Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
Exodus is a free open-source tool that supports migrating local applications to Google Cloud Run via coding agents such as Claude Code, Codex, or Cursor; it inspects applications, compares current costs with Cloud Run costs, lists more than 50 real-world migration playbooks, and executes transformations item by item, or redesigns applications for Cloud Run.
Huawei is deploying an Atlas 950 SuperNode cluster composed of 256,000 Ascend 950DT cards, aiming to bridge the single-card performance gap through system-level interconnects (such as the Lingqu bus); concurrently, AMD acquired World Labs for $8.2 billion, and NVIDIA obtained a license for Groq LPU technology for $20 billion, reflecting US companies' strategy of understanding next-generation model computing needs through acquisitions; while domestic chips are currently primarily deployed in inference scenarios, training still relies on NVIDIA clusters, with large-scale training applications of Ascend SuperNodes expected to begin in early 2027; it faces triple constraints of tight production capacity, Model FLOPs Utilization (MFU), and electricity costs.
Razorpay has released SlashDocs, a self-hosted, open-source AI agent collaboration platform that supports agents directly generating and updating interactive HTML documents. Teams can comment directly on the same real-time document; once feedback is read by the agent, the file is updated in place. The project is compatible with the MCP protocol, is not tied to specific agents, and currently supports approximately half of Razorpay's internal document creation.
Microsoft has introduced an interactive canvas feature for Excel, enabling users to convert existing table data into real-time, auto-refreshing interactive reports using natural language instructions; charts and layouts update as source data changes. Additionally, two companion tools, Skill Builder and Skill Recorder, have been released, supporting the solidification of Excel operations into reusable skills via description or demonstration respectively. This feature is currently available for testing to Microsoft 365 Insiders.
Independent developer Joker2Man released Vixel v0.4.1, a fully offline, free open-source (MIT) local media search tool. It supports read-only indexing of photos, videos, audio, and design files in existing folders, enabling natural language queries like "sunset by the sea" and direct jumping to hit segments in videos; v0.4.1 adds global shortcuts (⇧⌘Space / Alt+Shift+Space), automatically pauses batch indexing during low battery/overheating, implements a more restrained file crawling strategy, and fixes multiple Windows issues; the EmbeddingGemma 2 model is built into the installer, supporting Apple Silicon and Windows x64, with an installer size of approximately 600 MB.
KasuganoSoras released the thefuckplus command error correction plugin, based on OpenAI-compatible large models. It supports generating fix suggestions based on the context of erroneous commands, exit codes, and files in the current directory, and provides natural language to command conversion functionality. The project is open source on GitHub, supporting zsh plugin installation and custom configuration.
AI labs are purchasing operational data such as codebases, Slack chat logs, and Notion documents from defunct startups through tech "funeral" service providers like SimpleClosure and Sunset; SimpleClosure's Asset Hub has facilitated over 150 transactions and is expected to return more than $10 million to startups by the end of 2026; Google won Spirit Airlines' bankruptcy data with a $10 million bid, including 100 million emails and 500 million Teams messages; the legal community questions whether this practice infringes on employee privacy, labeling it "legal radioactive" data, noting that no federal law currently prohibits such sales.
Sablier团队发布一项强化学习研究：一个从未见过真实价格的RL智能体，在纯合成市场（无日期、无代码、无新闻的再标准化价格路径）中训练后，在2020年和2022年真实股灾窗口的期权对冲任务上显著降低最差10%损失，误差减半；所有实验预注册，全部数字和轨迹可复现；使用Qwen3-1.7B模型，LoRA微调，训练耗时约13小时；代码与痕迹已开源，合成市场数据集可应要求提供。
QingYunA has open-sourced an Agent skill named answer-me-with-html, enabling agents that support CLI invocation such as Claude Code, Codex, and Cursor to automatically generate offline-browsable HTML pages (including SVG sequence diagrams, flowcharts, tables, code blocks, etc.) when answering complex questions. The model only needs to output a lightweight Markdown draft (averaging 612 tokens), while the CLI tool handles layout and style rendering; it supports always-on mode, explainer video generation, ASD-STE100 writing checks, multilingual support, and theme switching; the project is hosted on GitHub under the MIT license, requires no npm install, and provides one-click installation commands and complete documentation.
Gvmux is a native terminal application running on macOS, built on libghostty, offering features such as a vertical tab bar, split screens, session restoration, live configuration reloading, search and command palette, clipboard protection, notifications, and a dropdown quick terminal.
The parent company of Character.AI filed for bankruptcy protection on October 12, 2026, with Disney bidding $20 million as a poison pill buyer for some of its intellectual property; the company had received a $2.7 billion technology licensing fee from Google in August 2024, but subsequently experienced core talent attrition, multiple lawsuits related to user suicides, and regulatory actions by the FTC and the Pennsylvania Medical Board; bankruptcy filings show over 20 million monthly active users and only about 139,000 paid subscription users, with monthly operating costs of approximately $30 million; based on this, the article proposes three observations: dedicated models struggle against frontier general-purpose models, C-end user scale has become a liability in the AI era, and human-machine emotional interaction will evolve toward localization and offline deployment.
Seekter is an open-source job search agent designed for Claude Code. It automatically searches for jobs daily, filters them according to user rules, fills out application forms in Chrome, and records each application and skip as readable and writable Markdown files. It does not handle CAPTCHAs, log into LinkedIn, or submit unverified content; all personal data remains local. Running it requires a paid Claude plan or an Anthropic API account, and the project code is hosted on GitHub.
