understanding of existing code and by generating test suites. He also noted that he trusts human review of AI‑generated code in production environments. The project currently lacks a graphical desktop, but a Gershwin‑based interface is being developed in the gershwin‑on‑nextbsd repository. The README outlines the components that are functional, including kernel module loading and launchd integration. Maloney described the architecture as a clean‑room implementation, with some code borrowed from ravynOS's libxpc library. He has organized the code into separate repositories for the kernel, kext conversion tools, and Darwin‑derived utilities, and uses automated testing to verify each change. Other recent projects such as helloSystem and ravynOS have also sought to create macOS‑like environments on BSD bases, often reusing components from one another. Maloney's effort follows this trend but emphasizes a minimal approach that builds on FreeBSD's mature kernel rather than creating a new OS from scratch. The project remains in an experimental phase, and it is not yet clear whether a complete operating system can be delivered for laptop use.
2P_Game_Arcade, a Korean games culture site, is hosting an open game development contest that requires entries to fit entirely within a 1.44 MB floppy disk.
The competition, which closes on September 4 at 23:39 Korean time, offers a prize pool of ₩1.14 million, roughly $750, for the top three submissions.
Entrants must create a self‑contained game that includes the engine, libraries and all resource files within the 1.44 MB limit, and may not rely on a browser or stream data.
2P_Game_Arcade said the contest invites developers of any experience level to participate.
Participants may use any programming language, game engine or development tool, provided they own or have permission for all images, audio and other IP used in the game.
The site said the contest is intended to challenge developers to work within tight storage constraints, a practice that reflects long‑standing demoscene traditions.
Similar size‑limited contests have been held previously, including itch.io’s FloppyJam in 2018.
Earlier examples of demoscene size competitions include a 16‑byte x86 assembly demo released at the Outline Demoparty in May 2026, a 56‑byte version of Snake that fit inside a QR code, and a 2,749‑byte Notepad‑style application by Dave W. Plummer.
The competition will close on September 4, after which the winners will be announced.
People discover that organizing social events accelerates community participation.
Demand for organized gatherings often exceeds the supply of available events.
Individuals who consistently attend others' events can build relationships, but organizing one's own activities tends to produce faster and stronger connections.
The author notes that many people adopt a passive stance, expecting community benefits without contributing effort.
Efforts required to plan and execute events deter most participants, according to the author's observations.
Organizers recognize those who take on responsibilities, while free riders remain less visible.
The author suggests that addressing social isolation at a local level involves providing the missing event supply.
By hosting events, individuals can fulfill the unmet demand for social interaction within their own groups.
Apple released the first public betas of iOS 27, iPadOS 27, macOS Golden Gate, and related software, allowing anyone with compatible devices to download and test the updates, an...
Apple released the first public betas of iOS 27, iPadOS 27, macOS Golden Gate, and related software, allowing anyone with compatible devices to download and test the updates, and simultaneously launched its annual Back to School promotion for education customers.
In the United States the promotion offers a $150 gift card with the purchase of a new MacBook Pro and a $100 gift card with a new MacBook Air, iPad Pro, or iPad Air, while the iPad mini, entry‑level iPad, MacBook Neo, desktop Macs, and other models are excluded; similar offers apply in additional countries with varying incentives.
Apple filed a lawsuit against OpenAI alleging that the company used stolen Apple trade secrets to develop its own AI hardware and to recruit Apple employees, to which OpenAI responded that it has no interest in other companies' trade secrets and is not aware of any merit in the allegations.
Analysts expect a refreshed iPad mini with an OLED display to arrive by October, and anticipate updates to the iPad lineup and new Apple Pencil models in early 2027; Apple is also said to be developing new home devices, including an updated Apple TV 4K, HomePod, and HomePod mini, to support the forthcoming Siri AI features in iOS 27, iPadOS 27 and macOS 27.
Bloomberg reported that Apple may forgo the M6 Pro and Max variants, accelerating the release of the M7 family, which could include up to 1.5 TB of unified memory for AI tasks; the company also announced price increases on several products ranging from 10 to 50 percent and issued security updates for iOS, iPadOS and macOS following a manufacturing‑related cyber incident.
Apple invites users to subscribe to its newsletter for weekly recaps of technology news and upcoming product announcements.
CMS has launched a demonstration project called WISeR that uses artificial intelligence to evaluate Medicare claims for potential waste, fraud, and inappropriate services. The pilot, which runs through December 2031 in six states, combines machine learning with clinical review to assess items such as skin and tissue substitutes, nerve stimulator implants, and knee arthroscopy for osteoarthritis. CMS says the model will ensure timely and appropriate payment for selected services.
Prior authorization, which insurers use to determine coverage before a service is provided, has been linked to delays and denials. A 2025 American Medical Association survey found that 61 percent of physicians worry AI could increase unnecessary denials. Medicare Advantage, covering about 55 percent of Medicare beneficiaries, issues millions of claim decisions each year based on prior authorization. Federal reports from June indicated that some plans reject requests for skilled nursing and rehabilitation even when coverage rules appear met.
