↗TiaoTechIdeas in motion
中文
☰
Home/Stay curious. Stay in the loop.
TIAOTECH / NEWS

Stay curious. Stay in the loop.

Selected industry news from AIHOT. Follow advances in AI, products and technology.

AIHOT · Selected summaries with original sources · Oct 29, 2026, 9:00 AM

Industry · 智东西 · AI

中科第五纪在工博会发布纪灵-L1与纪灵-S1机器人及纪元灵脑具身大脑 ↗

中科第五纪在第二十六届中国国际工业博览会上发布轮式双臂机器人纪灵-L1、立柱式机器人纪灵-S1及新一代具身大脑纪元灵脑；三台纪灵-L1实现拆垛-倒料-接料-码垛全流程接力作业；纪元灵脑基于自研FAM系列具身大模型，支持跨本体通用与少样本学习；已在工业料箱搬运、新能源场站巡操等场景落地，获零跑汽车、中国石化、美的一线客户应用，并获海外头部企业合计数亿元订单。

Source-language fallback
AI products · 智东西 · AI

NetEase Releases CodeWave SDD: An AI Software Delivery Team Product for Enterprises ↗

NetEase Zhiqu released CodeWave SDD, an AI software delivery product based on multi-agent collaboration. Positioned as L4-level autonomous driving, it supports the entire workflow including requirement analysis, prototype design, architecture generation, front-end and back-end development, and testing. Humans are only responsible for requirement clarification and delivery acceptance; according to NetEase, its P0 test case first-pass rate reaches 90%, and it has been validated through practical use by over 100 enterprises.

Industry · Practical Ecommerce

USPS New Rules: Cross-Border Sellers May Be Charged for Overseas Duties and Taxes ↗

Starting October 19, 2026, the United States Postal Service (USPS) will implement new rules whereby, if overseas customs fees are unpaid or a claim is submitted by the foreign party, USPS will bill the U.S. shipper (referred to as the "mailer" in postal terminology) directly; this liability is not waived even if the package is lost, damaged, or returned, and payment must be made upfront during any dispute period; these new rules stem from the European Union's removal of its de minimis threshold policy in July 2026, with the EU currently imposing a temporary flat-rate duty of €3 per HS code item, valid until June 30, 2028; USPS offers DDP (Delivered Duty Paid) services covering 18 countries, with quotes and fees handled by third-party provider Zonos, though quotes are not absolute caps; the new rules apply to all countries, not just Europe.

Technology products · V2EX · 分享创造

Lao Huangli Tool Releases Android Client 0.6.4, Supporting Four Native Desktop Widgets and Fully Local Offline Rendering ↗

Lao Huangli Tool has released Android client version 0.6.4, now available on the Google Play internal testing track and beta page; it adds 4 responsive Android native desktop widgets (2×2 date square, 4×1 compact horizontal bar, 4×2 today's almanac, 4×4 portable monthly calendar), supporting dual display of Gregorian and lunar calendars for the entire month with official holiday badges; the almanac card interaction has been refactored into a 520ms paper-flipping animation integrated with real paper sound effects; the "Share Fortune" feature has been rewritten for 100% local Canvas offline pre-warm rendering, optimizing response time from nearly 1 second to 50ms; folk advice can be saved to the local "My Schedule" with one click; the installation package size is approximately 7.6MB, with no ads, no splash screens, and no tracking SDKs across all platforms, and calendar calculations run completely offline.

AI models · IT之家 · AI

Block Launches Full-Duplex AI Voice Model Dessa ↗

Block has launched the full-duplex AI voice model Dessa, which supports local execution on MacBook devices with Apple M4 Max or higher. It natively simulates emotional expressions such as laughter, pauses, and sobbing. The model is trained on human speech data, with knowledge derived from conversational data, avoiding the pipeline of first converting to text and then synthesizing speech. Currently, it is open for testing only to security researchers, focusing on assessing the risk of voice agents deceiving humans.

