Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
A paper by Google researchers indicates that current recursive self-improvement methods, which automatically rewrite harnesses based on language models, tend to cause agents to memorize limited test tasks, thereby impairing their generalization ability to new tasks. To address this, they propose RRSI (Regularized Recursive Self-Improvement of Agent Harnesses).
DeepSeek Harness v0.2 preview was released on September 29, offering out-of-the-box installation packages for macOS and Windows desktops; a new plugin management page allows users to install, disable, or uninstall plugins by directly entering npm package names; the v0.2.1-alpha.1 version includes an experimental Claude Code Mods compatibility layer to verify whether its extension capabilities are a subset of DSH's 'everything is a plugin' architecture.
AI21 在其约 10,000 GPU 的共享 GKE 集群上，将 GPU 资源调度从内部消息频道手动协商升级为基于开源 Kueue 的自动化系统，实现了手动干预归零。
AirPlay Receiver for Android is a free, open-source application that turns Android TVs, Android TV boxes, or idle Android tablets into AirPlay receivers, supporting screen mirroring and audio/video casting from iPhone, iPad, and Mac. Features include H.264/H.265 decoding, AAC-ELD/AAC-LC/ALAC audio, HLS streams, picture-in-picture, PIN verification, and resolution/audio latency adjustment; DRM content is not supported. It is based on the open-source project UxPlay. Real-world testing by Qing Xiaowa showed good stability and synchronized audio and video.
Nightwatch is a free macOS menu bar application designed specifically for astrophotography enthusiasts. It predicts clear sky windows for tonight, sorts observable celestial objects, and provides setup timing alerts.
SCM is a local-first AI search tool for macOS that supports semantic search across every frame of photos and videos in any folder. It offers five search modes: Files (visual + text matching), Scenes (video segment-level localization), OCR (text recognition in 35 languages), Dialogue (precise script retrieval via Whisper), and LLMs (local Q&A using Qwen3 1.7B or Llama 3.2 3B). All models (CLIP/SigLIP/Whisper/Tesseract) run offline and are fully usable without internet access after initial weight downloads. Features include saving queries as tabs, smart folder monitoring, content hash deduplication, hot-switching between multiple models, and embedding version snapshots. Installation is via Homebrew, with the current version being SCM-0.2.4.
Starting October 1, 2026, Google has suspended its Open Source Software Vulnerability Reward Program (OSS VRP) and will no longer accept product vulnerability reports, as a large volume of AI-generated false vulnerability reports has overwhelmed maintenance teams; previously submitted reports remain valid, and vulnerabilities involving Google Cloud code repositories can be submitted through the Cloud Vulnerability Reward Program (Cloud VRP); Google plans to announce progress on OSS VRP optimization in Q1 2027.
QbitAI analysis points out that although GPT-6 Astra can generate structured 3D scenes such as buildings via Computer Use and Coding operations on Blender, it is significantly weaker than the specialized 3D model Meshy in tasks requiring high-detail reconstruction, such as characters and props.
Starting October 1, 2026, arXiv will implement new rules: each submission account may submit at most 2 papers per month, with no more than 3 manuscripts under review simultaneously; rejected papers also count against the quota and are attributed only to the account that clicked "Submit." This measure stems from 40,363 submissions received in September alone, with the cs.AI category growing more than sixfold over two years. The new rule is a temporary measure aimed at alleviating pressure on approximately 300 volunteer moderators facing low-quality AI-generated drafts, salami-sliced papers, and hallucinated content.
RobotUse is a robotic agent framework that selects targets and poses via vision, while the backend handles geometry, motion planning, and control. Sub-agents retain fine-grained interactions within sub-goals and return information required for subsequent decisions, supporting the continuous updating of persistent playbooks based on execution feedback. On RoboLab, it achieves a task success rate of 45%, which is 6.7 percentage points higher than CaP-X, and demonstrates the ability to learn from real-world executions under imperfect feedback and transfer this knowledge to subsequent tasks.
Kandinsky 6.0 Video is a set of diffusion foundation models for synchronized text/image-to-audio-video generation, including Lite (3B parameters) and Pro (29B parameters) versions. It supports 5-second Full-HD video and 44 kHz synchronized audio generation, employs a dual-stream CrossDiT architecture with cross-modal alignment mechanisms, and open-sources code, model weights, and diffusers integration.
