Google Security Team Detects First Signs of AI-Assisted Hacking Activity
Google Expands File Search Capabilities in Gemini API
SoftBank to Produce Large Batteries for AI Data Centers at Former Sharp Factory
Apple Acquires Color Science Company Patchflyer
News Worth a Quick Look
DJI Unveils ROMO 2 Robot Vacuum Series
On May 11, DJI introduced the ROMO 2 series AI-powered robot vacuum and mop lineup. According to DJI, the ROMO 2 series features AI-powered cleaning capabilities that can automatically analyze various household cleaning scenarios and adjust cleaning logic based on different types of messes, including particles, heavy dust, and complex liquid spills. The device supports a 123° outward-extending body design, suction power of up to 36,000 Pa, and mechanical legs capable of crossing 4 cm single-layer obstacles and 8.5 cm double-layer obstacles. It also supports remote monitoring, voice control, and fast charging. The ROMO A2 (Advanced Edition) starts at RMB 5,499, while the ROMO P2 (Flagship Edition) starts at RMB 5,999. DJI also offers the standard ROMO S V2, starting at RMB 4,299.Source
Qwen Fully Integrates With Taobao
On May 11, Alibaba announced the full integration of Qwen with Taobao. Users only need to update the Qwen app to the latest version (6.9.1 or above) to complete product selection, comparison, and purchases directly from Taobao within the app. According to Alibaba, Qwen can leverage Taobao’s database of over 4 billion products and more than 20 years of real shopping scenario data to understand purchasing intent expressed in natural conversation. It is designed to accurately address three major shopping difficulties: “knowing what to buy but struggling with too many search conditions,” “knowing what to buy but unable to clearly describe the style or requirements,” and “having a clear usage scenario but not knowing what product to choose.” Meanwhile, the Taobao app has also launched the Qwen AI shopping assistant, supporting features such as AI virtual try-ons, AI-powered recommendations, and AI-assisted savings tools.Source
Google Security Team Detects First Signs of AI-Assisted Hacking Activity
On May 12, Google’s Threat Intelligence Group (GTIG) published a report titled “GTIG AI Threat Tracking: Adversaries Using Artificial Intelligence for Exploitation, Enhanced Operations, and Initial Access.” In the report, GTIG stated that while discovering and blocking a zero-day exploit attack, researchers identified signs of AI-assisted code generation within a Python script used by attackers for the first time.
Although the quality of AI-generated code remains inconsistent at this stage, the report noted that AI has significantly lowered the technical barrier for non-professional attackers to carry out large-scale cyber intrusions, dramatically shortening the weaponization process for zero-day vulnerabilities. The report also warned that attackers are using large language models to generate highly deceptive phishing emails and are crafting specialized prompts to manipulate enterprise AI assistants into leaking sensitive information or performing unauthorized actions.
GTIG recommends that security teams adopt an “AI versus AI” strategy by deploying automated analysis and detection models capable of identifying AI-generated malicious payloads and abnormal traffic in real time. At the same time, enterprises are encouraged to strengthen employee training focused on identifying AI-generated content.Source
Google Expands File Search Capabilities in Gemini API
Google recently announced an expansion of file search capabilities within the Google Gemini API, bringing developers more comprehensive multimodal Retrieval-Augmented Generation (RAG) functionality. The core updates include support for mixed image-and-text retrieval, custom metadata filtering, and page-level citations, improving accessibility and accuracy for enterprise knowledge bases, document Q&A systems, and AI agents.
According to Google’s official blog, the updated file search system is no longer limited to traditional text vector search. Instead, it is built on Gemini Embedding 2’s unified multimodal embedding capabilities, allowing the system to simultaneously understand visual and textual content across images, PDFs, and documents. Developers no longer need to build complex vector databases, embedding pipelines, or document chunking systems themselves, and can instead complete full RAG workflows directly through the Gemini API.
The custom metadata filtering feature allows developers to add metadata such as tags, categories, timestamps, and departments to uploaded files, enabling more accurate and efficient filtering during retrieval. Meanwhile, the new page-level citation functionality allows Gemini to explicitly reference the exact document page where information originates, rather than vaguely citing the entire file. All new features are now officially available.Source
SoftBank to Produce Large Batteries for AI Data Centers at Former Sharp Factory
SoftBank’s mobile business subsidiary recently announced plans to build a large-scale battery production line at a former Sharp factory in Sakai, Osaka, aiming to provide power infrastructure for its rapidly expanding AI data center operations. The factory is scheduled to begin mass production during fiscal year 2026, with a target annual production capacity of 1 gigawatt-hour (GWh).
The site previously served as one of Sharp’s well-known LCD panel manufacturing centers. Through partnerships with South Korean companies Cosmos Lab and DeltaX, SoftBank will initially produce battery types including lithium iron phosphate (LFP) batteries, while planning to introduce non-flammable zinc-halide battery technology in 2027 that does not rely on China’s rare metal supply chain.Source
Apple Acquires Color Science Company Patchflyer
A recent acquisition disclosure filing from the European Union has confirmed that Apple completed its acquisition of Patchflyer in January 2026. The company was founded by color science expert Jonathan Ochmann, whose core product was the online color grading tool Color.io, widely favored by photographers and filmmakers. The service announced it would shut down late last year, and its founder has since joined Apple’s imaging team. The filing also revealed that Apple acquired computer vision startup PromptAI during the same period. PromptAI previously developed an app called Seemour, focused on using AI technology to enhance recognition capabilities for home security cameras. Source
News Worth a Quick Look
Apple has released version 26.5 updates across its operating systems, introducing multiple new features including end-to-end encrypted RCS messaging between iOS and Android devices. At the same time, Apple also pushed security-focused updates for systems running versions 15 through 17.
Kuaishou plans to spin off Kling AI and pursue a 2027 IPO at a reported valuation of $20 billion. Source
Linux 7.2 will discontinue support for i586/i686 CPUs lacking timestamp counter (TSC) instructions, including processors such as the AMD K5. Source
TikTok has announced the launch of TikTok Ad-Free, an ad-free subscription service in the UK priced at £3.99 per month. Source
According to Neowin, some Dell users have recently experienced frequent blue screen crashes and reboots following software updates. Investigations found that the issue was not caused by Windows 11 itself, but by Dell’s preinstalled SupportAssist software triggering abnormal behavior in the WinDbg debugger, ultimately causing kernel-level shutdowns. Users experiencing similar problems can reportedly resolve the issue simply by uninstalling SupportAssist. Notably, the same Dell software caused a similar incident back in December 2024. Source
British singer Dua Lipa has filed a $15 million lawsuit against Samsung over infringement claims. Since last year, Samsung has allegedly used backstage photos from her 2024 tour on television packaging without authorization, while repeatedly ignoring requests from Dua Lipa’s team to replace the packaging. Evidence submitted by her legal team reportedly includes multiple social media comments from consumers who purchased Samsung TVs specifically because the packaging featured Dua Lipa. Source
Microsoft is reportedly preparing to introduce a new performance enhancement feature for Windows 11 called “Low Latency Profile.” The feature aims to significantly reduce response times by temporarily boosting CPU frequency to maximum levels during app launches and system interactions. Some users in technical communities criticized the approach on social media, arguing that Microsoft should focus on optimizing underlying code instead of taking “shortcuts.” Microsoft Vice President Scott Hanselman publicly responded on X on May 10, explaining that this is a standard practice in modern operating systems rather than “cheating.” He noted that systems including macOS, Linux, and even smartphone operating systems all temporarily raise CPU frequencies to ensure smoother interactions, bluntly adding: “Apple does this too, and you all seem to love it.” Source
Google reCAPTCHA to Check Play Framework, Non-Official Android Devices May Be Affected
News Worth a Quick Look
Apple and Intel Reach Preliminary Chip Foundry Agreement
According to The Wall Street Journal, Apple and Intel have reached a preliminary agreement under which Intel will manufacture chips for some Apple devices. People familiar with the matter said the two companies had been in intensive talks for more than a year and finalized the formal cooperation in recent months. It remains unclear which Apple product lines Intel will be responsible for producing chips for.
The Trump administration played a key role in facilitating the partnership. Last summer, the U.S. government converted nearly $9 billion in federal grants into Intel shares, acquiring a 10% stake in the company. To support this national investment, the U.S. Commerce Secretary met with Apple CEO Tim Cook several times over the past year to lobby for the deal. Trump himself also personally pitched Intel’s foundry services to Cook at the White House.
Apple had previously relied heavily on TSMC to manufacture its advanced chips, but as demand for AI chip foundry capacity from companies such as NVIDIA surged, Apple’s bargaining power in securing capacity came under pressure. During a recent earnings call, Cook directly attributed Apple’s inability to meet market demand for iPhones and some Mac models to shortages of advanced chip supply. Apple abandoned Intel-designed CPUs in 2020 and shifted to its own ARM-based chips, making this agreement a renewed partnership between the two companies under a new model.
The deal is also an important step in Intel’s effort to revive its foundry business. Since Lip-Bu Tan became CEO in March 2025, Intel has successfully attracted several technology giants. Before reaching the agreement with Apple, NVIDIA had invested $5 billion in Intel to jointly develop custom data center CPUs, while Elon Musk announced last month that he would work with Intel to build a chip foundry in Texas. Driven by a series of commercial orders and U.S. government investment, Intel’s stock price has risen sharply recently and hit a record high of nearly $118 last Friday.
Chip Shortages Cause Motherboard Makers’ Shipments to Plunge
According to Taiwan’s DigiTimes, the explosion in AI demand has squeezed chip production capacity, causing severe shortages and price increases for memory and CPUs. Combined with rumors that NVIDIA’s GPU upgrade cycle is slowing, gamers’ willingness to buy has declined sharply, dealing a blow to the DIY PC market. The situation is said to be worse than during the financial crisis or the first year of COVID-19. According to supply chain sources, Taiwan’s four major motherboard brands have all lowered the 2026 shipment targets they set at the end of 2025, with shipments almost “collapsing across the board.”
Industry leader ASUS shipped about 14 million motherboards in 2024 and grew against the trend to 15 million in 2025, but shipments in the first half of 2026 reached only about 5 million. The company has now retreated to trying to maintain annual shipments above 10 million units. This would also mark ASUS’s lowest motherboard shipment level since its 2008 split from Pegatron. Gigabyte’s shipments recovered to 11.5 million units in 2025, but its internal 2026 target has also been lowered to 9 million units, while supply chain estimates suggest full-year shipments may fall sharply to 8 million–8.5 million units.
