A study published in Science shows that the tallest tropical Dipterocarp trees can transport water to their crowns without impaired hydraulic function, and that their growth rates were unchanged during a severe 2023‑2024 drought. Conventional theory held that greater height in trees would increase hydraulic resistance, limiting water transport and making tall trees more vulnerable to drought. The research examined Dipterocarp trees ranging from 7 to 71 metres tall in Malaysian Borneo, measuring water‑carrying vessels, leaf adaptations, and trunk growth before, during, and after the El Niño drought period. Findings indicated that taller trees compensated for height by widening vessels near the ground and modifying leaf traits to withstand greater water stress, resulting in no height‑related loss of growth. Professor Lucy Rowland of the University of Exeter said the hydraulic system of very tall Dipterocarp trees is perfectly evolved for its height, and should not suffer more than smaller trees exposed to the same drought conditions. Dr Paulo Bittencourt, now at Cardiff University, noted that the tallest 1% of trees store more than half of above‑ground carbon in forests and that predictions of higher drought mortality for these trees may be incorrect. The study suggests that existing climate‑change models, which include a weaker hydraulic system for tall trees, might need revision. Further research is needed to investigate the hydraulic systems and drought resilience of other tall tree species.
Models, chips, platforms, and the policy lines being drawn around them.
New AI‑Powered Browsers Challenge Chrome and Safari 2026 The browser market is evolving as several companies launch AI‑focused browsers that aim to perform tasks on behalf of users, challenging the dominance of Chrome and Safari. Chrome and Safari together hold the majority of global browser usage, and recent developments have shifted the competition toward AI capabilities rather than search engine features. Perplexity's Comet, a chatbot‑based browser, is available to users on the $200 per month Max plan and has a waitlist for broader access according to the company. The Browser Company introduced Dia, an AI‑centric browser that integrates a chat tool and is currently offered as an invite‑only beta to Arc members with a public waitlist for non‑members. Opera's Neon provides contextual awareness and can conduct research, shopping or code writing, including offline tasks, and costs $19.90 per month for macOS and Windows users the company says. OpenAI launched Atlas which lets users query ChatGPT for search results and browse the web within the chat interface; the macOS version launched in October with versions for Windows iOS and Android planned according to OpenAI. Aside a Y Combinator‑backed startup plans to offer browser‑native automation that can fill forms and manage data across services such as Gmail Notion Slack and banking platforms; a waitlist is open for early access. Jatter launched in June providing AI‑driven page analysis personalized recommendations and an integrated notes feature and is free with an optional $10 per month subscription the company reports. Brave known for ad and tracker blocking and a built‑in cryptocurrency reward system also offers a VPN AI assistant and video calling according to its developer. DuckDuckGo added a chatbot and expanded its scam detection to cover cryptocurrency exchange scams scareware and fraudulent e‑commerce sites while continuing to block trackers and ads and not collect user data. Ladybird led by former GitHub CEO Chris Wanstrath is developing an open source browser from scratch that will block ads and third‑party cookies; an alpha version is scheduled for 2026 on Linux and macOS. Vivaldi built on the Chromium project by a former Opera developer offers a customizable interface color‑changing windows ad blocking a password manager and no user data tracking the company states. Opera Air a mindfulness‑focused browser includes break reminders breathing exercises and binaural beat options for focus or relaxation the company notes. SigmaOS a Mac‑only browser with a vertical tab layout and workspace groups provides AI summarization and translation features and offers unlimited workspaces for $8 per month the company says. Zen Browser an open source project aims to create a calmer internet experience with tab workspaces split view and community‑made themes and extensions according to its developers. The market remains fluid as these services compete for users and the ultimate impact on browser usage and user privacy is yet to be determined.
The number of high‑ and critical‑severity Common Vulnerabilities and Exposures (CVEs) reported by 21 notable organizations rose to approximately 1,500 in June 2026, exceeding the prior monthly record by more than 3.5 times, according to a dataset released July 2, 2026.
The increase follows Anthropic’s April 2026 announcement that its Claude Mythos Preview model could autonomously identify software vulnerabilities, and that its Project Glasswing initiative, which includes Microsoft, Google, Apple, and Amazon Web Services, had been using the model to locate and remediate bugs before the model’s public release. Project Glasswing reported finding more than 10,000 high‑ or critical‑severity vulnerabilities since the program began, many of which have not yet been individually disclosed. OpenAI also launched its Daybreak product to harden software using similar AI capabilities.
The spike in disclosed CVEs was recorded after the Claude Mythos Preview announcement, with the monthly total surpassing the previous record set before the announcement. The data, compiled from monthly disclosures by the participating organizations, shows a sharp upward trend beginning in June.
The broader effect of AI‑driven vulnerability discovery on software security remains uncertain, and further monitoring will be needed to assess whether the elevated CVE count reflects a lasting change in the threat landscape.
Chevrolet unveiled the Silverado EV, a full‑size electric pickup, in 2026 with a claimed range of more than 400 miles and a 205 kilowatt‑hour battery. The vehicle joins a growing lineup of electric pickups as manufacturers seek to expand into the traditional truck market. GM reported selling about 14,000 units in the United States and Canada during the 2025 model year, compared with roughly ten times that number of gasoline‑powered Silverado trucks sold each quarter. The EV’s price for the LT Extended Range trim is listed at about $66,000, roughly $5,000 above the average transaction price of full‑size pickup buyers, according to industry data. GM says the battery delivers 410 miles of range on a full charge, while the LT Max Range model adds 68 miles at a higher price. Independent testing recorded an efficiency of 2.1 miles per kilowatt‑hour, about 10‑20 percent better than a comparable gasoline vehicle, according to the reviewer’s measurements. The truck includes a frunk, a rear‑mounted battery pack positioned mid‑ships, and a 205 kilowatt‑hour battery that GM describes as providing ballast and stability. Chevrolet equipped the cabin with a Google‑powered infotainment system that supports voice commands and offers a Super Cruise hands‑free driver‑assistance feature classified as Level 2. Super Cruise performed well in highway driving but exhibited occasional lane‑keeping challenges when surrounding vehicles changed speed rapidly, as reported during testing in the Detroit metro area. The vehicle’s exterior styling echoes the older Avalanche model, featuring four doors, a short bed that can extend into the cabin, and a “sail” panel intended to improve aerodynamics. Chevrolet noted plans to adopt a lithium‑manganese‑rich battery chemistry that could reduce costs by roughly $6,000 while maintaining range, potentially bringing the vehicle closer to price parity with gasoline versions later in the decade. The company has not disclosed a timeline for the battery update, and it remains unclear whether the price reduction will stimulate higher sales.
Google, Lumen, Shadowserver, the FBI and other partners disrupted the NetNut residential proxy network, which comprised roughly 2 million devices, mainly small streaming hardware, in July 2026.
Residential proxy services route traffic through ordinary home devices to hide the origin of user requests. They are legal but are frequently used by cybercriminals to conceal activity.
Researchers said the botnet was one of the most widely used residential proxy services. Google Cloud noted that the network enabled threat actors to mask IP addresses during attacks, including password spray campaigns. In a single week of June 2026, Google’s Threat Intelligence Group identified 316 distinct threat clusters using NetNut exit nodes, involving both criminal and espionage groups. The network also intersected with other botnets such as Badbox 2.0 and Mirai variants, according to public reports.
The takedown follows the January disruption of the IPIDEA proxy network. Google’s Threat Intelligence Group said that when a botnet is degraded, operators often purchase capacity from competitors, becoming resellers, which can sustain the ecosystem. The agency plans to continue monitoring the network and mapping how peers adapt. Google indicated that similar ad hoc actions may be taken in the future, but long‑term mitigation will require cooperation from ISPs, mobile platforms and other technology firms.
The residential proxy market continues to expand, and the effectiveness of targeted takedowns is limited without broader industry support.
Amazon said its low‑earth‑orbit satellite constellation Leo reached 396 satellites after a July 2 launch that added 29 spacecraft. The company said it plans to begin broadband service later this year.
The constellation, formerly known as Project Kuiper, is part of Amazon's plan to deliver internet connectivity from space. It competes with SpaceX's Starlink and operates under a Federal Communications Commission license for up to 3,232 satellites.
The July 2 launch used a United Launch Alliance Atlas V rocket to place the 29 satellites into low Earth orbit.
Melissa Wuerl, director of Launch Systems at Amazon, said the firm has hundreds of flight‑ready satellites at the Cape and a dedicated vertical integration facility to support future missions, which will increase launch cadence.
Amazon holds a licence for 3,232 satellites and intends to eventually operate more than 7,700, but it missed the FCC requirement to have half of the planned constellation in orbit by July 30 2026.
The company sought an extension in January and received a conditional approval last month.
The service is projected to deliver download speeds from 100 Mbps to 1 Gbps, depending on the antenna option selected.
