The Center for Security and Emerging Technology (CSET) at Georgetown University released a research report comparing Position Location Verification (PLV) based on signal round-trip time with on-site inspections as two methods for tracking AI chip flows. The study calculated detections per dollar across 10,500,000 simulated scenarios, finding that PLV was no worse than on-site inspection in all scenarios, with significant advantages when cluster sizes exceeded 11,000; PLV costs are amortized annually while on-site inspections are paid per occurrence, suggesting the two should be used in combination—PLV for broad coverage and on-site inspections for verifying suspicious points.
Starting October 9, Russia has opened a unified product card verification service for e-commerce platforms under the Platform Economy Law. Based on the National Unified Product Information System (ЕСНСИ), the service integrates data from 9 competent authorities and enables sellers and platforms to verify compliance requirements such as sales bans, labeling, and certification before listing products. The service will become mandatory after 90 days of operation.
Agendas and multiple roundtable discussions at TOKEN2049 Singapore indicate that the crypto industry's center of gravity is shifting from narratives to real assets, real utility, real revenue, and compliance execution. Institutions such as a16z, Fidelity, Hyperliquid, and Kamino noted that growth stems from practical applications like RWA tokenization and lending, with improved fund retention rates; institutions are increasingly focused on whether on-chain products can solve existing financial pain points and meet regulatory requirements.
oneisall announced the release of the independently developed app 'Tower', used to locally manage airport subscriptions and self-built nodes on iPhone and Mac. It supports traffic routing rule configuration and one-click generation of client-usable configuration files; it runs purely locally, is fully open-source, and has no servers; the Mac version is free to download, while the App Store version costs $1.99; the author stated that AI-assisted coding was used throughout, but emphasized that understanding requirements and promotion are the core barriers.
TypeSafe AI released Jev, a decision-making model focused on judgment tasks, on September 15, 2026. Within 3 days of launch, daily throughput reached 1 trillion tokens, and an $870 million Series A funding round was completed within 24 days, bringing the post-money valuation to $7.5 billion. Its input costs only $0.042 per million tokens, with output tokens permanently free; Microsoft, OpenAI, and Cloudflare launched Microsoft-Decision-1, Decisions API (powered by GPT-6 Luna), and the open-source model Clef respectively within 24 hours, all benchmarking against Jev's capabilities.
hashkitly tested the 24 browser-playable games that went viral on Twitter, confirming that versions of Quake trilogy, RTCW, Half-Life, Diablo, etc., compiled to WASM based on open-source engines can run directly; Halo CE and GTA:VC require users to import legitimate files; Skate 3 uses a Rust-rewritten engine with completely remade assets and relies on WebGPU; COD series and GTA5 browser versions are unreliable or non-existent.
Alexey Indeev released the open-source tool Chrome Relay, which supports coding agents such as Claude Code, Codex, Cursor, and Paseo to call the user's local Chrome browser via CLI (agent-browser), opening tabs with real login states silently in the background without stealing focus or popping up windows; the single-command latency is approximately 0.15 seconds. The toolchain includes a CLI, a local relay service (MV3 extension), audit logs, and identity management. It is limited to macOS, requires manual loading of the extension, and needs a one-time setup. The project is hosted on GitHub under the MIT license.
ShenNian has released an AI sharing feature that allows users to reuse locally logged-in AIs such as Codex and Claude Code. Without surrendering account credentials, users can generate dedicated agent links for visitors to ask questions; it supports login-free options and visitor conversation management, and is currently free.
One week before his Claude account was banned, whrss9527 independently developed and open-sourced 4 macOS utilities: Stox (a menu bar stock quote viewer supporting A-shares/HK stocks/US stocks, already listed on the App Store), Pop (a right-click radial quick action panel supporting Shell/JS/Shortcuts extensions), Proxi (a system proxy configuration and switching tool), and Meno (a menu bar icon organizer and scenario-based management tool).
The article analyzes the status quo of on-chain stock tokens: as of October 2026, the total exposure from tokenization paths ($3.21 billion) and synthetic paths ($3.01 billion perpetual contracts) exceeds $6 billion; it proposes a four-stage evolution model—from offshore claims, collateralization, native conversion within brokerage accounts, to direct on-chain listing of corporate shares; it points out that the current profit-making entities are front-end applications such as Robinhood, Binance, and xStocks, while issuers and custodians (such as Alpaca) barely make money; it emphasizes that price anchoring, reliability of weekend pricing, and acceptance of lending are key constraints for the next stage.
CLI Manager is an open-source (MIT license) macOS application that supports running Claude Code, Codex CLI, and Gemini CLI in parallel within the same sidebar; it provides project and git worktree management, session renaming, and one-click editor navigation features; a new AI Control API (REST + MCP, enabled locally only, disabled by default) allows AI agents to launch and drive terminal sessions, with users able to reclaim control at any time via Disconnect AI; free to use.
On October 10, Tencent Cloud open-sourced TeamAI (github.com/Tencent/teamai-cli). This tool uses Git to uniformly manage team Skills, work standards, tool configurations, and project knowledge, supporting 16 types of Agents including WorkBuddy, CodeBuddy, Claude Code, Codex, and Cursor. Submitting a Skill automatically creates a branch and initiates an MR; after merging, Agents with session start Hooks automatically pull updates when starting new conversations. Shared scopes include MCP tool configurations, Rules, and sub-Agents, with recently added capabilities for centralized distribution of model configurations. Official tests on 13 R&D tasks showed that introducing team knowledge improved the comprehensive score of low-cost models by 6.1% and reduced model invocation costs by 76%. TeamAI has been integrated into Tencent Cloud's intelligent agent management platform ClawPro.
