OpenAI has launched the Slides tool within ChatGPT, which supports generating presentations based on conversation content, uploaded files, or outlines, and offers multiple built-in templates. It features real-time collaborative editing, comment feedback, and automatic adjustments based on feedback. Presentations can be exported as PDF or PowerPoint files, compatible with PowerPoint and Google Slides. Currently, it is only available to Pro, Business, and Enterprise subscribers; ChatGPT Plus users do not yet have access.
Developer anichikage has released Open Resume, an open-source self-hosted resume tool featuring 156 templates. It supports content editing, real-time preview, auto-save, undo/redo, photo management, and module sorting. Exports are available in PDF, PNG, Word (.doc), and JSON formats. Data is stored locally in SQLite. Both the frontend and backend use TypeScript, supporting deployment via Docker or Node.js 24+. AI-powered job description polishing is an optional feature that requires users to configure a compatible service and API key themselves.
Meshy has released its local editing feature, Meshy Edit, based on a new proprietary AI model. Users can select a region of a 3D model and describe the change in one sentence (e.g., "remove the hat"), generating a new model in about 1 minute while keeping unselected areas unchanged; this feature is now open to all registered users, each edit consumes 20 credits, and it supports high-precision models generated by Meshy 6 and Meshy 7.1; the entry point is located at the bottom of the model preview window under the "Edit" button or in the post-processing area of the print module.
The article states that Muse's popularity stems from the US open web ecosystem and the "public corridor" formed by antitrust measures, whereas China has walled-off apps and a neglected web layer, preventing Agents from performing cross-platform tasks. ByteDance, Tencent, and Alibaba are each advancing their own Agents but are constrained by closed ecosystems; Doubao Phone, Huawei Xiaoyi, and StepFun are attempting to break down these walls from the system or protocol layers. Private deployment for B-end clients has begun charging fees, while C-end services remain stuck in a free-service dilemma.
Death Plan Switch is a self-hosted dead man's switch tool designed specifically for independent web developers. It runs on PHP 8.1+ without a database, is licensed under MIT, and is available for free download from GitHub.
After REA (Reverse Engineer Anything) went viral, product managers quickly formed a defensive consensus due to the vulnerability of client-side logic to AI-driven reverse engineering. The core countermeasure involves moving critical logic such as keys, permission verification, and trust decisions entirely to the server side, reducing clients to pure interface shells; for scenarios that cannot be migrated to the cloud, techniques like packing, obfuscation, and virtualization are used to mislead AI. Thales launched dedicated protection solutions such as Sentinel Envelope Plus, which testing showed can reduce AI Agent vulnerability detection rates to zero while increasing token consumption by 970 times. Simultaneously, numerous malicious repositories imitating REA have emerged, among which rea-x64.7z released by the rea-tools organization has been flagged as a Trojan by multiple engines and has been downloaded 172 times.
GeoD Global 0.2.0 is an open-source Windows geospatial data workbench that supports optical and radar imagery from Sentinel-2, Landsat, NAIP, MODIS, VIIRS, HLS, CBERS, and other sources, along with 17 types of data including DEM, land cover, and climate. It provides access to STAC/WMS/XYZ/WCS/WFS/OGC API/ArcGIS services and the ability to open local GeoTIFF/GeoJSON/GeoPackage/Shapefile/PMTiles/MBTiles files. Features include multi-select map preview, download plan verification, local layer management, and raster clipping/stitching. Users can integrate their own models to generate Agent-based retrieval and download plans. It requires WebView2, supports Windows 10/11 x64, and hosts source code on GitHub under the GPL-3.0-only license.
This paper proposes the Split-Intent Jailbreak Attack (SIJA), which uses text to specify the editing target while hiding dangerous operations through graphical auxiliary cues, achieving jailbreak attacks against multimodal image editing models. It constructs the SI2ESBench benchmark, reaching attack success rates of 68.28% on GPT Image 2 and 91.38% on Qwen-Image-3.0-Pro respectively. It also proposes defense principles based on editing intent reconstruction, emphasizing the need to jointly evaluate the overall transformation intent implied by multimodal inputs.
The paper proposes Long Video Meta-memory Harness (LVMH), an automated multi-agent framework that models memory strategy design as multi-objective code search around a fixed answer model; through this framework, the Self-Evolving Evidence Memory (SEEM) strategy is discovered, employing a 'text-first, vision-second' mechanism based on compact structured event memory and evidence experts, invoking local frames via selective visual gating; after evolving on EgoLifeQA, SEEM generalizes to Ego-R1 Bench and Video-MME (L), achieving an average accuracy of 65.7% using GPT-5 mini, a 7.7 percentage point improvement over previous SOTA, with approximately 32x reduction in average evaluation token usage and 91x reduction in average response time.
EMMA is an event-driven multimodal memory system for long video understanding, implemented through a three-stage design: event segmentation (fusing visual semantic drift and speech transitions), hierarchical memory construction (raw video → segment-level → event-level), and multi-agent retrieval (query decomposition, hierarchical search, evidence verification, and answer generation). Its effectiveness was validated on three benchmarks. Code, prompt templates, and reproduction scripts are open-sourced on GitHub.
Rouler is an independently developed network tool for Windows that supports specifying the egress network interface for each app, IP, or subnet, enabling parallel operation of multiple VPN clients such as WireGuard, Tailscale, and OpenVPN. It uses a lightweight Windows network driver written in self-developed Rust, without modifying the routing table, using virtual network adapters, or employing user-space proxies. It provides a real-time connection activity view, fine-grained rule configuration, and a lightweight system tray UI. Currently available as a free public beta, the official version will adopt either a one-time purchase ($79) or annual subscription ($39) model.