The Biden administration in 2024 set rules requiring insurers to decide urgent prior authorizations within 72 hours and non‑urgent ones within seven days, effective January 1 for most public health plans. The Trump administration pledged to streamline electronic prior authorization requests and reduce the volume of services subject to review, targeting common procedures such as colonoscopies and cataract surgeries.
The WISeR pilot will expand AI use in original Medicare, a system that has rarely employed prior authorization. Critics argue that automated decisions may lead to more denials and that vendors could profit from reduced expenditures. An HHS Office of Inspector General memorandum in 2022 noted that over 10 percent of Medicare Advantage denials initially appeared inconsistent with coverage rules, though most were reversed on appeal.
Lawmakers have introduced measures to block funding for the model, citing patient access concerns. CMS Administrator Mehmet Oz warned insurers that failure to ease prior authorization burdens could result in federal regulation, saying the government will act if insurers do not comply.
Industry surveys indicate a decline in prior authorization requests of 11 percent between June 2025 and April 2026, though the impact on denial rates remains unclear. Insurers have pledged greater transparency on clinical reasoning and stated they do not use AI alone to deny medically necessary care.
The grade‑school method multiplies each digit of one number by each digit of another, requiring n² single‑digit multiplications for two n‑digit numbers. Computer scientists measure algorithmic speed in computational steps rather than elapsed time. In 1960, a 23‑year‑old student named Anatoly Karatsuba showed that the process could be reduced to three multiplications by reusing intermediate sums, achieving O(n^1.585) complexity. Subsequent work in 2019 by David Harvey and Joris van der Hoeven introduced an algorithm with O(n log n) complexity, marking the fastest known method in theory. Both Karatsuba’s and Harvey‑van der Hoeven’s techniques outperform the grade‑school approach only when the operands are sufficiently large; for typical sizes, the simpler method remains faster due to lower overhead. Programming languages such as Python switch to Karatsuba’s method for numbers exceeding roughly 630 decimal digits, reflecting the practical threshold where the asymptotic advantage becomes measurable. Researchers continue to investigate whether O(n log n) represents the theoretical limit of multiplication, a question that remains unproven and central to the field. The open problem of establishing the optimal multiplication complexity persists, with potential impact on computer hardware, cryptography, and large‑scale data processing.
OpenAI's head of strategic futures, Dean W. Ball, said in a post on X that a future dominated by open‑weight models such as Kimi K3 could amount to "full AI communism," describing the shift as a threat to his company's investment in proprietary models. Kimi K3, an open‑source model with 2.8 trillion parameters and a context window of about one million tokens, was released by Moonshot and topped Arena's front‑end code rankings. The model uses a hybrid attention mechanism called Kimi Delta Attention (KDA) and WideEP optimization to reduce memory usage. Ball argued that open‑weight models represent a "decelerationist" approach to capital expenditure and that a market dominated by such models would make AI a public good provided by the state. He labeled Chinese open‑source models as "decelerationists." The comment followed Kimi K3's rise and coincided with remarks from other observers, including Ball's earlier tweet praising the model's performance relative to public benchmarks. Moonshot's release also highlighted technical details such as a novel KV cache architecture and expert distribution across GPUs. Anthropic responded by keeping its Fable 5 model available in major subscription tiers after Kimi K3's prominence. The situation highlights growing debate over the role of open‑source AI in the broader technology landscape.
San Francisco city attorney David Chiu sent cease‑and‑desist letters to Apple and Google, ordering the removal of 13 mobile applications that can generate nude images of individuals without consent. The letters cite California law that prohibits the creation and distribution of sexualized AI deepfakes, stating that the companies are aiding and abetting the sale of such content. Google said it has suspended the five Android apps referenced in the letters and reiterated that its Play Store does not permit sexual content. The action follows a report by Wired that detailed the cease‑and‑desist demands. The regulatory move highlights growing scrutiny of AI tools that can alter personal imagery, and it remains unclear how Apple will respond.
The TikTok US joint venture, TikTok USDS Joint Venture LLC, completed its formation on schedule in December, meeting the requirements of the 2024 divest‑or‑ban law signed by President Biden. ByteDance’s ownership in the venture was reduced to 19.9 percent to comply with the law, while the remaining 80.1 percent was allocated to a group of investors. The investors consist of Silver Lake, Oracle, and MGX, each holding a 15 percent stake, and include Michael Dell’s family office, Vastmere Strategic Investments, Alpha Wave Partners, Revolution, Merritt Way, Via Nova, Virgo LI, and NJJ Capital. The venture’s board comprises seven members, including TikTok US CEO Shou Zi Chew and former TikTok operations chief Adam Presser, who now serves as CEO. The joint venture will manage data protection, algorithm security, content moderation, and software assurance for TikTok, CapCut, Lemon8, and other affiliated apps, allowing U.S. creators and businesses to operate on a global scale. U.S. user data will be stored in Oracle’s secure U.S. cloud environment, and the venture will maintain a cybersecurity program that is audited against NIST CSF, NIST 800‑53, ISO 27001, and CISA security requirements. The recommendation algorithm will be retrained, tested, and updated using U.S. user data within the same cloud environment. The venture will apply software assurance protocols, reviewing and validating source code continuously with Oracle as the trusted security partner. It will enforce trust and safety policies and content moderation for the U.S. content ecosystem. The press release announcing the deal did not disclose the financial terms or any timeline for launching a new U.S. app, and the full operational integration schedule has not been specified.
