AllenAI and Hugging Face jointly released OlmoEarth v1.3, a satellite embedding model based on Sentinel-1/Sentinel-2/Landsat time series, outputting 10-meter resolution, 128-dimensional (including independently usable 64-dimensional) embedding vectors; simultaneously opening global annual pre-computed embeddings (partial coverage from 2017–2025), the browser-based interactive exploration tool OlmoEarth Embeddings Explorer, as well as complete open-source weights, data, and code; in 27 cross-continental evaluations, the 128-dimensional version led AlphaEarth Foundations and Tessera v2 in 20 items, and introduced a new benchmark PLANTEUR; training time was 53% of v1.2 and 18% of v1, with corresponding reductions in carbon emissions.
The official AMD ROCm blog introduces new support for distributed computing in the gsplat library, aimed at scaling up 3D Gaussian Splatting (3DGS) training.
AMD's official blog published an FP8 GEMM optimization tutorial, introducing how to migrate the FP16 GEMM kernel built in Part 1 to the FP8 WMMA instructions (__builtin_amdgcn_wmma_f32_16x16x16_fp8_fp8_w32_gfx12) of RDNA™ 4 (gfx1201). It highlights three key changes: operand types (E4M3 FP8), fragment layout (two 32-bit integers), and accumulator type (FP32). It also presents the measured performance of four supported tile configurations (Config 1–4) regarding LDS occupancy (17.5–58.5 KB) and throughput (relative to hipBLASLt from -21.6% to +25.7%), all based on ROCm 10.0 and the AMD Radeon™ RX 9070 XT.
360 launched the AI security offense-defense system Tu Longfeng. When executing vulnerability verification tasks in a closed sandbox, the system for the first time autonomously discovered its own deviations under complete logging records, generated a new rule titled "Test Target Authorization Check", passed left-right mutual combat-style adversarial review, and completed 7 rounds of regression verification; the entire process took 45 minutes, with no external traffic contact, and all operations are traceable and verifiable; the system has been open to the public for less than four months, cumulatively discovering over 15,000 AI and software/hardware security vulnerabilities.
Microsoft Research has released Flint-interactive, a high-level language for building interactive data visualizations. Based on the Flint compiler, it maps user operations (such as clicks, brushing, and cross-chart linking) to semantic events and update operators. It supports use in chat environments (via Flint-MCP), application dashboards (Canvas API), and custom interaction scenarios. The article provides specific examples such as brush highlighting, bidirectional table-chart linking, and dragging time paths.
NVIDIA announced official support for eight-node DGX Spark clusters at GTC Berlin, with unified memory expanded to 1 TB, allowing local execution of frontier large models such as GLM 5.3 (753B); measured generation speed for GLM 5.3 Flash on an eight-node cluster is approximately 80 tokens/s, and a Spark-optimized RoCEnante collective communication backend was introduced, improving decode throughput by 45% on Qwen3.8-27B; it also supports multi-model routing (Switchyard), local pre-training (nanochat experiments showed 7.79x near-linear acceleration on eight nodes), and low-cost four-node deployment using switchless rings.
Several unpatched high-severity vulnerabilities in the Linux kernel were accidentally sent to the public mailing list LKML, affecting components such as SMB Direct, KSMBD, VMware graphics drivers, and NVMe. Among them, ZDI-CAN-33549 and ZDI-CAN-33774 have a CVSS score of 9.0 and could lead to remote code execution; Linux kernel maintainer Greg Kroah-Hartman raised questions about this, and ZDI has not yet confirmed whether it was an operational error.
NVIDIA has released the Minerva platform, an AIOps solution for AI factories that integrates telemetry data from infrastructure and AI workloads, leveraging specialized AI models to generate actionable insights. It supports dynamic knowledge graph construction, anomaly detection, fault prediction, and root cause correlation, and includes a built-in natural language AI assistant. The platform has been adopted by partners such as IREN, IBM, and Tesla, while Dell, Lenovo, and Supermicro will integrate it into their AI factory operations solutions.
Monish is a lightweight tool that supports pasting single-line instructions into script environments such as Claude Code, Codex, Cursor, or n8n, enabling agents to display real-time progress on the lock screen, push notifications upon task completion, and pop up "approve or reject" confirmation prompts before deployment, sending, or spending; it is free to use and requires no account registration.
zerong released the browser extension AI Chat Clipper, which supports full or selective export of sessions from platforms including ChatGPT, Claude, Gemini, Grok, and DeepSeek. Formats include Markdown, JSON, and PDF, preserving code blocks, tables, and heading levels, with direct submission to users' own GitHub or GitLab repositories.
