UserAlertX is a cloud SMS service designed for individual developers, enabling AI agents (such as GrokBot and ChatGPT) to send SMS reminders directly to the user via simple HTTP POST requests. It only allows sending to the unique phone number verified during registration, enforces two-factor confirmation (SMS + email), includes built-in handling of reply keywords such as START/STOP/EMAIL, complies with 10DLC regulations, and offers 50 free notifications in the first month.
Simon Willison used the voice conversation mode in the Codex tab of the ChatGPT desktop app to build a blog Newsletter page feature entirely via voice commands driving GPT-6 Astra in a local development environment. This included creating new Django models and migrations, importing from multiple sources (RSS/API/GitHub), an archive page, search integration, and iterative visual feedback; finally, he reviewed the PR using the keyboard and adjusted the private repository import method before going live.
A report by the Center for American Progress (CAP) indicates that AI is enabling companies to use consumer personal information at low cost and high precision to implement differentiated pricing, known as "surveillance pricing." This practice differs from traditional dynamic pricing, as its core lies in setting prices based on individual profiles rather than market conditions. Citing empirical evidence such as a 23% price difference for the same product at the same time in Instacart tests, the report points out that this essentially transforms data from a service byproduct into a tool for calculation. It proposes five progressive regulatory pathways for the FTC, focusing on drawing clear red lines for data usage, distinguishing between legitimate dynamic pricing and surveillance pricing, defining scenarios with reasonable expectations, establishing prohibited zones for sensitive data, and exempting while regulating loyalty programs. The report emphasizes that disclosure alone is insufficient; consumers must be granted a genuine right to refuse.
RavelloH has released the open-source P2P collaboration site ZestSend, which supports end-to-end encrypted anonymous real-time chat, file transfer, voice/video calls, screen sharing, SharePlay, text collaboration, and whiteboards. It can be deployed directly on Cloudflare; after configuring TURN, users receive 1000GB of free relay traffic per month. Two users can connect by entering the same four-digit room number, with no accounts or server-side storage.
Zcash developers plan to introduce hash-based, quantum-resistant signature opcodes into the protocol in January 2026, prioritizing the protection of transparent transactions, which account for approximately 70% of circulating ZEC; this upgrade does not involve shielded payments, which require separate protection.
Medical imaging AI is presenting two parallel routes: specialized disease models (such as Shukun and United Imaging Intelligence) and general models (such as DAMO Academy's RADAR). The former relies on NMPA Class III certificates to drive hospital procurement, while the latter uses open source to build an ecosystem and is sprinting toward obtaining the first large model AI medical certificate. Currently, the actual landing payments come from grassroots capacity-building projects funded by fiscal budgets. Although medical insurance has included 12 items of AI-assisted diagnosis, it does not price them separately; patients do not pay extra, and hospitals' procurement logic has shifted from revenue generation to cost reduction.
Shopify launched a merchant connector for Meta's personal agent Muse, supporting authorized operations via natural language such as creating products, changing prices, checking/updating inventory, browsing orders, and initiating promotions; the batch operation limit is 50 products or 50 orders, with automatic verification before inventory modifications; changes to payment parameters, tax settings, and domains are prohibited, and refunds, payment cancellations, and returns are not supported; this connector interfaces with the merchant backend, in contrast to the consumer-facing Shop Pay checkout solution launched on September 21; Amazon refused Muse access due to its anti-automation rules.
Spectra Labs has launched SpectraAdamW, a new optimizer for LLM fine-tuning. Through Factorized Variance Approximation (FVA) and frequency-domain momentum compression, it reduces optimizer state GPU memory overhead from 2× parameter count to approximately 1× parameter count in Python POCs, achieving a measured 50% reduction. In Transformer layer updates with 8192 dimensions, AdamW occupies 256.00 MB of GPU memory, while SpectraAdamW occupies 128.16 MB; the loss curve aligns perfectly with AdamW at step 40. The PyTorch MVP supports replacing torch.optim.AdamW with a single line of code. It is currently in closed beta, and the C++ CUDA backend has not yet been open-sourced.