Castor is an open‑source command‑line tool that extracts video streams from web pages and casts them to DLNA‑compatible televisions in real time. It was created to enable casting web video when Chromecast or AirPlay is not available. Screen mirroring often suffers from lag and reduced resolution. Castor captures the video stream directly from the browser. The tool runs headless Chrome and uses the Chrome DevTools Protocol to monitor network traffic. It transcodes the stream with ffmpeg before sending it to the TV. It also accepts direct stream URLs or IMDB/TMDB identifiers and can embed subtitles generated by Whisper. Installation requires Chrome or Chromium, ffmpeg, and ffprobe on the system PATH. A native binary is recommended for most users. Docker offers an optional bundled package that works only on Linux hosts. Homebrew provides a macOS cask installation. The prebuilt Docker image from GitHub simplifies setup on Linux systems. Castor discovers DLNA devices on the local network. It supports most smart TV brands and media players such as Kodi, VLC, and Plex. Users create a config.yaml file specifying the target device name and source proxies. The command "castor cast movie tt12300742" casts a movie by its IMDB/TMDB ID. The command "castor cast player " casts the video currently playing on a web page. The command "castor scan" lists devices on the network. An interactive TMDB browser can be launched with "castor cast" when a TMDB API key is provided. Flags such as "--dry-run" display found URLs without casting. Castor does not host any video content. Users must ensure that the material they cast complies with the terms of service of the source sites. The Chromecast integration is marked as experimental and may not work with all sites. Future updates may improve compatibility with sophisticated bot protection mechanisms.
Adam Newbold published a guide on July 18, 2026 that describes a method for maintaining a personal website using only a text editor, an SFTP client, and a low‑cost web host. The guide, titled "Hardcore IndieWeb", states that a static website can be hosted for $0.01 per day at NearlyFreeSpeech.net. The recommended workflow involves authoring HTML files locally, previewing them in a web browser, and uploading the files to the host via SFTP. The guide lists three required components: a text editor, a file transfer tool, and a web host account. It notes that a domain name is optional but can be added through the host’s control panel. The IndieWeb movement emphasizes ownership of content and portability of web sites. The guide argues that relying on external blogging services can limit control because content is stored on a provider’s server and may be difficult to export. By storing HTML files locally, the site remains fully under the owner’s control at all times. The guide describes a simple file structure: an index.html file for the landing page, individual HTML files for blog posts placed in a designated folder, an archive page that lists posts, and an RSS or Atom feed file. It suggests using free website templates downloaded from public sources and modifying the HTML as needed. Uploading the files to the host updates the site instantly. The guide notes that the process mirrors early web publishing practices from the 1990s and does not require programming environments or monthly subscriptions. The method provides independence from service outages and software updates, but the longevity of the solution depends on the availability of the chosen host and the user’s continued management of local files.
AMD Medusa Point 10‑core Zen 6 engineering sample achieved Geekbench scores of 3,329 in single‑core and 16,555 in multi‑core at 5.4 GHz, a 5 % increase over its previous single‑core result of 3,174. The sample, identified by CPU ID 100‑000001713‑33_N on the AMD Plum‑MDS1 platform, also showed a 9 % improvement in multi‑core performance compared with its earlier multi‑core score of 15,092. In comparison, the Ryzen AI 9 HX 370 recorded 2,605 single‑core and 13,396 multi‑core points at 5.1 GHz, while the Ryzen AI 9 365 posted 2,470 single‑core and 12,407 multi‑core points at 5.0 GHz. The benchmark, which is synthetic, provides a rough indication of relative performance and may not reflect real‑world usage. The sample operated at a maximum clock speed of 5,370 MHz, about 300 MHz higher than the fastest Strix Point APUs recorded in the test. Further evaluation on platforms such as PassMark or Cinebench is expected, but the chips are not expected to be widely available before the official announcement, which is scheduled for several months in the future.
Fabien Sanglard, a software engineer at Google, compiled a detailed inventory of the computers that appear in the 1993 film Jurassic Park. The inventory lists six functional workstations, a personal digital assistant, two large CRT monitors, and a mechanical keyboard. According to Sanglard, the equipment was displayed in the movie’s control room and was sourced from real manufacturers. His website provides the full specifications and indicates that the hardware was valued at $1.725 million in 1992, equivalent to more than $4 million today. Most of the items were functional, with the exception of the Thinking Machines CM‑5 supercomputer mockup, which consisted only of front panels and red LED lights. The special effects coordinator Cory Faucher noted in the book The Making of Jurassic Park that the set was built with genuine hardware because audiences were expected to be familiar with computers. Sanglard also pointed out that the on‑screen meteorology software was the same program used in newsrooms at the time. Eagle‑eyed viewers have observed that a scene in which the character Nedry appears to speak with his getaway driver via a security feed was actually a prerecorded QuickTime video, not a live feed. Unlike many Hollywood depictions that show hackers typing gibberish before unplugging a terminal, the Jurassic Park equipment operated as intended. Fans interested in seeing the vintage machines in action can re‑watch the film or explore related 3D‑printed retro PC projects.
