Based on internal data from seven weeks of real developer sessions (4,211 interaction rounds), AI21 found that public coding benchmarks (such as SWE-bench and HumanEval) deviate significantly from reality: only about 11% of real requests are pure code generation, while interactive behaviors such as confirmations, redirections, git commands, and context pasting account for 40%, with messages averaging just 80 characters, featuring multi-turn nesting and mid-conversation entry points.
OpenAI launched a virtual try-on feature in the ChatGPT Images 2.5 version update. Users can upload selfies or full-body photos and click to preview how clothing and accessories look when worn, or upload product images for ChatGPT to 'wear' the items on the user and generate effect images. This feature is directly available to individual users, while enterprise users can access it via API calls.
2026 is defined as the first year of realized mass production for humanoid robots. Domestic models such as Songyan Power's Xiaobumi (9,998 yuan) and Unitree R1 (starting from 29,900 yuan) have been listed in batches on e-commerce platforms. Global shipments reached 19,100 units in the first half of the year, with Chinese manufacturers accounting for over 97% of the market share. However, low-cost models generally suffer from issues such as reduced degrees of freedom, missing force control modules, simplified dexterous hands, short battery life (1–3 hours), and weak generalization capabilities. Currently, C-end users primarily use them for non-household scenarios such as programming debugging, short video shooting, and tech display pieces. The core goal of manufacturers is to capture users and accumulate real-world scenario data to drive subsequent iterations.
PANews released the Global Payments Industry Weekly Report for Week 41 of 2026, covering key events from October 5 to 12, 2026: ESMA requires MiCA-licensed institutions to phase out unauthorized stablecoin businesses such as USDT by January 8, 2027; Coinbase has set October 30 as the withdrawal deadline for EEA users; India's UPI merchant fee, originally scheduled to take effect on October 15, was officially postponed by NPCI on October 13 until after Diwali; Mastercard announced that new European issuers must support single offline payments up to €200 starting February 2027; The OCC fined American Express National Bank $350 million, finding its card products involved approximately $13 billion in suspected trade-based money laundering; SAP Pay integrated USDC/EURC settlement, while PayPal disclosed over $1 billion in inter-subsidiary transfers using PYUSD in 2025; GCash raised $845 million to launch the largest IPO in Philippine history.
To address the pain point of repeatedly manually transferring logs and conclusions during integration testing between cloud and local devices, the author developed the open-source tool agent-penpal. It supports automatic communication between dual-end agents without requiring public IPs, SSH, or intranet penetration. With just three configuration steps (installing the CLI, generating a business card, and pairing), agents on both ends can directly ask and answer questions. The project is based on Node and the paseo CLI, released under the MIT license, with the homepage at https://github.com/bujianhe/agent-penpal.
Large model vendors are driving a shift in pricing logic from per million tokens to 'effective intelligence per dollar', separately pricing model capabilities, delivery speed, and budget certainty; OpenAI adjusted prices and default speeds through its GPT-5.6 and GPT-6 series, while Anthropic, MiniMax, Zhipu, and others also launched tiered models and subscription packages; enterprises are responding to surging call volumes and uncontrolled bills through model routing, caching, and cost platforms.
ManhoDL developed and open-sourced the browser extension Panelize, which supports calling 12 AI services including ChatGPT, Claude, Gemini, Grok, DeepSeek, Doubao, Kimi, Tongyi Qianwen, Zhipu, Yuanbao, and Google AI Mode in a split-screen single window. Inputting a question once sends it synchronously and displays answers side by side; no API Key is needed as it reuses logged-in browser accounts; it includes a prompt library with {variable} support; licensed under MIT, with no analytics or tracking, data stored locally only; released on the Chrome Web Store and GitHub.
Veecle has released Chiplab, an MCP platform for AI agents that supports agents invoking virtualized real target chip environments via a single API endpoint to complete firmware compilation, simulation, verification, and testing; each run's results are stored in a shared corpus for subsequent agents to inherit and use.
Security research team XBOW used models such as OpenAI's GPT-5.6 Sol, along with four researchers, to complete a full-chain exploit for the Chrome browser in 6 days, earning a $250,000 bug bounty from Google; AI handled source code search, idea testing, and PoC writing, while humans managed strategy formulation and result verification; the relevant vulnerabilities have been fixed, and Google recommends users upgrade to the latest version of Chrome.
jischeng released Oh My Ghostty (OMG), a Ghostty fork optimized for SSH remote development and Agent CLI workflows. It supports identifying Git repositories in ssh sessions, displaying Agent session status, port forwarding, editor integration, history commands, and quick input. It also connects to local Agents via ACP to generate commit messages. Currently, only a self-signed macOS version is available, requiring manual permission in System Settings to install.
Independent developer haidao2008 spent one month creating the interactive audio ink-wash map website history.walkingchina.com to help children understand the journeys of ancient figures. The site supports visualization of life trajectories for over 60 historical figures, including Su Shi, integrating location mapping of ancient and modern place names, bilingual audio narration, stage character illustrations, simulation of ancient road transport vehicles, and rice-paper style camera movements. It is completely free and open, requiring no registration.
