AWS SageMaker AI has launched a Multi-Turn Reinforcement Learning (MTRL) feature, supporting model training on Amazon Bedrock AgentCore or self-hosted infrastructure to complete multi-step agent tasks; using Wordle as an example, gpt-oss-20b improved its solution rate on the validation set from 61% to 84% after MTRL training, with training taking approximately 12 hours and 40 minutes, costing about $339 in training tokens and $79 in AgentCore runtime fees; all code is open-sourced in the aws-samples repository, and data and models remain entirely within the user's AWS account.
NVIDIA released the Global Performance Tuner feature in TensorRT 11.1, addressing challenges such as performance regressions, accuracy-performance coupling, and scalability issues faced by traditional heuristic scheduling across diverse models and deployment environments. This is achieved through three principles: a customizable build route, end-to-end accuracy constraints, and an expanded feasible solution space. The article defines the model space, solution space, runtime function, build algorithm, and search algorithm within a mathematical framework, modeling the compilation process as a composable sequence of IR transformations.
The AWS official blog released a reusable AWS CDK construct that resolves deployment failures caused by failing to deregister scalable targets before updating SageMaker AI real-time endpoint configurations, using three CloudFormation custom resources (Deregister, Register Scalable Target, Attach Scaling Policy) and explicit dependency ordering; this construct supports synchronized updates of endpoint configuration and scaling parameters in a single deployment, and includes a complete sample repository, configuration instructions, and integration methods.
Amazon QuickSight layer maps now support GeoJSON Point and LineString geometry types, allowing users to plot specific locations (such as stores) and paths (such as flight routes or delivery routes) on a map. GeoJSON files must be associated with QuickSight datasets via shared keys, supporting viewport locking, multi-layer overlays, and dynamic updates of GeoJSON files through S3 + API.
Koala is an open-source tool for Apple Silicon Macs that supports running AI agents, services, and development environments within ARM64 Linux virtual machines. Each VM has its own kernel, disk, and network, with permissions granted on demand. It supports installation via brew and provides commands such as koala run/create/start/exec. API keys are managed by the macOS login keychain and do not enter the VM.
AllenAI and the University of Pennsylvania collaborated to launch the LitPivot system, which supports researchers in dynamically revising proposal segments based on literature across three dimensions—problem, contribution, and evaluation—during the initial drafting of research proposals. Upon clicking any segment, the system retrieves relevant papers using Asta Paper Finder and groups them, offering citation-backed critical evaluations along with accept/edit/reject rewrite suggestions. A controlled experiment with 17 participants showed that proposals generated using LitPivot were rated by domain experts as more grounded in the literature, and participants self-reported stronger novelty, practicality, and understanding of related literature for their ideas. This work will be presented at UIST 2026.
The site owner released the video download tool xztool666.top/tools/video-dl, which supports platforms such as YouTube, Weibo, Xiaohongshu (RED), Bilibili, X, and Instagram. It can parse URLs from shared links or entire text posts; an accompanying iOS shortcut enables one-click downloading within Safari or clipboard-triggered downloads. The tech stack consists of an Astro frontend + Node + yt-dlp backend. It offers 3 free uses per browser; heavy users can top up on a pay-per-use basis, accessed via email.
"Sanxingdui: Future Past" will be released in theaters on October 23, 2026. It is China's first AI hyper-realistic theatrical film to receive a public screening license from the National Film Administration, with a runtime of 100 minutes and all characters being original digital avatars. Shengshu Technology's Vidu provides chief technical support, while Bona Film Group's AIGMS team is responsible for creative direction and execution of the film industry workflow.
An NVIDIA technical blog post released a tutorial demonstrating how AI agents such as Hermes, Nemotron 3.5 Bolt, and Astra collaborate to complete obstacle-walking training for the Microduck robot in Isaac Sim: first converting prompts into step-by-step plans, then having Astra generate USD assets and assemble the simulation environment, followed by training perception and motion policies using 4096 parallel environments in Isaac Lab, finally validating the policy on a local RTX Spark, and narrowing the sim-to-real gap through Better Actuator Models.
theHaagenDazs released Rectfly, an independently developed iOS double-pinyin input method. It supports Xiaohuo, Ziranma, Smart ABC, and Pinyin Jiajia, but does not support full pinyin. It relies on the locally running Wanxiang grammar model for full-sentence error correction; without downloading the model, the preferred accuracy is approximately 57%, which improves to approximately 97% after download (based on a standard double-pinyin full-sentence test of 30 sentences). All processing is completed on-device, with no backend or data collection. Currently, only a TestFlight beta version is available, requiring iOS 26 or higher and full access enabled.