Alibaba Qoder has officially launched Fast mode, available for desktop version 0.4.4 and above, covering Personal, Enterprise, Domestic, and International editions. The Domestic edition adds a Fast tier, while the International edition upgrades the former Performance mode directly to Fast mode without requiring user action. Initial response time is reduced from 8 seconds to 3 seconds, and credits consumption drops from 1.1x to 0.8x (a decrease of approximately 27%). In horizontal testing across 10 common developer task categories, Qoder Fast takes only one-third of the time compared to similar domestic products and one-fourth compared to similar international products.
Instinct是一款无需下载App、通过iMessage或WhatsApp注册即可使用的个人AI助理，用户以短信形式发起需求，Instinct在服务器端完成任务后以短信反馈；已官宣接入Shopify、Stripe等电商与支付平台，并推出I2I（Instinct to Instinct）功能实现AI间日程协商；2026年9月28日完成10亿美元C轮融资，投后估值达100亿美元，半年内估值增长约68倍；目前仍为邀请制内测，eBay上邀请码被炒至29美元。
Huawei Ascend AI chips are in short supply due to US restrictions and surging domestic demand. Xu Zhijun stated that Ascend's market share in China has exceeded Nvidia (based on Huawei's own statistics). Over 1000 sets of Ascend 910C have been deployed, and the 950 series is advancing commercialization by version. The indicative quote for the 950DT accelerator card has risen to over 250,000 yuan, and the quote for the 950PR has risen to over 80,000 yuan. Richard Yu acknowledged that advanced manufacturing processes still rely on DUV, while EUV equipment is still under development. On September 30, DeepSeek open-sourced software components for Ascend, including TileKernels and DeepGEMM-Ascend, supporting cross-platform reuse via Python interfaces and reducing migration costs.
REA (Reverse Engineer Anything) is an open-source reverse analysis toolbox designed for Agents. The latest version 6.1.0 supports the reverse analysis of Android APKs, iOS binaries, JavaScript/Electron applications, websites, .NET assemblies, EVM bytecode, etc.; it connects to Agents such as Claude Code, Codex, Cursor, and Gemini CLI via the MCP protocol; the project was developed by GitHub user morluto, with the original intent of providing callable reverse capabilities for Agents; the official documentation emphasizes that it is limited to legally authorized research purposes.
TikTok Shop plans to launch the 'Sell Across EU' feature on October 19. UK sellers will be able to sell to 12 EU countries: Germany, France, Italy, Spain, Poland, Netherlands, Belgium, Czech Republic, Austria, Greece, Portugal, and Hungary; multi-country EU sellers will also be able to sell to the UK.
Noah Dunnagan, a Customer Success Engineer at Railway, has open-sourced FSearch, a full-disk file search tool for macOS, claiming it is 17000 times faster than Finder. The tool is a command-line program accompanied by a daemon and API, requiring self-compilation, installation, and granting of Full Disk Access permissions. It supports advanced search syntax including filename, type, size, modification time, in-file text (grep/regex/sym), exact match, regex prefix/suffix, and exclusions.
Tesla has renamed Full Self-Driving (Supervised) to Tesla Assisted Driving on its official websites in multiple European countries including Germany, Spain, the Netherlands, and Slovakia, directly addressing concerns raised by the German Federal Ministry of Transport regarding misleading names; this move secured public support from the German government for FSD's entry into the European market, including recognition of its speed offset feature (not exceeding the speed limit by more than 10%) and promotion of approval at the EU level; during the same period, Tesla FSD has been approved in 8 EU countries, but major markets such as Germany, France, Italy, and Spain have not yet opened it to ordinary car owners; Musk thanked the German government in German on October 7, contrasting with his stance criticizing EU regulations in March of this year.
Shortly after Liu Huan's death, AI videos impersonating his wife appeared on social platforms. The first half simulated the widow mourning, while the second half promoted "Huangjing Wuhei Powder," receiving tens of thousands of likes; the account was actually forged, and Liu Huan's wife Lu Lu had previously stated that the couple had no social media accounts.
After upgrading to v2.87.10, Tencent Yuanbao PC launched a desktop pet feature, accessible by clicking the avatar in the bottom-left corner; this Beta feature supports proactive greetings based on time, drag-and-drop summarization of files/images/text (text dragging is being gradually rolled out), automatic digestion of long web pages with key point extraction, and anthropomorphic behaviors such as edge docking, silent companionship, and cold joke interactions, while providing adjustable parameters for activation conditions, reminder timing, and screen edge position.
Cloudflare has launched Clef-omni, an open-weight decision AI model built on Qwen3-Omni-30B-A3B-Instruct. It adds support for audio and full video inputs in wav, mp3, mp4, and webm formats, requiring no additional speech transcription or audio/video splitting. The model weights have been open-sourced to Hugging Face. Targeted at structured decision tasks, it does not generate conventional text. Additionally, multiple benchmark results and token pricing and context window adjustments for various models in the Clef series have been published.