Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
TomCad is an open-source tool released by an independent developer that supports defining Caddy server configurations via a single .toml file, including HTTPS listening, TLS automation, domain redirection, static file serving, reverse proxying, and subpath routing; usage is $ cargo run -- start "<PATH_TO_TOML_FILE>".
Aura is a self-hosted, multi-user AI agent platform written in Go, supporting per-person graph memory, tool calling, document retrieval, scheduled tasks, and an embedded web console; it defaults to calling DeepSeek-V4 Flash via OpenRouter, but also supports ChatGPT plans, local llama.cpp (Gemma 4 12B QAT), or Ollama; it uses Docker Compose for deployment, depends on Postgres and ArcadeDB, and requires a mini PC with 16 GB RAM; the project is open source under the MIT license, with v1.1.0 being the latest release.
xcrashlytics is an open-source macOS CLI tool designed to integrate Firebase Crashlytics and Xcode Organizer crash reports into AI coding agent workflows, supporting platforms such as iOS, Android, and macOS.
Amazon Bedrock AgentCore officially launched the AgentCore payments feature, supporting AI agents in making automatic, controlled micro-payments for services such as single model inference at runtime via the x402 protocol; Incarna and BlockRun implemented end-to-end pay-per-use processes based on this capability, reducing the integration cycle from months to three days, with each payment settled in USDC on the Base network and verifiable on-chain.
NVIDIA KGMON 团队在 KDD Cup 2026 数据智能体竞赛中获第二名，其核心方法是围绕固定小语言模型构建轻量、清晰、可验证的智能体运行时（harness）。
k10s is an open-source Kubernetes TUI tool written in Go and built on Bubble Tea. It supports dual-mode interaction via mouse clicks and keyboard shortcuts; it groups all 30+ types of Kubernetes resources by resource type, automatically discovers CRDs, supports lazy-loading informers, includes a built-in AI mode (connecting to OpenAI-compatible or Anthropic APIs), offers offline demo clusters, theme switching, plugin extensions, and one-click self-updates; it provides precompiled binaries for macOS/Linux/Windows and installation via Go, with source code hosted on GitHub.
A group of mathematicians led by Fields Medalist Terence Tao, known as AHM, released a statement criticizing OpenAI's release of over 700 AI-generated mathematical proofs at once as "a display of power rather than scholarship," violating the core recommendation of its advisory group AGMAI to "not use frontier mathematical problems for internal model testing." The statement questioned whether the results relied on training with unlicensed works by mathematicians, and pointed out that AI problem-solving is squeezing core values of mathematics such as conceptual understanding, collaborative dissemination, and exploration of new directions.
Pocketty is a native SSH terminal app for iPhone and iPad that supports real-time display of agent (e.g., Claude Code, Codex, Cursor) running status (working/needs you/done) via the herdr API, and pushes end-to-end encrypted notifications to users when an agent is blocked or completes; notifications are encrypted by a local daemon using HPKE/X25519/ChaCha20 and delivered via Apple Push Notification Service, decryptable only by the target device; the app features the built-in Ghostty terminal engine, supporting hardware keyboards, Nerd Fonts, file uploads, port previews, and pure SSH login; the macOS/Linux daemon can be installed with a one-line command; pricing is a one-time payment of $99 after a 14-day free trial.
Superwhisper has released S1-mini v1, a GGUF-format text normalization model based on the Qwen3-0.6B architecture, dedicated to cleaning ASR output. It supports three-axis control for styling/structure/context. The Q4_K_M quantized version is 462 MiB in size and achieves 94.8% token accuracy on 7,519 English samples. It requires disabling thinking mode and using greedy decoding (--temp 0). Licensed under Apache 2.0 with attribution constraints.
Viberuler is an open-source CLI tool that scans local AI programming session logs (such as Claude Code, Codex), Git commits, and project structures on the user's machine to calculate the VIBE SCORE. Metrics include tokens per dollar (6.5M tok/$), LoC shipped (312K), machine noise (163K, accounting for 33% of diffs), 47 projects, and a 212-day commit streak. All scans default to zero network calls; --submit sends only 14 aggregated fields, without uploading code, paths, or prompts. It supports --share to generate unverified share cards, --audit to analyze context waste (such as cold contexts, ghost tokens, dead weight MCP servers), and --wrapped to generate monthly recaps. Licensed under MIT, with source code hosted on GitHub.
Rinkata is a shared project context platform designed for teams using coding agents (such as Codex, Claude, Cursor, Grok, Gemini). It provides four modules—Knowledge, Ideas, Work, and Complete—and supports integration with existing toolchains via CLI. It does not generate code but serves as an alignment layer that unifies goals, specifications, decisions, and completion evidence. All project documentation and company materials are stored in a searchable knowledge base, and completed work is automatically written back to the knowledge base.
