Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
Apogee is an open-source (MIT licensed) AI browser extension for locally summarizing web pages, YouTube/Bilibili videos, PDF/DOCX files, and pasted text, requiring no backend, API keys, or cloud uploads.
Conduit sued Tether, alleging that on September 24, 2025, it unilaterally froze approximately $2.76 million in USDT in the company's treasury wallet, which is used for daily operations and customer settlements; over four months, this wallet processed 4,427 transactions involving 78 counterparties and totaling over $1.1 billion. Conduit claims the freeze lacked law enforcement basis, the involved wallet had no substantive connection to the Onix case in Brazil, and Brazilian police did not request the freeze; the lawsuit seeks to unfreeze assets, compensate for at least equivalent losses, recover reserve yields during the freeze period, and challenge the formation mechanism of the freeze decision and the lack of remedies.
xnjiang has open-sourced two MIT-licensed Rails applications: Aprove AI (an integrated project operations system) supports the entire process from order to delivery to invoicing, featuring a Chinese interface and seed data for a demo company; MuseFlow is a private PLM for manufacturing, focusing on CAD drawing versioning, BOM, and process document management, using a single-tenant design. Both support self-deployment, and their GitHub repositories are public.
PocketWebTools has launched PocketWebTools for Mac, a paid application that runs open-source models (up to 35B parameters) locally on Apple Silicon devices. It supports offline chat, audio/video transcription with speaker labels, and automatic segmentation of long videos into subtitled clips; all files remain local on the device. The web version offers free tools for voice cloning, TTS, transcription, video editing, image super-resolution, and background removal, requiring no registration.
Doubao Work launched the infinite canvas feature on October 9, supporting automatic import of images, videos, and HTML into the same canvas, and integrated the new image model Seedream 5.0 Flash; simultaneously, it officially released the DouBao 2.1 Lite (version 0921) model, optimized for response speed and quota consumption in high-frequency office scenarios, with stable performance in tasks such as table processing and multimodal understanding. Users can select this version from the model list and adjust inference intensity.
On October 8, Anthropic launched the OSS Scanner, an open source software vulnerability scanning service that uses top AI models such as Claude Mythos for fully automated scanning without manual verification; maintainers of core open source projects can apply for access by submitting a PR to the GitHub repository; manual review was conducted on 97 high/critical severity vulnerabilities detected across 48 projects, with 85 (88%) meeting Coordinated Vulnerability Disclosure (CVD) standards, 11 being known defects or duplicates, and only 1 being a false positive.
Sui推出Hashi协议，允许机构用户将比特币锁定在比特币主网地址中作为抵押品，无需跨链桥接；协议在Sui链上发行hBTC凭证代币用于借贷等DeFi活动，退出时销毁hBTC并解锁原BTC。该系统已获$500百万美元资本承诺，由超20家行业伙伴联合支持，并通过Certora和CommonPrefix完成智能合约与MPC密码学审计。
According to a Tiger Research report, sentiment in the crypto market turned greedy in Q3 2026, but corporate fundraising did not recover in tandem. Capital exhibited five structural changes: M&A shifted toward acquiring licenses and specific capabilities (e.g., Circle's acquisition of Tazapay); the lead investment power of established VCs weakened while strategic capital from CEXs (e.g., Binance Labs, Coinbase Ventures) became more active; early rounds contracted (seed round share dropped to 15.0%) while Series A-C focused on validated businesses; debt financing and IPO proceeds exceeded equity financing; and capital concentrated in sectors connecting to traditional finance—such as payments, stablecoins, tokenized securities, and AI-related infrastructure—rather than new public chains or DeFi.
Senba2024 has open-sourced Papers-Go, an intelligent paper browsing platform that supports Docker self-deployment. It features swipe-up/down browsing on mobile devices and feed display on desktops, covering all disciplines on arXiv as well as major AI conferences. It supports large language model integration via Codex login, API, and Ollama. It offers user interest memory, skills configuration, tool calling, and editable recommendation score calculation formulas. It also supports category browsing, search, research trend generation, and AI-powered in-depth reading.
Liquid AI has open-sourced two open-weight decision models in the d1 series: d1-3B (31.2B parameters, supporting text and image inputs) and d1-omni-600M (587M parameters, supporting combined text+image or text+voice inputs), both of which are available on Hugging Face; d1-3B scored 48.57 on Decision Index v0.2.1, claiming to lead all models under 10B; measured single-judgment latency is 8 milliseconds on RTX 4090, 50 milliseconds on Jetson Orin Nano, and 25 milliseconds on RTX 3060.
Request Eagle is an open-source API client written in Rust, featuring GPU-based rendering without reliance on WebView. It responds to the first keystroke within 0.36 seconds after launch. Requests are saved as TOML files, supporting Postman-style pm.* scripts, and it provides a CLI designed for AI agents.
OpenPilot is a free, MIT-licensed desktop AI agent that supports connecting to any OpenAI-compatible model (such as OpenAI, OpenRouter, Ollama, LM Studio). It can read and write files within specified folders and execute commands; sensitive operations require user confirmation. Currently, it supports Windows, while macOS is in early testing stages.
