Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 2, 2026, 9:00 AM
A new "Integrate with Claude / Cursor MCP" feature has been added to the Davflare settings page. It supports one-click generation of mcp.json (for Cursor) and claude_desktop_config.json (for Claude Desktop). The API Key is injected as an environment variable: ${env:DAVFLARE_API_KEY} for Cursor and ${DAVFLARE_API_KEY} for Claude. Restart the client and call tools/list to enable it.
The author's self-developed qBittorrent Android client received a half-year update, adding support for mixed management of Transmission (2.80+) and qBittorrent (4.x/5.x), multi-server homepage aggregation with drag-and-drop sorting, restrained dark/light interfaces, monospaced font number display, swipe-right to pause/swipe-left to delete torrents, peer lists and single-file priority in detail pages, per-server dashboards (including upload country distribution, Tracker proportion, share ratio distribution), global speed limits and scheduled backup speed limits, and adaptation for Android 15+ full-screen displays and predictive back gestures; technically using Kotlin + Jetpack Compose, open-source and ad-free, passwords encrypted via Android Keystore (AES-GCM) and excluded from backups, warnings issued when connecting to public networks via HTTP plaintext, support for RPC path detection and session id handshake, automatic fallback for qB 4.x/5.x pause interface, including 280+ unit tests; requires Android 8.0+, WebUI must be enabled for qB, RPC must be enabled for TR; provides GitHub signed APK download, Google Play version, and source code links.
rankinzheng developed and launched the AION 2 Pet Tool Site (https://aion2pets.app), offering a pet encyclopedia with 206 entries (filterable by genus and acquisition source), three regional soul farming maps (marking monsters and coordinates corresponding to 157 pets), calculations for souls required for Pet Insight and Genus EXP, and a manual collection checklist based on browser local storage (supporting JSON export and merge import).
PixelToKB is a browser-based image processing tool independently released by developer yyyh. It supports compressing JPEG/PNG/WebP images to target sizes (such as presets or custom units like 10/20/30 KB). All processing is completed within Web Workers, ensuring files are not uploaded to servers. After generating results, it reads the actual Blob byte count to verify compliance; if limits are exceeded, it provides clear prompts. It also offers scaling and cropping, batch lossless PNG optimization, metadata inspection, and a multi-language, multi-theme interface.
Author lhwu developed and released the native app LaunchHome in response to the removal of Launchpad in macOS 26. It supports full-screen grid display of apps and folders, drag-and-drop sorting, and type-to-search, emphasizing fast invocation and low resource usage; priced at $9.9 for a one-time purchase, supporting 3 Macs and all 1.x updates, with a 14-day refund policy; official website is https://launchhome.app.
The author launched the My Terminal Themes web app (terminalthemes.com), offering over 800 terminal themes with downloads in nearly 40 formats including iTerm, Ghostty, and Windows. Browsing and downloading require no registration; registered users can vote, favorite, submit themes, and claim authorship. The app includes accessibility features, a Find My Theme tool, and live terminal and code previews, replacing the ANSI screenshots from the original GitHub repository.
A new version of Amazon's brand protection tool "Report a Violation" has been launched. The reporting scope has expanded from IP infringement only to five major categories: IP infringement, product listing violations, product listings altered to other products, item not as described, and violations or safety risks, with an "Other" catch-all option. The "Report Infringement" module has added an AI feature called "Report Assistant," which supports real-time rule queries, guides ASIN searches, explains copyright registration numbers and evidence requirements, helping brand owners submit a complete chain of evidence in one go.
baisensenseng released a Chrome extension that runs entirely locally in the browser, supporting long screenshots, screen recording, simple video editing, and Chinese/English text recognition. It requires no account, uploads no data, and requests only permissions for the current tab; known limitations include inability to process chrome:// pages, text recognition limited to Chinese and English, and basic video editing features.
Sentient Foundation launched the open benchmark Fingerprinting Leaderboard to evaluate model fingerprinting techniques in terms of embedding effectiveness, performance impact, and robustness against adversarial removal attacks; concurrently, it open-sourced the mlprints toolkit, integrating ten fingerprinting schemes and their adaptive attack methods. The study found that almost all schemes, including Sentient's proprietary Perinucleus, can be bypassed by complex attacks. This benchmark aims to advance verifiable model provenance research.
Simon Willison released a URL fragment (#fragment)-based LaTeX renderer tool that supports inputting LaTeX source code, loading GitHub Gists or URLs, includes built-in examples, and allows switching between editing and rendering views; the tool uses MathJax 4.1.3 and faster-latex 0.3.1 for rendering, Pako 2.1.0 for compression, and DOMPurify 3.4.16 for XSS protection, inspired by Tiny Little Office, with the prompt sequence and iframe sandbox design fully disclosed in the body.
arcKode is a local-first AI coding agent that parses codebases via Tree-sitter to build a real-time symbol graph (including functions, classes, imports, and call relationships), dynamically trimming context per task; it supports Ollama-compatible models (default qwen2.5-coder:7b), runs the entire process offline with no data leaving the device; it provides two modes—single-step editing and multi-step planning—with built-in safety mechanisms such as snapshots, anchor editing, and path validation; the indexing structure is based on SQLite, supports 11 languages, and incremental updates rely on SHA-256 change detection.
silkpulse is a newly open-sourced remote web debugging tool that can be integrated via a script tag. It supports real-time reporting of console logs, network requests, and errors, and embeds official Vue and React devtools panels. It provides HTTP interfaces and accompanying skills, enabling AI to read page snapshots and logs, automatically locate errors, and validate forms. It also supports non-intrusive integration via bookmarklets and Tampermonkey.