PNY is offering its GeForce RTX 5070 OC Triple Fan graphics card for $579 on Amazon, the lowest price recorded in the past three months.
The graphics processing unit market has been constrained by DRAM shortages, which have driven prices for many components, including GPUs, 20‑30 % above historical averages. The RTX 5070 debuted at a manufacturer's suggested retail price of $549 and has been selling for $650‑$700 for several months.
The card is normally listed at about $629, making the current $579 price about $50 below the typical market rate. Amazon’s price history shows the GPU reached $659 earlier this year before declining. The RTX 5070, positioned in the mid‑range segment, is marketed for 1440p gaming and has been one of the more popular models among consumers.
The discounted price may be limited in availability, and pricing remains subject to supply constraints.
Q3Edit, a web‑based editor for Quake III Arena map files, lets users create, modify, and launch maps directly in a browser. The tool runs on TypeScript and WebGL2 and compiles .map files using WebAssembly built from id Software’s original q3map compiler. It includes OpenArena assets and requires WebGL2 support, enabling one‑click execution with ioquake3. The interface features a four‑view layout and tools for brushes, patches, constructive solid geometry, and terrain editing. No original game files are needed, and the compiled maps can be played instantly in the browser. Map editing for Quake III has traditionally required desktop software, but Q3Edit demonstrates how modern web technologies can replace those workflows. The project demonstrates the ability to edit and run Quake III maps in a browser, potentially expanding community map creation capabilities.
A researcher benchmarked Claude's Fable 5 and GPT-5.6 Sol on an unpublished NP‑hard optimization problem, with and without the /goal mode.
The KIRO problem asks solvers to connect distribution points and terminals in a fiber‑network design while minimizing total cable length.
The benchmark used three paired runs for each model under a 30‑minute limit and a 1,900‑second agent timeout.
An initial single run for each model provided a baseline.
The main comparison consisted of three runs for Fable 5 and three for GPT-5.6 Sol.
Fable 5's plain‑mode mean score was 1,875 points lower than Sol's mean.
In goal mode, Fable 5's mean was 1,984 points lower than Sol's, but the overall means for both models were higher than in plain mode.
Fable 5's plain‑mode scores varied by 319 points, while Sol's varied by 1,958 points.
Fable 5 achieved its best goal‑mode score of 31,934; Sol's best plain‑mode score was 32,703.
Negative differences were observed in four of six trials, indicating that /goal performed better in those cases.
Analysis showed that /goal works differently in Claude Code, where a session‑scoped evaluator monitors the goal, and in Codex, where the goal is stored as persistent thread state with tool access.
Limitations include the use of a single unpublished task, limited model replication, and possible resource allocation differences due to container CPU allocations.
Reproduction scripts and data are available in the CLIArena repository, allowing others to run the benchmark and review detailed logs.
The results suggest that persistence features may improve individual trial outcomes but can reduce average performance on hard optimization problems.
Moonshine Micro, an open-source voice toolkit for microcontrollers, enables real time speech detection, transcription and synthesis on devices with as little as 470 KB of RAM, demonstrated on the Raspberry Pi RP2350.
The toolkit, released under the MIT License, includes components for voice activity detection, speech-to-text, custom word recognition and neural speech synthesis, all built on TensorFlow Lite Micro.
The demo runs on the RP2350 microcontroller, a low cost processor priced at 80 cents, and consumes approximately 3.6 MB of flash memory and 468 KB of provisioned SRAM. Voice activity detection uses about 89 KB flash and 36 KB SRAM with 0.8 million multiply-accumulate operations per frame. Speech-to-text requires 1.3 MB flash and 346 KB SRAM, operating at 36 million MACs per second. Neural speech synthesis consumes 1.8 MB of voice-pack data and 340 KB SRAM, delivering typical inference at 37 million MACs per reply (65 million MACs per second). The combined classification and speech pipeline completes in 0.7–1.0 seconds. These memory budgets fit within the RP2350’s 520 KB SRAM, with WiFi hardware using the remaining capacity.
The code repository provides a complete end-to-end example for establishing a WiFi connection using voice commands on the RP2350. Individual libraries for voice activity detection, speech-to-text, custom word recognition and neural speech synthesis can be used independently of one another.
Documentation sections cover voice activity detection, speech-to-text, custom word recognition, neural text-to-speech, and a WiFi setup example. The project is licensed under the MIT License, with third-party model licenses also under MIT.
Chase Bliss announced the Big Time pedal on May 1, 2026. The $1,000 unit is a hybrid analog‑digital delay that also serves as a looper, sequencer and modulation pedal. It is part of the company’s Automatone series and was developed in collaboration with Electronic Audio Experiments.
