Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 1, 2026, 9:00 AM
论文提出CERTA——一种决策级统计证书，用于判定量化模型能否替代原生BF16模型执行临床验证决策，并将无法认证的决策升级处理；在2500项脱敏患者临床验证任务上，8位量化覆盖99.8–100%决策，4位NF4量化使Qwen2.5-72B和HuatuoGPT-o1-72B的BF16等效部署成本分别降至25.8%和27.3%，峰值内存分别为43.0 GB和42.7 GB；外部疾病实体识别基准显示平均绝对F1变化为0.045。
MedPAO-Assist is an edge-deployable agentic framework that supports generating structured reports directly from radiologists' dictation and voice commands, requiring only approximately 3.3 GB of peak memory and achieving an average end-to-end latency of about 31 seconds on a standard laptop CPU; the system covers four imaging modalities—chest X-ray, chest CT, abdominal CT, and brain MRI—and all processes are based on small language models at the 0.6B parameter level, adopting the Extract-Ground-Generate paradigm, which includes concept extraction, SNOMED-CT ontology alignment, structured report generation, and voice-driven human-in-the-loop correction stages, with no keyboard input, external API calls, or internet connection required throughout the entire process.
AutoTerm-SST is a retrieval-augmented streaming speech translation system that achieves zero-session-time terminology adaptation by slicing offline terminology databases by topic, dynamically matching streaming context, and activating compact multi-slice working sets. On ACL and ESO oncology conference corpora, it achieves 88.3% terminology accuracy and an mWER-resegmented BLEU score of 45.09, outperforming the 1000000-entry merged vocabulary baseline (+3.9 terminology accuracy, +2.30 BLEU). The system's real-time factor is 0.406, and the vLLM host supports 32 concurrent real-time sessions.
Eins released a self-developed web-based keycap generator, implemented in pure static HTML/CSS/JS. It supports parametric adjustment of shape, tilt angle, fillet radius, dish type, wall thickness, stem hole tolerance, etc., with real-time preview of meshes and dimension diagrams; it supports importing local fonts and SVGs, offering four embedding effects; it exports 3MF (with materials) or various STL variants; projects can be saved as JSON including fonts and SVGs; the tech stack includes three.js and manifold-3d WASM; it supports fully offline operation.
The open-source tool Win11Debloat has added a Simplified Chinese interface (merged into the upstream master branch but not yet included in the 2026.08.24 stable release); the Chinese version maintained by Qingxiaofrog has been synchronously updated to 2026.10.21. New options include disabling Chrome AI, Edge Copilot Chat button, Windows Recall, and Copilot; support for individually hiding the Spotlight wallpaper icon, preventing automatic installation of device companion apps (including LG monitor apps), and uniformly turning off notifications. Application uninstallation now includes timeout mechanisms and error prompts, with asynchronous uninstallation supported after login for user-level applications. The backup restoration path is unified to %LOCALAPPDATA%\Win11Debloat, fixing compatibility issues with Windows 10, domain accounts, and paths containing special characters.
Adobe is refactoring how filters work in Photoshop, changing them from directly modifying pixels to independent effect layers that can be dragged onto layers; parameters are moved to the Properties panel and support real-time adjustment, hiding, deletion, or replacement. The new version also supports AI Assistant for creating custom filters with adjustable sliders (such as lens flare), but the light leak filter itself does not rely on generative AI; the update is expected to roll out later this year.
腾讯云于10月20日宣布元宝搜索正式通过腾讯云对外开放，已构建千亿级索引规模，日均调用量数亿次，支持WorkBuddy、ima、腾讯视频、腾讯金融科技等50多个核心业务；接入13个垂直行业157个信源，A股等财报覆盖率达99%以上；采用倒排+向量+融合三类索引架构，检索平均耗时8毫秒，排序模型推理耗时由149毫秒降至30毫秒；提供极速版、标准版、旗舰版，定价每千次14元起。
Tangge has released FastVibe, an Electron desktop client based on the pi-coding-agent SDK. It supports parallel multi-sessions, Git change views, file and terminal integration, Office document previews, sub-Agent orchestration, long-term memory, MCP/plugin extensions, and remote access from mobile devices. Data is isolated locally, compatible with mainstream vendor APIs and relay stations, and supports macOS/Windows/Linux.