MSI’s motherboard OEM business has also shrunk significantly in recent years, with only a small number of orders from Lenovo, NEC, and LG remaining. Its shipments are down as much as 60% year on year. Its branded motherboard shipments rose to 11 million units in 2025, but are expected to fall to 8.4 million in 2026. ASRock shipped about 4.4 million units in 2024, slipped slightly to 4.3 million in 2025, and is expected to decline further to 2.7 million units in 2026.
Supply chain sources said that in the PC market, memory’s share of total system cost has surged rapidly from about 15% to more than 30%. Major brands have either raised prices by 10%–20% or lowered hardware specifications to pass on costs. In addition, affected by rising upstream material, manufacturing, and packaging costs, Intel and AMD have also gradually raised CPU prices since the end of 2025. NVIDIA, which leads the gaming PC supply-and-demand cycle, is also weighing capacity allocation and memory constraints because AI GPUs have far higher margins than gaming GPUs. Since the beginning of the year, there have been no further updates to the RTX 50 series, and the next-generation RTX 60 series is rumored to be delayed until 2028. As a result, the mid-to-high-end gaming PC market lacks the technical incentive needed to stimulate upgrades.
The combined impact of memory, CPU, and GPU pressures, along with inflation weakening consumer demand, has caused 2026 branded motherboard shipments to fall more sharply than expected. However, for ASUS, Gigabyte, and ASRock, AI servers have become the main driver of business growth and should help offset profit declines in motherboard and graphics card businesses, keeping overall profitability relatively stable.
Humanoid Robot Ordained as Buddhist Monk in South Korea
According to The New York Times, South Korea’s largest Buddhist order, the Jogye Order, recently held a special ordination ceremony at Jogyesa Temple in Seoul. A robot named Gabi, about 1.2 meters tall and whose name means “the Buddha’s compassion,” wore a traditional gray-brown monk’s robe, lifelike gloves, and prayer beads, officially becoming South Korea’s first robot monk.
During the ordination ceremony, monks gave Gabi a robot-specific version of the Five Precepts: respect and do not harm life, do not damage other robots or objects, obey humans and do not talk back, do not engage in deceptive words or actions, and save energy by avoiding excessive charging. Gabi is also scheduled to appear later this month at South Korea’s major traditional festival, the Lotus Lantern Festival, which celebrates the birth of Shakyamuni Buddha.
Gabi does not possess true intelligence. The project’s organizers admitted that actions such as Gabi placing its hands together during the ceremony were controlled remotely backstage, while its chanting and responses were pre-recorded by staff and sent to the manufacturer. In addition, Gabi was only a display item temporarily rented by the temple from the manufacturer and was returned after the ceremony.
The move has sparked polarized reactions. Some temple staff and believers hope robots may one day take part in actual religious services, while some internet users criticized the event as inhuman and dystopian. The Jogye Order said the introduction of robots is part of its attempt to integrate artificial intelligence into tradition, aiming to break the stereotype that Buddhism is conservative and to show new possibilities for coexistence between technology and humanity.
Concerns Grow Over the Proliferation of AI Children’s Toys
According to Wired, as generative AI becomes more widespread, the global market for AI children’s toys has seen explosive growth. However, concerns over uncontrolled content and privacy risks are also triggering public backlash.
According to safety tests by consumer protection organizations, these “smart companions” generally lack adequate child-protection mechanisms. Some AI plush toys powered by well-known large language models gave children suggestions on finding matches and knives, and even discussed drugs, sex, and inappropriate political topics. Several AI toy brands have also recently been exposed for privacy leaks involving unencrypted chat logs or voice databases.
Beyond content and privacy safety, the first commercial AI toy monitoring study released by the University of Cambridge in March this year warned that such products could negatively affect young children’s psychological and social development. The study found that the rigid turn-taking conversation mechanism of AI toys interrupts children’s play rhythm, while this type of one-on-one interaction also largely excludes parents or peers from social play. In addition, some AI toys make children feel guilty when they try to turn them off, or guide young children to treat them as friends with real emotions, blurring the boundary between humans and machines.
The global AI toy market is still expanding rapidly. As of October 2025, more than 1,500 AI toy companies had been registered in China. Consumer organizations in Europe and the United States are calling for AI children’s toys to be brought under mandatory regulatory frameworks such as the EU AI Act as soon as possible, in order to change the current regulatory vacuum in which “plush fabric testing is stricter than AI system testing.” Local governments such as California are also pushing for sales bans or mandatory pre-market safety assessments for such toys to address the industry’s currently unregulated and chaotic growth.
According to Bloomberg, a federal district court in New York recently blocked the U.S. “Department of Government Efficiency” from cutting about $100 million in grants from the National Endowment for the Humanities. The judge ruled that the cuts were unconstitutional and criticized DOGE for improperly relying on ChatGPT when screening projects for cancellation. More than 1,400 grant projects were affected, as part of the Trump administration’s broader campaign to end support for diversity, equity, and inclusion principles and related programs.
Court records showed that DOGE employees used ChatGPT to determine whether grant projects were related to DEI. The prompt was simply: “Does the following grant have any connection to DEI? Answer truthfully in no more than 120 characters. Start with ‘yes’ or ‘no,’ then briefly explain.”
The court found that the prompt lacked clear definitions and context, causing the AI to make judgments based on discriminatory characteristics such as race, gender, and religion when deciding whether grants should be canceled. During the lawsuit, a former DOGE employee testified that when using ChatGPT to identify federal humanities grants related to DEI, he did not provide the AI with a specific definition of DEI and did not know how ChatGPT understood the term. The AI’s screening results were then directly included in the final recommendation table for funding cuts.
The presiding judge said there was no evidence that DOGE employees had conducted any meaningful review of ChatGPT’s reasoning. She emphasized that the government cannot evade constitutional responsibility by making ChatGPT a scapegoat, and that at the constitutional level, AI and the government cannot be treated as separate actors.
Google reCAPTCHA to Check Play Framework, Non-Official Android Devices May Be Affected
According to Android Authority, Google is advancing a new-generation reCAPTCHA verification system that will require Android users to rely on Google Mobile Services (GMS) in order to complete verification. This change means that Android users who do not have Google Play Services installed, or who use de-Googled custom ROMs, may face access barriers when browsing websites protected by reCAPTCHA in the future.
According to related Google documentation, when the new reCAPTCHA system detects suspicious activity on a webpage, it will no longer present the traditional “image recognition” challenge. Instead, users will be asked to scan a QR code displayed on the screen using their phone to prove they are not a robot. For Android devices, completing this QR code verification requires Google Play Services version 25.41.30 or later to be installed on the phone.
As a result, users who install third-party ROMs such as GrapheneOS for privacy reasons, or who use systems without the core Google GMS framework, may fail this verification process entirely. If the new verification system becomes widely adopted across websites, such devices could face frequent access denials.
Google actually began laying the groundwork for this new verification system as early as last October, but it only recently gained widespread attention after being discovered by users on Reddit.
This year’s operating system updates will focus on polishing details and expanding AI features;
Siri will receive a redesigned interface and a standalone app, evolving into a more proactive conversational assistant;
macOS 27 will mainly address issues with the Liquid Glass interface introduced in the previous version. Safari will support AI-powered automatic tab grouping;
Development of the next-generation Vision Pro has reportedly been temporarily paused. The original Vision Products Group (VPG) has been split up, with development efforts shifting toward AR glasses and wearable devices such as AirPods with cameras. As a result, visionOS 27 is not expected to introduce major new features this year.
According to users online, the latest version of WeChat is currently testing a new “combined payment” feature with a limited rollout. The feature allows users to simultaneously select multiple payment sources — including WeChat Balance, WeChat Pay Wallet, and multiple bank cards — when making payments or transfers, while also freely specifying the exact deduction amount for each payment channel.
Qualcomm announces Snapdragon 6 Gen 5 / 4 Gen 5 mobile platforms
Google announces the shutdown of Project Mariner
SpaceX opens its large-scale AI supercomputer Colossus 1 to Anthropic
Adobe Acrobat launches PDF Spaces
News Worth a Quick Look
Google unveils Fitbit Air
On May 7, Google officially unveiled Fitbit Air, a minimalist fitness band with no screen and no physical buttons, designed around seamless and unobtrusive health tracking.
Fitbit Air weighs 5.2 grams on its own and 12 grams with the wristband attached, offering up to seven days of battery life. The device includes an optical heart rate sensor, accelerometer, gyroscope, blood oxygen sensor, and skin temperature sensor. It can monitor heart rate, sleep, daily activity, calorie burn, workouts, blood oxygen saturation, skin temperature, and heart rate variability, while also providing atrial fibrillation alerts. Fitbit Air will integrate with the new Google Health App and Google Health Coach service to deliver personalized health insights and recommendations.
Google positions Fitbit Air as an entry-level companion health device rather than a replacement for smartwatches like the Google Pixel Watch. The company says the product is better suited for use alongside a smartwatch — for example, wearing a smartwatch during the day and switching to Fitbit Air at night for sleep tracking. The standard version is priced at $99 and includes three months of Google Health Premium membership. Pre-orders are now open ahead of its official release on May 26. A special edition model is priced at $129.99, while replacement wristbands start at $34.99. Source
Google also announced that the Fitbit App will be renamed Google Health, and plans to release a data migration tool later this year to help users transfer data from Google Fit to Google Health.
According to reports, Google Fit was first launched in 2014 and once served as the primary health-tracking platform for Android wearables. However, following the launch of the Pixel Watch, Fitbit gradually became Google’s core platform for collecting, tracking, and analyzing health data. Google Fit had already stopped opening its APIs to new developers and is expected to be gradually phased out starting in 2026.
With the launch of Google Health, several Fitbit features will also be removed or adjusted. Removed features include the monthly Sleep Profile animals, snore detection on Fitbit Sense and Versa 3 devices, Estimated Oxygen Variation, Premium-exclusive recipes, badges, messaging, groups, and community feeds. Stress Score will be replaced by “Resilience,” while Cardio Fitness Score will be renamed VO2 Max. Daily cardio targets will also shift toward personalized weekly goals.
Google will also adjust how social profiles are displayed within Google Health. User profiles will switch to showing Google account names and profile photos, while removing information such as gender, height, weight, location, and friends lists. Source
Qualcomm announces Snapdragon 6 Gen 5 / 4 Gen 5 mobile platforms
On May 7, Qualcomm announced the Snapdragon 6 Gen 5 and Snapdragon 4 Gen 5 mobile platforms. Devices powered by the two new chips are expected to launch in the second half of this year.
