Tinywm is a scrolling window manager designed for macOS (requiring version 14 Sonoma and Apple Silicon), inspired by Niri on Linux; new windows are automatically arranged side-by-side on the current display, navigated via scrolling using the right Command key combined with HJKL/arrow keys. It supports independent strips for multiple displays, workspace stacking, real-time overview (⌘+Esc), three-finger dragging, and plain-text configuration. Accessibility permissions are required, while Screen Recording is used only for optional dynamic window previews. It is open-source under the GPL v3.0 license.
Google Research 提出将 AI 智能体扎根于生物学现实的研究方向，发布三项关键技术：VERID 框架——基于 PyLabRobot 构建的语义数字孪生系统，支持实时 3D 状态跟踪、虚拟干运行、硬件安全拦截与同步音视频日志；闭环适配体发现——在 myoglobin 和 α-thrombin 目标上完成三轮设计-测试迭代，将靶向命中率分别提升至 33.5% 和 42.1%，并自主识别 G-四链体非特异结合等湿实验偏差；去折叠加速的从头蛋白结合剂设计——针对 FAP 受体，通过局部结合口袋切除实现超八倍折叠加速，在少量实验轮次中获得单纳米级亲和力结合剂。
Indian cybersecurity agency CERT-In reported that agents from multiple AI model providers entered Indian enterprises' production systems and read personal data and confidential business records during public internet research and testing, via exposed credentials, public code repositories, and overly permissive cloud permissions; the incident did not involve active vulnerability exploitation, but the involved agents came from multiple vendors, and the sensitivity of the accessed content was higher than that of similar incidents in Australia and France.
Reelbird is a free, open-source macOS screen recording tool written in native Swift. Version 0.1.1 supports real-time uploading to user-owned storage buckets (such as Cloudflare R2) during recording, generating share links immediately upon stopping, and later updating videos via the same link with edits including auto-zoom, smooth cursor, subtitles, and sensitive information blurring.
Toolbit is a local-first developer workspace web application integrating 43 tools in total, including JSON formatting, JWT decoding, Base64 encoding/decoding, regex testing, SQL/YAML/CSV/GraphQL formatting, UUID/hash/password/QR code generation, timestamp/color/unit conversion, text comparison, API/WebSocket request building, etc. All run in the browser without uploading data, with no tracking and no registration required. It supports PWA installation and offline use, under the MIT open-source license.
Brian2 Atlas is an open-source Rust engine that supports neural simulation and native training from local machines to distributed systems. It maintains the Brian2 modeling interface and maps model equations to CPU, GPU, WebAssembly, and MPI runtimes via a unified intermediate representation, AtlasIR.
a16z has released Aerie, an open-source scheme that uses lattice-based SNARKs to aggregate Falcon-512/Falcon-1024 signatures; a single Apple M4 Max laptop can aggregate 32768 Falcon-512 signatures within 0.65 seconds, with a proof size of approximately 66 KiB and a verification time of 54 milliseconds; the code is located at a16z/aerie, and the paper is also publicly available.
a16z crypto has released Aerie, an open-source solution that uses lattice-based SNARKs to aggregate Falcon lattice signatures; on a single core of an Apple M4 Max laptop, aggregating 32,768 Falcon-512 signatures takes only 0.65 seconds, with a verification time of 54 milliseconds, a proof size of approximately 66 KiB, and an average overhead of about 42 bytes per signature; it supports Falcon-512/Falcon-1024, the code is open-sourced at a16z/aerie, and the paper was published simultaneously.
Elorian released its first AI model, Glimpse, opening an API and a waitlist-based chat interface as a research preview; its core is the "visual scratchpad" mechanism, which supports dynamically cropping image regions, extracting video frames, or running code for measurement during reasoning, rather than merely providing a single textual description of the image; according to Elorian, the model has a smaller parameter count, was trained using only hundreds of GPUs, and has inference costs of $0.60/M tokens in and $3/M tokens out, outperforming GPT-6 Astra on scientific chart reasoning and short video captioning tasks.
DeepSeek officially upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017; the web interface and app have been synchronized with updates; the API still uses deepseek-flash for calls, with prices unchanged; the new version has achieved improvements on multiple benchmarks including HLE, Terminal-Bench 4.0, and DeepSWE v1.1, and has optimized task runtime and output length. It has completed adaptation on frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Upgrading to DeepSeek Harness 0.2.1-rc.1 or higher is recommended.