The pedal features programmable faders that control delay time from 500 milliseconds to two seconds. It includes a built‑in sequencer for rhythmic looping and a modulation section for real‑time effect shaping. The design combines analog circuitry with digital processing to provide both warmth and flexibility.
The announcement follows a series of high‑end effects releases from Chase Bliss, which has positioned itself as a leader in boutique guitar technology. No controversy was reported; a walkthrough video was posted to demonstrate the pedal’s functions.
The product exemplifies a trend of blending analog and digital technologies in professional guitar equipment. It is expected to ship later in 2026.
Waymo announced that its robotaxi service in San Francisco has resumed after a one‑hour pause caused by a power outage that affected roughly 7,000 PG&E customers. The company said it made temporary adjustments while monitoring local conditions and coordinating with officials. A spokesperson stated that riders depend on the service and that normal operations would return as soon as possible. The pause followed previous incidents in which Waymo vehicles stalled during blackouts in December and during a Fourth of July fireworks event on the Golden Gate Bridge. San Francisco Mayor Daniel Lurie called for stronger state regulations on autonomous vehicles during major incidents. Waymo confirmed that service has resumed.
The Amiga Freeware Archive provides thousands of classic Amiga titles, including games, applications, demos, graphics, music and tools, for free download.
The collection aggregates content from 14 public domain libraries, scene groups, disk magazines and user group compilations that were active during the 1980s and 1990s.
17 Bit Software, a UK‑based library operating from 1987 to 1997, catalogued and distributed thousands of disks covering the breadth of Amiga productions.
American programmer Fred Fish began curating floppy disks in 1986, creating a library that became widely distributed in the pre‑Internet era.
Scope, a Texas‑based series, grew to 175 numbered volumes in the early years of the platform.
The Slipped Disk series, also U.S.-based, released at least 20 volumes of curated software.
The Apprentice and Journeyman library, another U.S. collection, contained 23 numbered volumes focused on utilities, tools and source code.
The Assassins group, based in northern England, produced a games compilation series distributed throughout the UK.
The LSD Compendium CD‑ROM series, compiled by a demo group, gathered thousands of demos, modules, graphics and utilities contributed worldwide.
The A.M.I.G.O.S. disk series, issued by a Miami, Florida user group in 1990, ran through the early 1990s.
TBAG, the Tampa Bay Amiga Group, began its monthly series in 1986 and remained active for many years.
Arug, an Amiga amateur radio user group, specialized in radio‑related applications.
The New Zealand Amiga Users Group issued a newsdisk series beginning in December 1986, eventually producing at least 29 disks.
The Southern Nevada Amiga Group released a Disk of the Month series in 1990, numbering at least 14 releases.
JAM, a UK subscription magazine published from 1991 to 1995, issued 58 issues.
The archive remains accessible online, allowing preservation and study of software that might otherwise be lost.
World Cup theme composer Zach Golden showed his studio equipment during a July 18, 2026 interview, noting that a modest collection can support music production. Golden has been tasked with creating the official theme music for the 2026 FIFA World Cup. He presented a Tascam Portastudio 424 MK III, a Telepathic Instruments Orchid, two Chase Bliss pedals called Lost & Found and Mood MKII, a Hologram Electronics Chroma Console, a Moog Sub37, and an Ableton Move. A review of the Orchid is forthcoming. Golden said the gear demonstrates that extensive equipment is not required to make music. The interview is available online. The footage remains accessible, and the listed gear may inform future production approaches.
Apple and the U.S. Department of Justice are in discussions to settle an antitrust lawsuit filed in 2024 that alleges the company maintains a smartphone monopoly. The suit claims the monopoly affects five aspects of the iPhone experience, including super apps, cloud gaming, third‑party messaging, smartwatches and digital wallets. Bloomberg reported that Apple has made multiple settlement offers this year, though the negotiations remain private and no agreement is guaranteed. The department under the Trump administration has prioritized settlement to reduce litigation costs and provide quicker consumer relief, according to Stanley Woodward, the agency’s antitrust deputy. Apple filed a joint status report with the DOJ and recently won a discovery dispute over access to federal agency documents. The company has implemented several changes in the past year, such as adding RCS support to the Messages app, launching a Mini Apps Partner Program and expanding NFC access. Whether these measures satisfy the agency’s concerns is still unknown. The talks could have implications for Apple’s upcoming leadership transition in September.
The Department of Justice issued a memo stating that federal employees may download TikTok onto agency‑issued devices, subject to each agency’s discretion.
The guidance follows a 2024 settlement reported by Reuters that required ByteDance to limit its ownership stake and to store United States user data in American data centers.
Agency heads will decide whether to permit the app, and security officials must assess compliance with federal data protection requirements before any installation.
The policy change leaves open questions about the app’s continued compliance with government security standards.
A 2025 AAA survey found that just over half of prospective EV buyers consider public charging a concern. Last week a 600‑mile road trip from New Hampshire to Montreal was completed in an Audi e‑tron, which has a range of about 220 miles per charge, with three charging stops and no reported delays.