Project Arena is an open-source AI SRE benchmark that evaluates AI agents in Kubernetes environments through fault injection, investigation logging, and configurable judge scoring. It supports deployment on local kind or AWS EKS clusters, offering a full set of 21 scenarios and a smoke test suite of 6 scenarios. The benchmark has been used to compare the native capabilities of Edge Delta and Grafana with Claude's investigation performance on the edx/gcx platforms, publishing scores across dimensions such as detection rate, root cause analysis, impact scope, and mitigation recommendations.
lsearch is an offline desktop search tool developed by iamzhaopengx. It supports building local full-text indexes for documents in 27 formats, including Word, Excel, PowerPoint, WPS, PDF, HTML, source code, plain text, and EPUB/MOBI/AZW3/FB2. It uses CJK bigram segmentation and full-width/half-width normalization to achieve precise hits for Chinese two-character words. It supports Google-style query syntax, regex search (including AI-assisted generation, compatible with DeepSeek / Tongyi Qianwen / Ollama / OpenAI / Anthropic), and direct jumping to edit results. It features incremental indexing + real-time monitoring; no cloud, no account, no telemetry; supports Windows/macOS/Linux (including domestic systems); does not support OCR or DRM e-books. The tech stack is Go + Wails v2 + React 19, using Bleve + SQLite, packaged as a single executable file. Free download address: https://github.com/cmzhaopeng/lsearch-releases/releases/latest.
Odyssey-3 is a foundation world model that generates interactive environments from prompts and predicts in real time how user or agent actions change the environment; its Pro version scored 66.1 (highest score across 8 runs) on the Physics-IQ Verified video-to-video evaluation; the same model has been adapted for controlling robotic arms and humanoid robots, autonomous driving, and agent training; a research preview is available, and API access requires contacting the team.
Jevman is an open-source Pac-Man decision benchmark that supports models with any HTTP interface: at each junction, the game sends the maze state in JSON format and requests direction probabilities, requiring the model to respond within 2 seconds; six models each completed 100 games against classic ghosts, with scores ranked by mean and 95% confidence interval, and all games can be precisely replayed for verification.
DrawFig is a scientific drawing website developed by the author based on draw.io. It supports generating horizontal flowcharts from a single sentence, with draggable boxes and arrows for editing. Users can add or delete steps and modify labels later, and export to TikZ. The tool is currently open for use: https://drawfig.com/editor.html
Amazon SageMaker HyperPod has released an EKS-based reference architecture for multi-team shared GPU clusters. It uses AWS IAM Identity Center to implement unified identity authentication, and achieves team isolation, resource fairness, and cost attribution through SageMaker AI domains, Kubernetes namespaces, HyperPod Task Governance, and namespace-level cost allocation.
openzine is a tool that supports converting PDF or image pages into 3D manuals with page-turning animations, paper shaders, and textures. It supports mouse dragging, clicking, keyboard page turning, and dual-page mode on mobile devices; converting a 60-page NASA graphics standard manual PDF took 19 seconds. The demo address is https://openzine.pen-ine.workers.dev/, and the repository address is https://github.com/JosssphZhou/openzine.
亚马逊云科技宣布 Amazon Bedrock AgentCore 在由光环新网运营的中国（北京）区域和由西云数据运营的中国（宁夏）区域正式可用。
Goodfire has launched a new internal activation monitoring system that reads intermediate neural activation values during the model's forward pass in real time via embedded probes, rather than relying on independent review models. In tests with Kimi K3, monitoring 1500 sessions cost approximately $51, significantly lower than the $233 to $10000 required by traditional review models. The system achieved a 94% accuracy rate in detecting malicious hacker behavior, an 8.7% false positive rate, and less than a 2% increase in latency. The system is integrated into the Baseten platform, supporting customer-defined risk types and response strategies.
As of September 29, 2026, global VC-backed AI companies completed a total of 195 AI startup acquisitions, representing a year-over-year increase of 14%. OpenAI led with 20 deals (10 this year), followed closely by Anthropic, Legora, Harvey, and others. Only 12 deals disclosed their transaction amounts, with the largest being Nscale's $1.65 billion acquisition of Anyscale. Crunchbase statistics exclude acquisitions initiated by major tech giants such as Tencent and Zhipu.
Decrypt has launched Money Accounts on Solana, supporting the use of USDT balances uniformly across Earn, Swaps, Predictions, and Perps functions without requiring bridging or product-specific deposits; the platform is self-custodial, with Decrypt not holding private keys; Gas and rent are sponsored via co-signing by fee payers, so users do not need to hold SOL; Myriad wallet users can directly upgrade and use it without migration.
The European Securities and Markets Authority (ESMA) issued guidance requiring authorized crypto-asset service providers (CASPs) to cease offering services related to stablecoins that do not comply with the Markets in Crypto-Assets Regulation (MiCA), including trading and accumulation, to EU clients within three months; unauthorized mainstream stablecoins such as USDT and PYUSD were specifically named; existing client holdings must be disposed of by January 8, 2027, during which time only selling, converting, withdrawing, transferring, or custody are permitted, while new purchases or promotions are prohibited.
Diffle is an open-source, keyboard-first Git diff viewer that supports reviewing diffs in the browser, adding line-level/block-level/file-level comments, and exporting comments as GitHub pending reviews or as prompts for agents.