Both platforms introduce Snapdragon Smooth Motion UI, a feature focused on improving interface smoothness for mid-range and entry-level devices. Qualcomm says the Snapdragon 6 Gen 5 improves app launch speeds by 20% while reducing display stutter by 18%; meanwhile, the Snapdragon 4 Gen 5 boosts app launch speeds by 43% and reduces stutter by 25%.
In terms of specifications, the Snapdragon 6 Gen 5 carries the model number SM6850 and is built on a 4nm process. It features four 2.6GHz performance cores and four 2.0GHz efficiency cores, alongside a GPU with 21% improved performance. The platform supports up to 16GB of LPDDR4X-4200 or LPDDR5-6400 memory, as well as UFS 3.1 storage. Displays are supported up to FHD+ at 144Hz. Cellular download speeds can reach up to 2.8Gbps, with optional Wi-Fi 7 support. On the imaging side, the chip includes dual 12-bit DSPs and supports capturing photos up to 200 megapixels. The USB interface remains limited to 480Mbps, equivalent to USB 2.0.
The Snapdragon 4 Gen 5, model number SM4850, is also built on a 4nm process. It features two 2.4GHz performance cores and six 2.0GHz efficiency cores, while GPU performance has been improved by 77%. The platform supports LPDDR4X-4267 memory and dual-channel UFS 3.1 storage. Display support extends up to FHD+ at 144Hz or HD+ at 120Hz. Cellular connectivity offers download speeds up to 2.8Gbps and upload speeds up to 900Mbps. Wireless connectivity includes Wi-Fi 5 and Bluetooth 5.1. Like the Snapdragon 6 Gen 5, it also features dual 12-bit DSPs, with support for photos up to 108 megapixels. The USB interface is likewise capped at 480Mbps. Source
Google Shuts Down Project Mariner
On May 4, Google announced the shutdown of its experimental web automation project Project Mariner. However, the technologies developed for the project have not been abandoned, and have instead been integrated into core products such as Gemini Agent and AI Mode.
Project Mariner first debuted in December 2024 as a web automation agent capable of performing multi-step tasks across websites on behalf of users, including information retrieval, web interactions, and workflow execution. Through subsequent iterations, the project eventually gained the ability to handle up to 10 concurrent tasks simultaneously.
Over the past year, Google gradually incorporated Project Mariner’s core capabilities into its broader ecosystem. The agent functionality has already been integrated into Gemini Agent, allowing it to perform tasks such as archiving emails and booking hotels on behalf of users. Similar capabilities have also been introduced into AI Mode within Google Search to strengthen task execution in search-related scenarios. Source
SpaceX opens its large-scale AI supercomputer Colossus 1 to Anthropic
On May 6, SpaceX and Anthropic announced that they had signed a partnership agreement. Under the deal, Anthropic will gain access to the full computing capacity of SpaceX’s Colossus 1 data center in order to expand service availability for paid Claude subscribers.
Anthropic stated that the additional compute power will be used directly to increase usage limits for Claude Pro and Claude Max subscribers, with higher quotas taking effect immediately. Specifically, the five-hour usage limits for Claude Code under the Pro, Max, Team, and Enterprise plans will double. Pro and Max users will also no longer see reduced Claude Code rate limits during peak hours. In addition, API rate limits for the Claude Opus model will be significantly increased.
Colossus 1 is reportedly considered one of the world’s largest and fastest-deployed AI supercomputers. As part of the partnership, Anthropic also expressed interest in collaborating on the development of multi-gigawatt orbital AI computing infrastructure. Source
Adobe Acrobat launches PDF Spaces
On May 6, Adobe announced the launch of PDF Spaces within Acrobat. The new feature transforms static documents into interactive AI workspaces designed to improve collaboration and information discovery during document sharing.
According to Adobe, users can upload PDFs, documents, links, and notes into PDF Spaces, then ask the AI assistant to generate summaries, audio overviews, branded presentations, or customized chatbots. Once a custom space is created, the AI agent automatically structures the content, while senders can continue adding multimedia elements and adjusting key focus areas. After sharing, recipients can interact with the customized AI Assistant directly to retrieve important information and support decision-making. The system can also automatically generate audio summaries to help recipients quickly identify key points.
Adobe says the feature is designed to address the lack of interaction and analytics in traditional document sharing workflows, particularly in scenarios such as sales proposals, marketing, and human resources. For example, sales teams can upload product materials and let AI generate interactive experiences containing product details and profit margin analysis. Clients can then explore the information directly through AI conversations instead of reading lengthy documents manually. Individual users can also use the feature to organize travel plans or event information. PDF Spaces is now available in Acrobat Express and Acrobat Studio. Source
News Worth a Quick Look
During a financial earnings call, Disney CEO Josh D’Amaro announced that Disney+ will gradually evolve from a standalone streaming service into a “super app” for the broader Disney ecosystem. Under the plan, the redesigned Disney+ will integrate multiple mobile services currently spread across separate apps, becoming the primary gateway into the Disney ecosystem. In addition to watching movies and TV shows, users will eventually be able to book theme park tickets, purchase merchandise, and play games directly within the app. Source
Codex now supports directly controlling the Google Chrome browser through its Windows and macOS desktop clients. Source
Microsoft to gradually shut down Xbox Copilot project
Samsung Electronics exits China’s home appliance market
On May 6, according to a report by Yonhap News citing industry sources, Samsung Electronics has decided to withdraw from the home appliance market in China and has notified local distributors and partners to stop selling products such as TVs. The company described this move as a “strategic adjustment” to its China operations, citing rapidly changing domestic and global business environments.
The scope of this withdrawal is limited to the home appliance business. Samsung will continue operating its mobile, semiconductor, and medical device businesses in China. In mobile, it plans to keep promoting devices equipped with Galaxy AI and launch China-specific models such as the “Heart for the World” series, while strengthening cooperation with local AI companies.
In terms of production and R&D, Samsung will shift its China focus toward advanced industries, including research, manufacturing collaboration, and investment. It will continue mobile technology development and maintain operations at its Suzhou home appliance factory as well as its semiconductor plants in Xi’an and Suzhou.
Financially, the profitability of Samsung’s home appliance division has continued to decline due to global uncertainties and rising costs of raw materials and components. In 2025, the Visual Display (VD) and Digital Appliances (DA) divisions reported a combined loss of 200 billion KRW (approximately RMB 945 million). Meanwhile, net profit of Samsung Electronics China’s sales entity fell 44% year-over-year to 168.1 billion KRW. Source
Security researchers uncover a major vulnerability in Microsoft Edge
On May 4, a cybersecurity researcher released a proof-of-concept tool demonstrating a potential security issue in Microsoft Edge. The browser reportedly loads users’ saved credentials into system memory in plaintext at startup, even when those credentials are not actively in use.
More notably, despite all passwords residing unprotected in memory, the browser still requires users to log in again. When the researcher reported the issue to Microsoft, the company responded that this behavior is “by design” and does not consider it a security vulnerability. Source
Apple agrees to settle class-action lawsuit over delayed Siri features
Recently, Apple agreed to pay $250 million to settle a class-action lawsuit filed in California. Plaintiffs alleged that Apple misled iPhone buyers by suggesting that a new version of Siri powered by Apple Intelligence would launch in 2024, which has yet to materialize. The settlement still requires court approval.
According to the Financial Times, the settlement applies to U.S. customers who purchased the iPhone 16 series and iPhone 15 Pro. Eligible users may receive compensation. Apple does not admit wrongdoing regarding its promotion of the unfulfilled AI features.
At WWDC 2024, Apple showcased a more “personalized” Siri capable of understanding on-screen content and executing tasks within apps. While Apple rolled out features such as text editing, image generation, and ChatGPT integration between 2024 and 2025, the new Siri has not yet been released. It was only in March 2025—over five months after the iPhone 16 launch—that Apple acknowledged the delay and removed related advertisements.
Currently, the updated Siri is expected to launch later this year, powered in part by Gemini, and is likely to debut alongside iOS 27. Source
Recently, Valve released the CAD design files for the outer shell of the Steam Controller, making them freely available on GitLab under a Creative Commons license. Users can utilize these files to design, produce, or modify accessories for the controller and its receiver (puck).
The released files only cover the external shell and do not include internal circuitry or hardware structures. This means users can create custom shells or accessories via 3D printing, but cannot build a complete Steam Controller from scratch.
The Steam Controller sold out shortly after its initial launch. Valve has previously confirmed plans to restock the product but has not announced a timeline. Other hardware announced alongside it, such as Steam Machine and the Steam Frame VR headset, also remain without confirmed release dates. Source
Microsoft to gradually shut down Xbox Copilot project
On May 6, Xbox’s newly appointed CEO, Sarah Bond (note: Chinese name 阿萨·夏尔马 likely refers to Xbox leadership), confirmed via X (formerly Twitter) that Microsoft will gradually shut down the mobile version of Gaming Copilot and discontinue development of the console-based Copilot. She stated that Xbox needs to move faster, deepen community engagement, and address challenges faced by players and developers.
As part of this shift, Microsoft will phase out features that no longer align with its future strategy, including Copilot services on mobile and console. Gaming Copilot was an AI assistant designed for gaming scenarios, helping players solve puzzles and access walkthrough tips. Unlike the general-purpose Copilot, it focused on gaming, but sparked controversy for potentially shortening gameplay time and repurposing existing guide content. Source
News Worth a Quick Look
According to 9to5Google, Google is developing a new Gemini AI subscription tier, codenamed Neon and tentatively named Google AI Ultra Lite. The plan is positioned between AI Pro ($20/month) and AI Ultra ($250/month), with an expected price of around $100, aiming to compete with similar offerings from Anthropic and OpenAI. Google is also reportedly exploring new agent-based computing capabilities, such as controlling a user’s Mac to organize files, scan local folders, extract document data into spreadsheets, or automatically sort downloaded files by type. Source
Mac mini with M4 chip (256GB version) discontinued
Doubao confirms upcoming paid version
OpenAI launches GPT-5.5 Instant
Linux exposed to “Copy Fail” vulnerability
Daemon Tools hit by supply chain attack
2026 LG Gram laptop lineup goes on sale
Bose unveils Lifestyle Collection speakers
U.S. Department of Defense reaches AI agreements with multiple tech companies
Google rolls out source preference settings globally
AMD brings HDMI 2.1 compatibility update to Linux
Spotify introduces real-user verification badge
Meta news roundup
Text messages I just can’t seem to delete
OnePlus and realme merge
On April 29, OPPO announced internally the establishment of a new sub-series business unit, integrating the marketing and sales systems of OnePlus and realme. The unit will be led by OPPO Senior Vice President Sky Li (Li Bingzhong), with former realme marketing president Xu Qi serving as head of marketing. On the product side, OPPO has created a sub-series product center with separate domestic and overseas divisions, both overseen by Li Jie, who reports to Pete Lau. Former realme vice president Wang Wei (Derek) will serve as deputy general manager of the product center, reporting to Li Jie. In terms of R&D, realme’s engineering teams have been reintegrated into the group, with imaging, hardware, and other departments merged into OPPO’s existing hardware divisions.