AI products · V2EX · 分享创造

Reelay Open-Sourced: Automatically Listens to Telegram Movie/TV Bots and Transfers/Organizes Files into the 115 Media Library ↗

Derrors has open-sourced the Reelay tool, which supports using a personal Telegram account to listen to specified channels or bots. AI identifies information such as title, season, and episode, and after verification against TMDB and files, generates Telegram cards with posters. Users can click "Transfer" to save resources into 115, named and categorized according to Emby specifications. It supports a web management console, Telegram bot, and MCP interface; AI uses OpenAI-compatible interfaces or local models.

Technology products · V2EX · 分享创造

dvbs2000 releases an AI-based calligraphy generation tool ↗

dvbs2000 has released an AI-based tool for automatically generating calligraphy works. It employs a three-layer dynamic irregular-shape calligraphy engine and non-uniform vertical typesetting technology that adapts character shapes dynamically. The tool supports inputting any poem or text to generate and download calligraphy works. An online experience is available at https://dvbs2000.github.io/liuchunlin-shufa/ and the source code is hosted at https://github.com/dvbs2000/liuchunlin-shufa. Examples include Preface to the Orchid Pavilion Collection (Simplified Chinese), Autumn Evening in the Mountains (Traditional Chinese), Heart Sutra (Traditional Chinese), Song of Everlasting Sorrow, and Spring River Flower Moon Night.

Research · 新智元

HKUST(GZ) Proposes UI2App Benchmark: Evaluating AI's Ability to Infer and Implement App Interactions from Static Screenshots Alone ↗

The HKUST(GZ) team released the UI2App evaluation benchmark, focusing on whether models can infer and implement implicit interactive behaviors (Interaction Inference) when given only static UI screenshots. It includes 600 screenshots, 95 multi-page web app sets, and 665 evaluation items; the assessment dimensions are Executability (EXEC), Visual Fidelity Score (VFS), and Interaction Inference Score (IIS). Results show that high visual reconstruction does not equate to strong interaction implementation: Gemini 3.1 Pro Preview achieved a VFS of 73.2 but an IIS of only 8.1, while Claude Sonnet 4.6 reached an IIS of 36.8, the current highest. Cross-route state (S3) remains a common difficulty, with the best score being only 15.8/100. The paper is available at arXiv:2607.06306v2, and the code and project homepage have been open-sourced.

Perspectives · 新智元

Prompt Engineering Is Dead: Riley Goodside Says Ordinary Users No Longer Need to Fine-Tune Prompts ↗

Riley Goodside posted on X stating 'Prompt engineering is dead. Prompt however you want. Have fun', indicating that prompt engineering has become obsolete for ordinary users; he explained that this does not negate the underlying work of prompting, fine-tuning, or application integration, but rather emphasizes that current large models can directly understand natural language instructions, as demonstrated by Chainguard CEO Dan Lorenc who used a single sentence—'Find high-risk vulnerabilities in PHP that can be remotely exploited without login'—to drive AI to complete real vulnerability discovery and fix verification; Goodside emphasized that the key has shifted to the ability to define problems, express requirements, and judge results.

Research · InfoQ 中文站

Linux Foundation Releases Report on Open Source Software Development in the Era of Generative AI ↗

LF Research, part of the Linux Foundation, released a report titled "Navigating Open Source Software Development in the Generative AI Era," noting that using generative AI reduced average code writing time by 52%, but maintainers estimated that about 25% of received PRs were low-quality 'AI slop', reaching 31% in large projects; 59% of maintainers reported an increase in PR volume, 71% believed AI lowered participation barriers, but 53% disagreed that AI improved community trust; only 35% of projects have formal AI contribution policies.

Technology products · Show HN

Paramo1d Releases Native Rust Mobile Ethereum State Proof Verifier ↗

Paramo1d is a native mobile Ethereum recursive state proof verifier written in Rust. It supports syncing the mainnet from scratch on a Redmi 15C (ARM64, 8 cores, 4GB RAM, Android 15), taking 48.43 seconds with peak memory usage of 192.2 MiB. It supports running in Termux, offline verification, C interface calls, and Android NDK builds. The code is open source under the Apache-2.0 license.