Three years earlier the same route required multiple service calls and resulted in several hours of lost travel time.
The trip used the A Better Route Planner app, owned by Rivian, to schedule stops. The first stop was a Rivian charger in Lebanon, New Hampshire, offering six 300‑kilowatt chargers, all operational, and no wait time was observed. The charger accepted a credit card and delivered more than 140 kilowatts, close to the vehicle’s maximum rate. The same chargers were used on the return trip with similar results.
At a Circuit Électrique station near Montreal a card‑reader failure required the driver to download the operator’s app and add $20 Canadian before charging could continue; the session then proceeded without issue. The stop was not essential for range but was made for a break and a coffee.
Each charging session lasted about 20 minutes and was combined with a meal or rest, so the car was never idle for an extended period.
The Joint Office of Energy and Transportation reported that the United States had about 32,000 DC fast chargers in July 2023, a number that has more than doubled to over 64,000 today. Expansion by Tesla and other companies has increased access to Tesla’s network, which now allows most EV drivers to use its stations.
Paren’s reliability index, which measures successful charging sessions and station downtime, improved from 85 in 2023 to the mid‑90s in 2026. Tesla’s network remains the largest, but competition from other operators has driven overall reliability gains.
Some stations still malfunction and broken chargers are repaired more quickly than in the past, but the network continues to grow each month.
The experience shows that fast charging has become considerably more dependable, though continued expansion and maintenance will determine its long‑term consistency.
The author of the Handy project announced the release of transcribe.cpp, a C++ transcription library built on the ggml framework. The release comes as developers seek a single, reliable implementation for speech‑to‑text tasks that can run locally on a variety of platforms. Existing options such as whisper.cpp and ONNX require separate model adaptations for different engines, and CPU‑only inference often falls short of performance targets. transcribe.cpp supports 16 ASR families and more than 60 models, with acceleration options for Vulkan, Metal, CUDA and TinyBLAS. The author reports that the library has been numerically validated against reference implementations and that word‑error‑rate tests confirm equivalence with the original models. It provides streaming and batch transcription capabilities and can run on CPUs, GPUs or specialized hardware such as the RK3566. The library is presented as a drop‑in replacement for whisper.cpp, allowing use of existing model files while adding support for GPU‑accelerated inference. It includes bindings for Python, JavaScript/TypeScript, Rust and Objective‑C/Swift, and the author invites community contributions for additional language interfaces. The author notes that local inference reduces reliance on cloud services and can operate in low‑power environments, citing an RK3566 device that processes models faster than real time with minimal wattage. While the library currently covers most publicly available models, the author plans to add missing models and expand language bindings. The project will be maintained as part of the Handy ecosystem. The author thanked Mozilla AI, ggml contributors, Modal, Blacksmith and Hugging Face for support and resources that facilitated testing, CI/CD and model distribution. The release represents a step toward making high‑quality speech transcription more accessible on consumer devices, though broader adoption will depend on further development and community participation.
Apple increased hardware prices in June and has since raised costs for several services and protection plans. The hardware adjustments, announced on June 25, raised the price of the 13‑inch MacBook Air by $200 to $1,299, the base iPad by $100 to $449, and the MacBook Neo by $100 to $699. Additional products, including the MacBook Pro, iPad Air, iPad Pro, iPad mini, Mac mini, iMac, Mac Studio, Apple TV 4K, HomePod, HomePod mini and Apple Vision Pro, also saw price changes.
Apple attributed the increases to higher costs for memory and storage components, noting that demand from AI data centers has strained supply. CEO Tim Cook said the company is facing unprecedented pressure on memory availability.
Price changes have also affected refurbished devices; a 256GB refurbished MacBook Neo is now listed at $599, matching the pre‑increase price of a new model.
iPhone prices in Japan rose 8% to 11% (¥8,000 to ¥20,000) depending on model, a change Apple linked to the recent weakness of the yen. No price changes for iPhones have been reported in the United States yet, but the upcoming fall lineup is expected to be more expensive.
AppleCare+ fees rose by $5 per year for new plans covering current Macs and iPads, while existing subscribers retain original rates. Apple Music individual plans are now $11.99 per month and family plans $19.99, with Apple citing higher licensing costs. Apple One Family and Premier plans each increased by $2, while the Individual tier remains unchanged.
Overall, the price adjustments reflect component cost pressures, licensing expenses and currency effects, and Apple has not indicated which products may be affected next. The company has not specified future price changes, leaving uncertainty for consumers about upcoming cost increases across the Apple ecosystem.
