Own 1 is a fanless personal AI computer priced at $3899, with no subscription fees or per-token charges. It can run two large language models and eight agents locally. Equipped with the AI-native operating system OwnXOS, it assigns each agent an identity, budget, and audit credentials, and integrates an x402 USDC payment wallet. It connects to OwnXCloud only when authorized by the user. Pre-installed private AI colleague Nexy can orchestrate hundreds of agents. All data and memory remain local.
Karakeep v0.34.0 has been released, replacing the web scraping engine from Playwright with Patchright and integrating full Google Chrome to enhance Cloudflare verification bypass capabilities; new features include automatic offline downloading on mobile devices and saving complete lists for offline access; manual video saving via yt-dlp is now supported; expected server resident memory usage is reduced by approximately 40MB; the authentication system has migrated from NextAuth to Better Auth, requiring web users to log in again.
NASDrop, developed by littleweirdlab05, is an SPK package compatible with Synology DSM 7.1/7.2. It supports direct downloads from sharing links on GigaFile, GoFile, Pixeldrain, 1fichier, Buzzheavier, etc., featuring download queue control, specified folders, extraction passwords, waiting for and automatic extraction of split archives. A companion Chrome extension handles sites requiring browser verification. It is open-source and free, currently having only several hundred downloads.
InstantJSONFormatter is a 100% client-side JSON formatting tool that supports formatting, validation, minification, side-by-side comparison, and streaming parsing of files up to 10MB. All operations are performed in browser memory with no data uploads, no accounts, and no tracking. It provides native keyboard shortcuts like Ctrl+Enter and is compatible with dark mode and WCAG accessibility standards.
In an interview, Typesafe CEO Diogo Almeida officially released the new large model Jev, positioning it as "machine-native, system-first, large programmable model," with the goal of allowing code to directly invoke intelligence. It provides three types of programming primitives—choice, score, and newel—to support conditional judgment, ranking, and branch execution. The emphasis is placed on reliability calibration and prioritizing intelligence per dollar over deterministic output, while rejecting the embedding of safety alignment mechanisms at the model layer.
JSynth is a visual programming editor for Julia that supports connecting input, function, and chart nodes on a canvas, running in real time and displaying the values of each node. Designs can be exported as pure Julia scripts. It supports Windows, Linux (including WSL), and macOS (Apple Silicon), automatically installing Julia (via juliaup) and dependencies. It provides parameter panels, breakpoint debugging, grouping/frames, library management, multi-tab support, streaming data processing (folder monitoring, scheduled triggers, button events, etc.), and various examples. Built on Avalonia 12, it is MIT licensed, with source code hosted on GitHub.
vLLM v0.32.0 has been officially released, with core updates including: the addition of the structured decision interface /v1/systemone supporting choice/noul/score; low-precision KV caches such as NVFP4/FP8/ultraquant_4bit implemented across NVIDIA/ROCm/XPU platforms; Model Runner V2 becoming the default backend for ROCm; startup and service acceleration (e.g., shared-memory tensor arena, FlashInfer autotune preload); added support for models like EmbeddingGemma 2, MiniCPM-V 4.7, and Nemotron 3.5 ASR; and fixes for multiple GHSA security vulnerabilities along with tightened default policies for CORS, image formats, and JSON schema depth.
A programmer dad independently developed the English learning app "Ai Kaikou English" for his elementary school child. It is now available on iOS, domestic Android stores, and HarmonyOS. Features include learning vocabulary by textbook unit with memory curve review, graded video point-reading with original audio and word lookup, AI character voice dialogue with sentence-by-sentence pronunciation correction and Chinese assistance support. The dictation workflow involves standard phone pronunciation → child handwriting → photo capture → AI error recognition → automatic addition to the mistake notebook. The front-end, back-end, and operations dashboard were all self-developed by the author, who heavily utilized AI programming tools. Currently, the author's family uses it daily while continuing to iterate.
Sphinx is an open-source Git hook tool that can ask developers questions based on code diffs before pushing, verifying understanding of key logic (such as Stripe retry handling); it supports coding agents like Codex, Claude Code, and Cursor, runs locally without uploading .env files or secrets; install via npm install -g sphinx-understand, supporting sphinx install for model configuration and sphinx yolo for quick skipping.