In March 2026, Bo Rui Kang's NEO-ONE SCI received the world's first registration certificate for an invasive brain-computer interface medical device. Two days later, it obtained a medical insurance code, and within one week, it was listed on Shanghai's procurement platform, with the invasive implantation fee set at 6000–6600 yuan per procedure. Its listed price is 475,000 yuan per set; as of disclosure, 13 sets have been sold and 8 terminal implants performed, with over 70 orders in hand and intended. However, more than 70% of its revenue in 2025 came from non-invasive EEG acquisition systems. Industry financing is concentrating among leading players, making it difficult for second- and third-tier companies to support complete clinical trials. Long-term safety, real-world efficacy, and hospital operational costs remain key hurdles not yet covered by institutional frameworks.
Cifnews partnered with Changfeng Smart Tax to launch the "9810 Compliance 100 Questions" series. Addressing 28 specific issues in the 9810 export model—such as discrepancies between customs declaration amounts and platform sales revenue, calculation of platform fees, order ledger fields, mapping of FBA Shipment IDs to SKUs, split accounting for overseas warehouses in multiple countries, sharing a single export entity across multiple platforms, the relationship between foreign exchange collection and tax refunds, use of third-party payments, risks of receiving payments via personal accounts, and VAT exemption conditions for cross-border services—the series provides direct conclusions and professional explanations for each, emphasizing that the core principle is reconstructing the actual business flow rather than mechanically matching numbers.
Yana is a local-first, human-governed AI runtime (yana-rt) and desktop IDE (Yana Studio) developed by Vietnamese independent developer Vũ Văn Tâm. Written in Rust, it supports 19 model providers (including Claude, GPT, Gemini, DeepSeek, and local Ollama). It implements action authorization control through capability manifests, policy rules, manual approvals, and hash-chain audit logs. Supporting components include an independent daemon Giám Thị (which verifies core file integrity every 6 hours), adapters (Claude Code/Codex/Cursor/Antigravity), a skill library (2026 skills), and a command set (170 Claude Code slash commands). All code is open source, with no Python dependencies, and provides installation packages for macOS/Windows/Linux as well as source code build options.
Zihao is an iOS input method released by independent developer Hare. It is based on the Rime (Zhongzhou Yun) engine, featuring built-in Wusong Pinyin and a self-developed ~5 MB context-aware word selection model. It supports Xiaohé Double Pinyin auxiliary codes, position-based error correction, Rime configuration file import, and iCloud synchronization. Typing features are free and ad-free; AI features require users to configure their own API keys for OpenAI/Claude/Gemini/DeepSeek, etc. The initial release was not listed on the Mainland China or EU App Stores.
loethen released the CaptionCat browser plugin, which adds pixel cow-cat animations to YouTube English podcast subtitles; the cats pace around, wash their faces, stretch, chase red dots, and react to words (such as cute/wow/fish). It also offers two optional learning features—hovering over text to see full-sentence translations and clicking words to save them—which are disabled by default. The official website is captioncat.app, and the GitHub repository is at https://github.com/loethen/captioncat.
The Ministry of Finance, General Administration of Customs, and State Taxation Administration jointly issued Announcement No. 16 of 2026, extending the tax-exempt policy for cross-border e-commerce export returns until December 31, 2027; it applies to four supervision codes: 1210, 9610, 9710, and 9810. Returns must be re-entered in their original condition within 6 months, excluding food products. Goods that have already received tax refunds must first repay the refunded amount to qualify for tax exemption. The policy covers four channels: bonded e-commerce, direct purchase, B2B direct export, and overseas warehouse exports.
Sign-in Assistant is a browser extension supporting Chrome and Edge that can batch automatically open web pages and simulate clicking sign-in buttons; users record tasks via "Select Sign-in Button", with support for one-click execution, multi-device synchronization, and local export backups, requiring no account registration and collecting no data.
Developer suxiaoshao released "Yuedu," a local TXT novel reader for iOS. It supports file import via the Files app, automatic encoding detection (UTF-8/GBK/Big5), streaming loading for large files, custom typography, table of contents search, bookmarks, annotations, reading statistics, and iCloud sync. The AI assistant is an optional feature, disabled by default, requiring users to configure their own API Key. The app is free, open-source (MIT), not listed on the mainland China App Store, and requires iOS 18+.
QiHang1217 released the open-source browser extension All API Hub, which supports centralized management of multiple AI relay station accounts and API keys, automatic check-ins (e.g., AnyRouter, AgentRouter), cross-site model price comparison (including scenario-based cost estimation), aggregation of site announcements, usage statistics (by time/request volume/Token/cost), and AI gateway integration (e.g., Magpie, New API). Sensitive information is stored locally by default, with optional encrypted WebDAV synchronization. The project is open-source and supports Chrome, Edge, and Firefox.
Ant Ling has open-sourced the Ming-3.0 speech model series, comprising three categories with a total of five models: Text-to-Speech (Ming-3.0-tts series), Audio Understanding and Dialogue (Ming-3.0-audio-flash / tiny), and Automatic Speech Recognition (Ming-3.0-asr-tiny). The TTS models support voice description via text and director mode for controlling performance details; the audio understanding models can directly process long audio while preserving emotion and speaker information; the ASR model covers Mandarin Chinese, English, dialects, and domain-specific hotwords. All models have been open-sourced on Hugging Face and ModelScope, with online experience available via Ling Studio and API access provided.