OpenAI announced a series of usage limit resets for its Codex and ChatGPT Work services for paid subscribers. The resets occurred over the course of a week and included full resets, banked resets, and temporary 2X limits. In total, 35 resets have been recorded, with an average interval of 8.9 days and the longest gap lasting 67.7 days, according to the company's public data. The announcements came amid reports that usage limits had been consumed more quickly than expected for some users. OpenAI said investigations identified specific technical issues, such as a rolled-back optimization that affected cache hit rates and a bug that prevented a promotional 2X increase from applying to a subset of plus and pro users. In response, the company reset limits for all affected accounts and credited additional resets to user balances. OpenAI noted growth in active user numbers, reaching 9 million for Codex and ChatGPT Work combined. The company said the frequent resets were part of ongoing efforts to maintain system stability and to acknowledge milestones such as the one‑year anniversary of Codex and surpassing three million weekly users. Mitigation steps included a system rewrite to track usage and backfill usage data, as well as temporary capacity expansions to support higher demand. OpenAI said it would continue monitoring performance and adjust limits as needed.
Francesco Cicala and a team of researchers released a preprint on arXiv on July 10, 2026 describing how self-replication can emerge spontaneously in digital primordial soup environments and co-evolve with problem‑solving abilities.
Evolutionary algorithms normally embed reproduction mechanisms, but the authors note that self‑replication can develop from randomness alone.
The paper initializes a population of random 32‑byte Z80 assembly programs and applies mutation and pairwise interaction operators. A task‑based validation step raises a program’s interaction probability when it correctly evaluates a polynomial. Across multiple experiments the authors observed that self‑replication and mathematical problem solving co‑evolved from the initial randomness.
The computational pressure to solve the polynomial promoted the emergence of compact, robust reproductive structures that retained memory for task execution. Introducing metabolic constraints increased the frequency of conditional halting, allowing programs to terminate early during validation while still replicating during interaction. Partitioning programs into spatial task niches caused self‑replication to generate an implicit learning curriculum, using simple solutions as steps toward more complex polynomials.
These results show that environmental task demands actively shape the physical architecture of self‑replication, while spontaneous replication alters the evolutionary trajectory of functional problem solving, opening new avenues for research in evolutionary computation.
Apple confirmed it has removed three of the apps that San Francisco city attorney David Chiu had asked it and Google to take down, saying the apps violate its App Review Guidelines that prohibit software designed to generate, distribute or consume pornography. The company said it has also begun terminating the developer accounts of the removed apps. Chiu’s office identified 13 apps capable of creating non-consensual nude images; Apple’s statement did not specify which of the 13 were removed beyond the three. The policy applies to any app that produces synthetic nudity without consent. Apple said it rejects such apps during the review process and removes them when flagged by users.
Google announced that it will open‑source its 3D emoji set, making raw .OBJ files available to the public. The company said the files will let developers integrate the emojis into virtual reality experiences, independent applications, or other projects. The announcement came alongside a description of the design process used to create the emojis for World Emoji Day. Google explained that creating 3D versions required deciding on basic shapes such as spheres, masks, or flat discs. The Noto Emoji 3D collection debuted in May and attracted a range of reactions. By releasing the files, Google joins other firms that share assets to encourage community use. The practical impact of the open‑source release has not yet been quantified.
Federal employees can now download TikTok on government devices, the Department of Justice said.
A 2022 law barred the app on official equipment, but the DOJ said the restriction no longer applies after TikTok’s U.S. operations were transferred to a joint venture owned by Oracle, Silver Lake and MGX.
Oracle serves as the security partner for the venture, and ByteDance retains a 19.9% stake.
The DOJ memo said President Donald Trump authorized “employees of Executive Branch agencies” to download TikTok on official devices, subject to agency discretion and applicable workplace policies.
The ban had previously taken effect early last year, after which the app experienced a brief outage before Trump repeatedly delayed enforcement and urged providers to restore service.
Agency officials must weigh the decision against internal security guidelines.
Amazon lists the GoPro Max 2 accessory bundle at $369, a $1 higher price at Best Buy and GoPro. The Max 2 is a 360-degree camera that records 8K video at 30 frames per second or 4K at 60 frames per second and includes a six‑mic array that reduces wind noise. It features a 1.82‑inch touchscreen for reviewing footage and settings, and it is waterproof to a depth of 16 feet. The bundle adds a four‑foot extension pole, two protective lens caps, two high‑capacity batteries, a 64 GB microSD card, a curved adhesive mount, a mounting buckle for GoPro mounts, a mounting finger adapter, two thumb screws, a wrist lanyard, a microfiber cloth and a USB‑C cable. The standalone camera ships with fewer mounts, a single battery, no microSD card and no extension pole. The pole and an additional battery are sold separately for about $75. The six‑mic array records in all directions and can be paired with Bluetooth microphones. Battery life is estimated at 66 minutes at full 8K resolution or 90 minutes at 5.6K, and the lenses are replaceable if damaged. Potential buyers may compare the bundle with other 360‑camera options as GoPro continues to update its product lineup.
LG monitors have been found to automatically install software through Windows Update without prompting users. Gamers Nexus reported that the installer appeared on 31 of 32 consecutive system boots, showing a McAfee trial prompt that converts to a paid subscription; on the remaining boot it promoted an LG monitor utility. The installation occurs without a consent dialog and the Microsoft Store listing for the application includes internet and system resource permissions. Dell also employs a similar automatic installation method for Alienware Command Center when compatible devices are detected, which can be blocked via Windows Group Policy settings. The practice raises privacy concerns for monitor owners and highlights the need for greater transparency in automated software distribution.