Siemens Numerical Control (Nanjing) Co., Ltd. piloted wheeled embodied AI robots on its SMT production line to handle tote box transportation, using human teleoperation demonstrations combined with large model virtual simulations to complete skill deployment within three weeks; its core breakthrough lies in reducing the cost of teaching robots to work, while addressing trial-and-error tolerance and multi-vendor equipment coordination issues through Process Simulate simulation and the GB/T 47859—2026 standard. The article points out that what factories truly need is not talent shows, but stable, generalizable, schedulable, and reusable system capabilities.
Independent developer zishubuding developed the iOS player ddPlayer to meet high-intensity video-watching needs for film commentary work. It supports mounting Emby, WebDAV, and NAS, specifically addressing the issue of stuttering at 4x speed under WebDAV mounting common in mainstream players; the app is now available on the App Store with a current limited-time buyout price of 0 yuan; the author notes that it is currently tested and maintained by themselves alone and welcomes user feedback on bugs.
Independent developer JollyJ released the self-developed email workbench Zhiliao OnIt, integrated into DingTalk. It supports intent recognition and strategy suggestions based on enterprise profiles, real-time bidirectional synchronized editing in a left-right dual-column layout, native invocation of DingTalk APIs for one-click meeting scheduling, and decoupled HTML translation for foreign-language emails. The backend uses Python 3.10 + FastAPI + SQLite, while the frontend uses React 19 + TypeScript + Tailwind CSS, with all data stored locally. A pure-memory Mock mode is provided; after cloning the repository, it can be launched locally for trial with a single command.
GitVolt 0.7.0 has been officially released. It is an open-source Git GUI tool built with Rust (using gitoxide for reading and command-line git for writing) and Tauri/CEF, licensed under MIT, and supports private repositories. Features include a command palette (Ctrl+P), independent state management for multiple worktrees, code review and stacked MR/PRs for GitHub/GitLab (including self-hosted instances), file previews rendering Markdown/images/PDF/audio/video, interactive rebase (including fixup/edit/conflict prediction), secure sandboxed WebView, and local keyring token storage. It lists complete capability coverage differences compared to GitKraken. Downloads are available for macOS (Apple Silicon, 12+) and Windows, requiring system Git 2.40+.
Google officially launched Gemini 4 Argon globally. Pro users can experience it via the Gemini app and Search AI mode, while Ultra users additionally get support for AI Studio and Antigravity; developers can access it through paid API calls.
YouTube Music has launched an AI music labeling feature, displaying an oval "AI" label below the name for songs "primarily created using generative AI"; clicking it pops up a menu explaining the specific ways AI was used; labeling is based on voluntary declarations by music partners (categorized into Fully Gen AI / Partly Gen AI / No Gen AI) and YouTube's own AI detection tools, with pitch correction software such as Auto-Tune not considered AI usage.
Linus Torvalds created a new branch named noref-filemap-read-with-llm in his official Linux kernel repository, explicitly noting AI involvement; he personally guided Claude Code to complete code writing, test case generation, review assistance, and documentation drafts for the memory management module, improving 8-thread file read throughput to approximately 3 times; additionally, he used large models to sort through legacy assoc-array code from 2013, reproducing and fixing 4 bugs that had existed for over a decade.
At its Ecosystem Partner Conference, Houmo.ai announced that its entire AI ecosystem software stack for the M50 chip will be fully open-sourced. It covers four layers of tools: model quantization compilation, inference frameworks, domain suites, and deployment operations. The first batch includes projects such as postmo-modelzoo, postmo-quantflow, vllm-postmo, and an adapted version of llama.cpp. This software stack is derived from practical experience with over 150 terminal products equipped with the M50, supports Linux, Windows 11, and Android platforms, and aims to lower the barrier to developing on-device large model products.
The paper proposes the Intact method, which runs on an unmodified browser, constructs a GenTree representation to capture web page provenance and semantics, passes pages to the LLM for decision-making element by element, restricts execution scope, and requires user confirmation for third-party content operations. In tests on 1270 real web pages, GenTree increased the attribution rate of third-party elements to 2.3 times that of the browser; the browser mislabeled 58.7% of third-party elements as first-party. On 1091 web pages containing injection attacks, the Intact planner allowed only 12.8% of attacks to pass (all limited to user confirmation requests), whereas conventional planners reached 73.1%. Functional isolation affected only 3.7% of tasks.
The paper proposes a frozen 16 MB, 512-bit encoded vocabulary to replace the large trainable embedding tables in small language models. Validated on the Gemma 3 270M architecture, this approach compresses the overall model to 46 MB, achieves CPU inference speeds of 615–740 tokens/s, uses only 62 MB of memory, matches floating-point models on tasks such as LAMBADA and PIQA, improves rare-word modeling by 1.7–2.3 nats, and supports adding 10,000 new words post-training.
OpenGCUA 是一项旨在构建并开源通用计算机使用智能体（GCUA）全栈的社区协作项目，首版发布包含支持代码执行与harness-to-model决策迁移的智能体设计；在OSWorld-v2上提升8个模型任务得分，平均部分得分+19.6，步骤减少53.0%；同步开源实现、提示词、全部模型评测轨迹，以及初步的真实设备harness和跨设备SDK；在macOS真实任务中，Qwen3.8-Max推理token用量为GPT-6 Astra的19.4倍。
CarryOn is a dual-loop framework designed to enhance frozen VLA policies, comprising a planning loop (continuously maintaining and updating user-level and scene-level contexts to support personalized task execution across dialogues and dynamic scenes) and an execution loop (proactively blocking predicted dangerous contacts and invoking control-based reusable recovery skills upon execution failure); in both real robots and simulations, it reduced the violation rate of safety-fine-tuned VLAs from 52.6% to 6.1% and increased the recovery success rate from 13% to 80%.