After being fired, three OpenAI safety researchers—Tomek Korbak, Mikita Balesni, and Jasmine Wang—denied leaking confidential information, stating they had not received written explanations detailing the specific content of their alleged violations. They questioned whether the dismissals were related to their advocacy for transparency in model reasoning processes, noting that OpenAI's safety team had launched an investigation citing Signal chat logs, while Signal is considered standard security practice in Silicon Valley. The core of the controversy lies in whether AI companies should allow external review of models' chain-of-thought capabilities; the three had co-signed a 2025 paper calling for continued openness of this capability. The incident originated from a September 1 report by The Information stating that OpenAI's new model Astra had less transparent reasoning processes, after which OpenAI initiated an internal leak investigation. Nearly half of the on-site audit reports conducted by METR and Redwood Research on OpenAI were redacted.
Screen Fasting is a focus tool for iPhone and iPad that leverages the iOS 17+ Assistive Access feature to generate random passcodes that users cannot memorize, switching devices to a simplified mode with only a few preset apps and preventing exit before the countdown ends; the free version supports up to 5 sessions per month with a maximum duration of 8 hours, while the Pro version supports durations up to 30 days and includes weekly Screen Time reviews by real people; it is not compatible with Android.
Shotstash is a self-hosted media cloud tool designed for creators and small teams. It supports deployment on PCs, NAS devices, or idle hard drives, offering video-first storage, sharing, and collaboration features. Licensed under MIT with no paid tiers, it supports local disks, NAS mounts, or any S3-compatible storage backend (such as AWS S3, Cloudflare R2, MinIO). It includes a reference worker for generating 720p video proxies and exposes standardized job contracts for users to integrate custom AI workflows (such as transcription, scene detection). The tech stack consists of Next.js 16, React 19, TypeScript, PostgreSQL, and Docker. The project code is hosted on GitHub, has been running in production environments, and is continuously maintained.
The open-source project deepseek-web-retrieval by hugo54 implements engineering integration for the DeepSeek Search&Fetch API, supporting four major Coding Agents: OpenCode, Pi, Codex, and Claude Code. It offers capabilities such as domain allowlists/blocklists, automatic compilation of language preferences, body length clamping with independent timeouts, retrying transient failures, preview and rollback of configuration changes, and keeping keys off disk. The implementation has been validated through real end-to-end sessions and automated releases.
zkitkevin launched the English AI fanfic generator fanficwizard.com within 2 days. It supports generating complete scenes by inputting work, characters, tropes, tone, and point of view; it offers 3 free generations per day without registration. The tech stack is Next.js 16 + Postgres (Neon) + Prisma + Vercel. Models are called via an OpenAI-compatible gateway with tiered routing. Payment processing has been switched to a MoR proxy. Issues regarding www redirection, content safety, and anti-abuse have been resolved.
HSBC has launched RedCoin (RCHKD), a HKD stablecoin licensed by the Hong Kong Monetary Authority (HKMA), on its PayMe app. Over 3.3 million PayMe users can soon activate wallets for peer-to-peer transfers or payments at more than 100,000 PayMe merchants; this stablecoin runs on HSBC's proprietary private ledger, currently does not support interoperability with the HSBC Mobile Banking App, and is not protected by the Hong Kong Deposit Protection Scheme; HSBC plans to integrate it into the HK Mobile Banking App later this year and support redemption of tokenized shares of its HKD money market fund.
Amazon CEO Andy Jassy stated on the a16z podcast that AWS capital expenditures are projected to reach $220 billion by 2026. However, compute capacity for 2026 is already sold out, 2027 is largely booked, and some allocations for 2028 have also been locked in. He refuted misconceptions such as high water consumption by data centers, emphasizing that AI demand follows a 'dumbbell shape,' noting that core enterprise production systems have not yet deeply adopted AI primarily because inference costs have not decreased by an order of magnitude. AWS is reducing costs and reshaping its organization through self-developed Trainium chips, the Nitro system, and rewriting the Bedrock engine with a team of 6 people over 76 days. Simultaneously, it is partnering with Anthropic and OpenAI to drive the coordinated implementation of models and compute resources.
The article explores the identity crisis of 1024 Programmer's Day against the backdrop of widespread AI adoption, noting that programmers are shifting from hand-coding to managing AI agents, verifying outputs, and signing off on results. Citing Anthropic's employment data, DHH's abandonment of hand-coding practices, and the historical analogy of NASA computers, it illustrates that the essence of the profession has fundamentally migrated, with the festive atmosphere masking the dissolution of the original group's identity.
Developers discovered that Tensor Cores in Ampere architecture GPUs such as the RTX 3090 and 3080 have artificial throughput limits in AI matrix operations like BF16/TF32. They achieved software-level unlocking through two new open-source projects for Linux and Windows. Measured results show RTX 3090 TFLOPS increased from 75 to 135 (+80%), while a modified RTX 3080 20GB version saw approximately double performance in specific matrix calculations. The Strata inference engine showed a 16% speed increase for processing 16K prompts, but no significant change was observed for llama.cpp quantized models. Unlocking on Windows requires UEFI-level operations and has currently been verified only on limited hardware.