In April, Amazon agreed to purchase Globalstar and its 24‑satellite network, which supports Apple's direct‑to‑device satellite feature, to gain a foothold in that market.
Starlink currently operates about 10,400 satellites, far exceeding Amazon's current fleet.
The UK regulator Ofcom granted Amazon a license to provide broadband services to British consumers in 2023.
Amazon has not disclosed pricing for its service, and it remains unclear how the company will meet its orbital deployment targets to launch the broadband service.
A KPMG survey of 2,145 senior leaders across 20 countries found that 29% of corporate executives have difficulty understanding and controlling AI operating costs as they scale deployments, a challenge that has grown alongside a shift to usage‑based pricing models.
The survey, released in July 2026, indicates that many organizations are still building the capabilities needed to forecast, monitor, and manage AI spending effectively.
Nearly half of the respondents said they have rephased AI projects when costs exceeded expected value, and lower‑cost, high‑fidelity models have become the fastest‑growing influence on AI strategy, up seven percentage points from the first quarter.
Amazon plans to spend about $200 billion this year on AI capacity in its AWS data centers, a 50 % increase over the prior year, while Microsoft expects total capital expenditure of $190 billion, a 61 % rise.
Both companies are investing heavily in forward‑deployed engineering: Amazon has committed $1 billion to an AWS Forward Deployed Engineering team to help customers adopt AI agents and shorten deployment timelines, and Microsoft is providing $2.5 billion to establish Microsoft Frontier Company, which aims to let customers amplify their intelligence with AI.
KPMG notes that governance remains a concern, with questions about who is responsible for decisions made by statistical models that can produce erroneous outputs, commonly referred to as "hallucinations".
Executive accountability is important, but the consultancy says governance ultimately depends on day‑to‑day operating practices, including clear rules for employee intervention, cost ownership, output review, and incident handling; few organizations describe these practices as fully embedded.
Additionally, GPTZero claimed that a forensic review of KPMG's October 2025 report "Total Experience: Redefining Excellence in the Age of Agentic AI" found only five of 45 citations accurately referenced the intended source, with the remainder containing misleading or vague information; KPMG subsequently removed the report from some websites and issued a statement reaffirming its commitment to content accuracy and responsible AI use.
As AI spending accelerates, the ability of companies to manage costs and establish clear governance frameworks will determine the broader success of enterprise AI initiatives.
An informal guide to threat modeling published on a personal blog provides a beginner‑level framework for identifying and evaluating security risks. The article states that threat modeling is a formal process used by security professionals, but it can also be applied informally during the design of a product or service without restriction.
The guide outlines seven key questions to answer: what is being protected; who or what might want to harm that asset; how an attacker could exploit the asset; what assumptions are being made; which threats are intentionally excluded; how assets are related in a system; and how the model will be updated. It recommends drawing the system architecture as a graph, noting component relationships, and iterating the analysis until deeper layers are examined.
The article critiques an existing threat model for the Matrix messaging platform, noting that it lists attack vectors without specifying underlying assumptions or asset relationships and has not been revised since 2021. In contrast, the guide emphasizes that threat models should be living documents that are updated as designs change.
The piece also discusses authentication choices, recommending passkeys over passwords to reduce credential‑stuffing attacks, and describes a proposal to standardize non‑hybrid post‑quantum key encapsulation mechanisms in the IETF. The author explains that the proposal supports compliance with government cryptography requirements and that the security implications depend on the chosen algorithm rather than the presence of a backdoor.
Modern smartphones and laptops perform billions of calculations each second while presenting a seamless user experience. The device contains a silicon chip made of purified rock, etched with microscopic circuits that manipulate electron flow. These circuits are built from transistors that act as switches, forming logic gates that enable binary arithmetic. The design evolved from mechanical calculators and vacuum‑tube computers to solid‑state transistors and integrated circuits. Process nodes measured in nanometers allow billions of transistors to occupy a few square millimeters. Software is a set of instructions stored in memory that the CPU fetches, decodes, and executes. Early computers required rewiring for each task; the stored‑program architecture lets a single processor run many programs by loading different instruction sequences. A processor’s control unit, arithmetic logic unit, and registers coordinate this process, while a clock signal synchronizes operations at gigahertz frequencies. Binary digits, or bits, represent 0 and 1 using two‑state switches. Logic gates—AND, OR, NOT—combine bits to implement any logical rule; an XOR gate produces a sum bit, an AND gate generates a carry in binary addition. Half‑adders and ripple‑carry adders cascade these circuits to add multi‑bit numbers. A flip‑flop stores a single bit by feeding its output back into its input; a register groups many flip‑flops for faster storage. RAM is volatile because it requires continuous power to maintain the stored state, whereas hard drives and SSDs retain data without power. Modern CPUs contain multiple cores that execute instructions in parallel, and specialized units such as GPUs and neural accelerators handle specific workloads. Clock speed, measured in gigahertz, determines how many fetch‑decode‑execute cycles occur each second. High‑level languages such as C, Python, or JavaScript translate human‑readable code into machine instructions via compilers or interpreters. Functions encapsulate repeated tasks, and compilers convert loops and conditionals into assembly instructions that the CPU executes directly. Interpreted languages incur additional overhead because each operation passes through a runtime interpreter. Operating systems enable multitasking by switching CPU time among processes using timers and context switches. Virtual memory isolates each process’s address space, translating virtual addresses to physical RAM. The file system abstracts storage devices, providing hierarchical paths instead of raw disk blocks. The Internet routes packets using IP, while TCP ensures reliable delivery through acknowledgments and retransmissions. HTTPS encrypts traffic with a handshake that establishes a temporary key, preventing eavesdropping on intermediate nodes. Web browsers act as sandboxes that render HTML pages and execute scripts without direct access to the operating system. Web browsers and virtual machines add abstraction layers that increase memory use and latency compared with native binaries, but they simplify distribution and cross‑platform compatibility. Software engineers must balance these trade‑offs when designing systems for performance, safety, or scalability. As AI tools lower the barrier to writing code, a solid grasp of the underlying hardware and software concepts remains essential for debugging, verification, and securing systems.
According to the Searxng GitHub repository, the project is an open-source metasearch engine that aggregates results from multiple search providers. The software is available at no cost and can be self-hosted.
The initiative was created to provide a search engine that emphasizes user privacy. It supports queries from Google, Bing, DuckDuckGo, and other services, and allows users to customize result ranking and filtering options.
The repository contains more than 9,500 commits and includes documentation, configuration files, and source code primarily written in Python. Features include support for multiple search backends, language detection, and custom query parameters. The project is licensed under the Apache License 2.0 and is maintained by a community of contributors.
Recent commits show continued development and active maintenance of the codebase.
Sony plans to end physical PlayStation 5 game production by 2028, and the Austrian plant that prints Blu‑ray discs for the titles is being converted to manufacture micro‑optical lenses. The €30 million investment to retool the facility reflects a marked decrease in demand for disc manufacturing. ORF.at reported that DADC’s CEO said order volume has decreased, with PlayStation accounting for roughly 50 % of the plant’s output today and about 10 % expected by 2028. Employees were briefed that the change will not result in job losses. DADC intends to shift production capacity toward micro‑optical components for Sony, indicating a continued focus on non‑disc technologies. The transition suggests physical game distribution will be phased out, with no indication of a reversal.
Amazon updated its 2023 Fire HD 10 tablet, adding 4GB of RAM to the 32GB model and raising the price to $154.99. The tablet originally shipped with 3GB of RAM across all storage options, and the 64GB version continues to include 3GB. The update retains the 10.1‑inch 1920×1200 display, 2GHz eight‑core processor, 13‑hour battery life, and microSD card expansion. The device is sold only with lock‑screen ads, which can be removed for an additional fee. Amazon still sells the previous 3GB/32GB model for $139.99, though stock may be limited. Amazon has not disclosed the reason for the change, but the source said the update is tied to a shortage of memory chips caused by increased demand from AI data centers.
A father and engineer visited a primary school class to discuss how products are made, using his AI‑controlled clock as a case study. He showed sketches, CAD models and photos of the device’s evolution from initial concept to 3‑D printed prototype, then to injection‑molded parts and final assembly on a production line. He explained that testing includes vibration machines to ensure durability and that packaging design plays a role in the final product. The presentation highlighted that manufacturing takes place in ordinary facilities, not the stylized media depictions often shown, and encouraged students to consider careers in design, engineering and invention. He said that seeing the step‑by‑step process helps demystify manufacturing and shows that complex products are created by people. He urged other professionals to share similar experiences in schools, noting the curiosity and engagement it generates among young learners.
