During the National Day holiday, ZhouAI broke down the Xiaohongshu content production process into 11 AI video production Skills, covering stages such as resource indexing, story scripting, storyboard planning, motion graphics, concept shots, voiceover-subtitle synchronization, cover design, and final quality inspection. These Skills support Agents compatible with SKILL.md, such as Codex and Claude Code, and have been open-sourced under the MIT license on GitHub.
Claude now supports direct operation within Google Docs, Sheets, and Slides, enabling document drafting and editing, spreadsheet analysis, formula writing, chart creation, and slide production. New connectors have also been added, allowing users to create and edit Workspace files directly from Claude. This feature is currently available for testing across all paid Claude plans.
Update to the LAYA local open-source decision model training workbench. New features include decision model training (supporting creation or fine-tuning of LAYA-format models using Hugging Face encoders) and object detection training (supporting SSDLite / Faster R-CNN). Applicable scenarios include fruit picking, weed removal, pest monitoring, steel pipe quality inspection, and textile quality inspection. Training can be performed on a GPU with 4 GB VRAM or on CPU; the model service interface remains /v1/systemone.
On October 6, WordPress released version 7.1.3, fixing 7 core security vulnerabilities and 4 program defects. Anthropic reported 3 of these, including a delayed-trigger backdoor in the WXR exporter, a denial-of-service vulnerability in WP_Http::make_absolute_url(), and an access control issue allowing authors to set posts as sticky.
agentc is a lightweight agent written in freestanding C23. It has no libc dependency, uses static linking, and supports multiple platforms including Linux, macOS, and Windows. Its cold start time is only 0.3 ms, resident memory is approximately 0.7 MB, and the stripped binary for release version 0.5.0 is 916 KiB. It includes built-in support for MCP tool calls, provides a versioned C ABI for C/Rust extensions, and implements TLS via vendored mbedTLS. The project is open source on GitHub under the MIT license.
free-llm-intel is an open-source static site that automatically inspects official pages of over 30 domestic and international LLM vendors daily, including OpenAI, Anthropic, Gemini, Mistral, xAI, Tongyi, Kimi, DeepSeek, Zhipu, MiniMax, Doubao, and SiliconFlow. It generates custom feeds for sites without RSS and synthesizes a master subscription stream. English content is precisely translated paragraph by paragraph by an agent (over 2,000 articles), with bilingual Chinese-English retention of brands/models/terminology/code/links as-is. It also compiles a comprehensive table of free API quotas, permanently free models, and limited-time activities for 70 vendors; each entry includes an official link for verification, and unverifiable items are marked 'not accepted'. The project is MIT open-source and supports one-command self-deployment.
FlashTiger has launched Builderoom (https://builderoom.com), an independently developed maker showcase platform. It supports publishing in all formats including Web, App, Mini Programs, Desktop, CLI, Browser Extensions, AI Agent/Skill, and Design Works. After entering a link, built-in AI automatically extracts features, descriptions, tags, and other content. The platform offers semantic search based on natural language descriptions. Projects are directly linked to creator profiles; the platform is completely free with no paid ranking options.
Within 60 days of proposing the Everyday Finance strategy, Bitget Wallet continuously supplemented its chain across four layers: the entry layer (Pay/Apple Pay/mobile top-ups), the execution layer (high-performance API/DEX aggregation/Agentic Wallet), the payment and settlement network layer (Stellar/Tempo/physical cards/Onchain Payments Matrix), and the Agent/AI layer (Skills/social login/AI analysis). Data covering more than 50 countries, tens of millions of transactions, hundreds of millions of dollars in transaction volume, an average of 15 days of payment behavior per user per month, and an average transaction value of around ten-plus dollars indicate that it has become a real high-frequency small-value payment path, but it has not yet proven that users treat it as their default account.
MoeXian developed and launched PartLab, a web-based component inventory management tool. It supports importing orders from LCSC with automatic parsing of part numbers, packages, parameters, and datasheet links; enables batch check-in via QR code scanning; provides BOM shortage comparison features; and includes a companion JLCPCB EDA plugin (search for 'PartLab' in the extension marketplace to install).
Semwright is an open-source Rust runtime designed specifically for AI agents to interact with desktop and professional software. It supports access to application objects and project data through structured drivers. A shared runtime uniformly handles permissions, execution, dependencies, artifact passing, and validation. It supports the MCP protocol but is not bound to a single protocol, model, or agent. The goal is to reuse the same infrastructure to build cross-application workflows.
On October 8, OpenAI rolled out GPT-6 to all global ChatGPT users, gradually replacing GPT-5.6 Sol and GPT-5.6 Luna. The core change in the new version is the introduction of a Smart UI interaction layer, which supports displaying images, charts, buttons, forms, and interactive mini-interfaces directly within the chat box. Paid users run GPT-6 Sol, while free and Go users run GPT-6 Luna. The model employs a 'think-and-output' mechanism, reducing the first-response latency for online search in GPT-6 Instant by 44% compared to GPT-5.6 Instant.
Michael Smith, a man from North Carolina, USA, used AI to generate hundreds of thousands of songs between 2017 and 2024. He uploaded them to streaming platforms through over ten thousand fake accounts and conducted fraudulent plays using a bot matrix to defraud copyright revenues. On October 6, 2026, the US Department of Justice announced that he was sentenced to 18 months in prison and forfeited $8 million in illicit gains.