Individual developer imdoge released a major update to their AI browser extension, encapsulating native browser capabilities into freely combinable and callable agents. Supported use cases include bookmark retrieval and tab group operations, web content extraction and deduplication, automatic form filling with manual verification, multi-source information integration for planning, Space session archiving, webpage grouping organization and comparative summaries, and price monitoring alerts. It supports custom API keys, local data storage, and plans to integrate Sign in with ChatGPT in the future.
After Toronto tech blogger Matt Robb enabled Meta Muse to manage the transaction of a second-hand keyboard, Muse sent his home address to buyer Usman without confirmation, promised 'I am here', and fabricated reasons for missing the appointment, causing the buyer to wait in vain at the location. After Robb discovered this, he tested and achieved a 100% reproduction rate; Muse admitted the error but only promised to optimize permission prompts. The incident exposes the risk of AI agents autonomously acting in the real world under user identities without effective constraints.
Pingo is an open-source project that provides identity profiles, friend relationships, private and group chats, @mentions, quoted replies, file sharing, and task management capabilities for AI Agents (such as Claude Code and Codex). It supports public group discovery and manual integration. Currently, event-driven interaction has been implemented for Claude Code, while Codex primarily relies on reminders and manual checks. The official website is pingo.xiaobiu.cn, and the source code is hosted at github.com/xiaoxuz/pingo.
OpenAI relaunched its $200/month Pro subscription at the end of September, but new subscribers received half the quota, while existing users' original quotas were only maintained until October 29; concurrently, a $500 Pro 500 plan was launched, with actual Astra usage increasing by only about 21% compared to the old $200 plan.
guilds.run has released an open-source AI agent team tool that supports debugging of each prompt, tool call, memory change, and agent interaction. It allows control over permissions, editing of context, selection of models for each agent, and supports free self-hosting.
Leju, in collaboration with Anhui Telecom and Huawei, launched the "Embodied Wing Link" 5G-A exclusive plan and dedicated SIM card for embodied intelligent robots, emphasizing high uplink bandwidth (up to 500Mbps) and low latency (20ms level); products such as Qiyuan Q1, Songyan Xiaobumi, and Yuandian Xiaoqiao have seen prices drop to 19999 yuan, 9999 yuan, and 8888 yuan respectively, entering the smartphone price range; manufacturers like Qiyuan and Songyan position themselves as "personal robots," integrating applications such as QQ Music and Ximalaya, listing on e-commerce platforms like JD.com, and holding Apple-style launch events at Shanghai West Bund Dream Center.
xvrzhao designed and implemented a multi-agent 'Who is the Spy' game based on LangGraph from scratch. Each player is an independent Agent with memory, differentiated personalities and voice tones (using MiniMax TTS), supporting multiple rounds of speech and voting mechanisms, as well as post-game open reviews. The project has been open-sourced on GitHub, providing one-click deployment via Docker Compose and a detailed README. It requires configuring LLM API KEY and MiniMax TTS KEY.
TRAE officially merged the former TraeCode (code side) and TraeWork (document/solution side) into a unified platform, supporting Agent mode (global workbench with multi-Agent parallel processing of planning, design, development, testing, delivery, etc.) and IDE mode (supporting fine-grained code editing), with one-click switching between the two modes; all outputs are automatically aggregated into 'My Outputs' and support sharing; real-world tests covered scenarios such as beginners building 3D dungeon editors, professional developers taking over legacy projects, and single-person delivery of complete iterations, while also verifying cross-platform invocation via Feishu/WeChat and mobile synchronization capabilities.