Dune, a startup, announced a three‑key aluminum keypad that plugs into a MacBook’s USB‑C port and can be programmed for different applications. The $119 device, currently available at an introductory price, will be priced at $149 after the introductory period ends. It is compatible with M2 Air or later and M1 Pro or later MacBook models running macOS 15 Sequoia or newer. The keypad draws power directly from the laptop and contains no battery. Each of the three buttons changes function based on the active application, for example toggling microphone and video in meeting software, copying and pasting in spreadsheet apps, or refreshing and navigating in a web browser. Developers can assign the keys to keyboard shortcuts, commands, or links through a companion app. The app also lets users set system‑wide shortcuts and syncs with the calendar to provide meeting reminders. Users can create custom shortcuts with a Python script or, without coding, describe a desired action in plain language to Claude Desktop, which generates the code. A marketplace will eventually host user‑created skills, though currently only a few skills are available. The startup notes that the keys have low resistance, making accidental presses possible; users have reported unintended muting or camera shutdown when the device contacts a nearby object. Competing products such as MuteMe provide only mute functionality, while Stream Deck offers macro‑focused controls for business use. Dune is marketed as a hardware‑based controller for productivity tasks.
Eric Berger, senior space editor at Ars Technica, wrote a reflective piece on the United States’ 250th anniversary, noting both historical milestones and current challenges.
He recalled his childhood memory of the 1976 bicentennial parade in Michigan, where his father, a veteran and city councilman, spoke about democracy. He observed that the nation had achieved major accomplishments such as leading Allied efforts in World War II, advancing civil‑rights activism, and landing on the Moon, while also experiencing economic growth in the 1980s and 1990s when the United States was the sole superpower.
Berger described a series of events that have affected public confidence, including the September 11 attacks, the wars in Afghanistan and Iraq, the 2008 financial crisis, the spread of misinformation online, and increasing political polarization. He said surveys indicated optimism about the nation’s direction had fallen to a record low in 2026, lower than during the pandemic. He also noted that the White House had proposed cuts to science funding and that Congress had resisted those cuts in a bipartisan manner.
Despite these challenges, Berger said he continues to see reasons for optimism. He described his work with Space City Weather, which provides factual weather information without sensationalism, and his coverage of space exploration, where private companies and government agencies develop satellites for environmental monitoring, renewable energy, and global communications. He added that scientific progress continues in areas such as cancer research, aging, and genetics, and that technology has enabled rapid knowledge sharing and global collaboration. He concluded that the United States retains the capacity to remake itself through elections, immigration, free speech, and economic mobility, and that individual citizens can influence change.
He urged readers to engage in the democratic process and to support evidence‑based information, suggesting that collective effort offers a path forward as the nation approaches its 250th birthday.
The European Commission may amend its proposed datacenter environmental rating system after lobbying from industry groups, potentially allowing facilities to offset emissions with clean‑energy certificates from any EU member state.
The original March draft set an A‑to‑G rating based on energy and water efficiency and allowed emissions offsets only for projects located in the same region as the data centre. A revision disclosed by the Financial Times would broaden the geographic scope of eligible certificates.
Industry groups, including datacenter operators and technology companies, requested the amendment, arguing that the stricter regional requirement could increase operating costs. The Climate Neutral Data Centre Pact, the Cloud Infrastructure Service Providers in Europe, and the European Data Centre Association have offered feedback, with the latter reaffirming its commitment to climate‑neutral datacenters while noting that Europe must also address grid expansion, permitting processes and stable access to low‑carbon electricity to meet AI and cloud capacity goals.
The Commission plans to discuss the proposal with member‑state representatives on Thursday. Adoption is scheduled for the third quarter of this year, though the debate may delay the timeline.
Steam Controller Auto-Charge is an open-source web application that automatically guides a Steam Controller to its magnetic charging puck using optical flow computer vision and WebHID telemetry. The application tracks selected points on the controller and puck with OpenCV.js from an overhead camera, then uses the WebHID API to send 70 Hz asymmetric haptic pulses through the controller’s linear resonant actuators, steering it toward the puck. When the controller comes within 150 pixels of the puck, the haptic pulse frequency is reduced to ensure a smooth magnetic connection. Battery status is monitored by intercepting report IDs for charging and battery voltage, allowing the interface to display live battery percentage and voltage. Setup requires the Nix package manager and a Chromium-based browser that supports WebHID, plus an overhead webcam. A single command installs dependencies and launches the development server. After positioning the camera, the puck, and the controller, the user connects via WebHID, activates auto-track, and the application navigates the controller autonomously. Manual tracking can also be used by selecting points on the interface. The project is released under the MIT License.
On the 35th anniversary of Terminator 2, several founding members of Industrial Light & Magic’s original computer graphics department gathered to discuss the technical challenges of creating the film’s liquid metal T‑1000 character.
The computer graphics group at ILM, established in the early 1980s, pioneered digital techniques that shaped modern visual effects. Terminator 2, released in 1991, featured the T‑1000, a shape‑shifting metallic entity that required advances in fluid simulation, shading, and rendering to appear realistic on the hardware of the era.
The participants described the need to develop custom fluid dynamics algorithms and specialized shaders to simulate a surface that could flow and reflect light like mercury while maintaining the character’s articulated form. They noted that rendering times were constrained by the limited processing power and memory of contemporary workstations, prompting iterative testing and optimizations. The group also highlighted how the project contributed to later ILM work on fluid effects in films such as Titanic and the Pirates of the Caribbean series.
A documentary titled Light & Magic, now streaming on Disney Plus, provides a detailed look at these efforts and the broader evolution of computer graphics at the studio.
John Baez and Simon Burton created a web-based course on applied category theory using the textbook Seven Sketches in Compositionality: An Invitation to Applied Category Theory. The course comprises 77 lectures covering topics such as ordered sets, preorders, Galois connections, logic, category theory fundamentals, databases, and collaborative design. Lecture titles are listed sequentially from Lecture 1 through Lecture 77, with chapters organized around ordered sets, resource theories, databases, and collaborative design. The material is hosted at https://math.ucr.edu/home/baez/act_course/index.html.
Wordgard is an open‑source JavaScript library that provides an in‑browser rich‑text editor with control over the content types it supports. The library, created by its developer, offers a set of tools for building content editors and includes a programmable interface for custom editor implementations. It is released under the MIT license and is hosted at code.haverbeke.berlin. Users are encouraged to report bugs through the project’s issue tracker, while pull requests are not accepted. Commercial users are asked to contribute voluntarily to help maintain the project. Discussion and questions are best handled on the forum.
A researcher discovered a vulnerability in MSI Center that allows local privilege escalation via a named pipe used by the Notebook Foundation service.
A researcher discovered a vulnerability in MSI Center that allows local privilege escalation via a named pipe used by the Notebook Foundation service.
MSI Center ships with MSI laptops and desktops, and the researcher examined the software after finding similar issues in AMD and ASUS OEM products.
The Notebook Foundation service creates a named pipe named \\.\pipe\MSI_SERVICE_2 with security settings that permit interaction from any authenticated user. Commands such as PC:REXE and PC:KEXE can be sent to execute code with LocalSystem privileges. A proof‑of‑concept script triggered cmd.exe, and remote code execution over SMB is possible when valid credentials are present.
The researcher emailed the vulnerability report to MSI PSIRT on May 10, 2026. MSI responded that a patch was developed and would be incorporated into MSI Center version 2.0.70.0 released June 1, 2026. The researcher requested a CVE identifier through VulDB, and the request remained under review as of July 1, 2026. The researcher also noted that no bug bounty payments have been issued for the reported issues.
The researcher continues to monitor the VulDB review process while awaiting distribution of the patch.
Leanstral 1.5, a free Apache-2.0 licensed model with 6 billion active parameters out of 119 billion total, was released to improve formal verification performance. The model solves 587 of 672 PutnamBench problems and achieves 87 % on FATE‑H and 34 % on FATE‑X, saturating the miniF2F benchmark at 100 % on validation and test sets.
Leanstral was created to make formal proof engineering practical in Lean 4. Since the initial release, the project has offered an open approach to verifying mathematical statements and code properties.
Training proceeds through three stages: mid‑training, supervised fine‑tuning, and reinforcement learning using the CISPO algorithm. The model was evaluated on miniF2F, PutnamBench, FATE‑H/X, and FLTEval, which tests proof engineering on real pull requests. In the multiturn environment, Leanstral iteratively submits proofs, receives compiler feedback, and refines attempts until compilation succeeds or the token budget is exhausted. In the code‑agent environment, it edits files, runs shell commands, and queries the Lean language server to manage long‑horizon tasks. Test‑time scaling shows a monotonic increase in Pass@8 as the token budget per attempt rises from 50 k to 4 M tokens, reaching 587 solved problems at the highest budget.
Leanstral achieves a new state‑of‑the‑art on FATE‑H and FATE‑X while costing roughly $4 per solved PutnamBench problem, compared with estimated costs of $300 or more for competing provers such as Seed‑Prover. Other provers that rank higher either require natural‑language guidance or incur substantially higher per‑problem expenses.