Broadcom is raising over $50 billion in financing for its custom AI chip project in collaboration with OpenAI, with Apollo Global Management and Blackstone Group participating in loan discussions. Meanwhile, Oracle is also negotiating with Apollo and Goldman Sachs for financing to purchase chips. Multiple top-tier financial institutions are simultaneously intervening on both the supply and demand sides of the chip market.
On October 7, 2026, Meta released a dedicated iPad app for Muse. This AI assistant completed its large-screen adaptation just one month after its initial iOS and Android launch on September 8. Sensor Tower data shows that downloads across both platforms have exceeded 6.6 million. The new version added Connectors for Asana, Canva, Dropbox, Figma, QuickBooks, GitHub, Klaviyo, Zoom, and Meta Ads Accounts, and integrated Notion and Granola. Previously, it had already connected with retail platforms such as Best Buy, Gap, Sephora, Walmart, and Wayfair. On October 6, Meta collaborated with industry partners to initiate the development of open technical standards for AI agent identity authentication.
OpenAI has released 722 mathematical manuscripts on its GitHub repository openai/math, generated by an internal unpublished model. These cover 372 problem families, including the zero region of the Riemann zeta function and Mahler's conjecture. Each manuscript includes a PDF preprint, Lean 4 formalized proofs, and partial reasoning summaries. OpenAI stated that each problem took an average of 3 hours of ChatGPT Pro compute time, while also noting in the README that "some non-formalized results may contain errors." This batch release sparked criticism from 25 Fields Medalists, including Terence Tao, who questioned the bypassing of peer review and treating mathematics as a benchmark.
Microsoft introduced the Surface Laptop Ultra at a launch event, featuring the Nvidia RTX Spark SoC as its first application. It is equipped with a Blackwell GPU offering 5120 or 6144 cores, up to 128GB of LPDDR5x unified memory, starts at $2599, and ships on October 16; the device emphasizes unified memory allocation capabilities, supporting local AI development and AAA game execution, such as "Gears of War: E-Day".
The paper analyzed 127,833 queries issued by 6,342 clinicians over an 8-month deployment period, using the RCQ-Map framework to characterize their task types, intents, and answerability. It found that document management (36.2%) and knowledge retrieval (28.9%) were dominant, while diagnosis accounted for only 3.7%; more than 1/3 of queries could not be directly well answered. Compared against the Clinical AI Benchmark Atlas comprising 58 public benchmarks, the median benchmark did not include document-related requests, and the task distribution overlapped with real-world usage by only 31%.
Perplexity has migrated its core search service layer from Amazon DynamoDB to CobbleDB, a self-developed distributed key-value store written in Rust. The median batch read latency decreased from 31.4 milliseconds to 5.60 milliseconds, and storage costs dropped by at least 20%. CobbleDB utilizes the RocksDB engine, memory-mapped caching, and local NVMe, supporting asynchronous replica updates and cross-availability zone hedged reads. The project was completed within two months by two engineers and a cluster of AI coding agents, comprising approximately 40,000 lines of Rust code, with plans to open-source it.
OpenAI has launched Dots, a personal agent platform positioned as an agent that runs continuously, understands user goals, and learns from feedback, without relying on model selectors or complex configurations. Built on the Codex agent harness and the Astra model, Dots emphasizes safety and alignment. It currently supports a single primary Dot, with multi-Dot virtual teams to be opened up later. Additionally, OpenAI announced a revenue-sharing mechanism for its plugin ecosystem, sharing earnings from user quota usage with partners such as Notion and Figma.
US company Reflection released its first open-weight MoE model, Beam, with 501 billion total parameters and 23 billion active parameters, focusing on programming, reasoning, and Agent workloads; weights are planned for public release in October. French company Mistral simultaneously launched Mistral Large 4, with 1.05 trillion total parameters and 52 billion active parameters, supporting multimodal capabilities and over 160 languages, with a context window of 1 million tokens. An API preview is currently available, and full weights will be released at the end of October.
The author used the ML-intern feature in Hugging Chat to autonomously build 6 specialized models within a few days: a citrus disease recognition VLM, a Huggy character LoRA, a camera perspective LoRA, a graffiti replacement LoRA, a lightweight prompt rewriter (0.8B/2B), and an Agate 4-step text-to-image model; each example includes a clear prompt structure (containing idea, verified facts, baseline requirements, smoke test, deliverables, and budget cap), training data construction logic, hardware resource consumption (USD 1.90–37), and publicly released results; all original prompts are hosted on GitHub at yvrjsharma/ml-intern-prompts.
US company Reflection released its first open-weight model, Beam, with a total parameter count of 501 billion, using an MoE architecture focused on coding, reasoning, and agent workloads; French company Mistral simultaneously launched Mistral Large 4, with 1.05 trillion total parameters, 52 billion active parameters, supporting multimodality and a 1 million token context window.
Google's security team used Gemini to automatically translate the C image library giflib (approximately 3000 lines) into an ABI-compatible Rust implementation. Through a three-stage process (prompt-driven translation, FFI safety modeling, and differential fuzzing feedback iteration), the project became immune to the heap write zero-day vulnerability CVE-2026-26740 before its public disclosure. The project was validated for semantic equivalence through regression testing on 30 million real GIF files and 200 million side-by-side fuzzing runs, and has been open-sourced as giflib-rs.