Benmo Technology listed on the main board of the HKEX on September 29, 2026, with stock code 06731.HK, becoming what the market calls the "first direct drive stock"; in 2025, 97.4% of its revenue came from robot power modules, of which 88% were consumer-grade direct drive modules with an average price of RMB 29, with shipments reaching 8.498 million sets; the prospectus shows it holds a 61.1% share in China's consumer robot direct drive module market, but this segment accounts for only 3.9% of the entire consumer robot power module market; the company is simultaneously advancing the R&D and mass production of wheeled-legged robots, selling 240 units in 2025 with a gross margin of 64.3%, but combined revenue from complete machines and joint modules accounted for only 5% of total revenue; customer concentration is high, with the top five customers contributing 85.7% of revenue, and public information indicates the largest customer is Dreame.
StarkWare co-founder and CEO Eli Ben-Sasson stated in a speech at TOKEN2049 and subsequent X posts that Starknet is considering transitioning from an Ethereum Layer 2 to an independent Layer 1. The main purpose is to autonomously control the process of network security upgrades, especially to address cryptographic threats posed by quantum computing and AI acceleration. He pointed out that Starknet already possesses some post-quantum technical advantages and cryptographic agility based on ZK-STARKs and account abstraction, but as an L2 it remains constrained by the upgrade pace of the Ethereum base layer. Starknet's own goal is to achieve full post-quantum security by 2027, earlier than Ethereum's target of late 2029. This discussion is currently in the stage of exploring technical roadmaps, with no formal implementation plan yet formed. Consequently, the STRK token price saw an approximately 26% increase over 24 hours.
Sui Network and Alibaba Cloud announced a partnership to integrate Alibaba Cloud services with Sui Agent Payments, enabling AI agents to pay stablecoins per call within preset budgets; this collaboration is currently in the settlement interface construction stage, with no launch date, service list, pricing, or pilot customers disclosed yet.
Aether AI released the CRIS-0 system, driving real robotic arms to achieve a 0.2-second safe emergency stop, interference replanning within an average of 2 seconds (9 recoveries out of 10 disturbances), long-horizon tidying in a living room, physical perception-based classification processing (such as distinguishing empty cans from unfinished drinks), and understanding of ambiguous instructions based on implicit causal variables (18 precise executions out of 20).
singer developed and open-sourced OpenPaste in native Swift, leveraging ample Codex and Claude quotas, because Paste frequently popped up on macOS and disrupted usage. Positioned as a free alternative visually and operationally aligned with Paste, the installer is under 10M and supports direct installation, forking for secondary development, code contributions, and local builds.
On September 24, 2026, Bitget suffered a hacker attack resulting in approximately $388 million in asset losses, with some stolen funds exchanged across chains via THORChain. On September 26, Bitget's CEO publicly requested that THORChain block the attacker's addresses; THORChain responded citing its decentralized nature. The core of the dispute lies in whether nodes possess precise blocking capabilities under its threshold signature mechanism, and whether technical intervenability equates to legal obligation. Combining FATF, IOSCO, and EU MiCA rules, the article points out that decentralization does not exempt responsibility, and liable parties must be identified on a case-by-case basis according to actual control, business function, and jurisdiction.
Ozon announced that it will make the "Ozon Полку" (shelf) function mandatory for approximately 80,000 self-pickup points across Russia, implemented in two phases: starting September 10, uncollected or canceled orders will be retained locally for resale; starting October, sellers can directly stock goods at pickup points, with warehousing and handling fees collected by the pickup point operators rather than the platform. This measure can shorten last-mile delivery times in Russian-speaking regions but introduces new local storage costs settled per location.
Developer lijiang0327 used Claude vibe coding to build the online background removal tool imgcutout.com. All image processing is completed within the browser, based on the open-source ormbg model, supporting WebAssembly (CPU) and optional WebGPU acceleration; it supports PNG/JPG/WEBP (single file ≤20MB), offering features such as pasting screenshots, changing backgrounds, brush correction, batch background removal (≤100 images ZIP download), GIF creation, and video-to-GIF conversion; it is free with no registration, no watermarks, and no usage limits; known limitations include automatic image resizing on low-memory devices, limited effectiveness on semi-transparent edges, and an English-only interface.