Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
AWS released a tutorial introducing how to use Amazon Bedrock AgentCore Evaluations to assess the explainability and usability of multi-agent systems, proposing a three-tier evaluation framework.
Huawei and Qualcomm announced on October 5 that they have reached a multi-year, broad-scope patent licensing agreement covering areas such as 5G, computing, artificial intelligence, and networking. The deal includes an asset transaction where Qualcomm acquires some of Huawei's US patents. The agreement makes Qualcomm a net payer to Huawei for the first time, involving technologies such as Huawei's LogicFolding chip architecture. As of the end of 2025, Huawei holds approximately 165,000 valid patents globally, and the total value of its existing patent licensing agreements is expected to exceed USD 6.9 billion.
Brnch is a code hosting platform designed for current human-AI collaborative development models, supporting the coexistence of humans and coding agents in the same repository; it currently uses Git as its underlying technology, preserving existing toolchains; each agent has an independent identity and operates under user sponsorship, with actions traceable down to the line-of-code level; every merge is tested on the actual resulting state and accompanied by a publicly verifiable signed receipt; users can provision services for agents via brnch.io.
Aleph Alpha released the Kolibri model, featuring 78 billion parameters, employing a MoE architecture (activating approximately 3 billion parameters per token), with German data accounting for 21.3% of the training set, and supporting a million-token context window.
The Android screen mirroring tool scrcpy 5.0 has been officially released, introducing a new hardware decoding feature enabled by default. It uses D3D11VA on Windows, VideoToolbox on macOS, and VA-API on Linux.
Cohere has released North 2, a full-stack agent platform designed for industries with high compliance requirements. It supports private deployment (including air-gapped environments), SOC 2 Type 2 / ISO 27001 / ISO 42001 certifications, fine-grained permission management, pre-built enterprise connectors such as Slack, GitHub, and Jira, and efficient token consumption optimization based on NVIDIA Blackwell and Hopper.
zhugecaomao has released ALTRun, an open-source Windows launcher developed with AutoHotkey v2. With a size under 1 MB, it supports invocation via Alt+Space, pinyin initial matching, Everything file search, clipboard history, and text snippet expansion. It is portable, does not write to the registry, stores data locally, and offers Simplified Chinese and other multilingual interfaces along with 20 themes. The project is hosted on GitHub under the GPL-3.0 license.
A report released by MIT in June 2026 indicates that widespread AI use has led to decreased student attendance at office hours, reduced participation in online discussions, and shrinking study groups in dormitories and libraries; faculty find it difficult to assess genuine learning outcomes, and mutual trust between students and faculty has been undermined.
Minigraf is a single-file embedded graph database written in Rust, supporting bitemporal (transaction time + valid time) queries and the Datalog language. It is suitable for AI agents, mobile devices, and browser environments. It provides multi-language bindings for WASM, Python, Node.js, Android, iOS, etc., and can be enabled without configuration via Minigraf::open("data.graph"). The current version is 2.0.3. Known issue #371 affects write consistency for multi-valued attributes; a fix will be introduced in v3.0.0 with a new file format.
Google Search Console has launched a new feature allowing account holders to add YouTube, TikTok, X, and Instagram professional accounts as "platform properties" to track their performance in organic search. It supports two main views: Insights and Performance. The latter covers text/multimodal web search, image search, video search, and Discover visibility. Video search data is particularly useful for YouTube content, showing rankings and clicks for specific videos in Google Search. Instagram properties support a "multimodal search" filter to monitor whether product images appear in image search. While it does not include a brand search filter, manual filtering is supported. Data becomes visible within 36 hours after verification.
vLLM v0.31.0 has been officially released. Key updates include enabling FlashMLA mega attention and NVFP4 compressed KV cache by default for the DeepSeek-V4.1-Flash model on SM100.