JetBrains 发布 Mellum2.1 编程 AI 模型，延续 12B 混合专家架构与 2.5B 活跃参数，重点增强智能体编程能力，支持代码库探索、文件编辑与修改验证；高负载推理吞吐接近 Qwen3.5-9B 的两倍，已在 Hugging Face 开源，后续将推出 llama.cpp、Ollama、LM Studio 的 GGUF 版本及 vLLM 的 MTP 推测解码组件。
Porch is an open-source CLI tool that automatically aggregates locally running Claude Code and Codex coding sessions into daily, weekly, and monthly summaries organized by project, including estimated task duration and blockers; it supports viewing session status across terminals/editors, tracking token usage and API equivalent costs, and exporting summaries as Markdown to Obsidian or any folder; no Porch account is required, all session status and usage data remain on the user's Mac, and summary generation relies on the user's installed and logged-in Claude Code or Codex.
Solana plans to reduce target block time from 250 ms to 200 ms at epoch 1053 (expected around October 9, 23:00 Beijing Time), completing the third reduction since August from an initial 400 ms; maximum compute units per block will decrease from 37.5M to 30M, with theoretical throughput remaining largely unchanged; this upgrade has been deployed to testnet and devnet, while mainnet implementation depends on network conditions such as validator miss rates.
During the National Day holiday, over 300 humanoid robots from companies such as Galaxy General, Zhiyuan, and Cross-dimensional Intelligence were scheduled to work in cultural tourism and commercial scenarios including Jingdezhen Taoxichuan, the Summer Palace, Hengqin Chimelong Spaceship Paradise, and Shanghai White Magnolia Plaza, undertaking tasks such as vending, inspection, interaction, and store management; the daily rental price of humanoid robots has dropped from the ten-thousand-yuan level at the beginning of 2025 to less than 1000 yuan, with some Xianyu merchants listing prices as low as 289 yuan/day; the industry is shifting from 'performance economy' to service implementation, but faces structural challenges such as seasonal profitability, high operation and maintenance costs (current manpower ratio approximately 3:1), data closed loops, and scenario migration.
Huxiu analysis indicates that Musk, through SpaceX and Tesla jointly preparing to build the Terafab super chip manufacturing base, aims to reduce dependence on Nvidia GPUs and master computing power production capacity; Zuckerberg, through leading the $1.8 billion virtual biology program via Biohub, has united Meta, DeepMind, etc., to invest in cell culture, spatial transcriptomics, and experimental data production, seizing the new training data entry point in the AI era. Although they superficially belong to different fields, they actually both point to the contest for core means of production—wafers and cells—as AI extends into the real world.
On October 8, XRP Ledger activated the PermissionDelegationV1_1 feature, which allows account owners to authorize other accounts to perform specific operations without sharing their master keys; this feature supports restrictions based on action types (such as compliance approvals) rather than amounts, but currently disables PaymentBurn permissions to prevent accidental token creation; the upgrade requires support from at least 29 trusted validators for two consecutive weeks.
iwant is a tool developed by Oskar Kocol that supports automatically finding available GPUs in GCP regions, provisioning resources, downloading models such as DeepSeek, Kimi, Minimax, MiMo, and Step, and configuring vLLM through a single command. It ultimately provides OpenAI-compatible API endpoints, API keys, and SSH access.
Sharpa unveiled three new products at IROS: the integrated tactile-sensing dexterous manipulation robot D01, featuring upper-body electronic skin with a 100Hz tactile sampling rate; the next-generation fully tactile ultra-compact dexterous hand W02, with 21 active degrees of freedom, an overall hand size approximately 30% smaller than W01, and zero grasp radius capability; and the high-fidelity exoskeleton haptic data glove AE01, equipped with 22 encoders, supporting bidirectional teleoperation and synchronized collection of tactile feedback.
加州大学伯克利分校杰·基斯林院士团队与王子龙博士等在《自然·生物技术》发表论文，提出基于可编程合成酶（聚酮合酶）与专用小模型协同的AI for Science平台，实现塑料单体的可设计生物合成、性能预测、解聚条件优化及技术经济分析四维闭环；该平台使酶产量提升数个数量级，支持从克级实验室产出迈向百吨级中试，目标2028年前验证电子拆解等高价值小用量场景；研究强调高质量实验数据与AI模型的闭环互训是突破'死亡谷'的关键。
excali-mode is an open-source plugin developed for Emacs 32 (master), supporting native opening, editing, and saving of standard .excalidraw files. It fully replicates all Excalidraw tools, including hand-drawn style, arrow bindings, and layer operations.
This paper proposes the agent-controlled forgetting method: tool-using agents autonomously select historical observations, replace them in-place with brief notes, and store the original content in a recoverable archive; the Python implementation requires no task-specific training and protects user instructions and assistant messages.
The paper proposes Synthesis Through Simulation (STS), a schema-free data synthesis paradigm: LLM agents call policy-enforced APIs in simulated enterprise environments to generate data, thereby naturally ensuring structural validity; its general-purpose generator GP achieves an average marginal fidelity of 0.88 and a 100% constraint satisfaction rate across all 10 environments, without requiring access to the DB schema; the framework, 10 simulated environments, and synthetic datasets have been open-sourced.