In a verification of AVL‑tree time‑complexity, Leanstral produced proofs spanning over 2.7 million tokens across 22 compactions, establishing an O(log n) bound for insertion and deletion. A separate bug‑discovery pipeline used Leanstral to infer correctness properties from Rust code translated to Lean, flagging 47 property violations and identifying 11 genuine bugs, including five previously unreported issues in the zigzag decoding library of the datrs/varinteger project.
The model is released under an Apache‑2.0 license, with weights hosted on Hugging Face and a free API endpoint at leanstral‑1‑5. Users are encouraged to access the model through Mistral Vibe, set up an API key, and install the accompanying tools as described in the project documentation. The release illustrates that rigorous formal methods can be both effective and affordable for real‑world software verification. Its long‑term impact on industry and research adoption remains to be measured.
AdaptHealth disclosed a data breach after attackers used social engineering to gain access to its cloud environment through a third‑party contractor. The company, which provides home medical equipment and serves more than 4.2 million patients nationwide, reported the incident to the Securities and Exchange Commission on Thursday. Investigators determined that the attackers accessed internal patient management systems, document storage platforms, and external electronic health record portals. They obtained a password file associated with insurance billing and other personally identifiable information and protected health information. Social Security numbers and payment details were not believed to be compromised. AdaptHealth disabled the contractor’s user account, reset affected credentials, and implemented additional access controls. The company said the breach is now contained and it has taken steps intended to mitigate the risk of data dissemination. No extortion demand was reported and no cybercrime group has claimed responsibility. AdaptHealth continues to investigate the full scope of the breach and will provide further updates as the investigation proceeds.
Cooler Master announced the MasterFan A Series and MasterFan M Series, 120 mm fans that received the highest Cybenetics Diamond rating for static pressure. The A120, A120 FC and M120 ARGB models are the latest 120 mm releases from the company. Cybenetics Labs tested the fans and recorded the Diamond rating for each model. Full evaluation reports for the A and M series are publicly available. Because of the high static pressure, the fans are suited for liquid‑cooling radiators and cases with limited airflow. The M series adds ARGB lighting for visual customization. The new models are being shipped worldwide beginning this month. Pricing is listed at €12.99 for the A120, €14.99 for the A120 FC and €17.99 for the M120 ARGB. KitGuru asked how much users typically spend on new PC fans.
Sony will announce a new RX10 camera on July 9 at 10 AM ET, marking its first major update in almost nine years. The teaser was posted on Sony’s Alpha Instagram account and shows no specific details about the model. The previous RX10 IV launched in 2017 with a 24‑600 mm 25× optical zoom lens, a 20.1‑megapixel 1‑inch sensor and up to 24 fps shooting with autofocus and tracking. The company has not disclosed technical specifications or a price for the upcoming version.
AMD's MI355X GPU delivers inference throughput that is up to 2.75 times cheaper per GPU than NVIDIA's B200 while offering comparable hardware specifications. The growing demand for tokens in frontier models such as Claude Fable, GLM5.2, and Minimax M3 has driven up NVIDIA prices and token costs. Wafer AI reported that the MI355X achieves aggregate throughput of 2626 tokens per second per node on a workload with 20,000 input tokens and a 60% cache hit rate, representing 80% of the performance measured on a B200 at a fraction of the cost. Single‑stream tests on GLM5.2 using 10,000 input tokens and 1,500 output tokens achieved 213 tokens per second on AMD hardware, according to standards set by Artificial Analysis. The study used quantization of GLM‑5.2 from bf16 to MXFP4 with AMD Quark, a method that proved lossless compared with official FP8 quantization. To improve performance, the team selected sglang as the inference engine, because it supported the quantized weights and avoided compatibility issues observed with vLLM and ATOM. Speculative decode was enabled after two code modifications: aligning the MTP head's module prefix with the decoder stack and adding a guard for deep speculative decode on the ROCm image. These changes increased single‑stream decode throughput to 213 tokens per second. For aggregate throughput, workload characteristics showed the MI355X was prefill bound, limiting performance relative to the B200, which reached 3192 tokens per second at 3.0 requests per second. The study noted that software support, rather than raw silicon, has become the primary factor affecting AMD's competitiveness, as the CUDA moat erodes. While the MI355X does not yet match NVIDIA's peak performance, the measured performance per dollar demonstrates a viable, cost‑effective alternative for inference workloads. Continued optimization of kernel selection and framework support is expected to narrow the performance gap.
Costco’s business model differs from Amazon’s, with a limited SKU count, high member loyalty and lower operating costs that have driven revenue growth of more than 10 percent annually for five years. The retail sector is debating the future of e‑commerce and delivery, as Amazon pursues near‑infinite assortment and rapid home delivery while Costco emphasizes fewer products and in‑store shopping. Costco carries about 4,000 SKUs compared with roughly 130,000 in a typical Walmart Supercenter and limits its selection to improve turnover and reduce logistics complexity. The company’s narrow assortment supports a short or negative cash conversion cycle, allowing it to sell inventory faster and negotiate better terms with suppliers. Its distribution network moves full pallets directly between suppliers and stores without breaking down pallets or using automation, keeping overhead low. Selling, general and administrative expenses represent about 10 percent of sales, compared with Amazon’s 40 percent delivery costs for non‑AWS sales. These efficiencies enable Costco to pay front‑end associates an average wage of $21.29 an hour, resulting in a 6 percent annual employee turnover rate, far below the 60 percent industry average and Amazon’s 150 percent. In May, New York City Mayor Zohran Mamdani announced plans to open a public grocery store in each borough, using a Costco‑style low‑SKU, high‑volume approach and a $2.12 halal wrap as a potential loss leader. The proposal aims to provide cheaper groceries while avoiding inefficiencies that critics fear. The article concludes that although Amazon excels in complex logistics, the simple, high‑volume Costco model offers a practical blueprint for everyday retail and public grocery initiatives.
Texas authorities have charged a 44-year-old man with manslaughter after a June crash in which his Tesla Model 3, operating with driver-assistance software, struck a home in Katy and killed 76-year-old Martha Avila.
Full Self-Driving (FSD) is a Tesla software suite that allows the vehicle to steer, change lanes and navigate traffic without driver input, though the driver must remain attentive. Police said the accelerator pedal was fully depressed moments before impact, overriding the system’s usual controls.
Butler was performing DoorDash deliveries at the time. The vehicle entered the residence, causing fatal injuries to Avila. A lawsuit filed by Avila’s surviving family references a 2023 Washington Post investigation that documented multiple deaths and injuries involving Teslas using driver-assistance features.
Attorney Ryan Zehl said responsibility for the incident could lie with Tesla, Butler, or both.
The case underscores ongoing concerns about the safety of driver-assistance systems and may affect future regulatory review.
On April 23, 2026, the Wikipedia article for the Odin programming language was removed following an articles for deletion (AfD) process. The discussion involved multiple contributors, including editor Helpful Raccoon, who initiated the delete nomination, and other editors such as Stepwise Continuous Dysfunction and GearsDatapacks. After a vote in which five of seven participants supported deletion and one of four supported retention, an administrator deleted the page.
GingerBill, the creator of Odin, said he was not concerned about the deletion itself but criticized the notability criteria applied to programming languages, arguing that reliable sources are scarce and that Wikipedia's policies are ill‑suited to a field dominated by informal, user‑generated content.
Casey Muratori, a game programmer, argued that Odin is notable because companies such as JangaFX use it, and suggested that Wikipedia's rules may be outdated for modern software development.
Jimmy Wales, co‑founder of Wikipedia, responded that the deletion appeared to be based on policy compliance and that reliable sources were needed to demonstrate notability.
Subsequent tweets highlighted disputes over the reliability of corporate statements, the role of primary sources, and broader concerns about the ability of Wikipedia's guidelines to accommodate contemporary programming knowledge. The episode illustrates tensions between Wikipedia's sourcing standards and the evolving landscape of software development.
