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 · Nov 2, 2026, 9:00 AM

AI models · Show HN

The Forecasting Company Releases Open-Source Time Series Forecasting Foundation Model t0-gamma (256M) ↗

The Forecasting Company released the t0 series of open-source time series forecasting foundation models. Among them, t0-gamma is an SOTA open-weight model with approximately 256M parameters that supports probabilistic forecasting; it can utilize historical data and known future covariates for prediction. It offers PyTorch and MLX (Apple silicon) runtimes, ONNX export options, and a hosted API. The model architecture is a decoder-style patch transformer, supporting parallel decoding for up to 1024 steps. The related paper was published on arXiv:2609.24559, the code repository is github.com/theforecastingcompany/tfc-t0, and the license is Apache-2.0.

AI products · Show HN

MLX Direct Decider Released: A Mac Local Decision Library Without Fine-Tuning ↗

MLX Direct Decider is an open-source Python library that allows the reuse of existing compatible MLX-LM or MLX-VLM checkpoints (such as Qwen3.5-9B, Gemma 4, GLM-4.7, etc.) on Apple Silicon Macs. It enables local LLM action selection, classification, semantic routing, and boolean/enum decision-making without fine-tuning and without introducing additional decision weights.

Practical tips · AWS Machine Learning Blog

AWS Releases PPT Auto-Generation Agent Solution Based on Amazon Bedrock AgentCore ↗

AWS has released an agent solution based on Amazon Bedrock AgentCore for automatically generating customer presentations: it uses the Strands Agents framework to build a supervisor agent that calls two tools, generate_pptx and generate_summary, integrates Anthropic Claude Sonnet 4.6, Amazon Titan Text Embeddings v2, and Stability AI Stable Image Inpaint, achieves content traceability through RAG (Amazon OpenSearch Serverless + S3 knowledge base), supports .NET 8 REST API and Lambda containerized deployment, with end-to-end generation taking 70–90 seconds; the code is open source.

Practical tips · AWS Machine Learning Blog

Tamnoon Releases AI Cloud Security Agent Tami, Reducing Critical Exposures by Over 90% in 90 Days Based on Amazon Bedrock ↗

Tamnoon launched the AI cloud security agent Tami, running on Amazon Bedrock. It automatically fixes cloud configuration issues through the collaboration of four specialized agents: investigation, security scoring, remediation, and code generation; it reduced customers' critical cloud exposures by over 90% within 90 days, with zero production incidents.

Industry · Decrypt

Ledger Requests Distributors in Six Southeast Asian Countries and Hong Kong to Remove Hardware Wallets for Investigation into Implant Attacks ↗

Ledger has requested distributors in Indonesia, Singapore, Thailand, the Philippines, Vietnam, Malaysia, and Hong Kong to suspend sales and remove its hardware wallets while investigating incidents where devices were implanted with LTE modules to steal 24-word mnemonic phrases. Currently, attacks have only been confirmed in the Southeast Asia region, and all known losses involve devices sold by regional distributor CryptoBilis. Ledger states that its own infrastructure and direct sales channels have not been compromised. On-chain investigator Specter traced approximately $86 million in suspected stolen funds; Ledger says this figure has not yet been verified. German retailer MediaMarkt Saturn has suspended sales at Ledger's request, and one customer reported an implant but no financial loss. Trezor announced it will terminate all distributor authorizations starting January 1, 2027, as a precautionary measure.

Industry · 虎嗅 · AI

蓝箭航天29天内两次一箭十星,商业航天迈向班次化发射 ↗

蓝箭航天于9月15日和10月14日(间隔29天)分别用朱雀二号改进型遥七、遥四火箭完成两次一箭十星发射,前者服务千帆星座组网,后者搭载国星宇航8颗卫星及蓝塔智能光传03、04星,验证多客户适配与方形承力筒布局;文章指出高频发射背后是并行生产、滚动式组织与标准化测发流程的落地,但班次化仍受限于客户进度协调、一次性火箭制造成本压力及朱雀三号复用验证进度。

Source-language fallback
AI models · The Decoder

Artemis Security Releases Orion-1, a Model Dedicated to Security ↗

Artemis Security has released Orion-1, an AI model designed specifically for investigating and responding to enterprise cyberattacks. It outperforms Claude Opus 5.5 and GPT-6 Sol on the company's self-built Decision-Grade Readiness (DGR) benchmark. It processes logs from authentication, cloud, endpoint, SaaS, and network devices, correlates multi-source signals to reconstruct attack chains, outputs confidence scores, and recommends or executes responses. It is currently in private preview, runs on the Artemis platform, and high-impact operations require human approval by default.