Rill Browser is an AI-native browser built around the Claude Code and Codex tools already used by users. Pressing ⌘E on any web page converts the current page into an agent task, automatically matching projects, passing context, and executing in the background. It also supports categorizing tabs by topic, organizing browsing history by day, and aggregating all agent projects. It is free and requires no new subscription.
iOS 27 新增通知自动化功能，允许以任意 App 的系统通知为触发器运行快捷指令；作者谷雨G2rain原创设计 10 个可复用案例，包括重要联系人加重提醒。
Marv is a newly released screen-interactive agent tool. Users can press a hotkey to ask about any content on the screen; Marv responds via voice and directly guides operations by drawing arrows, circling buttons, and writing annotations on the screen. It supports tasks such as debugging code, deriving spreadsheet formulas, and locating video editor controls, all without switching from the current application.
Meta has launched the AI assistant Muse. This agent automatically generates and updates personal profiles for all of a user's contacts (family, partners, friends, colleagues, etc.) every hour. The profiles include fields such as address, occupation, important dates, past experiences, relationship foundations, interaction patterns, and maintenance suggestions. The profiles are built solely based on user-authorized data and public information, without fabrication. The system runs in a user-specific virtual machine and supports manual memory clearing and operation auditing.
TypeSafe AI released the decision model Jev on September 15. This model does not generate text but instead performs binary judgments, numerical scoring, or selection from limited options for inputs, positioning itself as a calibrated decision tool for automation scenarios.
OpenAI announced that GPT-5.5 will be completely discontinued across all plans in ChatGPT, ChatGPT Work, and Codex on October 14, 2026, with no grace period or legacy retention channel; the official migration paths are specified as GPT-5.6 Sol or GPT-6 Astra. This decision has sparked widespread concern among developers regarding Prompt Drift, Agent fallback circuit breaking, and rising costs.
Cloudflare refactored the memory representation of its 1.1.1.1 DNS cache, reducing per-entry memory usage from 953 bytes to 420 bytes and freeing approximately 100 TB of working set memory across the cluster; cache insertion throughput increased by 43%, and query latency decreased by 19%; the optimization is implemented in Rust, using a contiguous byte buffer storage format based on DNS wire format, with full production deployment completed between May 18 and July 6, 2026.
Under its 'Trillion Dollar Security' initiative, the Ethereum Foundation has proposed a native transaction assertions mechanism aimed at resolving issues where signatures are valid but execution results deviate from user intent. This mechanism allows for read-only logic checks of all net state changes (balances, storage, events, newly deployed contracts, etc.) after transaction execution but before submission; if preset rules are violated, the entire transaction is rolled back. EIP-7906 is the specific proposal currently being advanced, defining the POST_TX frame and three new EVM instructions: TXTRACE, TXDIFF, and EVENTDATACOPY. It is currently in the 'considering inclusion' stage and has not yet been confirmed for entry into an upgrade.
Apple has introduced the user-enabled Apple Reference Image mode on the main cameras of the iPhone 18 Pro and iPhone 18 Pro Max, utilizing secure boot signing at the sensor level.
Meta has open-sourced the Muse Gadgets project, releasing Apache 2.0 licensed ESP32 firmware and a Linux device SDK, with support including the ESP32-C5 DevKitC-1.
Arcadeia is an open-source, self-hosted media library application that supports browsing, searching, and managing personal photos and videos in a browser. It features dynamic video thumbnails with automatically extracted keyframes, drag-and-drop import, favorite marking, real-time transcoding playback, and network storage access. Deployment requires Docker; as public network deployment security has not yet been fully tested, it is recommended for use only on local or trusted networks.
AMD officially released ROCm 10.1, with key optimizations targeting data movement bottlenecks, including the hipFILE asynchronous Fast-Path backend (bypassing host memory staging), Batch I/O API, and Multi-Tier I/O statistics.
AstroHelm is a mobile application designed for astronomy enthusiasts. By capturing real-time images of the night sky with a smartphone camera and performing plate-solving, it enables sub-arcminute precision telescope pointing guidance without requiring internet access, GoTo devices, or specialized hardware. The app supports offline operation and integration via ASCOM Alpaca/INDI/LX200 protocols, offering free basic features and one-time paid upgrade options starting at $9.99.