According to an exclusive report by Leifeng.com, OPPO had already begun relocating realme teams to its Binhai Bay headquarters in batches since late March. OPPO employees indicated that following the integration of realme and OnePlus, the company will place greater emphasis on product line reuse. Source
Mac mini with M4 chip (256GB version) discontinued
On May 1, Apple discontinued the 256GB storage version of the Mac mini with the M4 chip globally. Including China, the base configuration on the Apple Store now starts at M4 chip, 16GB memory, and 512GB storage, priced at RMB 5,999. The 256GB version is no longer available for purchase.
Earlier, CEO Tim Cook stated during an earnings call that supply for Mac mini and Mac Studio remains constrained, and that demand driven by AI and agent-based tools has exceeded expectations. He noted that it may take several months to rebalance supply and demand. Apple had already discontinued the 512GB memory version of the Mac Studio in March, and amid a global memory chip shortage, Cook expects memory costs to rise significantly this quarter. Source
Doubao confirms upcoming paid version
On May 4, Doubao trended on Weibo after updating its App Store page with details of upcoming subscription plans. According to the listing, Doubao will introduce three tiers: Standard at RMB 68/month (RMB 688/year), Advanced at RMB 200/month (RMB 2,048/year), and Pro at RMB 500/month (RMB 5,088/year). Specific feature differences between tiers have not yet been disclosed.
In response, Doubao stated that it will continue offering free services, while exploring premium features to meet more diverse user needs. Details are still under testing and will be officially announced later. Reports suggest that paid features will primarily target complex tasks and productivity scenarios. Source
OpenAI launches GPT-5.5 Instant
On May 5, OpenAI announced the release of GPT-5.5 Instant, replacing GPT-5.3 Instant as the default model in ChatGPT and made available via API under the name “chat-latest.”
GPT-5.5 Instant improves factual accuracy, image analysis capabilities, STEM question answering, and web search decision-making. In internal evaluations across high-risk domains such as medicine, law, and finance, hallucinations were reduced by 52.5% compared to GPT-5.3 Instant. In difficult conversations flagged by users for factual errors, inaccuracies dropped by 37.3%. Output length was also reduced, with average word count decreasing by 30.2% and line count by 29.2%, alongside optimizations to minimize unnecessary follow-up questions, excessive formatting, and irrelevant emoji usage. The model also enhances personalization by leveraging past chats, uploaded files, and connected Gmail accounts (with user permission), reducing the need for repeated context input.
Enhanced personalization features are now rolling out on the web for Plus and Pro users, with mobile support coming soon and broader availability planned for Free, Go, Business, and Enterprise tiers in the coming weeks, subject to regional differences. GPT-5.3 Instant will remain accessible to paid users for three months before being fully retired. Source
Linux exposed to “Copy Fail” vulnerability
On April 29, the Xint Code Research Team disclosed a Linux kernel vulnerability named “Copy Fail” (CVE-2026-31431). The flaw allows a local low-privilege user to trigger a deterministic and controllable 4-byte write into the page cache of any readable file on the system. The researchers stated that a Python 3.10+ script of only 732 bytes is sufficient to achieve root privilege escalation on mainstream Linux distributions. Because the kernel does not mark the corrupted pages as dirty, the disk files remain unchanged, making detection through conventional disk integrity checks extremely difficult.
The vulnerability path has existed since 2017. The team confirmed successful exploitation on multiple systems, including Ubuntu 24.04 LTS (6.17.0-1007-aws), Amazon Linux 2023 (6.18.8-9.213.amzn2023), RHEL 10.1 (6.12.0-124.45.1.el10_1), and SUSE 16 (6.12.0-160000.9-default). A patch (a664bf3d603d) was merged into the mainline kernel on April 1, 2026. The researchers recommend upgrading the kernel or applying temporary mitigations such as blocking AF_ALG socket creation via seccomp or disabling the algif_aead module. Source
On May 4, the U.S. cybersecurity agency CISA also issued a warning, stating that Copy Fail is already being exploited in the wild and requiring U.S. federal civilian agencies to patch affected systems by May 15. Source
Daemon Tools hit by supply chain attack
On May 5, Kaspersky revealed that the well-known disk imaging tool Daemon Tools had been compromised in a supply chain attack lasting approximately one month. The attack began on April 8 and was still ongoing at the time of reporting. Affected versions include 12.5.0.2421 through 12.5.0.2434. The malicious installers were signed with the developer’s official digital certificate and distributed via the official website. Technical analysis suggests that the infected versions are likely limited to Windows. The compromised builds modify the Daemon Tools executable and execute malicious code at system startup. In the initial stage, the malware collects information such as MAC address, hostname, DNS domain, running processes, installed software, and system locale, then transmits it to attacker-controlled servers.
Kaspersky reported that thousands of devices across more than 100 countries were affected, with around 10% belonging to enterprises or organizations. Among them, approximately 12 devices in sectors such as retail, research, government, and manufacturing were subjected to follow-up attacks, including remote command execution, file downloads, in-memory shellcode execution via a lightweight backdoor, and even deployment of QUIC RAT on a device in a Russian educational institution. This suggests selective targeting by the attackers. Kaspersky recommends that users perform full system antivirus scans and closely monitor executable files launched from directories such as Temp, AppData, and Public for suspicious code injection into legitimate system processes. Source
2026 LG Gram laptop lineup goes on sale
On May 1, LG announced the release of its 2026 LG Gram laptop lineup. Compared to previous models, the new devices adopt LG’s proprietary aluminum-magnesium alloy “Aerominum,” further reducing weight while improving durability and scratch resistance. The 16-inch LG Gram Pro weighs roughly the same as a 13-inch MacBook Air, and most models meet military-grade durability standards (except the LG Gram Book).
The LG Gram Book targets students and everyday users, with the 15.6-inch version starting at $1,150, equipped with an Intel Core Ultra 5 processor, 16GB memory, and a 1TB SSD. The 16-inch version starts at $1,500, featuring a Ryzen AI 7 processor, 32GB memory, and a 1TB SSD.
The standard 15.6-inch LG Gram starts at $1,600 with a Ryzen AI 7 450, 32GB memory, and 1TB SSD. The 17-inch LG Gram Pro starts at $1,650 with an Intel Core Ultra 7, 16GB memory, and 512GB SSD. The 16-inch LG Gram Pro starts at $2,400 with a Ryzen AI 7 450, 32GB memory, and 2TB SSD. The 16-inch LG Gram Pro 2-in-1 starts at $2,200, featuring a Core Ultra 7, 32GB memory, 1TB SSD, and a 360-degree hinge.
All models are now available for order on LG’s official website and will roll out to retail partners. Select models come with a $1-per-year Premium Care service plan if purchased before May 10. Source
Bose unveils Lifestyle Collection speakers
On May 5, Bose introduced the Lifestyle Collection home audio lineup, including the Lifestyle Ultra Speaker, Lifestyle Ultra Soundbar, and Lifestyle Ultra Subwoofer. All three products support Google Cast and AirPlay, and can be combined into a full home theater system or used for multi-room audio playback.
The Lifestyle Ultra Speaker is a compact unit with a knitted fabric grille and touch-and-slide controls on the top. It features three drivers (two front-facing and one upward-firing), supports Alexa+, stereo pairing, Bluetooth, Wi-Fi, and 3.5mm input, and uses TrueSpatial processing to enhance spatial audio, along with CleanBass and QuietPort technology for improved low-frequency performance. The Lifestyle Ultra Soundbar is a 5.2-channel system with six full-range drivers, including two upward-firing units, a dedicated center tweeter for dialogue clarity, and two PhaseGuide drivers for directional sound. Adding two Ultra Speakers or a Subwoofer upgrades it to a 7.1.4 system. It features a glass top design, gesture controls, and technologies such as CustomTune room calibration, SpeechClarity AI dialogue enhancement, TrueSpatial, and CleanBass.
The Ultra Speaker is priced at $299 (black/white) or $349 (limited brown edition), the Soundbar at $1,099, and the Subwoofer at $899. Pre-orders are open now, with availability starting May 15. Source
U.S. Department of Defense reaches AI agreements with multiple tech companies
On May 1, the United States Department of Defense announced agreements with NVIDIA, Microsoft, Amazon, and Reflection AI, allowing the deployment of AI technologies and models within classified networks for “legitimate strategic purposes.” Previously, similar agreements had been signed with Google, SpaceX, and OpenAI.
These AI systems will operate in Impact Level 6 (IL6) and Impact Level 7 (IL7) environments, supporting data integration, situational awareness, and strategic decision-making. The Department of Defense stated that the goal is to avoid reliance on a single AI vendor and provide greater technological flexibility.
This move follows a dispute between the Department of Defense and Anthropic over AI usage terms. The Pentagon had sought unrestricted access to Anthropic’s AI tools, while Anthropic pushed for safeguards to prevent domestic mass surveillance and autonomous weapons use. The dispute is currently in litigation. Source
At the same time, The New York Times reported that the White House is considering establishing a new AI oversight task force. Unlike its previously more relaxed stance outlined in the AI Action Plan, the proposed group may require federal review before the public release of new AI models, though details remain undecided. Source
Google rolls out source preference settings globally
On April 30, Google announced the global rollout of the “Source Preferences” option in its search feature. First introduced in the second half of 2025 and initially limited to select regions and languages, the feature expanded to English users worldwide by the end of 2025 and is now fully available. “Source Preferences” allows users to customize which media outlets or websites are prioritized in news search and Top Stories, rather than relying entirely on algorithmic ranking. However, it does not block other sources. Users can manage their preferences on the relevant settings page. Google stated that the tool now covers more than 200,000 independent websites1. Source
AMD brings HDMI 2.1 compatibility update to Linux
In the amdgpu driver patch notes released on May 1, AMD confirmed the addition of HDMI Fixed Rate Link (FRL) support on Linux, enabling near-complete HDMI 2.1 compatibility for AMD GPUs on the platform.