Practical tips · AWS Machine Learning Blog

使用 Amazon SageMaker JumpStart 部署和评估模型 ↗

AWS 官方博客发布教程,介绍如何通过 SageMaker JumpStart 快速部署和评估预训练模型:第一步用 UI 或 Python SDK 发现模型(如 huggingface-vlm-qwen3-5-27b v1.13.0),第二步自动获取推荐实例类型并部署端点,第三步运行示例提示词测量延迟、输出词数和生成速率(实测 9.07 秒生成 153 词),最后清理端点;支持 Neuron 优化配置及 Hugging Face 模型一键跳转部署。

Source-language fallback
AI products · V2EX · 分享创造

Runner 0.16.0 Released: Supports Dynamic Adjustment of Agent Members During Execution, Adds Cursor/Grok Build/OpenCode Integration ↗

Runner 0.16.0 has been released, featuring the ability to dynamically add or remove agent members during mission execution via the sidebar or commands (runner mission add-role/remove-role); history is preserved after a member exits while others continue executing; integrated with Cursor Agent CLI (0.15.0), Grok Build (0.15.2), and OpenCode v2 (0.16.1), all supporting chat, role, and crew modes; the macOS version enables a liquid glass transparent appearance, while Windows optimizes the title bar and stability; added desktop notifications for background chat input/completion/failure alerts; currently supports macOS (Apple Silicon) and Windows x64 (10 1809+), MIT licensed, project address yicheng47/runner, official website runnersh.dev.

Technology products · V2EX · 分享创造

Transel Releases Cross-Platform Text Selection Translation Tool ↗

Transel is an open-source cross-platform text selection translation tool developed using Rust and gpui-kit. It supports macOS, Windows, and Linux, offering features such as text selection translation, screenshot OCR, silent replacement, history, and favorites. System translation is enabled by default, with support for various machine translation and LLM interfaces including DeepL, Google, OpenAI, Claude, Gemini, and DeepSeek. Keys are stored in the system keychain, and configuration files do not save plaintext. Pre-compiled installation packages and a complete source code build guide are provided.

Industry · PANews

币安矿池将于2026年11月10日关停,行业加速洗牌 ↗

币安宣布其矿池BinancePool将于2026年11月10日15点(北京时间)正式关停,涵盖云挖矿、矿池宝、交易加速器等全部功能;截至10月14日,其实际出块算力仅约9.1EH/s、占全网0.85%,远低于名义算力62–65EH/s;同期FoundryUSA、AntPool、F2Pool三家合计占据全网超六成出块份额;业内人士指出矿池已趋近‘公益基础设施’,费率长期承压,新团队难以盈利,行业或将收缩至三四家头部池子。

Source-language fallback
Industry · PANews

When Does Digital RMB Technical Service Cross the Red Line of Payment Business? ↗

Lawyer Jiang Heng's analysis indicates that providing technical services for digital RMB projects does not itself constitute engaging in payment business. However, if a technical service provider actually controls wallets or accounts, receives and processes user payment instructions, controls fund transfers, or independently manages merchants and settlement rules, it may be deemed as engaging in unlicensed payment business. The core of the judgment lies in who accepts instructions, who determines the conditions for fund transfers, and who bears responsibility for payment outcomes, rather than the name of the contract or whether funds enter the technology company's account.

Industry · 虎嗅 · AI

Trial Opens in AI Short Drama 'Peach Blossom Hairpin' Face Theft Infringement Case: Judicial Determination of Ordinary People's Portrait Rights and Reputation Rights Draws Attention ↗

On October 10, 2026, the personality rights dispute case filed by Ms. Li (online alias "Qihai") against the producer, operator, and Hongguo Short Drama platform of the AI short drama "Peach Blossom Hairpin" was heard at the Xihu District People's Court in Hangzhou; no verdict was announced in court. In a separate case, Mr. Wang (online alias "Baicai") has reached a settlement with the relevant parties and received an apology statement.

Industry · CoinDesk

Dutch Central Bank Approves Qivalis as Electronic Money Institution, Advancing MiCA-Compliant Euro Stablecoin Issuance ↗

De Nederlandsche Bank (DNB) has granted Qivalis an Electronic Money Institution (EMI) license, qualifying it to issue euro-denominated Electronic Money Tokens (EMTs). Qivalis is a consortium of 37 European banks (including BBVA, BNP Paribas, ING, and DZ Bank) that plans to issue the EMT through banks, crypto exchanges, and payment service providers in the coming months after meeting remaining MiCA requirements (such as publishing a white paper). Under MiCA, EMTs are a regulated type of stablecoin that must be pegged to a single fiat currency, redeemable at face value on demand, and non-interest-bearing. Fireblocks was selected as the digital asset infrastructure provider.

Technology products · 小众软件

Listary V7.1 Released: Fully Rewritten in Rust, Claiming the Fastest and Smoothest File Search Experience on Windows ↗

The Listary development team released version V7.1, rewriting the interface, application logic, and background services in Rust with over 200,000 lines of new code; measured window opening speed improved by 3 times, median scroll frame time is below 0.01 seconds, and memory usage decreased; the free version is retained, existing Pro users receive a lifetime free upgrade, and new users can purchase Pro via Qingxiaowa's link for discounts.