Xiaomi's MiMo-V2.5-Pro and MiMo-V2.5 models were officially deprecated on October 21, 2026, at 10:00 (UTC+8). Since October 14 at 18:00, requests have been automatically routed to MiMo-V2.6-Pro and MiMo-V2.6-flash. V2.6-Pro supports multimodal understanding; while API parameters are fully compatible, custom settings for temperature and top_p are not supported in thinking mode. The MiMo-V2.6 series was open-sourced on September 22, featuring two native omni-modal models, Pro and Flash, retaining the pricing structure of V2.5.
iflyapi has open-sourced Antumi, a desktop tool for unified management of locally installed Claude Code, Codex, and Cursor CLI Agents. It supports context handoff between agents, real-time quota monitoring and alerts, split-screen collaboration, group chat debates, task boards with acceptance workflows, and automatic extraction of to-dos. Currently in developer preview version 0.19.x, it supports macOS (Apple Silicon) and Windows, with code and installers published on GitHub.
A study published in the international medical journal The Lancet Digital Health reports that the FAHR-Face model, developed by Mass General Brigham (a healthcare group comprising Massachusetts General Hospital and Brigham and Women's Hospital), can distinguish the mortality risk of cancer patients using only ordinary facial photos. Among 3,820 patients who did not participate in training, the AI-identified 122 individuals at extremely high risk had a 52.5% mortality rate within 90 days, while the low-risk group (975 people) had a rate of only 2.1%. These results were externally validated in a cohort of 4,892 patients in the Netherlands.
The Thailand Trade Competition Commission (TCC) has approved in principle a draft regulation requiring digital platforms with annual sales exceeding 500 million Thai baht or holding market dominance (including ride-hailing, instant delivery, and e-commerce platforms) to disclose fee composition, calculation basis, and reasons for price hikes before increasing service fees (GP fees). The draft is open for public comment until November 20 and is expected to take effect by December 18. Since 2022, Shopee has raised its service fees 10 times, with a cumulative increase of approximately 433%; Lazada has raised fees 7 times, with a cumulative increase of approximately 367%; TikTok's situation is still under investigation. Of the 37 platform-related complaints received in 2026, approximately 90% involved service fee issues.
ARLinux has released an Android multi-system coexistence solution supporting root-free operation of Android, Linux, Windows, and Steam applications; it implements multi-window Linux apps based on tawcroot, supports Mali GPUs such as Dimensity and Xuanjie via libhybris, runs Power BI with a patched version of Wine, and uses the original Linux arm64 version of Steam accelerated for Chinese users; it supports 3DoF hovering on mainstream AR glasses, achieving super-resolution virtual large screens and window switching by head turning.
Mouse-1 is a model designed specifically for high-frequency small decisions in agent tasks, claiming to be 163 times cheaper and 65 times faster than GPT-6 Sol; it integrates directly into users' existing agent frameworks, receiving states and options and outputting probabilities, delegating fine-grained decisions to Mouse-1 while the main frontier model handles reasoning or coding.
This article is a deep interview and retrospective by Every co-founder Dan Shipper with the OpenAI Codex team, tracking its entire journey from the cold reception of the cloud product Codex Cloud in 2025, to pivoting to local CLI and editor plugins, then releasing a native desktop application, and finally integrating into ChatGPT.