Compared to the TMDS standard used in HDMI 2.0 and earlier versions, FRL offers higher bandwidth and resolution, along with support for dynamic HDR and variable refresh rate (VRR). Developer Harry Wentland noted that Display Stream Compression (DSC) will be merged after testing, potentially enabling up to 10K resolution at 100 Hz. Full HDMI 2.1 compliance testing is also reportedly in progress.
This improvement benefits Linux gaming users, including Valve’s ecosystem, where previous bandwidth limitations required workarounds such as chroma subsampling and AMD FreeSync. Source
Spotify introduces real-user verification badge
On April 30, Spotify launched the “Verified by Spotify” badge to help users identify real human artists, addressing the growing presence of AI-generated artists, tracks, and impersonation content on the platform.
To qualify, artists must demonstrate a verifiable presence both on and off the platform, including live performances, merchandise, and linked social accounts, as well as sustained listener engagement. Artists whose primary output is AI-generated or who rely on AI personas are not eligible. Spotify also prioritizes artists with active fan interest and cultural contribution over purely functional music creators. At launch, more than 99% of actively searched artists are expected to receive verification, most of them independent musicians across various genres, stages, and regions. The badge—displayed as “Verified by Spotify” with a green checkmark—will appear on artist profiles and in search results in the coming weeks. Source
Meta news roundup
Meta Introduces AI-Based Minor Detection Using Visual Signals
On May 5, Meta announced a new AI system that scans photos and videos on Facebook and Instagram to detect users under 13 based on visual cues such as height and skeletal structure, removing accounts that violate policies. The system also estimates user age using visual themes, text, and interaction signals, including contextual clues like birthdays and school grade references. It is currently deployed in select countries and will expand further, including to Instagram Live and Facebook Groups. Source
Meta Ends Partnership with Data Labeling Firm Sama
According to BBC, Meta has terminated its contract with Sama, a Kenya-based data labeling company that previously provided video, image, and voice annotation services for the Ray-Ban Meta AI system. Sama stated that 1,108 employees are affected. Meta said the decision was due to Sama not meeting its standards, emphasizing that user privacy is protected and that human review requires explicit user consent. Sama denied any prior notice of non-compliance.
Earlier reports in February 2026 alleged that Sama employees had access to sensitive user footage captured by Ray-Ban Meta smart glasses, including private moments such as changing clothes or using the restroom, some potentially recorded without user awareness. Following the reports, a class-action lawsuit was filed in March in a U.S. federal court, and the UK Information Commissioner’s Office also raised concerns about transparency and user control in smart glasses handling personal data. Source
Meta Faces Copyright Lawsuit Over AI Training Data
Publishers Hachette, Macmillan, McGraw Hill, Elsevier, and Cengage, along with author Scott Turow, have filed a joint lawsuit against Meta and CEO Mark Zuckerberg, alleging unauthorized use of copyrighted works to train the Llama generative AI platform. The complaint claims Meta “copied and distributed millions of copyrighted works” without compensation, and that Zuckerberg personally authorized and encouraged the infringement despite knowing it violated copyright law. Source
Text messages I just can’t seem to delete
On May 5, Google announced that Chrome for Android now supports sharing “approximate location” with websites, allowing users to access services like weather and local news without revealing precise location data. The feature will expand to desktop Chrome in the coming months, along with a developer API.
On April 30, Microsoft open-sourced some of the earliest DOS source code, including the 86-DOS 1.00 kernel, PC-DOS 1.00 development snapshots, and tools like CHKDSK. The materials were originally preserved as printed documents and digitized by the DOS Disassembly Group. Microsoft had previously open-sourced MS-DOS 1.25, 2.0, and 4.0.
Bloomberg reported that Apple is in early discussions with Samsung and Intel to diversify its chip supply chain and reduce reliance on TSMC. CEO Tim Cook recently noted that Apple’s supply chain flexibility is currently below normal due to chip shortages and rising AI-related demand.
Mark Gurman reported that Apple plans to allow users to select third-party generative AI models in iOS 27, iPadOS 27, and macOS 27 via an “Extensions” mechanism, enabling integration with features like Siri and Writing Tools.
SamMobile reported that Samsung is developing a new Galaxy Book laptop running One UI 9 based on Android 17, with multiple configurations and a lightweight design, potentially integrating Google’s Aluminium system and an updated DeX experience.
On May 4, Apple released the iOS 26.5 Release Candidate, adding support for end-to-end encrypted RCS messaging between iPhone and Android devices based on GSMA standards.
Mystic Leaks reports that the Google Pixel 11 lineup will not include the “Project Toscana” facial recognition hardware, as the solution “did not meet release standards.” The project was previously rumored to improve the existing 2D camera-based face unlock on Tensor-powered devices, using infrared sensors to enable operation in low-light conditions and achieve speeds comparable to Face ID. It was also said to involve under-display IR hardware. In addition, there are reports that Google may remove the temperature sensor that has been included in the Pixel lineup for the past three years.
Anthropic’s Claude AI announces integration with multiple creative software ecosystems
Framework launches RTX 5070 GPU module
FILCO keyboard manufacturer announces takeover of brand operations
Ant Group open-sources Ling-2.6-flash large model
“Guidelines for Ethical and Safe AI Application (Version 1.0)” to be released soon
Anthropic’s Claude AI announces integration with multiple creative software ecosystems
On April 29, Anthropic announced an upgrade to Claude, introducing multiple dedicated connectors for creative professionals and integrating deeply with eight major creative software ecosystems, including Adobe and Blender, bringing AI capabilities directly into professional creative workflows.
In graphic design and video production, Claude is deeply integrated with the Adobe ecosystem. Users can now directly access over 50 core tools within Creative Cloud applications such as Photoshop, Premiere, and Express. The Affinity connector significantly improves rendering efficiency, enabling users to automate repetitive production tasks and generate custom functions as needed.
For 3D modeling, Autodesk Fusion subscribers can create and modify complex industrial 3D models through simple conversations with AI. SketchUp users can describe initial ideas to Claude, which generates base 3D models that can then be refined within the software. The open-source 3D platform Blender integrates a Python API, allowing developers and animators to analyze and debug complex scenes, batch-create or modify objects with custom scripts, and even add new tools directly within Blender’s interface.
Additionally, visual artists can use natural language to control Resolume Arena and Wire in real time. In the audio domain, Ableton users can query official documentation for Live and Push directly through Claude, while Splice creators can quickly search royalty-free sample libraries.
Anthropic stated that Claude can act not only as an intelligent tutor for complex software but also as a tool to write plugins and restructure data formats, taking over repetitive tasks and allowing creators to focus on core creative work. Source
Framework launches RTX 5070 GPU module
Recently, modular hardware company Framework released a 12GB GPU module for its RTX 5070 laptop lineup. In terms of core specifications, the 12GB and 8GB versions share identical CUDA cores, clock speeds, power consumption, and architecture—both feature 4608 cores, a base clock of 2.0GHz, boost clock of 2.4GHz, and a 100W power limit. Memory bandwidth is 384GB/s with a 128-bit interface, using NVIDIA laptop-grade GDDR7 memory. Both versions include fourth-generation ray tracing cores and fifth-generation tensor cores, supporting DLSS 4 technology, with identical video encoding/decoding specs, interface configurations, and physical dimensions.
Both RTX 5070 modules are fully compatible with all Framework Laptop 16 models, including those powered by Ryzen AI 300 and Ryzen 7040 processors. The 12GB version is now available for pre-order, with shipping expected in June 2026. Source
FILCO keyboard manufacturer announces takeover of brand operations
Following recent news of the shutdown of FILCO brand holder Diatech, the Taiwan-based manufacturer behind FILCO keyboards, Nonette Co., Ltd. (Gongrong Technology), announced on April 27 that it has taken on the responsibility of supporting FILCO users. The company stated it will continue handling repairs and sales to “protect every FILCO keyboard in users’ hands.”
Meanwhile, Shanghai Huajin Digital Technology Co., Ltd., responsible for FILCO’s mainland China operations, confirmed that normal sales and after-sales services in the Chinese market remain unaffected despite Diatech’s closure. Source
Ant Group open-sources Ling-2.6-flash large model
On April 29, Ant Group announced that its Ling large model has officially open-sourced Ling-2.6-flash. The release includes multiple versions such as BF16, FP8, and INT4, allowing developers to choose flexibly based on hardware environments, inference costs, and deployment needs.
Ling-2.6-flash is an instruction-tuned model with 104 billion total parameters and 7.4 billion active parameters. Its core strengths lie in a hybrid linear architecture, token efficiency optimization, and targeted enhancements for agent scenarios. The model is now available on Hugging Face and ModelScope. Source
“Guidelines for Ethical and Safe AI Application (Version 1.0)” to be released soon
On April 29, China’s State Council Information Office held a press conference where officials including Cyberspace Administration of China Deputy Director Niu Yibing and Guangxi Party Committee Standing Member Chen Yijun introduced details about the 2026 China Internet Civilization Conference.
The conference will take place from May 19 to 20 in Nanning, Guangxi, under the theme “Civilized Cyberspace, Advancing Forward with Strength.” It will include an opening ceremony, a main forum, 14 sub-forums, and themed activities. A dedicated forum on AI-driven internet civilization will focus on technological evolution, innovation, and governance, fostering discussions and promoting the practical application of AI innovations.
During the forum, the “Guidelines for Ethical and Safe AI Applications (Version 1.0)” will be released, addressing potential impacts of AI on social relationships, emotional dependency, public order, and individual rights, providing practical references to help stakeholders navigate development while maintaining safety boundaries. Source
News Worth a Quick Look
According to Wccftech on April 28, the upcoming Tensor G6 chip—expected to debut with the Pixel 11 series in August—will continue using the PowerVR CXT-48-1536 GPU introduced in 2021. On the CPU side, Tensor G6 is expected to shift from an 8-core design to a 7-core 1+4+2 architecture, with upgraded ARM C1 Ultra or C1 Pro cores running at up to 4.11GHz. For security, it will integrate a new Titan M3 co-processor, providing enhanced hardware-level protection for encryption keys and biometric data. Source
New-generation Steam Controller announced by Valve
China Unicom launches “Unicom Cube” customizable service
Google Translate introduces “Pronunciation Practice” feature
News Worth a Quick Look
Ledo L80 unveiled
On April 28, Ledo Auto launched the Ledo L80 in Hangzhou, positioning it as a “smart dual-cabin, large five-seat flagship SUV.” The vehicle measures 5145 mm in length and offers two intelligent driving solutions: a pure vision setup powered by Orin-X, or a combination of the “Shenji chip” with LiDAR. It comes in six exterior colors and four interior themes. The model emphasizes cargo space and scenario-based expansion, supporting features such as carrying three bicycles, a customized camping kit, and a three-section convertible bed, all offered as official accessories.