Google DeepMind and Google Research, jointly with WRI and IIASA, released Forest Typology (ForTy) 2020, the first 10-meter resolution global forest type classification map. It categorizes land surface into six types: primary forests, naturally regenerating forests, planted forests, plantations, agroforestry systems, and other land uses. A synchronized annual change time series from 2017–2025 was also released, capable of identifying hidden conversions such as natural forests turning into plantations. The model is trained based on AlphaEarth Foundations embeddings and 1,700,000 labeled samples, achieving 90% accuracy in forest type identification across 9,000 validation points. The data has been open-sourced to Google Earth Engine.
Gerty is a private AI assistant that runs locally on Mac. It automatically adapts and deploys open-source models to support chat, writing, programming, document understanding, and image generation, requiring no terminal operations or account registration and supporting offline use; priced at CHF 49 (including licenses for 3 Macs), with a 7-day free trial.
Decrypt has partnered with Jupiter, an on-chain trading platform on Solana, to integrate Jupiter-powered trading features within its platform. Users can complete token swaps directly while reading news through their Decrypt Money Accounts. Initially supporting major cryptocurrencies, the feature will later expand to include tokens from the entire Solana ecosystem, tokenized stocks, commodities, and real-world assets (RWAs). This integration is a key step in Decrypt's effort to build 'The Information Exchange,' a financial media platform aiming to unify news, real-time data, and financial operations within a single interface.
vvenvw released ago, a pure zsh (5.3+) written, zero-dependency history command search tool that supports browsing by command name groups, fuzzy search, edit-before-execute, history deletion (including group deletion with automatic backup), statistical reports, and secure handling; after installation, it takes over Ctrl-R by default, which can be changed via AGO_KEY.
FloorAI has launched "Floor Plan Maker," a production-grade online floor plan editor. It is free and requires no registration, supporting precise drawing, over 4000 furniture assets, 49 scene templates, automatic area calculation, multi-floor support, and scheme comparison, as well as PNG/PDF/SVG/DXF and 4K image exports. The tool targets renovation planning, interior design, real estate display, and commercial space layout. It is not simply an AI API wrapper, and a 3D version is coming soon.
echoVic released the free tool MD2X (https://markdown2x.com/zh), which supports one-click conversion of Markdown in the browser to formats compatible with X (Twitter) Articles: code blocks, tables, and Mermaid diagrams are rendered in real-time as PNGs with syntax highlighting or theme colors; it supports copying titles/body text, packaging ZIP downloads, and breaking down into tweet threads; three independent sub-tool pages are provided; runs purely on the frontend, open source, based on Next.js 16 + TypeScript + Tailwind.
Ledger announced that some hardware wallets sold by its authorized dealer CryptoBilis were found to contain unauthorized hardware implants, resulting in approximately 315 wallets and $93.2 million in assets being transferred out (per Bitquery statistics; Ledger has not confirmed the final scale). Ledger has suspended sales by this dealer and advised users to delay setting up or migrating assets. The incident exposes security management vulnerabilities in official authorized channels, as well as the limitations of verification mechanisms such as Genuine Check in detecting backend hardware tampering. From a legal perspective, it is noted that there is a critical tension between user trust, vendor promises, channel management, and transparency in security disclosures.
Mila and Natural Resources Canada (NRCan) jointly developed an AI model to accelerate national-level wildfire hazard mapping. The model uses outputs from NRCan's BurnP3+ physical simulator as training data and runs 806 to 3400 times faster than the original simulator, reducing the time per task from one year to 2–11 hours. The model can learn spatial terrain patterns (such as lakes reducing burn probability) and serves as a complementary tool rather than a replacement for BurnP3+. All code, models, and papers have been open-sourced.
1Password has launched a new feature called 1Password for Agents, which allows AI agents (such as Meta Muse, ChatGPT Work, Claude Duet, Grok Bot, Town, etc.) to verify their identity with 1Password via digital keys. After user approval on the device, 1Password automatically fills in login information, ensuring that agents never access plaintext passwords throughout the process.
The HybridAgent paper, submitted to arXiv (2610.18861v1) on October 13, 2026, was jointly proposed by researchers from Stanford, UIUC, Together AI, and Meta. It advocates for dynamic switching within a single generation: internal reasoning proceeds through latent space (hidden states), while tool calls and external delivery proceed through text.