BeatBrowser is an MIT open-source Chrome extension combined with a local Node bridge tool. It supports MCP agents such as Codex, Claude Code, and Cursor operating user-logged-in Chrome profiles via WebSocket. Each task runs in an independent window, supporting pause, takeover, cleanup, and local persistence of task state. It offers 24 MCP tools (including batch steps for 'act' and site notes for 'learnings'). Cloud agents can connect via Cloudflare Tunnel. Requirements include Node.js 20+ and Chrome 116+. Currently, the extension must be loaded manually and is not listed on the Chrome Web Store.
Payward announced that it has signed a final agreement to acquire Saldo Bank of Lithuania (whose Finnish parent company is Saldo Finance Oy). The transaction is expected to be completed in 2027, pending regulatory approval. Saldo Bank holds an EU banking license and operates in markets including Lithuania, Finland, Sweden, Germany, and the Netherlands, offering retail term deposits, consumer loans, SME credit, credit cards, and embedded financial services. Payward plans to leverage this acquisition to integrate banking deposit and loan services, payments, credit cards, and digital asset services, building a unified financial system.
Davflare supports pasting screenshots via Ctrl/Cmd+V in the file manager and image hosting settings page. It automatically uploads through the existing image hosting interface and generates links in the format https://<SITES_HOST>/i/{id}. Pasted images are also saved to the current directory. The 'Text Quick Transfer' feature in the top bar does not process screenshots to avoid conflicts. The project's open-source address is https://github.com/fanchenggang/Davflare.
lucasx3 developed and launched the backlink integration directory site everybacklink.com, which includes 509 sites accepting backlink submissions, categorized by AI tools, SaaS, communities, etc., of which 214 are free or offer free plans; it provides bot-accessible interfaces such as /llms.txt, /llms-full.txt, /api/sites.json, and individual site .md files; the data is open under the CC BY-NC 4.0 license, with pricing and dofollow/nofollow attributes for approximately half of the sites currently being verified.
Independent developer f16a18 released a compute simulator tool that supports dragging four sliders (energy supply, chip efficiency, parallelism, data quality) to simulate throughput, energy consumption, and cost curves in real time, with automatic bottleneck diagnosis; it offers scheme archiving and 10 skins; free use without registration is available, while login enables archiving and comments.
Utah has approved Norwegian startup Nolla Health to launch a pilot for AI-based first-time diagnosis and acne prescriptions at a monthly fee of $4.99; the pilot will proceed in three stages, requiring an agreement rate between AI and physician judgment of no less than 95%, zero serious adverse events, and written approval from the state government; the liable party is explicitly identified as Nolla Company, but its user service agreement has not been made public, whereas similar product Doctronic has already included explicit disclaimers in its agreement; at the federal level, AI has not yet been defined as a 'practitioner', and the '2025 Health Technology Act' has still not passed.
NVIDIA released the Nemotron 3.5 Bolt model; after structural compression and quantization-aware distillation, the 74000000000-parameter model's size was reduced to 22.1GB, supporting local execution on laptops equipped with RTX Spark (up to 128GB unified memory). The model has been integrated into Ollama and is compatible with agent frameworks such as OpenClaw and Hermes. It can perform tool calling, code modification, and multimodal understanding in terminal operation tasks (Terminal-Bench 2.1 score: 50.98%) and web browsing tasks (BrowseComp score: 66.75%). Accompanying RTX Spark laptops will be available globally on October 16, and the Nemotron 3.5 Bolt E4B 4GB version is scheduled for release on October 23.
On October 15, DeepMind published the paper "A team of Gemini 4 Argon agents beats Minecraft" on arXiv. It reports that 4 Gemini 4 Argon agents defeated the Ender Dragon in a random survival world in 27 hours and 14 minutes, dying 605 times, under conditions with no human division of labor and only shared chat channels and notes. Throughout the process, characters were controlled via code written using the Mineflayer library, generating 2783 programs; all operations were confirmed as non-cheating by an independent Observer agent and manual sampling audits.
The DeepMind team's AlphaProof Nexus system published a paper in Science, successfully proving 9 Erdős problems (including two sub-problems proposed in 1970 that had remained unsolved for 56 years) and 44 OEIS integer sequence conjectures. Its core is a multi-agent architecture where Gemini 3.1 Pro collaborates with the specialized prover AlphaProof, supporting draft sharing, review filtering, and sub-goal caching. The system also found new parameters for the Anchored GDA optimization algorithm and provided a rigorous proof of O(1/t) convergence.