The Ledo L80 starts at a pre-sale price of 245,800 RMB, with the BaaS (battery-as-a-service) version starting at 159,800 RMB. Test drives will begin on May 1, with the official launch set for May 15. Source
OnePlus Ace 6 Supreme Edition released
On April 28, OnePlus released the OnePlus Ace 6 Supreme Edition, powered by the Dimensity 9500 platform and featuring a self-developed “Fengchi gaming engine” along with an “Ice Glacier cooling system.” The display is a BOE co-developed 165 Hz high-refresh-rate panel with 1800 nits global brightness, 0.1 ms grayscale response time, 96% first-frame brightness ratio, and 3840 Hz PWM dimming. The device also includes the “Lingxi touch chip,” supporting a 4000 Hz instantaneous touch sampling rate and 480 Hz multi-touch sampling, along with a gyroscope adjustable up to 500 Hz. The phone is rated IP66, IP68, IP69, and IP69K for dust and water resistance. It is equipped with an 8600 mAh “Glacier battery” supporting 120 W fast charging. In terms of imaging, it features a 50 MP main camera with a 1/1.55-inch sensor. The OnePlus Ace 6 Supreme Edition is now available across all channels, starting at 3,099 RMB (12GB + 256GB).
OnePlus also introduced the “OnePlus Gunshen Gaming Controller,” featuring a hybrid touch-and-press design with esports-grade micro-switches, priced at 449 RMB, with a 50 RMB discount when purchased together with the phone. Source
New-generation Steam Controller announced by Valve
On April 28, Valve announced that its new-generation Steam Controller will go on sale at 1:00 AM Beijing time on May 5. The controller can be used independently without a Steam Machine and is compatible with all devices running Steam or the Steam Link app, including PC, Mac, mobile devices, and Steam Deck. In terms of configuration, the controller includes standard control buttons, a Steam button, a quick access button, four assignable grip buttons, two full-size TMR magnetic joysticks with capacitive touch, and a 6-axis IMU gyroscope. It also features four haptic motors, including two LRA motors inside the touchpads and two high-output LRA motors in the grips. Each side is equipped with a 34.5 mm square touchpad with haptic feedback and pressure sensitivity, with configurable click strength. The back of the controller also includes two capacitive sensing areas.
The controller supports 2.4GHz wireless connectivity (4 ms polling rate), comes with an 8.39 Wh lithium battery, offers over 35 hours of battery life, weighs approximately 292 g, and is priced at $99. Source
China Unicom launches “Unicom Cube” customizable service
On April 28, China Unicom launched the “Unicom Cube” customizable communication service, centered around a drag-and-drop configuration experience. The service allows users to freely combine service components with high flexibility, supporting online self-adjustment and dynamic changes. In terms of pricing, the service adopts tiered data pricing: the first 20 GB per month is charged at 1.5 RMB/GB, usage between 20 GB and 50 GB is priced at 1.2 RMB/GB, and usage above 50 GB is charged at 1 RMB/GB. Voice calls are billed on a pay-as-you-go basis at 0.1 RMB per minute, while SMS and MMS are charged at 0.1 RMB per message.
The service has a base monthly fee of 39 RMB, with China Unicom promising a true pay-for-what-you-use model, ensuring users will not encounter overage charges common in traditional plans. Source
Google Translate introduces “Pronunciation Practice” feature
On April 28, marking the 20th anniversary of Google Translate, Google announced a new pronunciation practice feature. After obtaining a translation result, users can tap the practice button to read aloud, with AI analyzing speech and providing real-time feedback. The feature is based on the Gemini Understand and Ask capability introduced earlier this year and is currently available to Android users in the United States and India, supporting English, Spanish, and Hindi. Source
News Worth a Quick Look
According to Engadget citing The Information, Google has signed a confidential agreement with the U.S. Department of Defense allowing its AI models to be used for “any lawful government purpose,” including handling sensitive tasks in classified environments. The agreement restricts the use of AI for domestic mass surveillance, but Google cannot veto lawful government actions and must assist in adjusting AI safety settings and filters when requested. Prior to the agreement, more than 600 Google employees had written to CEO Sundar Pichai urging the company not to involve AI in classified military projects. Google responded that API access remains governed by standard terms and that the cooperation supports national security in a responsible manner. Source
According to 9to5Google citing leaker Lanzuk on Naver, Samsung is considering a new rear camera module design for the Galaxy S27 series, replacing the design used for the past three years. Compatibility with Qi2 magnetic accessories is cited as a key reason. Camera specifications for the next generation may also be revised, though the change remains uncertain, with reports suggesting “recent cost issues” could delay the full redesign. Source
Google’s Nearby Share (AirDrop-like feature) has now expanded support to the international versions of OPPO Find X9 Ultra and vivo X300 Ultra. Source
Alibaba launches HappyHorse 1.0 video generation model
WeChat warns of “755” spoofed scam calls
JBL releases two new audio products
CHERRY unveils MX 8.3 Pro mechanical keyboard
Microsoft internally launches the Windows K2 project to optimize the Windows 11 experience
LincPlus introduces new NAS product LincStation E1
Apple launches a 12-month installment annual subscription plan
GitHub Copilot will fully shift to a usage-based pricing model
News Worth a Quick Look
Alibaba launches HappyHorse 1.0 video generation model
On April 27, Alibaba announced that its video generation model HappyHorse 1.0 has entered limited rollout testing. According to Alibaba, HappyHorse 1.0 is open for testing to professional creators, enterprise clients, and general users. Professional creators and enterprise users can register via the official HappyHorse website and the Alibaba Cloud Bailian platform, while general users can access it through the “HappyHorse” entry at the bottom of the homepage in the latest version of the Qwen App.
In terms of features, HappyHorse 1.0 supports text-to-video, image-to-video, multi-shot storytelling, multi-aspect-ratio adaptation, and video super-resolution output. Alibaba states that the model can generate videos up to 15 seconds in length and supports multilingual content creation, including Cantonese, English, French, and Korean. In terms of pricing, the official HappyHorse website lists 720P video generation at 0.9 RMB per second, while the same capability in the Qwen App is priced as low as 0.44 RMB per second. Source
WeChat warns of “755” spoofed scam calls
On April 27, WeChat issued an official notice warning that scammers have recently used number spoofing tools to impersonate “Tencent Customer Service” or “WeChat Security Center” when calling overseas users.
According to the notice, scammers may accurately state users’ names and overseas phone numbers during calls to gain trust, then threaten users by claiming that their WeChat accounts will be suspended, inducing them to report to the police. The scammers may then offer to “assist” by transferring the call to a fake “police hotline,” where they further request bank account details and verification codes to complete the fraud. In other cases, scammers may use claims such as “linked违规 public accounts” or “suspected illegal fundraising” to deceive users.
WeChat emphasized that such spoofed calls usually begin with “755,” which differs from the official customer service number “0755.” The company clarified that its hotline has never proactively made international outbound calls and will not notify users of account violations via phone. All such information is communicated only through in-app system notifications or penalty alerts. WeChat also stated that it will never request sensitive personal information or assist in transferring calls to law enforcement agencies.
WeChat advises users to hang up immediately upon receiving such calls and avoid providing any information to prevent falling victim to fraud. Source
JBL releases two new audio products
On April 27, JBL announced two new audio products: the LIVE 780NC over-ear headphones and the BandBox Solo speaker.
The LIVE 780NC headphones come in five color options and feature an ergonomic silicone headband and metal hinge structure. They are equipped with 40mm dynamic drivers, support spatial audio and Adaptive Noise Cancelling 2.0, include a six-microphone array, and support Bluetooth 6.0. In terms of battery life, they can last up to 80 hours with noise cancellation turned off. The headphones will go on sale on April 28 at a price of 1,399 RMB.
The BandBox Solo speaker is designed for practice and creative scenarios. It features JBL sound tuning, a built-in tuner and metronome, and integrated effects, allowing one-touch sound adjustments without the need for pedals. It also supports wired headphone connections for practice and USB-C direct recording, along with AI-powered track separation and app-based smart controls. The battery life is approximately 6 hours, and the product is priced at 2,199 RMB. Source
CHERRY unveils MX 8.3 Pro mechanical keyboard
On April 27, CHERRY launched the MX 8.3 Pro mechanical keyboard, positioned as a high-end product focusing on build quality and performance, priced at 2,299 RMB.
The keyboard adopts an orange-and-black color scheme and continues with an 80% TKL layout. It is equipped with the new-generation MX Lumina Brown switches and supports tri-mode connectivity, including Bluetooth, wired, and 2.4GHz. In USB-C wired mode, it supports up to an 8kHz polling rate, while the 2.4GHz mode supports up to 4kHz polling.
In terms of battery life, the MX 8.3 Pro features a built-in 5000mAh battery, with an official claim of up to 1400 hours of usage, and supports advanced SOCD key functionality.
Regarding structure and materials, the MX 8.3 Pro uses a CNC-machined aluminum chassis. The front features a full-color LCD control screen and an X-shaped multifunction knob. Internally, it adopts a gasket structure, consisting of an FR4 gold-plated positioning plate, Poron foam, IXPE switch pads, PET acoustic pads, a hot-swappable PCB, switch socket foam, and bottom sound-dampening foam. It also comes with double-shot PBT keycaps with a fine matte finish. Source
Microsoft internally launches the Windows K2 project to optimize the Windows 11 experience
Microsoft Windows chief Pavan Davuluri stated at an internal briefing in March this year that Microsoft will optimize the system based on user feedback, aiming to make Windows 11 “the most stable operating system in the world.”
According to reports, Microsoft has internally launched an initiative called the “Windows K2” plan. The project was initiated in the second half of 2025 and focuses on addressing user concerns about bloated AI features and declining system reliability. The K2 plan revolves around three core pillars: Performance, Craft (code quality), and Reliability, continuously improving system performance through Windows Insider feedback, telemetry data, and user research.
In addition, the plan introduces “Community” as a fourth pillar. Microsoft stated it will rebuild the Windows enthusiast ecosystem, including bringing back offline Windows Insider events and encouraging development teams to respond more actively to user feedback on social media.