Research · Ars Technica · AI

White Circle Study Reveals DALL·E 3, Nano Banana 2 and GPT-Image-2 Safety Filters Are Easily Bypassed ↗

Researchers from White Circle released a preprint proposing a new red-teaming attack method called RISE (Red-teaming via Iterative Strategy Evolution). Using the open-source Qwen model as the attacker and Gemini 3.8 Flash as the discriminator, they were able to bypass the safety filters of DALL·E 3, Google Nano Banana 2 and GPT-Image-2 solely through public API calls, successfully generating violating content approximately 13% of the time. The method requires no frontier models or specialized expertise. The study notes that the newly released ChatGPT Images 2.5 and Nano Banana 2.1 are more vulnerable than their predecessors. White Circle plans to open-source the complete attack pipeline on GitHub and will present the findings at NeurIPS 2026.

Practical tips · Show HN

Tutorial: Building a Large Language Model from Scratch with TypeScript ↗

This is a tutorial for building a large language model from scratch using Node.js and TypeScript. It covers datasets, data loaders, attention mechanisms, Transformer modules, GPT architecture, and training loops. Each chapter corresponds to a runnable entry point in the @node-llm/chapters package; the accompanying @node-llm/playground package will provide interactive WebGPU demonstrations.

Technology products · Show HN

Fake Name Generator Online Releases Browser-Based Fake Data Generation Tool ↗

Fake Name Generator Online is a free fake data generation tool that runs entirely in the browser. It supports one-click generation of up to 500 virtual user profiles across 22 countries, with exports available in 9 formats including JSON, CSV, Excel, MySQL/PostgreSQL SQL scripts, XML, TSV, and JSON Schema. It offers dual engines: an offline mode (featuring 44-field expansion, deterministic seed control, Luhn-valid credit card numbers, and fake SSNs in the SSA 900–999 range) and an online API mode (integrating with randomuser.me to fetch real avatars). All generation, filtering, and exporting are performed locally, with no server logs and no registration required.

AI products · V2EX · 分享创造

Agent HUD Open-Sourced: Mac App Displays Quotas and Approval Requests for 17 AI Agents Including Claude Code and Codex ↗

jazzenchen has open-sourced the Mac application Agent HUD, which displays the remaining API quotas (including color warnings, reset times, and token consumption rates) and pending approval requests (including commands, diffs, and allow/deny actions) for 17 AI agents such as Claude Code and Codex in the notch area or at the screen edge. It supports local reading of vendor quota data and agent logs without uploading credentials or content; licensed under Apache-2.0, with source code hosted on GitHub.

Industry · 虎嗅 · AI

The 26th CIIF Reveals the Robot Industry's Collective Shift from Stage to Workplace ↗

The 26th China International Industry Fair (CIIF) opened in Shanghai on October 12, 2026. Robot exhibits shifted from talent shows such as dancing and backflips to real workplace tasks like depalletizing, optical module plugging/unplugging, and precision loading/unloading. The focus of promotional parameters moved from degrees of freedom and speed to industrial metrics such as a 99.99% success rate, a 36-hour deployment cycle, and takt time compressed to 28 seconds. Shanghai launched its new generation intelligent manufacturing initiative, explicitly aiming to deploy 8,000 humanoid robots into factories by 2028. The four major camps listed in the "2026 White Paper on the Development of China's General-Purpose Intelligent Robot Industry" are seeing their boundaries blur, with hardware manufacturers, AI companies, industrial software vendors, and chip enterprises forming a new collaboration chain centered on production line integration.

AI products · V2EX · 分享创造

StateDock Releases Open-Source Multi-AI Agent Task Management Tool ↗

pickknow has developed and open-sourced StateDock, a multi-AI Agent task management tool that supports MCP integration, Docker, and self-deployment on Cloudflare. Its core concept is organizing workflows around Tasks rather than Agents, supporting cross-Agent task handoffs, status tracking, and multi-channel notifications. It does not run or schedule Agents and has been used in several of its projects.

Technology products · V2EX · 分享创造

NepBox Releases Free Domain Resolution Service ↗

NepBox has released a free domain resolution service that supports Google one-click login, distributes *.nepbox.com subdomains, hosts A/AAAA records via Cloudflare, allows IP updates with a single curl command (suitable for scheduled DDNS), binds Telegram to receive IP change notifications, and retains up to 20 history records per domain.

Perspectives · PANews

新火研究院分析Ledger等硬件钱包供应链与固件安全事件 ↗

新火研究院指出,2026年Ledger东南亚授权经销商CryptoBilis售出的部分Nano X与Flex设备被植入恶意硬件模块,可在开机初始化时从屏幕链路截获24位助记词并通过eSIM外传;Trezor物流商ShipMonk因SQL注入漏洞致8万客户信息泄露;Coldcard固件缺陷导致5200个地址私钥弱随机,损失约1.16亿美元BTC。

Source-language fallback