Starting October 14, Ozon notified sellers in the backend to discontinue commercial freight forwarding dedicated lines such as CEL and GUOO; newly opened stores and existing stores that have never bound partner channels can no longer add them. Bound channels can be used until October 31, with a uniform switch to self-operated OGL starting November 1, supporting one-click transfer. Archived existing channels cannot be restored, and new entry points are closed. The postal WH channel is unaffected, and the entry point for lightweight items under 500 grams will be retained long-term. The XY and OYX delivery methods were taken offline early on October 21.
Tencent WorkBuddy has integrated Moonshot AI's Kimi-K3.1-Pro model, simultaneously supporting the high-speed version Kimi K3.1 Pro Highspeed, with output speed approximately 2-3 times faster; this model features 3 trillion parameters, supports a 1 million context window, increased reinforcement learning compute investment, scoring 65.7% on Terminal-Bench 4.0, 92.7% on BenchCAD, and achieving SoTA on both OfficeQA Pro v2 and SpreadsheetBench v2.
The Strata terminal version 0.2.4 update supports running the DeepSeek V4.1-Flash model locally, with minimum hardware requirements of an RTX 3060 12GB graphics card + 46GB memory + 239GB high-speed NVMe SSD; it uses a "hot expert caching" mechanism to handle the MoE architecture, offering three quantized versions: IQ1_M (239GB), IQ2_XXS (257GB), and IQ2_XS (274GB); measured generation speed on an RTX 3060 is 8–10 tokens/s; limited to the terminal version, the GUI version does not yet support this; the accompanying Terminal 0.2.4.1 has fixed installation lag issues.
Ethereum Foundation researcher Justin Drake warned that AI might crack Elliptic Curve Cryptography (ECC) earlier than quantum computers, calling for the industry to enter "bunker mode," i.e., transferring assets to new wallet addresses where public keys have never been exposed; Vitalik Buterin expressed support but emphasized that there is no need for panic.
hisea released the open-source Rust project Radian, a lightweight, headless Agent tool with built-in orchestration capabilities, supporting plan/task separation, automatic review, and rework mechanisms; the first unattended task cost $0.33; it provides brew installation, and the project address is https://github.com/hisea/radian.
Developer RPD1986 created an Agent Skill based on McDonald's open MCP capabilities to solve group ordering coordination issues. It supports mechanisms such as proposal-first, dual-pool quotas, dietary restriction grading, and three-level fallback. The entire process calls official MCP for real price calculation and coupon claiming; after confirmation, orders can be placed directly. The project is open-sourced on GitHub and supports usage in any MCP-compatible AI client by loading SKILL.md.
Huxiu's analysis points out that global AI data centers (AIDCs) are undergoing an unprecedented trillion-dollar investment wave. NVIDIA's DSX POD modular solution compresses deployment time to 14 days, but this solution only covers pre-integration; civil engineering, power supply approvals, etc., still require several years. Third-party data shows that total computing power investment in 2026 will reach $830 billion, a year-on-year increase of 79%, but more than 95% of Grace Blackwell GPUs have not yet been put into use. Among four business models—colocation, GPU leasing, scheduling services, and model invocation—currently, only some data center operators have seen rising renewal rates; CoreWeave has substantial revenue scale but a net loss of $626 million; overall capital expenditure by cloud providers far exceeds cloud revenue; and AIDCs generally have not yet achieved cash payback. The article argues that the only certain winners at present are equipment and factory suppliers like NVIDIA.
Retypit is a browser-based AI image text editing tool that supports uploading PNG/JPG/WebP files (up to 20MB). It automatically recognizes each line of text in the image; users can select text and enter new content, after which the AI generates new text matching the original font, color, and texture while keeping the background and layout unchanged. Currently, free use is available upon logging in with Google or email.
Huxiu analyzes the version change logic of the Artificial Analysis (AA) model evaluation leaderboard, pointing out that multiple updates from v4.1 to v4.3.2 led to a reduction in GDPval weights, recalibration of the question bank, and discontinuous Elo benchmarks. This caused models like Kimi K3 to score 57 points in July but only 44 points currently, and DeepSeek V4 Pro to drop from 44 points to 36 points. The leaderboard's exam syllabus is highly focused on US workplace scenarios, disconnected from Chinese workflows, yet it is widely cited domestically as a general report card.
margrop released the open-source mistake book project Smart Wrong Notebook, forked from wttwins/wrong-notebook. It supports photo recognition with multi-question box selection per image, AI redrawing of SVG illustrations, step-by-step SVG animation explanations, GeoGebra demonstrations, single-question extension exercises, AI paper generation and printing, 6-digit source ID tracking, and a closed loop for answer backfilling. It directly connects to domestic large models such as MiniMax-M3.1-Flash, deepseek-flash, glm-5.3-flash, and kimi-k3. It offers Docker, desktop versions, fnOS, and an Android APK standalone version (75MB) with built-in server, as well as a client version (2MB), all under the MIT license.
The article points out that the development of embodied AI has spawned numerous humanoid robot data collection centers, but their core model involves robot companies selling equipment to government-enterprise co-built data collection centers and then repurchasing the data, creating a closed loop of inflated book revenue. With stricter regulatory inquiries, the rise of body-less collection, halved data prices, and the bursting of industry valuation bubbles, this model is unsustainable. Currently, data collection is shifting from heavy-asset simulation fields to distributed real-scenario collection, but faces issues such as data duplication, questionable quality, labor chaos, and ethical concerns regarding human labor data.