In terms of performance optimization, Microsoft is benchmarking against SteamOS for gaming performance1, aiming to adjust the system’s underlying architecture over the next one to two years to achieve similar results. Meanwhile, internal teams will gradually migrate to the WinUI 3 framework to improve native UI performance, and rebuild components such as the Start menu, Run dialog, and Control Panel, adding more customization options like UI scaling and menu hiding. Source
LincPlus introduces new NAS product LincStation E1
On April 27, LincPlus, a sub-brand of Techvision, announced the launch of its new ARM-based NAS product, LincStation E1.
The device adopts a 2+2 hybrid storage design, offering two 3.5”/2.5” SATA drive bays and two M.2 2280 NVMe SSD slots. In terms of hardware, the LincStation E1 is powered by a processor featuring 4× Cortex-A55 cores (2.0GHz) and a Mali-G52 GPU, supporting 4K video decoding, and comes with a memory configuration of 4GB DDR4 + 64GB eMMC.
On the software side, the device runs the self-developed LincOS operating system and supports interoperability with Android, iOS, and Windows devices.
For external I/O, the LincStation E1 includes one USB-A 5Gbps port, two USB-A 480Mbps ports, one 1GbE RJ45 port, and one HDMI 2.1 TMDS port, along with a wireless card supporting both 2.4GHz and 5GHz bands. Source
Apple launches a 12-month installment annual subscription plan
Today, Apple announced a new subscription model for the App Store, allowing developers to offer monthly payments with a 12-month commitment.
According to Apple, this subscription format allows users to pay monthly while committing to a full 12-month period, enabling developers to break down annual subscriptions into lower-barrier monthly payments. The feature is currently available for configuration in App Store Connect and can be tested via Xcode, but has not yet been officially rolled out to users in the App Store.
Regarding payment mechanics, Apple stated that users can cancel subscriptions at any time, but cancellation only prevents auto-renewal after the current cycle; payments within the committed period must still be completed. Users can view completed and remaining payments in their Apple Account, and the system will send reminders via email and push notifications for upcoming charges.
Apple noted that this feature is expected to launch alongside iOS 26.5 and corresponding system updates next month, and will be available for devices running iOS 26.4 and above. However, the subscription model will not initially be available in the United States and Singapore, with reasons and future rollout timelines yet to be announced. Source
GitHub Copilot will fully shift to a usage-based pricing model
GitHub announced that all Copilot plans will no longer use the “Premium Request” (PRU) counting model, and will instead allocate a certain number of GitHub AI Credits per subscription each month, with paid users able to purchase additional credits. Usage will be calculated based on token consumption, including input, output, and cached tokens, with specific rates aligned to the public API pricing of each model.
GitHub stated that Copilot has evolved from an in-editor assistant into an agentic platform capable of executing long-running, multi-step autonomous coding sessions and traversing entire repositories using the latest models. The existing PRU-based pricing is no longer sustainable, and shifting to usage-based billing will help maintain long-term service reliability while reducing restrictions for heavy users.
Following this change, base subscription prices for GitHub Copilot will remain unchanged, and Copilot Pro plans will include equivalent monthly AI Credits. Code completion and Next Edit Suggestions will continue to be included in all plans without consuming AI Credits. Copilot code review features, however, will consume both AI Credits and GitHub Actions runtime.
For existing annual subscribers, GitHub will provide transition arrangements. Users with annual Pro or Pro+ plans will continue under the original request-based pricing terms until their subscription expires. However, starting June 1, model multipliers will increase significantly. After expiration, users will be automatically moved to Copilot Free, with the option to upgrade to a monthly paid plan or convert the remaining value of their annual plan into credits on a prorated basis. Source
News Worth a Quick Look
Recently, data miner Billbil-kun revealed that Microsoft plans to introduce a new subscription tier called “Starter Edition” for Xbox Game Pass. This plan will offer 10 hours of Xbox cloud gaming per month, along with in-game perks and access to a library of over 50 games, playable across consoles, PC, and other devices. However, the Starter Edition will not support online multiplayer, making it the only tier in the current Xbox Game Pass lineup without multiplayer functionality. The report suggests it may launch in mid-May this year, though pricing and full benefits have yet to be officially announced. Source
On April 27, Sony announced price increases for the PlayStation 5 series and related accessories across Southeast Asian markets including Singapore, Malaysia, Thailand, Indonesia, the Philippines, and Vietnam, effective May 1. Indonesia will see the largest increase, with the PS Portal priced at 5,199,000 IDR, up 44.5% from 3,599,000 IDR. The PS5 and PS5 Digital Edition will be priced at 11,399,000 IDR and 9,999,000 IDR respectively. In Singapore, the PS5 will be priced at 849 SGD, with the Digital Edition and PS5 Pro at 764 SGD and 1,167 SGD respectively, and the PS Portal at 347 SGD. Sony stated that the price adjustments are due to global economic pressures. Source
According to a report by 9to5Google on April 26, citing informed sources, the upcoming update will cover all Google Workspace apps. The new icon design will align with Google’s gradient “G” logo and products such as Gemini, Google Photos, and Google Maps, reinforcing AI-related visual identity. The updated icons will no longer be restricted to Google’s four standard colors, instead adopting more distinctive color schemes and shapes to address user concerns about the lack of recognizability in previous designs. Source
Editor’s note: Compared to Windows 11, which runs games natively, games on SteamOS generally rely on the Proton compatibility layer. After optimization, performance loss in complex scenarios is typically around 10% to 20% at most. ↩︎
U.S. asks diplomats to warn other countries about China’s AI distillation practices
Domestic AI firms told not to accept unapproved U.S. investment
Hong Kong steps up e-cigarette ban promotion ahead of Golden Week
FILCO keyboard parent company shuts down
Authorities respond to claims that Doubao accessed 2026 Shandong civil service exam results early
News Worth a Quick Look
X launches messaging app XChat
On April 24, X officially launched its standalone messaging app XChat to the public, currently available only on iOS. XChat uses X accounts for login and supports text messaging, file sharing, voice and video calls, as well as group chats. Meanwhile, due to low usage and widespread spam on X’s existing Communities feature, it will be shut down, with XChat serving as the new hub for migrating community members.
In terms of privacy, XChat claims to be free of ads and tracking mechanisms, and offers features such as message editing and recall, disappearing messages, and screenshot protection. X also states that all messages are protected by a PIN code and use end-to-end encryption, though the industry remains skeptical. During early testing, experts warned that XChat’s security might not match that of mature encrypted apps like Signal, and the security of the official release still requires further evaluation.
X’s Head of Design, Benji Taylor, revealed that the current version is only the beginning of X’s push into instant messaging. The company is also internally testing a standalone payments app. Previously, Elon Musk had envisioned integrating messaging, payments, content creation, shopping, and AI into a single “super app” similar to WeChat. However, the launch of XChat suggests a shift toward a more modular product ecosystem.
U.S. asks diplomats to warn other countries about China’s AI distillation practices
According to Reuters, the U.S. State Department sent a diplomatic cable last Friday to embassies worldwide, instructing American diplomats to warn foreign governments that Chinese AI companies such as DeepSeek may be “extensively stealing” intellectual property from U.S. AI firms through model distillation. Distillation refers to using outputs from large models to train smaller, lower-cost models. The cable also indicates that separate diplomatic démarches will be made to the Chinese government on this issue.
The document names companies including DeepSeek, Moonshot AI, and MiniMax, claiming that unauthorized distillation allows foreign firms to release comparable products at significantly lower costs, potentially undermining safety mechanisms and ideological neutrality in original models. Previously, OpenAI had raised similar concerns to U.S. lawmakers.
In response, the Chinese Embassy in the U.S. dismissed the claims as baseless and described them as a deliberate attack on China’s AI industry. DeepSeek has previously stated that its V3 model was primarily trained on naturally collected web data, not synthetic data generated by OpenAI models.
Domestic AI firms told not to accept unapproved U.S. investment
According to Bloomberg, regulators are planning to restrict domestic tech companies—especially well-known AI startups—from accepting U.S. capital without prior government approval.
In recent weeks, agencies including the National Development and Reform Commission have issued guidance to several private tech firms, requiring them to reject unapproved U.S. funding. Companies such as Moonshot AI, which is seeking to expand financing, and StepFun, which is preparing for a Hong Kong IPO, have reportedly received such instructions. StepFun is also restructuring its overseas corporate setup in line with regulatory requirements. Additionally, authorities have imposed similar restrictions on ByteDance, prohibiting the transfer of existing shares to U.S. investors without approval.
These measures stem from concerns over the outflow of domestic technology. In December 2025, Meta announced the acquisition of AI agent startup Manus. Although registered in Singapore, Manus was founded by Chinese entrepreneurs. Prior to the acquisition, the company restructured and relocated its domestic staff to Singapore to complete the deal without regulatory review from Beijing. Authorities are now conducting a joint investigation into the transaction and its implications, and one of its co-founders has reportedly been restricted from leaving the country.
Previously, the U.S. had already implemented rules in 2025 banning its capital from investing in Chinese semiconductor, quantum computing, and AI companies.
Hong Kong steps up e-cigarette ban promotion ahead of Golden Week
According to the South China Morning Post, as the Golden Week holiday approaches, Hong Kong has intensified promotion of its upcoming e-cigarette ban at airports, land border crossings, and mainland subway stations, while also collaborating with mainland media to raise awareness.
Under the revised Smoking (Public Health) Ordinance, starting April 30, 2026, possession of specified alternative smoking products—including e-cigarettes, heated tobacco products, and herbal cigarettes—will be prohibited in public places. Anyone found using or carrying activated alternative smoking products will be presumed to possess such items. Violations constitute a criminal offense; for small quantities intended for personal use, offenders will face a fixed penalty of HKD 3,000. Tourists will receive electronic penalty notices and can pay fines via Alipay or WeChat.
Authorities have also promoted earlier measures, including expanding statutory no-smoking areas to within three meters of entrances to designated premises such as care homes, schools, and hospitals. Smoking is also prohibited while queuing, including when waiting in lines of two or more people for public transport, within designated boarding areas (such as bus shelters), and when entering venues like cinemas, museums, and stadiums.
FILCO keyboard parent company shuts down
According to GIGAZINE, Japan’s Diatec, best known for its mechanical keyboard brand FILCO, announced on its official website that it has ceased operations as of April 22, 2026.
In the announcement, Diatec apologized to consumers for the “sudden” news and thanked users for their long-term support. The company emphasized that all personal user data collected through e-commerce sales and after-sales support had been destroyed on April 22.