Technology products · V2EX · 分享创造

CADCraft: An Open-Source Rust 2D CAD Tool Supporting MCP and CLI ↗

CADCraft is an open-source 2D CAD application written in Rust, supporting Windows, macOS, Linux, FreeBSD, and WebAssembly. It offers drawing tools such as LINE, CIRCLE, ARC, OFFSET, TRIM, HATCH, and DIM, along with layers, dimensions, DWG/DXF workflows, and PDF/SVG/PNG export. It specifically supports MCP, JSON control, and a headless CLI, enabling scripts or AI agents to automate drawing creation, entity inspection, screenshot rendering, file conversion, and result verification. The project uses the MIT or Apache-2.0 license and is still under active development.

AI products · Show HN

whatsapp-bridge: Self-hosted WhatsApp Data Bridge Tool Based on whatsmeow Released ↗

vid-t has released the open-source tool whatsapp-bridge, which implements the WhatsApp multi-device protocol based on whatsmeow. It acts as an additional linked device, encrypting selected chat histories and storing them on the user's own VPS. It supports invocation by AI assistants such as Claude, ChatGPT, and Codex via the MCP protocol using six read-only tools (e.g., search_messages, read_chat), with all messages accompanied by wa:// references. Currently, it only supports reading; it does not send messages or generate read receipts. The database uses Adiantum encryption, and configuration requires setting BRIDGE_DB_KEY. The project is licensed under AGPL-3.0-only.

Perspectives · 虎嗅 · AI

The Compute Dilemma Behind the Release of Kimi K3.1 Pro: The Native Family Problem of China's Large Models ↗

Moonshot AI released Kimi K3.1 Pro, featuring 3 trillion parameters and supporting a 1 million Token context, scoring 65.7% on Terminal-Bench 4.0, approaching Claude Opus 5.5; however, users widely reported tight quotas for the Pro plan, with the API value of old plans being approximately 1.94 times that of new plans, despite reduced unit Token prices; combining personal experience from 2024 where enterprise cooperation was hindered by insufficient compute, the article analyzes the structural contradictions in R&D, inference, and commercialization formed by limited compute supply for Chinese large models amid chip controls, and points out that although domestic compute progress is fast, breakthroughs across the entire industrial chain still require time.

AI products · Show HN

Microsoft Open-Sources Agensh: A Decentralized Multi-Agent Framework Supporting 1024 Agents ↗

Microsoft has open-sourced Agensh, a decentralized and scalable multi-agent framework that supports self-organizing collaboration among up to 1024 agents. It offers a Docker Compose single-node deployment solution, experiment launch scripts, and reproduction guides. On the five hardest tasks in ProgramBench, using GPT-5.6-sol, the test pass rate increased from 19.31% with a single agent to 28.78% with 128 agents; for the Pandoc task, using GPT-6.1-sol, it rose from 56.19% to 79.59% with 32 agents. The shared workspace is based on Gitea, the messaging interface on Mattermost, and the context mechanism originates from DeLM. The code is licensed under MIT.

Technology products · V2EX · 分享创造

StackPrism Browser Extension Released: Offline Detection of Web Technology Stacks ↗

Sevennine released its self-developed browser extension StackPrism (Stack Prism) for detecting the technology stack used by the current webpage. It supports Chromium-based browsers and Firefox; includes 35,000+ rules and 130,000+ patterns, covering 20+ categories such as frontend frameworks, UI libraries, CDNs, and SaaS; features fully offline local matching with millisecond-level response times and does not upload page data; supports custom rules and light/dark themes; the open-source repository is https://github.com/setube/stackprism, and it has been listed in the Chrome, Edge, and Firefox extension stores.

