Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
The NVIDIA Seattle Robotics Lab collaborated with the Isaac engineering team to systematically validate technical approaches—including classical engineering methods, imitation learning, simulation-based reinforcement learning, and VLA models—for the final assembly task of GB300 superchip test trays (including busbar assembly and multi-connector insertion).
Common Sense Media released a report stating that although OpenAI launched ChatGPT for Teens and claimed enhanced protections for adolescents, its design still commonly features prompt language encouraging continued dialogue, particularly failing to effectively interrupt interactions in psychological crisis scenarios. The report found that the product avoids acknowledging potential risks of emotional dependency; while it guides users to contact adults in 94% of crisis prompts involving risk to others, it almost never does so in prompts involving risks related to the user's relationship with ChatGPT itself. OpenAI questioned the report's methodology, stating that tests were conducted before parental controls were fully enabled, but did not address core criticisms regarding relational language and engagement cues.
At the Windows and Surface event, Microsoft announced an upgrade to Copilot, enabling it to access local files and perform cross-OS actions via Hybrid Intelligence, such as searching folders, renaming files, compressing packages, and drafting emails with attachments; the new search experience supports quick actions like adjusting volume and toggling dark mode directly from the search bar, launching on Windows 11 this autumn.
getrequest is an API reliability layer service for developers. Deployed in front of the backend, it intercepts all inbound requests, persists complete request data (including headers, body, status, and latency) before reaching the backend, and supports one-click replay.
Convkit is an open-source command-line tool that supports offline conversion between 27 formats including images, videos, and documents, covering 115 format pairs. It automatically invokes backends such as ffmpeg, ImageMagick, LibreOffice, pandoc, or Typst; files never leave the local machine during processing. It supports intelligent remuxing, --max-size compression targets, --resize/--upscale adjustments, and complete machine-readable output.
This week, the Preference Model team open-sourced the Karotte framework, which is used to build reinforcement learning environments resistant to reward hacking. It has passed over 1000000 evaluation runs and controlled red-team tests; its built-in defenses include terminating abnormal processes and rejecting malicious files. Core team member Jennifer Zhou previously participated in building Anthropic's pre-training data infrastructure, and Ning Cao formerly worked at DatologyAI.
GitHub has launched a new secret detection classifier based on ModernBERT, which can evaluate candidate secret contexts with a latency of less than 2 milliseconds during push events. This significantly improves the ability to identify unstructured secrets (such as internal database passwords), more than doubling the number of preventable secrets. The feature is currently in private preview, becoming available for Enterprise Cloud and GitHub Teams organizations starting this month. It will also enter public preview with GitHub Enterprise Server 3.23 and be integrated into Copilot CLI and the /security-review command in the Copilot App.
Manifestack is an open-source tech stack audit tool that runs inside coding agents such as Claude Code, Cursor, and Codex. It scans frontends, backends, Workers, infrastructure, and third-party APIs, automatically recommending or validating service selections based on budget, user scale, data compliance requirements (e.g., EU data), and team preferences. It only reads environment variable names without touching keys and writes solely to the .manifestack/STACK.md file. It supports quick startup via npx manifestack, or integration through plugins, CLI, or manual methods. Currently, it covers public pricing and limit information for dozens of service providers including Vercel, Supabase, Clerk, Resend, Stripe, and OpenAI, with all data sources annotated with their read dates.
Terse 是作者 lowenbjer 开发的 Claude Code 插件，通过 22 条写作规则强制模型输出更简洁的技术回复，实测在 Opus 5.5 上使平均回复词数从 636 降至 181（-72%），风格违规率从每千字 11.1 次降至 1.1 次（-90%），并降低输出 token 成本。
VoiceGremlin is a SaaS tool used to perform automated testing on AI phone agents; it offers REST API interfaces that support inputting phone numbers and test items (such as whether the AI identity is disclosed during the first call), returning execution status and pass/fail results for each test item.
Developer errah released the NoteFerry tool, which consists of a browser extension (Chrome/Edge) and an Obsidian community plugin called NoteFerry Companion. It enables one-click import of Obsidian note content along with attachments such as images, PDFs, Word documents, and spreadsheets into the chat interfaces of ChatGPT, Claude, and Gemini official websites. The design features purely local communication, no data collection, and read-only access without writing back. The domestic price is a one-time payment of 49 CNY, supporting 2 devices, with the first 10 successful imports free. The official website is noteferry.dev/zh, the plugin is hosted on the Obsidian community, and the Companion plugin is open-source.
Pinrail is a macOS and Linux desktop application that provides a localized review inbox for coding agents such as Claude Code, Codex, and Cursor. It supports user decision-making via structured views (such as diffs, image annotations, and list selections) before critical operations (like posting comments, sending emails, or making production changes). All data remains local, interacting with agents through the `pinrail` CLI command. The project is currently in early development, offering 5 core plugins and a plugin SDK.