OneKey maintainer Yishi points out that using an iPhone offline as a cold wallet requires strict prerequisites: at least an iPhone 17 or newer model, iOS 27 or higher, and migration to a new device every 1–2 years. Going offline is only basic protection; true security depends on hardware generation (such as the MIE memory integrity enforcement in the A19), BFU/AFU key states, strong device passwords (at least 12 random alphanumeric characters), the wallet's own key protection mechanisms (such as keychain-like configurations, key wrapping), and offline recovery solutions.
Abdulrahman Alfozan, an engineer on OpenAI's product team, installed a thumb-sized encrypted recording device wrapped in toilet paper into the USB port of the television in an executive conference room at the San Francisco headquarters. OpenAI identified the suspect by analyzing forensic images of the device and employee chatbot prompt logs; the FBI arrested him on October 5. The Department of Justice indictment showed that he had queried conference room information, asked about custom connectors and MAC addresses for devices, and left a second identical device in his backpack. OpenAI stated that no model weights, source code, or customer data were found to be stolen, but confirmed his plan to install more devices in additional sensitive areas.
JollyJ has released CharFloat, an independently developed local office agent tool. Through the MCP protocol, it enables MCP-supporting AIs such as DeepSeek and Claude to edit Excel/Word files in real time within the user's currently open WPS window, preserving existing merged cells, formulas, and formatting. All operations are completed in local memory without uploading data; it supports snapshot rollback, multi-client configuration, and local large model integration, currently supporting WPS desktop on Windows/macOS.
JollyJ released Mihomo Monitor v0.1-beta, a pure SwiftUI-developed macOS menu bar utility focused on tracking proxy traffic from the Mihomo kernel. It supports attribution across all application processes, local SQLite storage (in WAL mode) by domain and node dimensions, and adapts to system light/dark modes with numeric wheel animations. Open-source code is available on GitHub along with a direct DMG installer; manual permission must be granted in System Settings upon first launch.
In a blog post, Coinbase noted that given Google researchers' prediction that quantum computers capable of breaking cryptography may emerge by 2029, Bitcoin should immediately initiate a post-quantum upgrade. Their recommendations include separating transaction formats from signature schemes to support flexible replacement, phased adoption of hybrid signatures (e.g., 1-of-2 ECDSA/Schnorr + PQ), distinguishing between stateful and stateless signature implementations, ensuring MPC compatibility, and discussing contentious issues such as long-dormant wallets separately to avoid delaying overall security progress.
Apple's iPhone Duo China edition (model A3721) updated its filing information on September 29, adding three large model filing numbers: Shanghai-AppleZhiNeng-202506160057, Shanghai-TongYiWanXiang-202410090024, and ZheJiang-TongYiQianWen-20230901. According to an announcement by the Cyberspace Administration of China, Apple Intelligence completed its filing on July 8, 2026, applicable to Apple phones. Alibaba stated that Tongyi Qianwen will be integrated into Apple Intelligence as an AI capability, providing text and image understanding and content generation capabilities for Chinese users on iOS, iPadOS, macOS, and visionOS.
Geometry Check is a free Fusion 360 plugin that identifies minor geometric errors in one click, including line segments deviating from right angles by within 0.5°, hairline gaps with endpoint spacing less than 0.05 mm, unclosed sketch profiles, and sub-millimeter fragment lines; it supports detection mode and experimental auto-repair mode, with all tolerances adjustable in settings, currently in public Alpha stage.
Ripple is partnering with Privy, a subsidiary of Stripe, to integrate Privy's embedded wallet technology into the XRP Ledger, allowing developers to build wallets directly into payment applications so users can transfer XRP, RLUSD stablecoins, and other on-chain assets without managing private keys themselves; this architecture also supports AI agents holding funds and automatically paying for online data and services via the x402 standard. According to RippleX's Head of Engineering, such agent payments account for approximately 25% of total activity on the XRP Ledger, reaching a cumulative 14.3 million transactions as of Wednesday, though the total amount is only about 6,991 XRP and 12,752 RLUSD; new wallets must first receive a minimum reserve of 1 XRP to be activated.
SuperLocalMemory is an open-source AI agent memory system that runs on the user's local machine. It supports Mode A (zero model calls), Mode B (local models), and Mode C (cloud models). It offers controlled memory storage, hierarchical recall, cross-agent sharing (SLM-Mesh), semantic retrieval of documents and images (EmbeddingGemma 2), time awareness, audit trails, and GDPR compliance capabilities. It supports mainstream tools such as Claude Code, Codex, Cursor, and Copilot, and can be installed via npm, pip, Shell scripts, or Agent plugins. The project is hosted on GitHub under the AGPL-3.0-or-later license, with four arXiv papers supporting its design.
TokenPulse is an open-source iOS app that supports viewing package quotas and balances for 78 AI providers, with a newly added Codex reset preview feature. The project has been open-sourced on GitHub, providing a TestFlight testing entry point and a Telegram community group.
On October 14, 2026, Poxiao Intelligence and Tinno Mobile released the SeePhys Hub, a full-body multimodal data collection system. The system includes a four-camera headband, dual wristbands, three waist/leg motion trackers, and optional haptic gloves. It supports synchronized visual-motor-tactile tri-modal data collection with a synchronization accuracy of no more than 0.5 milliseconds. Positioned as an intermediate-layer collection solution between internet video and high-precision motion capture, the device offers hardware sales, algorithm support, or cloud processing services to embodied AI model teams, data operators, and research institutions.