This skill plugin designs a structured reference point coding system (Qxx/Dxx/Fxx/Rxx/Sxx/Axx/Oxx) for Claude Code to mark issues, decisions, findings, risks, steps, actions, and options, generating numbered status updates at the end of each response; it injects rules via Node.js SessionStart hooks and maintains continuous numbering across context compression; it relies on a local Node environment and requires manual plugin activation; the design is inspired by IndyDevDan's MIT-licensed project.
Joakim Eriksson released an FM-1 synthesizer web emulator, fully reverse-engineering the pi32v2 dual-core CPU of the JieLi AC791N chip. It runs the original firmware (such as Felucca, M-VAVE official firmware, etc.) in the browser using WebAssembly JIT, supporting keyboard, touch screen, MIDI input, and real-time audio output; the source code is open-sourced at github.com/joakimeriksson/fm1-emu.
Veria Labs' AI security agent discovered a 64-bit integer overflow vulnerability in the BookStep processing logic of the XRPL payment engine within the rippled code. This could cause the XRPNotCreated supply check to fail, theoretically generating 18.45 trillion XRP; the vulnerability was not exploited on the public mainnet. The fixed version xrpld 3.4.1 has been released, patching the issue by adding overflow checks and adjusting counter logic, while also introducing an attack testing regression verification process.
EditUI is a Chrome extension that allows users to click any element on live websites or applications, leave brief instructions (such as abnormal spacing, overly prominent buttons, mobile display anomalies, etc.), and after continuous review, send the entire batch of annotations to Claude Code at once; it requires no code repository configuration, has no backend dependencies, needs no registration, and is free and open-source.
On October 11, according to a report by The Wall Street Journal, Amazon is in talks to acquire Israeli AI company Decart, with a potential deal price of about $7 billion. Decart's core product, DOS, is a compute 'translation layer' compatible with GPUs, TPUs, and Trainium, capable of boosting model execution speed up to 8 times the industry average; its early video tests showed that Trainium frame throughput was 4 times that of GPU solutions, with cost efficiency reaching 2 times.
Microsoft has launched a new version of the Copilot app, integrating complete creation, editing, and real-time collaboration features for Word, Excel, and PowerPoint. Users can directly handle files within Copilot, supporting AI rewriting, adding tables, and manual editing; a new Library section is added for unified management of files and AI content; all files are still stored in OneDrive, retaining formatting, comments, sharing permissions, and version history; users can specify an AI model for Office tasks or enable Auto mode.
Shita Robot has released a flexible photonic crystal electronic skin with a thickness of less than 1 mm and a sensing density exceeding 10,000 points/cm². It uses cholesteric liquid crystal elastomer (CLCE) materials to achieve pressure-induced color-changing optical sensing and integrates a flexible TFT photosensitive array. The device has completed 1,000,000 tests with an error rate of less than 5%. Simultaneously, it is advancing a 'tactile cerebellum' control model, which improves robot success rates by nearly 100% in tasks such as wiping blackboards and inserting plugs.
xAI于10月2日推出X订阅计划，整合Grok（支持聊天、语音、图像和视频生成）、Grok Bot（AI协作伙伴）、Cursor（本地与云端编码AI智能体）及X Premium（含认证徽章、创作者工具与无广告体验），提供从思考、创作、构建到分享的完整AI工作流；采用共享配额池机制，编码用量至少为独立Cursor方案的两倍，并在Cursor高级版中支持通过API调用第三方模型；定价分Free、Plus（8美元/月）、Premium（30美元/月）、Super（100美元/月）、Ultra（200美元/月）五档；面向团队与企业用户提供统一管理后台、单点登录、管理员控制、合并账单及SOC2合规支持；现有SuperGrok、Cursor或X Premium用户可保留原计划或迁移并享限时切换优惠。
Bifrost is a high-performance AI gateway that unifies access to over 23 service providers, including OpenAI, Anthropic, AWS Bedrock, and Google Vertex, through a single OpenAI-compatible API.
Tp 开发并发布 Pikko 1.0 版本，一款面向 MBTI 中 Perceiving 型人格及 ADHD 人群的精力管理 App；它用精力、设备、场景、心境四个维度替代传统任务属性，通过本地运行的自研精力算法（结合作息、Apple Watch 心率/睡眠数据）动态推荐最多3条适配当前状态的任务；支持小票打印机式任务合集、专注围栏、任务卡片存档与印章仪式感功能；基础功能免费，Pro 版一次性付费解锁高级功能；目前仅支持 iPhone。
Author testing found that Jev can low-cost screen factual errors in AI-generated content. In 32 tests, both Jev and GPT-5.5 detected 16 issues with no false positives; median time was 0.87 seconds (Jev) vs 4.10 seconds (GPT-5.5), with lower call costs. The author proposed using it for scenarios such as customer service records, meeting minutes, CRM commitments, RAG answer decomposition and verification, and knowledge graph relationship validation, emphasizing the need for manually provided candidate facts, while noting that labor-saving effects in production environments have not yet been verified.
Supernodes—servers consisting of multiple computing nodes connected via high-efficiency interconnect protocols with cross-node unified memory addressing capabilities—have become the hottest track in AI infrastructure for 2026. Huawei's Ascend 950 supernode has achieved large-scale commercial use, with a single set containing 1024 NPUs priced at over 100 million yuan; production capacity for 2027 is already fully booked. The industry is shifting from pursuing scale in card counts to optimizing Token throughput efficiency, with liquid cooling, optical modules/CPO interconnects, HBM bandwidth, and AI-driven cluster operations becoming key bottlenecks. IDC data shows that 94% of enterprises consider supernodes an important capability for AI infrastructure, but high initial investment, difficulty in software ecosystem adaptation, and the need for network upgrades remain major pain points. Currently, most domestic supernodes have not yet achieved profitability.