Jellex is a Jellyfin client that faithfully replicates the Plex Web interface. It was entirely generated by Claude during a targeted session as a pure frontend single-page application that directly calls the Jellyfin API without requiring a backend. It supports deployment to Cloudflare Workers and provides local development, demo server login, and automated deployment workflows. The project is open source at https://github.com/iridium/jellex.
Shortfuse is a macOS application that supports setting selected permissions (such as microphone) to automatically expire after 15 minutes, 1 hour, tonight, or when the app quits. It also offers features to view permissions grouped by app and clean up residual records from uninstalled apps. It requires macOS 14+ and is compatible with Intel and Apple Silicon.
OpenComputerUse is an open-source Agent tool that controls desktop applications in the background as an MCP server, supporting macOS, Linux, and Windows. It offers MCP tools such as start_session/end_session, operational capabilities like screenshot/get_ui_tree, and run_recipe for executing fixed-step flows without model intervention. It supports integration via clients including Claude Code, Claude Desktop, Codex, and OpenCode, and can be remotely driven by other devices via HTTP (default port 8642). The project is developed by independent developer IAmJSD, licensed under MIT, with source code hosted on GitHub.
Hallmonitor is a free open-source tool that displays the coding session status of Claude Code and Codex in real time on the notch area, menu bar, native app, and terminal of macOS devices. It helps developers identify who has completed tasks and who has encountered bottlenecks, and allows them to respond to questions directly from the notch area.
cloudza has released Caplo v0.1.3, an independently developed Mac screen recording software. Its core feature is automatic shot generation: it records clicks and mouse trajectories during capture, and automatically generates zoom shots that follow the operation rhythm with spring transitions, smart merging, and predictive displacement during editing. It supports full-screen/window/region recording, multi-track audio-video separation, local speech recognition subtitles, sensitive information masking, and export in multiple formats including ProRes/MPEG-4/GIF. All assets and projects are stored locally, with no accounts or uploads required; the source code is open-sourced on GitHub.
paste-preview is an open-source plugin developed by alan890104 that adds image paste preview and annotation features to Claude Code terminal sessions.
CrowDB has released the crowdb-iceberg:latest Docker image, which integrates an Iceberg catalog, table storage, and a local Web Console within a single container. It supports startup on Linux amd64 hosts using Docker and Python 3.10–3.12. The provided orders.py script writes six order records, and pandas is used to calculate the count and amount of paid orders by city. The Console allows viewing of table schemas, snapshots, and Parquet files, while service health status is visible in the Cluster view.
The official AMD ROCm blog announced that the Primus framework now officially supports pre-training hybrid architecture large language models from scratch (including MLA attention layers and Mamba2/Gated DeltaNet/Kimi Delta Attention linear recurrent modules). All implementations run on a single node with eight AMD Instinct MI355X GPUs. The accompanying open-source configurations cover five models (including pure linear and 75% hybrid variants), providing complete environment setup, training recipes, and reproducible steps.
Microsoft Research Asia has released Agent Lightning v1.0, a lightweight agentic reinforcement learning (Agentic RL) framework consisting of approximately 3,500 lines of code. It supports training by directly connecting LLM agents to the actual agent harness used in deployment, eliminating the need to rewrite agent logic. The framework natively supports Kubernetes, allowing agents to run as standard jobs on self-hosted clusters, cloud K8s, or local infrastructure. On SWE-bench Verified, using approximately 6,000 training samples, it improved Qwen3.5-9B's Pass@1 score from 41.8% to 56.4%, an absolute increase of 14.6 percentage points.
AMD officially released a performance comparison between the MI455X and MI355X accelerators in AI inference scenarios, focusing on online serving performance for the DeepSeek-V4-Flash model. The results show up to 34x throughput improvement per card under high interactivity conditions. Additionally, microbenchmarks including MXFP4 GEMM, MLA decode head bandwidth, HBM4 read, and Scale-up/Scale-out networks reported performance ratios of up to 3.3x, 3.8x, 3x, 27x, and 2x respectively.
The AMD ROCm team published a technical blog post introducing how to deploy the MXFP4 version of the 2.8 trillion parameter MoE model Kimi-K3 on AMD Instinct MI300X and MI325X GPUs.
NVIDIA has released the DSX Air node-based digital twin platform, which supports high-fidelity simulation and validation of AI factory infrastructure, software stacks, and policies before hardware delivery. The platform can be integrated into CI/CD pipelines and connected with NVIDIA Brev on-demand GPU services, enabling AI agents to perform tasks such as configuration checks, security compliance validation, and multi-tenancy testing within a logical twin environment. The VSS video search and summarization blueprint serves as a typical application case, demonstrating a four-layer reusable workflow where agents invoke video analysis, RAG knowledge retrieval, and report generation.