Apple describes how its four iPad models are intended for different user categories. The lineup includes the base iPad, iPad mini, iPad Air, and iPad Pro, each with distinct specifications and price points. Apple states that the base 11‑inch iPad is positioned for everyday tasks such as streaming, video calls, and light productivity, and it is priced at $449 for a 128 GB configuration. The device includes an 11‑inch Liquid Retina display, an A16 chip, support for Apple Pencil 1st generation and a Magic Keyboard Folio, and it does not support Apple Intelligence. The iPad Air is offered in 11‑inch and 13‑inch versions, starting at $749, and features a 13‑inch model weighing 1.36 lb. It uses the M4 chip with Apple Intelligence support, a 11‑inch or 13‑inch Liquid Retina display with P3 wide color and True Tone, and it supports Apple Pencil Pro and Magic Keyboard. The iPad Pro, beginning at $1199, includes 11‑inch or 13‑inch Ultra Retina XDR displays, the M5 chip with Apple Intelligence, 6K external display capability, and data transfer speeds up to 40 Gb/s. Apple markets the Pro as a high‑performance tablet for professionals requiring advanced processing and display features. The iPad mini features an 8.3‑inch Liquid Retina display, an A17 Pro chip with Apple Intelligence, a 12 MP rear camera, and support for Apple Pencil Pro and Bluetooth keyboards, emphasizing portability and single‑handed use in settings such as medical rounds and aviation. Apple notes that while the mini can run the same apps as larger models, its smaller screen limits productivity tasks. Storage options range from 64 GB to 2 TB across the lineup, and pricing reflects those configurations. The article concludes that selecting a model depends on intended use, budget, and required performance levels.
Axel Rydby, a director on Suicide Squad: Kill the Justice League, said the game’s development shifted from game design to following monetisation spreadsheets.
Rocksteady, the studio behind the Batman Arkham series, moved to a live‑service multiplayer model after years of development and multiple delays; the game launched in 2024 and resulted in a reported $200 million loss for Warner Bros. Games, leading the studio to abandon the project.
In an interview with Bloomberg, Rydby explained that executives increasingly prioritized recouping investment over improving gameplay, noting that the shift began around 2022 and continued through the game’s delayed release. He said he felt he was no longer making games but was following a spreadsheet that was not clearly presented to the team. Rydby left Rocksteady in 2024 and founded an independent studio developing a new title called Secret of Circadia.
Rocksteady has not announced its next project, though industry speculation suggests a return to a single‑player Batman game may be in development.
ProtoTypist has announced a compact mechanical keyboard inspired by Nintendo's Game Boy Advance. The keyboard includes programmable shoulder buttons with pre‑installed microswitches to mimic the handheld's feel. Users must provide their own switches and keycaps. Pricing is set at £210.83, approximately $282, with shipments expected in the fourth quarter of 2026.
Google Wallet now shows transactions made with Wear OS smartwatches in the same history as phone purchases. The update, reported by 9to5Google, combines NFC payments from the watch with existing transaction data on Android devices. Users with compatible Wear OS watches and Android phones running recent operating system versions will see a unified transaction list. The change does not affect the ability to make payments; it simply adds watch activity to the existing view. Google has not disclosed a specific rollout date, but the feature appears to be live for most users.
Jamesob's GitHub repository outlines a step‑by‑step guide for deploying state‑of‑the‑art large language models on local hardware, including recommended system components and Docker‑based deployment configurations.
The article discusses cost thresholds, suggesting configurations around $2,000 for smaller models and $40,000 for systems capable of running models approaching the size of Claude Opus.
The guide specifies a base system built around a last‑generation EPYC processor and DDR4 memory sourced from second‑hand markets, paired with four NVIDIA RTX PRO 6000 workstation GPUs providing 384 GB of VRAM. An independent PCIe Gen4 switch enables peer‑to‑peer communication between GPUs, reducing latency compared with routing traffic through the motherboard’s PCIe root complex. Custom BIOS settings—including PCIe link width, forced Gen4 operation, disabling ASPM and SR‑IOV, and enabling Resizable BAR—are described as necessary for stable operation. Kernel parameters such as iommu=off and NCCL configuration prevent NCCL hangs in multi‑GPU environments, while ACS is disabled via a runtime script to keep peer‑to‑peer traffic within the switch fabric. Power management limits each GPU to 350 W, allowing the system to operate on a standard 110 V circuit. Performance metrics report line rates of 27.5 GB/s upstream and 50.4 GB/s bidirectional with sub‑microsecond latency. Docker‑compose files in the runners directory provide ready‑to‑run serving configurations for models such as GLM‑5.2‑594B and Whisper‑large‑v3 for speech‑to‑text, with reported throughput of approximately 80 tokens per second for the 594 B model. Additional tools include a GPU speed and latency benchmark script. The author notes that model weights are stored locally on a ZFS‑replicated filesystem and accessed via read‑only mounts in individual Docker containers.
The guide acknowledges alternative approaches, such as using a DGX Spark cluster with greater VRAM, and states that the presented configuration reflects personal experience rather than a definitive recommendation.
Summer of Ludd Festival Teaches Offline Living in New York City. The week‑long event took place in Tompkins Square Park and ran from July 1 to July 5. Organizers, operating under the name Gowanus, described the gathering as a loose coalition seeking to reduce reliance on digital platforms. The program included a play about the historical Luddite movement, workshops on mending and analog communication, a shortwave radio and walkie‑talkie session, and a discussion on locating events without social media. A 2024 Pew Research study found that 48 percent of teen respondents viewed social media negatively, up from 32 percent in 2022. Attendees comprised students, Pride participants, families and former technology employees. While participants expressed hope that the experience could encourage reduced technology use, Andrew Maynard, a professor at Arizona State University, said the movement may have limited effect on overall digital habits.
Smithsonian’s “Starstruck” VR exhibit opened in Washington, DC, in May, offering a 40‑minute virtual tour of the universe. Adult tickets cost $29 to $35, and group tickets for four or more are $18 each, reflecting a 15 percent discount; the exhibit will open in Denver, Orlando, Florida, and San Antonio, Texas, later this year.
Visitors wear an HTC Vive Focus 3 headset, complete onboarding that includes caption preferences and a waiver, and are guided by a virtual narrator, James Seawood, whose avatar resembles a SpaceX spacesuit. The exhibit initially uses a 2021‑vintage headset but will transition to the newer Vive Focus Vision in the DC location.
The tour begins at the Multiple Mirror Telescope at the Smithsonian Astrophysical Observatory’s Whipple Observatory, then proceeds to visualizations of the Big Bang, a stellar nursery such as the Pillars of Creation, and close approaches to the Sun, the Parker Solar Probe, and NASA’s Hubble, Chandra, and James Webb Space Telescopes, each with an interactive picture button and a small model for manipulation.
One segment transports visitors to the exoplanet Janssen (55 Cancri Ae), which completes an orbit in about 17 hours; the visual depicts a rocky surface with lava flows, while scientific analyses indicate the planet may be entirely molten.
The experience also shows the red supergiant Betelgeuse becoming lumpy as it fuses heavier elements, followed by a simulated supernova, and it places visitors near the event horizon of Sagittarius A*, the Milky Way’s central black hole, demonstrating light bending in its gravity.
The tour concludes with a depiction of the Atacama Desert in Chile, the future site of the Giant Magellan Telescope, which the exhibit renders as a completed facility with its seven mirrors aimed at the sky.
The Australian Securities Exchange (ASX) has been fined A$20.5 million by the Federal Court after an investigation found it misstated the status of a blockchain‑based platform project. In 2017 ASX announced plans to replace its CHESS trading platform, built on COBOL and OpenVMS, with a distributed‑ledger system. The regulator, the Australian Securities and Investments Commission (ASIC), sued in 2024, alleging the ASX’s claim that the project was “progressing well” was misleading. The Federal Court said ASX admitted its errors and ordered the fine, plus A$3 million in costs. A parliamentary report identified three causes: unclear objectives, overlapping planning and implementation phases, and insufficient identification of scalability risks. The settlement marks a setback for ASX’s distributed‑ledger ambitions and underscores regulatory scrutiny of public company statements.
TechCrunch released an AI glossary that defines common artificial intelligence terms for readers, investors, and developers. The guide, updated regularly, provides clear explanations of acronyms and concepts that appear in product meetings, investment discussions, and public commentary. AGI is described by OpenAI CEO Sam Altman as a system comparable to a median human employee, while OpenAI’s charter defines it as highly autonomous systems that outperform humans in most economically valuable tasks; DeepMind characterizes AGI as AI matching human capability across most cognitive tasks. An AI agent is a tool that uses AI to perform multiple tasks on a user’s behalf, such as filing expenses or writing code, though its precise meaning varies. API endpoints function as interfaces that allow software to invoke other services, enabling integrations and autonomous actions by AI agents. Chain‑of‑thought reasoning involves breaking a problem into intermediate steps to improve answer accuracy, a technique used in reasoning‑optimized large language models. A coding agent is a specialized AI agent that can write, test, and debug code with minimal human oversight. Compute refers to the hardware resources, such as GPUs and TPUs, that enable AI model training and inference. Deep learning is a subset of machine learning that employs multi‑layered neural networks to identify patterns in data. Diffusion models, used in generative AI for images, audio, and text, learn to reverse a noise‑adding process. Distillation extracts knowledge from a larger model to train a smaller, more efficient model. Fine‑tuning involves further training a model on specialized data to improve performance for a specific task. GANs are neural network frameworks that pit a generator against a discriminator to produce realistic data, commonly applied to deepfake creation. Hallucination describes instances where models generate incorrect information, raising concerns for applications such as medical advice. Inference is the execution of a trained model to produce predictions, requiring hardware ranging from smartphones to cloud GPUs. Large language models are neural networks with billions of parameters that generate text based on prompts, forming the basis of many AI assistants. The Model Context Protocol provides a standardized way for AI models to connect to external tools and data sources. Mixture of Experts architecture activates only a subset of specialized sub‑networks for each task, improving efficiency in large models. Neural networks are layered algorithms that underpin deep learning and modern generative AI. Open source AI models, such as Meta’s Llama series, are publicly available for inspection and modification, contrasting with closed‑source offerings like OpenAI’s GPT. Parallelization enables simultaneous processing across multiple chips, which is important for training and inference at scale. A shortage of RAM chips has driven up prices across consumer electronics, gaming, and enterprise computing. Recursive self‑improvement refers to AI systems that modify their own architecture without human input, a capability explored by several startups. Reinforcement learning trains models through trial and reward signals, and is used to enhance reasoning in large language models via methods such as RLHF. Tokens are discrete units of text processed by language models, and token throughput measures the speed at which a system can handle these units. The glossary continues to be updated as the field evolves, offering a reference for ongoing developments in artificial intelligence.