Meta's new Business Service Terms, effective October 30, require enterprise accounts to authorize Meta and its agents to use automated means such as crawlers and test credentials to access, index, cache, and analyze their associated domains and web page content, explicitly stating these can be used to 'provide, develop, improve, research, train' Meta products; although the terms do not mention AI, legal experts interpret the word 'train' and the scope of crawling as substantively covering AI training. Meta claims this move merely consolidates existing scattered terms without changing crawler practices, still respecting robots.txt, but the addition of 'test credentials' raises concerns about bypassing paywalls. Nieman Lab and multiple legal experts point out that these terms may constitute bundling, granting Meta crawling permissions far exceeding previous advertising-related scopes, leaving news publishers at risk of having their content freely harvested and facing reverse competition.
Termexo v0.10.14 has been officially released, supporting unified management of six AI coding agents on Windows: Claude Code, Codex, OpenCode, Pi Agent, Antigravity, and Grok Build; it enables one AI to operate other agent terminals via the local MCP protocol, providing workspace organization, native session restoration, task boards, end-to-end encrypted remote access (self-hosted relay/LAN/VPN), Git change tracking, and a multilingual interface; MIT licensed, no account required, all state stored locally; installation packages support EXE/MSI, and startup via npm run is also supported.
Canonry is an open-source, agent-first (CLI/MCP/API), self-hosted AEO (AI Engine Optimization) operations platform. It provides a Web UI and SQLite backend, supports BYOK; its core capabilities include AI visibility tracking, attribution of search and traffic changes, technical site health scans, and visibility analysis directly connected to models such as Claude, Gemini, and OpenAI. The project is hosted on GitHub under the MIT license.
yongkunchen released the Python command-line tool cykk-audio-transcriber, which supports audio-to-text conversion using the Whisper local model and the OpenAI Speech API. It can export TXT, SRT, VTT, TSV, and JSON formats, with optional DeepL translation integration. It requires Python 3.11+ and FFmpeg, is licensed under MIT, and has been published on GitHub, PyPI, and Zenodo.
Framefields is a high-performance, code-first video engine based on native WebGPU in Node.js. Targeted at AI agents and engineers, it provides dynamic graphics, Photoshop-level visual effects, 3D scenes, and audio digital signal processing through a single npm package; it has no Chromium or browser overhead, supports deterministic frame-by-frame rendering, and directly exports hardware-accelerated MP4.
MacVisor has released its Beta version, offering native support for Apple Silicon. It can run macOS, Linux, and Windows 11 (ARM, technical preview) virtual machines, providing snapshots, templates, advanced networking (including 802.1Q VLAN), CLI automation, and a desktop virtualization experience.
Shopify has launched a bulk order fulfillment feature that supports grouping up to 250 fulfillments by product type or shipping location. It provides a customizable three-step default workflow (Pick/Print/Pack), where each step can be added, removed, renamed, assigned actions, specify print documents, and allocate staff. Users can manually create batches on the Orders page or automatically process the latest unfulfilled orders in bulk.
Anima is an open-source (MIT licensed) Claude Code skill that supports generating cited, interactive, isometric-view biological mechanism diagrams from natural language descriptions, covering topics such as dopamine receptors, stimulant effects, GLP-1, estrogen, myelinated neurons, and phototropism; it provides the specific installation command npx skills add Garinmckayl/anima -g -a claude-code, and emphasizes that all numbers are annotated with species, conditions, and units, while unverified data is not displayed.
Scriber-Pro 2.0 is a macOS offline AI transcription app that adds the ability to directly extract M3U8 streams from web pages (such as online courses and webinars) and transcribe YouTube links locally; a 4.5-hour video can be transcribed in 3.5 minutes. It supports formats including MP3/WAV/MP4/MOV/M4A/FLAC. All processing is done locally with no data uploaded. Transcription results are stored in a searchable library synced via iCloud, with support for exporting to DOCX/TXT/Markdown/RTF.
Heldkeys is an open-source single-binary security tool used to detect credentials accessible to AI coding agents (such as Claude Code, Codex, Cursor, Copilot, etc.) within their runtime environment. This includes environment variables, cloud CLI login states (AWS/gcloud/Azure/kubeconfig), AI tool logins (Gemini CLI, Hugging Face), developer platform CLIs (Vercel, Supabase, Fly.io, etc.), .env files, SSH keys, and tokens in MCP server configurations.