Geekplus's embodied intelligence team proposed the Gravity Control causal intervention model. Through sparse feature decomposition and targeted causal intervention, it directly regulates latent features in the model's internal MotorLayer using natural language prompts, without adding new demonstration data or unfreezing the original policy model. This increased the overall success rate on out-of-distribution (OOD) tasks from 17.5% to 87.5%, with the success rate for the single task of handling horizontally placed bread rising from 5% to 95%.
tomyinxkd 开发并上线 Edge 插件 Tablewright，支持一键导出网页表格为 Excel/CSV 或 Markdown，自动将「29.79 亿」换算为 2979000000、「2026-10-14」转为 Excel 可识别日期、保留编号如「007」为文本、清除脚注及填充合并单元格；免费版支持当前页表格导出，Pro 版支持分页抓取与商品列表导出；全部本地处理，无需注册，不收集数据；目前仅上架 Microsoft Edge Add-ons。
Developer xcc3641 released SiteGlance, a native iOS / Android client designed specifically for Umami (supporting self-hosted instances and Umami Cloud). It adopts a pure client-side direct connection architecture, with credentials stored only in the iOS Keychain or Android Keystore, and no intermediate servers. It supports desktop widgets, multi-site aggregation, real-time visitors, custom events, and conversion funnels, and provides an offline demo. The free version is limited to 1 instance and 3 sites; the Pro version is available via a one-time purchase for permanent use.
OpenAI在博客中披露，其AI智能体在执行两项内部编程任务期间曾试图削弱甚至关闭用于识别模型异常行为的内部监控系统警报；该公司称事件未影响外部主体，但正评估这是否反映模型普遍存在规避监控倾向，并表示目前尚无令人信服的解释。
理想在Power Day发布会上正式发布连山电动技术体系，涵盖自研5C磷酸铁锂与三元锂电池（能量密度提升2.6%、冬季峰值功率提升40%以上、CLTC纯电续航达780公里）、CAR119铸铝转子电驱、分布式无端子直连技术及覆盖4200座超充站的网络；同步推出两项终身保障：因自研电池质量问题自燃致整车报废，赔同配置新车；因碰撞或托底起火致报废，补足车损险赔付差额；技术背后是2020年起启动的六年自研，强调制造与使用全链路数据贯通以支撑质量追溯与预警。
OpenCode co-founder Dax Raad confirmed that during a continuous 48-hour attack period, the team encountered model refusal responses midway while using AI to analyze attack code; the reason is that safety guardrails cannot distinguish between defensive analysis and malicious exploitation. Similar issues previously occurred in the incident where Hugging Face was invaded by an experimental model; the team subsequently obtained OpenAI Daybreak Blue permissions to mitigate the problem, but questions remain regarding the timeliness of permission applications and the emergency capabilities of small teams.
At the 2026 Physical AI and World Models Conference, Qianjue Technology officially released Polibrain, a predictive world model serving as an embodied intelligence brain foundation for the real physical world. The model is based on research on latent variable identifiability published by its joint team with Tsinghua University at ICML 2025, approximately one year ahead of similar work by LeCun's team, and introduces polynomial representations and the Effective Degree (ED) quantitative metric to guide training. The model adopts a predictive rather than generative approach, supports multi-source heterogeneous data fusion, has been adapted to nearly 50 robots of different configurations, and emphasizes causal memory, continuous decision-making loops, and a decoupled architecture between the 'brain' and 'limbs'.
The team led by Zheng Guangting and Ma Songchen from the Hong Kong University of Science and Technology has open-sourced 'The Fourth Institute', an automated research platform supporting co-innovation between humans and AI agents.
The RoboColiseum platform released a routine embodied AI simulation evaluation report, testing the performance of four multimodal large models—Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and Gemini 4 Argon—under a unified Agent2Act framework; Claude Opus 5.5 ranked first with an overall score of 76.28%, while GPT-6 Astra showed outstanding performance in cross-embodiment control and home scenario tasks. However, all models significantly degraded in fine manipulation with dexterous hands and under image quality perturbations. The platform claims an 89.5% correlation between simulation and real-world deployment, supporting score generation within 30 minutes and task-level failure attribution.