Noctua released an updated product roadmap that outlines launch windows for a low‑profile AM5 CPU cooler and Seasonic power supplies. The cooler, based on the NH‑L12 design, is 70 mm tall and uses six 6 mm copper heat pipes, dense aluminum fins and a 120 mm NF‑A12x25 G2 fan. It preserves 35 mm of RAM height clearance for Mini‑ITX compatibility. The company plans a Q2 2027 release for the cooler. The NT‑CP1 carbon nanotube thermal pad for AM4 and AM5 will launch in Q3 this year. The Pulsar Feinmann F01 Noctua Edition mouse is also scheduled for Q3 this year. A Noctua Edition of Thermal Grizzly’s WireView Pro II monitoring tool is set for Q3 2026. A 140 mm desk fan and a USB fan controller are slated for Q1 2027. Seasonic’s Prime PX Noctua Edition and Prime TX Noctua Edition power supplies, as well as next‑generation workstation coolers, are scheduled for Q2 2027. KitGuru asked which upcoming Noctua product fans are most excited about.
Steam Machine units have begun shipping to early buyers. An early adopter posted on Reddit that the device functioned normally at first but stopped working after a Steam update installation. The system displayed a red error light, which the user described as indicating GPU failure. The incident involved a single unit among thousands of shipped machines. A red LED, rather than the usual blue indicator, suggests the device requires return for repair or replacement. Valve has faced component shortages that delayed the launch multiple times. Limited availability of replacement parts could extend service times for failed units. KitGuru said hardware failures can occur with any new device and hoped the case remains isolated.
Former European Parliament member Stelios Kouloglou was found to have his iPhone infected with NSO Group’s Pegasus spyware on October 21, 2022, and again on March 6‑7, 2023, according to forensic analysis by the Citizen Lab. The infections occurred while he served as a substitute member of the PEGA Committee investigating Pegasus abuses. Kouloglou was a substitute member of the PEGA Committee from March 24, 2022 to July 18, 2023. The first infection preceded a series of committee hearings and the drafting of its initial report. The second infection coincided with intense deliberations before the committee adopted its first report in May 2023. The analysis showed multiple Apple threat notifications sent in March 2023, August 2023, and April 2024, indicating repeated targeting. Kouloglou said he did not recall receiving the notifications. The first infection aligned with committee discussions about spyware and its impact on fundamental rights, including hearings on “Big Tech and Spyware” and “Spyware and e‑privacy” held in late October 2022. It also preceded research trips to Greece and Cyprus. On the same day of infection, Kouloglou was hospitalized and visited by investigative journalist Thanasis Koukakis, who had previously been targeted with Predator spyware. The possibility that confidential medical information could have been captured raises legal concerns under Greek data protection law. The second infection occurred while Kouloglou was traveling from Athens to Brussels during continued committee deliberations. The timing overlapped with a mission by the Pegasus rapporteur and a LIBE Committee inquiry into the Greek spyware scandal. The report notes that other MEPs have been targeted with Pegasus or related spyware, including several Catalan MEPs and others, indicating a broader pattern within the European Parliament. Attribution remains unconfirmed; the Citizen Lab said no evidence links the Greek government or a Greek customer to the attacks, and the operator may have been active in multiple European countries, possibly linked to a campaign targeting Russian and Belarusian exiled journalists. Recommendations call for immediate forensic screening of MEPs and staff, annual reporting on surveillance threats, and improved response to company threat notifications.
Skyroot Aerospace aims to launch the first test flight of its Vikram‑1 rocket between July 12 and August 4 from the Satish Dhawan Space Center, India’s first private attempt to place a launch vehicle into orbit. SpaceX announced that it has produced its 1,000th Merlin 1D engine, a reusable engine that powers the Falcon 9 and Falcon Heavy rockets.
Skyroot, which has raised about $160 million and reached a valuation above $1 billion, will use a three‑stage solid‑propellant rocket with a liquid‑fueled upper stage to place up to 500 kilograms into low‑Earth orbit. The test flight will gather data on propulsion, stage separation, guidance, navigation and control, and overall vehicle performance. The launch pad was originally built for India’s government space program.
SpaceX has also launched a SiriusXM communications satellite to geostationary orbit, its third such mission this year, after completing 78 Falcon 9 and Falcon Heavy launches this year.
NASA selected Rocket Lab to launch three science missions on Electron rockets in 2027, including the Polarized Submillimeter Ice‑cloud Radiometer and the Total and Spectral Solar Irradiance Sensor‑2. The missions will depart from Rocket Lab’s New Zealand spaceport on dedicated Electron flights.
Isar Aerospace, a German startup, has signed a deal with Planet Labs to launch at least one of its Pelican Earth‑observation satellites on the Spectrum rocket, which has yet to reach orbit. The planned launch from Andøya Spaceport in Norway could occur as early as later this year. The final Atlas V launch for Amazon’s Leo broadband constellation took place on Thursday, marking the ninth flight for the rocket and concluding a series of missions that have used the vehicle for nearly 25 years. ULA plans to return to flight with its Vulcan rocket after a pause caused by booster issues.
Blue Origin is cleaning up its launch pad after an explosion on May 28 that destroyed the site’s lightning tower and transporter‑erector. The company says a return to flight is expected before the end of the year, using a crane to lift the New Glenn rocket rather than rebuilding the transporter. A schedule of upcoming launches includes a Long March 6A mission on July 4, a Falcon 9 Starlink mission on July 5, and a Falcon 9 Transporter mission on July 7.
The space industry continues to see a mix of first‑time private launches and milestones in reusable engine production, while launch schedules remain subject to weather and technical delays.
Cyberpunk 2077 has surpassed 40 million copies sold, according to a post from the official Cyberpunk account. The milestone reflects a resurgence driven by the upcoming second season of Edgerunners and recent updates. Launched in 2020, the game initially sold more than 13 million copies within the first two weeks but saw sales decline after launch issues, prompting CD Projekt Red to pause public updates for a period. By late 2022 the total reached 20 million, and the recent figure doubles that number. The company attributes the increase to next‑gen console versions, multiple free patches, downloadable content, and ongoing discount promotions. A small content update is expected with the release of Edgerunners Season 2, similar to the 1.6 patch released after the first season. CD Projekt Red indicated that major updates are currently on hold. Sales are expected to continue rising as interest in Edgerunners Season 2 grows, though no further major releases have been announced.
Increased surveillance is being deployed for the World Cup matches and the Fourth of July celebrations in Washington, D.C. The Department of Homeland Security has designated both the Fourth of July fireworks on the National Mall and the World Cup final in New Jersey as National Special Security Events. The designation, which also applies to the Super Bowl and a White House UFC event, signals heightened security measures. Law enforcement will use airport‑style checkpoints at the fireworks, prohibit folding chairs and coolers, and employ counter‑drone systems, bomb technicians, and countersnipers. New York City has allocated $6.5 million for counter‑drone technology, and Kansas City police have seized 16 drones since the tournament began. FEMA provided $250 million in grants to host cities, much of which funded counter‑drone equipment, and the FBI is training local officers on drone mitigation. Several cities plan to install facial‑recognition cameras on public buses, with images checked against missing‑person alerts and retained only when a match occurs. Thousands of officers, including National Guard troops and FBI agents wearing body cameras, will be present at the events. Privacy advocates note concerns about the collection and retention of biometric data and the potential for information shared between local and federal agencies to be used for domestic law enforcement. The long‑term storage and use of the gathered surveillance data are currently unspecified.
A Hacker News user investigated why memory usage reported by fastfetch, btop, and htop differed on a FreeBSD system after migrating from an Ubuntu server.