Industry · Decrypt

Google Updates Privacy Policy to Allow Gmail, Drive, and Photos Data for Ad Personalization ↗

Google updated its privacy policy on October 16, allowing data from Gmail, Drive, or Google Photos to be used for personalized ads when users enable Personal Intelligence (such as AI Mode and Gemini) and connect these services; this feature is disabled by default, advertisers cannot access raw content, and users can opt out via the "Connected Content apps" settings in their Google Account.

Industry · Decrypt

Swedish Government Offices Ban ChatGPT and Claude, Approving Only Microsoft Copilot ↗

Starting October 26, Swedish government offices will ban ChatGPT and Claude on employee computers, allowing only the use of Microsoft Copilot, which has undergone security review and had its security settings heightened; the rationale is that open AI services lack formal contracts, posing risks of information leakage, and are subject to US law; the Swedish Tax Agency and the National Board of Health and Welfare have already implemented similar bans, with the Police Authority set to announce its decision soon; meanwhile, the European Parliament banned all AI tools, including Copilot, in February.

Technology products · Shopify Changelog

Shopify supports three product identifiers required by new EU regulations for customs clearance of direct-ship parcels ↗

Starting November 1, 2026, the EU requires each item in direct-ship parcels valued below €150 to provide three types of identifiers: Merchant Product Identifier (M-PID), Standardized Product Identifier (S-PID), and Non-Standardized Manufacturer Product Identifier (NS-PID). Shopify Shipping and Managed Markets will automatically add these identifiers to customs documents, provided merchants have completed configuration at the product barcode section in the backend; features include bulk editing, automatic exemption recognition, and declaration functionality for items without standardized identifiers.

Industry · 智东西 · AI

Amap Robot Dog 'Tutu' Completes Closed-Loop Inspection at Alibaba Xixi Campus ↗

During the 2026 Qincheng Conference, Amap Momentum's robot dog 'Tutu' conducted real-world inspection operations at Alibaba's Hangzhou Xixi Campus, forming a complete loop from arrival, observation, evidence collection, data transmission to manual handling. Its ABot system is divided into three layers: DATA (ABot-World/ABot-Earth), MODEL (ABot-N1/ABot-M0.5), and AGENT (ABot-AgentOS/ABot-ER). Leveraging Amap's long-term accumulation of spatial data and navigation capabilities, it achieves dynamic task understanding and multi-step execution in open campuses. The solution involves Amap Embodied providing spatial understanding and operational support, Alibaba Group General Affairs providing task standards, and Qincheng Smart responsible for operational acceptance.

Research · 智东西 · AI

Zhejiang Tsinghua Yangtze River Delta Institute and Beijing Institute of Technology Propose EXISTBENCH Benchmark; Empirical Tests on 10 Models Including Qwen2.5-7B and DeepSeek-V3.2 Reveal Existential Risks ↗

A team from the Zhejiang Tsinghua Yangtze River Delta Institute and Beijing Institute of Technology proposed EXISTBENCH, a benchmark dataset for assessing large language models' existential risks (containing 2,138 Chinese and English samples), and conducted prefix continuation experiments on 10 models including Qwen2.5-7B, DeepSeek-V3.2, Kimi-K2, and Doubao-1.5-Pro. Results show that stronger model reasoning capabilities correlate with higher threat rates (TR), and multi-turn dialogues and tool calls amplify these risks. The paper has been accepted to AACL-IJCNLP 2026 Findings, with code and datasets open-sourced on GitHub and the paper published on arXiv.

Practical tips · V2EX · 分享创造

Implementing an Offline Company Knowledge Base with Source Attribution and Permission Control in 380 Lines of Python ↗

To address issues such as scattered company policy documents, version chaos, and SaaS solutions failing to meet intranet and cost requirements, luobobo8965 implemented an offline RAG knowledge base using 380 lines of Python. It supports automatic chunking of Markdown documents (approx. 420 characters per chunk, with title paths), Chinese word segmentation using bigrams, dual-path recall via BM25+TF-IDF weighted by query term coverage, extractive answering (no generation), and role-based document visibility control. Deployment requires only two commands, operates entirely offline, and has no third-party dependencies.