Microsoft and Hugging Face jointly released the ThinkingBox evaluation framework and the ThinkingBox-Bench benchmark. This approach measures agent reliability by checking the final state of the backend database and side effects across 20 repeated tasks, rather than relying on tool calls or natural language responses.
Thalia is a free native Mac application designed specifically for Meta's Muse Code CLI. It supports Plan mode (read-only restart until user approval), per-command preview and approval, snapshot review with tamper-proof mechanisms, built-in checkpoints, MCP connectors, voice mode, and an embedded terminal. It uses the user's own Muse access credentials, sends no telemetry data, and only supports Apple Silicon Macs running macOS 27.
Linda is a free private AI agent tool for Mac that supports voice or text input. It performs tasks such as research, price comparison, web scraping, spreadsheet processing, file renaming, and email replies within an independent browser and local folders. Users can configure expert roles like Researcher or Bookkeeper. All operational steps are visible in real-time in the Computer panel and support takeover at any time. It runs AI locally based on Apple Silicon or uses user-provided API keys. It proactively asks for confirmation before sensitive operations such as sending, purchasing, deleting, or publishing. No account registration is required.
TechCrunch outlines over ten current AI agents that can be interacted with directly via SMS (iMessage/RCS/SMS), including Caddy (schedule and reminder management) and Fambot (household coordination).
photoc is an open-source command-line toolset designed specifically for photographers. It supports metadata reading, blur detection (focus), duplicate file finding (duplicates), template-based renaming (rename), sorting by time or session (sort), contact sheet generation (contact), compressed copy creation (compress), and stripping of privacy metadata such as GPS (scrub) for JPEG and Sony ARW files.
Search is a newly released open-source Mac browser with an application size of approximately 6.1MB. Based on the macOS WebKit engine, it supports features such as Chrome extensions, tab groups, ad blocking, and reading mode, requiring Apple silicon and macOS 14 or later.
DeepSeek's elastic computing team is actively recruiting senior engineers, as its self-developed sandbox infrastructure, DSec, has already supported the training and evaluation of the entire DeepSeek-V4 series. Currently, a single shard comprises 160 servers and 30,000 CPU cores.
Meta has launched the Muse Gadgets open-source project, offering open-source firmware and a Linux SDK that allow developers to integrate Muse AI agents with hardware such as Raspberry Pi and ESP32, as well as peripherals including displays, buttons, and sensors. Additionally, Meta has introduced the Muse Home Link USB-C device, which connects home networks and smart devices, and is distributing 5,000 units free of charge to Muse subscribers.
Stability AI has released three new audio models and AI music editing software, supporting the generation of complete instrumental pieces or short clips from text, with plans to introduce features guided by humming melodies or tapping rhythms; it raised $76 million in funding, with Sony, Warner, and Universal participating and licensing their catalogs for training.
MiniMax has released the M3.1-Flash model and launched its public beta. In frontend tasks such as animated videos and interactive HTML pages (e.g., cocktail-making processes, pixel-art wizard Canvas animations), the model's generated effects are comparable to Claude Opus 5.5, supporting unified color schemes, rhythm control, playback interactions, and step decomposition. Its capabilities stem from the native multimodal training of the M3 series, featuring a context window of 1M tokens and deep fusion of semantic spaces across text, images, and video.
The AWS official blog introduces a design pattern called Adjudicated Query, which builds a compliance review system with completeness proofs and traceability using Amazon Quick; this pattern translates natural language queries into fixed tool calls (such as sweep_compliance).
AWS has introduced a managed Web Search feature within Amazon Bedrock AgentCore, supporting integration with Claude Desktop via the MCP protocol; this capability is based on AWS's proprietary web index, keeps all query traffic within AWS infrastructure, and achieves enterprise-grade identity governance through full-chain authentication using AWS IAM Identity Center + Amazon Cognito + JWT.
AWS has introduced the Multi-Turn Reinforcement Learning (MTRL) feature in Amazon SageMaker AI, supporting end-to-end fine-tuning of search agents on models such as Qwen3.6-27B. This feature offers a modular interface, serverless execution, asynchronous trajectory collection, a native algorithm library, and resumable training. It achieves up to a +23.7% improvement in nDCG@10 and reduces the failure rate from 22.89% to 0.68% on benchmarks such as BrowseComp-Plus.