FreeBSD uses a virtual memory system that divides physical memory into 4 KiB pages and organizes them into several queues, such as active, inactive, laundry, wired, and free.
Monitoring tools compute used memory with different heuristics; fastfetch adds inactive and cache to free memory, while btop adds active and wired memory.
In btop, the used‑memory calculation relies on a 32‑bit unsigned integer for page counts; values above 4 GiB wrap, causing an underestimate of used memory.
The cache count (v_cache_count) reported by FreeBSD is a dummy variable that always returns zero, so tools that depend on it show an empty cache.
To correct the issue, the author modified btop to use 64‑bit variables, subtracted the ZFS ARC size from wired memory, and added the ARC amount to the cache category, and submitted a pull request that has been merged.
A similar fix was applied to fastfetch, which now detects ARC cache on FreeBSD and other ZFS‑supporting systems.
The author tested the changes in virtual machines and reported that the revised memory figures more accurately reflect system state.
The investigation deepened the author’s understanding of FreeBSD internals and resulted in patches to multiple open‑source projects.
The author plans to continue researching memory management in DragonFly BSD.
Anthropic announced Claude Science, an AI workbench that aggregates scientific tools and creates figures and visuals, during its “The Briefing: AI for Science” event this week. The company also said it intends to develop its own drugs, focusing on neglected diseases, according to Head of Life Sciences Eric Kauderer‑Abrams. The platform is designed for scientists to access fragmented datasets and software in a single environment, and several biotech and pharmaceutical firms have already reported using it. Kauderer‑Abrams did not specify what the company would do with promising drug candidates, and The Verge has not received a response to requests for additional details on target diseases or partnership plans. AI is applied at every stage of drug discovery, from compound identification to clinical trial support, experts said. Namshik Han, a professor at the University of Cambridge and cofounder of CardiaTec, described the term “AI drug discovery” as broad, noting that AI can suggest new molecules, identify disease targets, and repurpose existing drugs. Matthew Todd, a professor of drug discovery at University College London, added that AI already assists in speeding up research and testing ideas. The drug development process remains dependent on laboratory experiments, clinical trials, and regulatory review. Frank von Delft, a professor at the University of Oxford, said that AI has not yet rendered experiments unnecessary and that developing a drug will require substantial investment in testing, especially human trials. He said Anthropic would need to spend heavily on experiments if it pursues drug development. Anthropic has increased its life‑science staff in the past year, hiring biologists and building wet labs, and is recruiting for several life‑science positions. Han said the company has recruited some candidates from Big Pharma and academic institutions. Industry analysts estimate that, even with AI assistance, an AI‑designed drug would likely require five to ten years to complete clinical trials and obtain regulatory approval. No AI‑generated drug has yet reached market, and candidates that have entered trials provide limited evidence of AI’s contribution. The announcement raises questions about the scope of Anthropic’s drug development program and its potential impact on the broader AI‑in‑pharma landscape.
Researchers from Canonical's dqlite team used the TLA+ specification language to model a Write-Ahead Log checkpointing bug in SQLite that has existed since 2010. The bug involves a race condition between WAL appends and checkpoints that can corrupt the database. Although the real-world impact is minimal, the issue persisted for 16 years and proved difficult to reproduce.
In the TLA+ model, the authors defined variables for the WAL, database, and checkpoint state, and created actions for appending pages and initiating checkpoints. By adding a write lock that must be held for both append and checkpoint operations, the model showed that simultaneous execution cannot occur. Model checking found no execution trace that violates the invariant that all WAL pages are applied in order, confirming dqlite is not affected.
SQLite addressed the bug in version 3.45.0 released on March 5, 2026, by adding a check that verifies the WAL salt has not changed since the checkpoint began. The model with this check also passes verification.
The research demonstrates that formal modeling can verify concurrency correctness in storage systems.
In the 1660s and 1670s Amsterdam introduced a new fire‑fighting system designed by brothers Jan and Nicolaas van der Heyden, which reduced fire damage and became the city’s standard method.
Amsterdam’s rapid trade and manufacturing in the 17th century created many fire hazards, including stores of pitch, tar, and other combustible materials, as well as workshops and ships that increased the risk of large urban fires.
Initially, the city purchased large water‑pumping engines from Hans Hautsch of Nuremberg, each equipped with a manually operated cistern, metal pipe, and nozzle, and distributed about 60 of these engines by 1670 along with ladders, hooks, tarpaulins, and more than 28,000 leather buckets. Members of four guilds — inland sailors, peat carriers, beer carriers, and grain weighers — were assigned to fight fires in their districts.
A fire at a sugar refinery on the Laurier Canal in December 1669 demonstrated the limits of the existing equipment; despite a ready water supply, the flames could not be reached inside the building, leading to total loss of the refinery and the owner’s house, valued at an estimated 195,000 guilders.
Another major fire in February 1672 at the printing office of cartographer Joan Blaeu destroyed two well‑equipped shops, including the space where the maps of the Atlas maior were printed, causing losses of 382,000 guilders.
In 1669, Jan van der Heyden proposed a system of 1,800 oil lanterns for street lighting, which the city installed and for which he was appointed director of the municipal lights, receiving a salary of 2,000 guilders per year. His work also included designs for improved fire engines.
The brothers’ new engine incorporated a leather suction hose that could draw water from a distance, a flexible leather hose — called a “snake” — that could reach deeper into buildings, and an air chamber that provided a continuous high‑pressure flow, addressing freezing issues in cold weather. In November 1672 the city council decided to retrofit half of the existing engines with these suction and pressure hoses.
A fire on January 12, 1672, at harbor buildings storing military supplies tested the new equipment. Although the new engine arrived late, it was able to save adjacent sheds and clear a path for hoses, demonstrating its effectiveness compared with older engines.
Following the January 1672 incident, the guild fire companies petitioned the city to replace the awkward retrofitted engines with lighter, fully new units. In 1681 the mayors ordered the Van der Heyden brothers to produce enough engines for each of the city’s sixty districts and to draft regulations for their use.
The new regulations introduced a system of fire alarms using trumpeters in city towers, monetary rewards for the first three fire companies to arrive, and fines for individuals who attempted to fight a fire before sounding an alarm. Fire fighters were not salaried; each district fire company consisted of 36 residents serving under two fire masters, who were required to serve a year or pay a fine.
From 1682 to 1687, Amsterdam recorded 41 fires with total losses of 18,353 guilders, a fraction of the losses recorded in the earlier decade under the previous system, indicating a marked improvement in fire control.
The reforms illustrated how civic institutions could adopt technical innovations and organizational changes to reduce urban risk, reflecting the city’s broader culture of paying for useful public improvements.
Philips has announced the M4 gaming monitor lineup, consisting of the 27M4N3500PT and 27M4N5500PT, which feature a triple‑mode capability allowing instant switching among three resolution and refresh‑rate configurations.
The displays use a Fast IPS panel and support three native settings: QHD at 275 Hz, FullHD at 360 Hz, and HD at 540 Hz. Both models include Low Blue Light and Flicker‑Free eye‑protection technologies, a 25 % haze anti‑glare coating, and a peak brightness of 350 nits. Connectivity options comprise HDMI 2.1 and DisplayPort 1.4.
Both models provide integrated AI functions such as Stark Shadow Boost, Smart Crosshair, Smart Sniper, and Smart MBR Sync. The premium 27M4N5500PT model adds a SmartErgoBase stand that offers height, tilt, swivel adjustments and a 90‑degree vertical pivot, along with a geometric cut base and an AI‑driven lighting system.
Philips noted that the ability to toggle between high‑fidelity and high‑speed settings without changing operating system configurations can benefit gamers who alternate between demanding titles and fast‑paced shooters; the company has not yet disclosed pricing or release dates.
Thermaltake announced a $50 voltage tester named Dr. Power III Pro, designed for use with modern ATX 3.1 power supplies. The device connects to the PSU’s SENSE0 and SENSE1 signals and includes multiple connectors to measure voltage output, allowing users to verify that the supply delivers power within expected parameters.
ATX 3.1, introduced by the PCIe Special Interest Group, adds new power connectors and requires precise voltage regulation. Testing PSU output has become more critical as components draw higher power.
The tester, released on July 3, 2026, provides digital readouts and can be used to diagnose potential issues before they affect other components. Thermaltake described the tool as a way to ensure PSU reliability without disassembling the system. Analysts noted that while the product targets enthusiasts and system builders, it may also be useful for warranty verification and troubleshooting.
The availability of a dedicated tester reflects growing attention to power delivery quality as hardware power demands increase.