Experimental Social Scene is developing Strandfall, a physical orienteering live action role‑playing game scheduled for September in a park in Edinburgh. The event will involve 30 participants who will use custom McNair‑Feldman Devices to locate missing scientists, forecast storms, and operate a long‑range mesh radio network over three hours. The devices are 3D‑printed enclosures that house a low‑power radio, GPS, accelerometer, compass and a small e‑paper display, and are intended to function as scientific instruments rather than consumer smartphones. Organizers say the hardware enables gameplay that would be difficult to simulate manually. Players will coordinate tasks, relay information, and may pursue storm “chasing” while others map exclusion zones. While smartphones could be used, organizers note that their typical functions conflict with the immersive, face‑to‑face nature of the experience. The event is low‑tech, not no‑tech, and will include maps, thermal printers, and role‑play elements. Veteran Nordic LARP designer Juhana Pettersson is consulted for design guidance. Strandfall is site‑responsive and can be adapted to other locations. Participation is free; details are available on the project website.
Over 80% of Samsung Foundry workers said they plan to leave their jobs, according to a union‑conducted survey of 8,297 employees. The survey, carried out by the Samsung‑based arm of the United Enterprise Labor Union, found that 81.5% of respondents expressed an intention to quit, a threshold the union says is "very high."
Foundry staff are divided from the memory division, which receives bonuses tied to operating profit targets of 200 trillion won for 2026‑2028 and 100 trillion won for 2029‑2035. Memory workers are projected to earn about 600 million won (roughly $400,000) in bonuses this year, while foundry employees receive only about 6 million won (about $4,000).
Kim Yong‑kwan, president of Samsung Electronics’ Device Solutions division, said at a town hall on July 3 that the division expects to meet its operating profit goal for the year, which analysts estimate at around 300 trillion won ($200 billion). He added that memory‑division profits are expected to exceed the cumulative earnings of the division since its entry into the semiconductor business.
The union noted that dissatisfaction is highest in the Foundry division, with 75.4% of workers in Samsung LSI and 60.6% in the Semiconductor R&D Center also indicating willingness to switch jobs, compared with 32.7% in the memory division.
Samsung has not publicly responded to the findings. The pay disparity remains unresolved, and its impact on talent retention at the company is uncertain.
Sony unveiled the Bravia 9 II RGB LED television, priced at $3,600 for the 65‑inch model. The Verge's review noted that the model adds more dimming zones, an anti‑reflective screen and higher peak brightness than the preceding Bravia 7 II. In HDR testing, the 65‑inch unit reached 3,800 nits, while full‑field brightness measured 885 nits. The 7 II reached 2,200 nits for highlights and 848 nits for full‑field brightness. The Bravia 9 II includes two HDMI 2.1 ports and two HDMI 2.0 ports, an ATSC 3.0 tuner and Google TV. The processor is the same as in the Bravia 7 II. The lenticular stand design is retained. The TV supports 4K/120Hz gaming, ALLM and VRR. Reviewer measurements were conducted with Calman color software, a Murideo 8K pattern generator, an X‑rite i1 Pro 3 spectrometer and a Konica Minolta luminance meter. The anti‑reflective screen reduced ambient light reflections, though the effect diminished at angles beyond 70 degrees. Off‑axis viewing showed some blooming around bright elements, especially at the edges of the screen. Sony reported 115‑zone dimming for the 115‑inch model, with finer control than the Bravia 7 II. The 65‑inch model uses fewer zones but still improves blooming control compared with previous RGB LED TVs. The review noted that colors such as sky blue, foliage and skin tones appear accurate, while SDR reds were slightly oversaturated. The TV’s gamma and EOTF tracking performed better than the Bravia 7 II in professional mode. Sony says the Bravia 9 II is the best bright‑room LED TV available. The Bravia 9 II demonstrates higher brightness and color accuracy than its predecessor, but its cost remains higher than competing OLED and mini‑LED models, and off‑axis viewing still exhibits some blooming.
Gleam, described as a language for building type‑safe and scalable systems, shows strong adoption with Rust accounting for 92.9% of its usage, according to a repository on Tangled.org.
The repository at https://tangled.org/gleam.run/gleam lists programming language distribution percentages and provides clone and download options. It reports 11k stars, 1 fork, and a 0 rating.
Language usage breakdown includes Rust 92.9%, JavaScript 3.6%, Gleam 1.8%, HTML 0.4%, CSS 0.4%, Erlang 0.3%, Shell 0.3%, TypeScript 0.1%, Makefile 0.1%, Dockerfile 0.1%, Elixir 0.1%, PowerShell 0.1%, and Other 0.2%.
Users can clone the repository via HTTPS or SSH, or download a tar.gz or .zip archive.
The statistics indicate Rust remains the dominant language for Gleam projects, while other languages occupy minor shares, with no further development milestones announced.