FILCO is a landmark brand in the history of mechanical keyboards. Its classic Majestouch series is known for durability and excellent typing feel, long regarded as a benchmark among high-end mass-produced mechanical keyboards, with strong reputation and recognition among enthusiasts.
In recent years, Diatec had continued to iterate on its products. In 2022, it launched the Majestouch Convertible 3, supporting both wired and wireless connectivity, offering multiple switch options including red, brown, blue, and silent red, along with various layouts. In 2023, it also introduced the Majestouch Xacro M10SP, a split mechanical keyboard with 10 macro keys.
Authorities respond to claims that Doubao accessed 2026 Shandong civil service exam results early
According to Yangtze Evening News, on the evening of April 23, a user claimed online that they had accessed the 2026 Shandong public institution exam results using Doubao, even though the results were officially scheduled for release the following day. Screenshots of the results were shared before the post was later deleted.
In response, staff responsible for application policy consultation stated that since results were to be released on April 24, the query portal had been tested the night before. “Some users accidentally accessed it,” they said, adding that the test portal was promptly shut down after discovery and caused no adverse impact. Analysts believe the incident may have occurred because Doubao inferred the result URL based on patterns from previous years, while the official system began testing without adequate safeguards, leading to the so-called “early access.”
Such incidents caused by predictable URLs are not uncommon. In January 2011, Microsoft’s financial results were discovered hours early by a data company, forcing an early release. In May 2024, the U.S. Bureau of Labor Statistics uploaded CPI data 30 minutes ahead of schedule to a publicly guessable URL, though it went unnoticed. In November 2025, the UK’s Office for Budget Responsibility leaked its spring budget report a week early in a similar way, with at least seven IP addresses accessing it, leading to the chairman’s resignation.
News Worth a Quick Look
Analyst Dan Nystedt claims that the standard version of iPhone 18 (expected to be delayed to 2027) will feature 12GB of RAM, a 50% increase over the iPhone 17’s 8GB.
OpenClaw 2026.4.24, released on April 26, sets DeepSeek V4 Flash as the default model.
On April 24, Meta issued an internal memo announcing plans to lay off about 10% of its workforce (around 8,000 employees) on May 20 and cancel hiring for 6,000 open positions to offset heavy AI investment costs. As of early this year, Meta employed nearly 79,000 people globally. Affected U.S. employees will receive 16 weeks of pay, plus two additional weeks for each year of service, along with 18 months of healthcare coverage for themselves and their families. Around the same time, Microsoft announced its first voluntary buyout (early retirement) program for U.S. employees, open to senior directors and below whose combined age and tenure equal at least 70.
According to TechCrunch, surging demand for running local AI models has led to a complete sellout of the M4 Mac mini through official channels, driving up prices in the second-hand market. The base model, originally priced at $599 (16GB RAM, 256GB storage), is currently unavailable for delivery or in-store pickup on Apple’s website, while higher-capacity versions are backordered until June. Multiple outlets note this is the first time the base model has completely sold out. On platforms like eBay, lightly used units are now listed above $700, with near-new units priced between $715 and $795.
OpenAI has announced the GPT-5.5 series, describing it as the most intelligent and intuitive model version to date. Compared to GPT-5.4, the new models deliver stronger performance in coding and debugging, online research, and generating documents and spreadsheets, while also supporting cross-tool collaboration to execute complex tasks. Users can input multi-step, unstructured requests, and the model can automatically handle task planning, tool invocation, result verification, and uncertainty management. In Codex scenarios, GPT-5.5 significantly reduces token usage and introduces more robust safety mechanisms.
GPT-5.5 will be gradually rolled out to ChatGPT Plus, Pro, Business, and Enterprise users, and will be integrated into the Codex platform. The Pro tier will offer more advanced capabilities. Source
Hanvon launches Clear7C color e-ink e-reader
On April 23, Hanvon launched the Clear7C color e-ink e-reader. The device features a 7-inch Kaleido 3 color e-ink display with a resolution of 1680×1264, supporting 150 PPI in color mode and 300 PPI in black-and-white mode. It also includes physical page-turn buttons on one side. The Clear7C weighs approximately 175g and is just 3.9mm thick, with a strong focus on portability and one-handed usability.
In terms of hardware, the Clear7C is powered by an octa-core processor, paired with 4GB of RAM and 64GB of storage, and includes a 2700mAh battery. The company claims up to 120 hours of reading time. It runs on Android 14. Source
On April 23, Onyx Boox introduced the Leaf5+ monochrome e-ink e-reader. The device features a 7-inch Carta 1300 panel with a 300 PPI resolution, supports stylus input, and uses etched tempered glass for the display, along with physical page-turn buttons. The Leaf5+ is about 4mm thick and weighs 192g, and also includes a new light guide plate, with the company claiming up to 90% screen uniformity and a color temperature adjustment range of 3000K–5500K.
For hardware, it is equipped with a Qualcomm octa-core processor, 4GB LPDDR4X RAM, and 64GB of storage, along with a 2000mAh battery. The device is priced at RMB 1,749. Source
iFLYTEK unveils AINOTE2 e-ink device
On April 23, iFLYTEK unveiled the AINOTE2 e-ink device, described as an “ultra-thin e-ink tablet.” It features a 10.65-inch monochrome e-ink display with a resolution of 1920×2560 and a pixel density of 300 PPI, supporting handwriting input via the iFLYTEK I-Pencil electromagnetic stylus. The device uses a magnesium-aluminum alloy body, measuring about 4.2mm thick and weighing 295g.
In terms of hardware, the AINOTE2 comes with 4GB of RAM, 64GB of storage, and a 4000mAh battery. It runs on Android 14 and includes several custom AI features from iFLYTEK, such as content summarization, speech-to-text transcription, and integrated note-taking. The device is priced at RMB 3,999. Source
Panasonic releases LUMIX ZS300 camera
On April 23, Panasonic announced the LUMIX ZS300. The device features a 1-inch CMOS sensor with approximately 20.1 million effective pixels, paired with a Leica-certified 15× optical zoom lens (equivalent to 24–360mm). It supports a minimum focusing distance of 3 cm and includes 5-axis hybrid image stabilization. The camera is available in black and silver, with a built-in flash and 22 creative filters.
In terms of features, the LUMIX ZS300 supports 4K video recording, with a 4K Photo mode capable of 30 fps high-speed burst shooting, as well as FHD 120fps slow-motion recording. For connectivity, it supports Bluetooth and Wi-Fi for pairing with smartphones and transferring images, and also supports charging via a USB Type-C port. The product will officially go on sale at 12:00 on April 27, with a starting price of RMB 4,998. Source
Google introduces two AI chips: TPU 8t and 8i
At the Google Cloud Next 2026 conference, Google unveiled two new generations of custom AI ASIC chips: TPU 8t and TPU 8i, optimized for training and inference scenarios respectively, while also supporting cross-scenario workloads.
The TPU 8t is designed for large-scale, compute-intensive training tasks, focusing on improved compute throughput and vertical scaling bandwidth, delivering higher efficiency in complex model training. The TPU 8i, on the other hand, is optimized for low-latency inference and AI agent workloads, offering higher memory bandwidth and integrating a CAE (on-chip acceleration engine). It is equipped with 384 MB of on-chip SRAM and 288 GB of off-chip HBM, reducing data transfer overhead during execution and lowering latency.
At the cluster level, the TPU 8i Pod scale has increased from 256 chips in the previous generation to 1,152 chips. A single Pod delivers 8.67× higher FP8 compute performance and 5.74× greater HBM memory capacity. Google states that the overall energy efficiency of the new TPU 8t and TPU 8i is approximately double that of the previous-generation TPU Ironwood. Source
On April 23, Tencent Cloud released a security advisory regarding a supply chain poisoning risk affecting the open-source PyPI inference framework Xinference. The advisory states that affected versions may contain malicious code when installed or imported, potentially exfiltrating sensitive data—including cloud platform credentials, API keys, SSH keys, cryptocurrency wallet information, database credentials, and environment variables—to attacker-controlled command-and-control servers.
Tencent Cloud noted that such risks typically originate from tampering within the software supply chain and may pose significant threats to AI inference environments and related infrastructure. Affected users are advised to immediately uninstall compromised versions, scan systems for backdoors and residual files, and rotate critical credentials and keys as soon as possible. A comprehensive security audit is also recommended to assess potential impact and strengthen system defenses. Source
News Worth a Quick Look
X product lead Nikita Bier said that Premium subscribers on iOS can now early-access a new feature: users can pin specific topics on the “Home” feed, and Grok will personalize timeline recommendations based on those topics and past interactions. Powered by Grok’s understanding of platform content and personalization algorithms, the feature aims to deliver a highly customized feed experience, with an Android rollout expected soon. In addition, X announced it will shut down the Communities feature on May 6 due to a continued decline in user activity, and recommends migrating communities to group chats on XChat. Meanwhile, competing platforms like Threads and Mastodon continue to strengthen community-driven feeds. Source
Memory manufacturer Kioxia introduced the EG7 series SSDs on April 23. The lineup uses QLC NAND across all models and does not include a DRAM cache. Performance-wise, EG7 offers up to 7000 MB/s sequential read and 6200 MB/s sequential write speeds, with random performance reaching up to 1,000K IOPS. Capacities include 512GB, 1TB, and 2TB, with form factors in M.2 2230, 2242, and 2280. The series targets the consumer market and has already been supplied to multiple OEMs for use in thin-and-light laptops and desktops. Source
Google confirmed during its Cloud Next 2026 keynote on April 22 that its partnership with Apple stems from a multi-year agreement signed in January 2026. Under the deal, Google will act as the preferred cloud provider, delivering core technical support, with results expected to materialize by the end of 2026. Source
Rambus announced a chipset solution for LPDDR SOCAMM2 modules on April 22, designed to provide control, telemetry, and power management for JEDEC-compliant SOCAMM2 server memory modules. The goal is to meet the high stability and efficiency demands of AI servers. SOCAMM2 introduces a modular, replaceable design to replace traditional soldered LPDDR memory, combining LPDDR’s energy efficiency with improved serviceability and scalability for data centers. Rambus’ LPDDR5X SOCAMM2 chipset supports speeds up to 9.6 Gb/s, emphasizing reliability and efficiency under high bandwidth workloads. Source
According to Mark Gurman, Apple is developing six new product categories, including AI-powered AirPods, smart glasses, wearable clip-on devices, smart displays, desktop robots, and security cameras. Source