Ubisoft announced pre‑load and launch times for Assassin’s Creed Black Flag Resynced, scheduled for release on July 9. Pre‑loads begin July 7 for PC and PlayStation 5, while Xbox Series X|S users can download the game immediately. Console launch occurs at midnight local time worldwide; PC access follows a coordinated global schedule. The remake revisits the 2013 title, offering updated visuals and performance improvements. In the United States, PC unlocks at 7 a.m. Pacific Daylight Time on the West Coast and 10 a.m. Eastern Daylight Time on the East Coast. In the United Kingdom, PC access begins at 3 p.m. British Summer Time, followed by 4 p.m. Central European Summer Time. Japan’s PC version unlocks at 11 p.m. Japan Standard Time, and Australia and New Zealand receive access at 12 a.m. Australian Eastern Standard Time, the same time as the console launch. The pre‑load schedule applies across Ubisoft Connect, Steam and the Epic Games Store. Ubisoft indicated that Black Flag remains a well‑received entry in the Assassin’s Creed series, and the remake is likely to attract a broad audience.
Apple Home does not provide a built‑in way to transfer full ownership of a smart‑home setup to a new resident. When the author moved, the buyer wanted to keep the existing accessories, so the parties used a temporary workaround: the buyer was added as a resident in the Apple Home app and the broadband service was left active for a short period. The resident role permits managing accessories, scenes, and other users, but cannot add HomePods, Apple TVs, or AirPlay devices, and it leaves the former owner with full control, raising security and privacy concerns. The author proposes that Apple introduce a third user class labeled Owner, which would issue a warning and then downgrade the previous owner to resident, allowing temporary access before the new owner removes the old account. The problem is likely to become more common as smart‑home technology spreads, and the lack of a clear transfer mechanism may hinder home sales. Comments are welcome.
NSW technology minister Anoulack Chanthivong’s office welcomed OpenAI’s first Australian office in Sydney after internal emails revealed staff jokes about a possible Skynet scenario. The phrase “absolutely thrilled” was removed from a press release after staff suggested more restrained language. A deputy chief of staff replied that a “golden era” comment might be used later, while another employee warned that a Skynet situation could occur within five years. The emails, tabled in NSW parliament, show the government courted OpenAI before its decision to open the office.
OpenAI announced its intention to open an Australian office in August 2025 and began operations in Sydney in December. The minister’s talking points highlighted Sydney as the southern hemisphere’s top startup location, attracting 65% of Australian venture capital and noting the presence of companies such as Atlassian, Canva and Afterpay. They also said the state hosts 45% of the nation’s AI businesses and that major US tech firms including Google, Microsoft, Meta, AWS and IBM have Australian headquarters in Sydney.
OpenAI has not responded to requests for comment. A government spokesperson said the administration is committed to embracing AI opportunities while ensuring emerging technologies are used responsibly and in the public interest. The state is developing a datacentre strategy to support the growing AI sector, though a release date has not been set.
Documents from a May 2026 meeting indicate that if all eight large Sydney datacentres operated their diesel generators simultaneously for one hour, air pollution would reach five to six times the level of all electricity generation or motor vehicle emissions in the state. The chief executive of Data Centres Australia described the model as a worst‑case scenario that is extremely rare. NSW Greens MLC Abigail Boyd, chair of the datacentre inquiry, argued that current air impact assessments underestimate operating hours, creating grid strain and increasing the likelihood of blackouts, especially near residential areas and schools.
The inquiry continues as the state balances AI investment with community concerns over datacentre environmental effects.
Midjourney, an AI image generation company, released a behind‑the‑scenes video of a prototype medical scanner that combines multiple ultrasound probes with a hot‑tub‑like platform and off‑the‑shelf computers.
The video, presented by engineer‑YouTuber Marcin Plaza, shows the hardware assembly and the team’s efforts, but does not address the technical challenges raised by experts. Those experts told The Verge that the scanner’s ability to produce clinically useful images is uncertain, given the well‑known limits of ultrasound technology.
Midjourney says the scanner will be marketed as a wellness product for body composition rather than as a diagnostic device, which would require FDA clearance and clinical trials. The company’s head of medical, Tom Calloway, said the focus on body composition allows the project to move quickly once testing is complete, and he promised frequent blog updates. CEO David Holz added that the absence of external investors gives the company freedom to pursue the scanner.
The video leaves unanswered questions about the scanner’s performance, its regulatory pathway, and whether it can meet the standards required for medical use.
Wing Commander IV, released in February 1996 after a delay from its original Christmas 1995 target, came on six CD‑ROM discs and included roughly four gigabytes of data dedicated to full‑motion video, according to reports. Origin Systems, under the direction of Chris Roberts, produced the game with a reported budget of $12 million, the majority allocated to filming rather than game development. Cast members included Mark Hamill, Tom Wilson, Malcolm McDowell and John Rhys‑Davies, who performed in live‑action scenes shot on 35 mm film with physical sets rather than digital backdrops. The 652‑page script generated multiple narrative branches, a length that exceeded typical film scripts of the era. The narrative follows Christopher Blair, portrayed by Hamill, who retires after the defeat of the Kilrathi in Wing Commander III before being recalled to duty to address a surge of piracy on the border worlds, according to the game's storyline. A pivotal decision point occurs midway through the campaign, and the story concludes with a dialogue‑driven confrontation between Blair and Admiral Tolwyn, played by McDowell, in the Terran Confederation's Grand Assembly chamber. Gameplay mechanics remained largely unchanged from Wing Commander III, featuring space combat against human pilots in human‑crafted vessels; minor lighting improvements were the only substantive changes to the engine, according to development notes. Contemporary reviews, including a PC Gamer assessment, described the experience as "a little hollow," noting the significant emphasis on cinematic content relative to interactive gameplay. Technical requirements included a Pentium MMX or faster processor for smooth FMV playback; systems with 486 processors were unable to handle the video data, according to system specifications of the time. A later DVD re‑release provided higher‑quality video but arrived after the original launch, according to release records. By 2026 the title is available for approximately four dollars, positioning it as a low‑cost experience that evokes nostalgia from its era.
Tesla reported second‑quarter vehicle deliveries that topped Wall Street forecasts, delivering a record for the period as demand in Europe rose while North America remained weak.
The company's automotive segment had posted two consecutive annual declines in sales, prompting concerns about its growth trajectory. The latest figures suggest a rebound in the European market.
Analysts said the rebound could support Tesla's investments in autonomous driving and artificial intelligence, which are key components of its valuation. The recent cooling of backlash toward Elon Musk was noted as the company continued to expand its product lineup.
Tesla's delivery numbers exceeded analyst expectations, indicating a potential stabilization after two years of decline. The company plans to maintain its spending on autonomous driving and AI, while the durability of the sales recovery remains to be seen.
The Vergecast podcast released its July 3, 2026 episode with suggestions for reducing screen time during the U.S. holiday weekend. As many Americans begin a three‑day weekend, digital wellbeing is a recurring topic in technology journalism. In the episode, Verge editors asked each other how they disconnect. Recommendations included disabling Wi‑Fi, enabling do‑not‑disturb modes, setting app limits on smartphones, using smartwatch activity reminders, and scheduling regular screen‑free periods. The hosts said the goal was not to eliminate screen use but to manage it responsibly. Listeners were invited to submit their own tips for a future episode. The segment ends with an invitation for audience contributions and a summer vacation.
The FIDE Ethics and Disciplinary Commission announced a two-year suspension for former world champion Vladimir Kramnik after finding he violated the organization’s ethics code through public statements targeting other players. The commission, an independent body within FIDE, reviewed complaints filed by the FIDE Management Board and the Fair Play Commission concerning Kramnik’s repeated social‑media posts and public remarks about grandmaster David Navara and the late Daniel Naroditsky. After examining evidence, the commission concluded that Kramnik’s conduct breached provisions on dignity, respect, bullying, psychological abuse, and his duty as a role model, as well as his failure to cooperate with the Fair Play Commission’s investigation. The panel dismissed allegations of cheating, dishonesty, and causing reputational harm to FIDE, stating that these claims lacked sufficient evidence. Kramnik received a two‑year worldwide ban from FIDE competitions and official duties, with the final twelve months suspended provided he completes a three‑year probation period without further violations. He also must perform twelve months of unpaid service for the chess community. The decision may be appealed to the FIDE Ethics and Disciplinary Commission Appeal Chamber within 21 days, and the full ruling is available on FIDE’s website.
Lee, a former systems administrator at a retail company's headquarters, recalled an incident from around the year 2000. He managed Novell servers, email servers, and supported roughly 1,000 desktop users. On a Friday afternoon, a vice president reported that he could not open files because a dialog box displayed 'General failure reading Drive C:'. Lee asked whether the message read 'General failure is reading Drive C:'; the vice president corrected that the exact wording was 'General failure reading Drive C:'. Lee concluded that the error indicated a failed hard disk, not an external intrusion using the pseudonym 'General Failure'. He arranged a support call, recommended a new disk and possibly a new computer for the vice president. The episode illustrates how error messages can be misinterpreted.