Vericto has released vericto-proxy v4.15.3, an open-source proxy operating at the database wire protocol layer that can intercept destructive SQL issued by AI agents, scripts, or humans (such as DELETE statements without WHERE clauses or SELECT queries on sensitive columns), with built-in rule sets enabled by default; it supports PostgreSQL and MySQL and can be integrated simply by modifying the connection string host/port, requiring no code changes; licensed under FSL-1.1-ALv2, it allows free use by for-profit companies but prohibits use by competitors; features include quick trial via Docker, sensitive column policy control, and agent access whitelisting.
sideOS.ai is an AI agent tool that runs in the Mac notch area. It supports waking up via the Control+Option shortcut keys, can read the screen, open applications, click, type, and provide voice feedback on completion status; it supports fully offline operation with local models, or using the user's own Claude, OpenAI, Gemini, or Grok API keys.
"Broomy" v1.0 is a free cleanup tool for macOS 14 and later systems, supporting both Apple Silicon and Intel chips, with no internet connectivity and no ads.
This article analyzes the reasons for the lag in on-chain option development, pointing out that the core contradiction lies in the lack of low-cost hedging tools and capital efficiency incentives for option sellers; by backtesting the liquidation probability of Bitcoin longs under different leverage perpetual futures (a 47% liquidation rate for holding a 10x leveraged position for 180 days), it contrasts the time tolerance advantage of options; it argues that perpetual futures platforms like Hyperliquid are becoming option infrastructure, with projects such as Derive, Hypercall, and Paradex alleviating the seller supply bottleneck through unified accounts, portfolio margin, and external market maker integration; simultaneously exploring new paths such as embedding option pricing into Metric liquidity pools and Papertrade-style token incentives.
Bitcoin Core updated its security policy to immediately fix and publish security advisories for vulnerabilities that can be reproduced by large language models using simple prompts, no longer enforcing a vulnerability embargo of up to 1 year; this adjustment stems from a recent surge in AI-assisted security reports, including a Bitcoin red team using AI to scan 390 projects in one day and discover 4,962 potential defects, of which 85 were critical vulnerabilities.
AI coding tools have significantly shortened the cycle for retro game decomp/recomp (decompilation and recompilation), spawning numerous cross-platform porting projects, such as the iOS version of "The Legend of Zelda: Ocarina of Time"; however, the automated process weakens understanding of the original code logic, impacting open-source decompilation communities centered on collaboration, transparency, and knowledge accumulation.
MovoHub is an Apple TV and iOS player independently developed by xiaoxiongbaba. It supports integrating Emby, Jellyfin, Plex, Feiniu Video, WebDAV, STRM, 115, Quark, Aliyun Drive, Baidu Netdisk, and M3U/M3U8 IPTV playlists; it provides features such as unified watch progress, resume playback, EPG program guides, and multi-line on-demand viewing; the app does not provide content or collect user data; the Pro version is a one-time purchase of $24.99, with all current features available for free.
Ethereum developers have postponed the Glamsterdam hard fork on the Hoodi testnet, originally scheduled for Monday launch, after audit firm ChainSecurity discovered security issues with its gas changes: modifications in EIP-8037 and EIP-8038 could cause smart contract operations intended for retry to be permanently recorded as failures; 13 real-world cases have been identified. ChainSecurity is conducting responsible disclosure, while the Ethereum Foundation is assessing the number of affected contracts.
Ismail Iusifli open-sourced an independently extracted version of the accounting module from his online education platform (300+ monthly active students). It is a framework-free TypeScript library running on pure PostgreSQL. It supports high-concurrency lesson billing, replay/tamper-resistant payment Webhook verification, and enumeration-preventing login rate limiting. Correctness is verified through 104 real PostgreSQL integration tests, randomized invariant checks, and mutation testing. The code has been published on GitHub.
Anthropic collaborated with Theorem to have Claude, based on crash tests from 123 open-source libraries in OSS-Fuzz, convert machine code analysis into mathematical problems and automatically generate approximately 180 million lines of proof code; the Rocq tool automatically verified these proofs, requiring almost no manual review. While verifying PDF readers, Claude discovered security vulnerabilities equivalent to the annual report volume in one week and caught a logic error that had existed for over ten years. However, verification was limited to specific processor architectures, single-threaded programs, and inputs within 64MB, with each proof consuming hundreds of millions of tokens.
huangzhenmingsz has open-sourced sshfs-win-tray, an approximately 600KB single-file Rust-written graphical frontend for SSHFS-Win. It supports connection management, three authentication methods, specifying drive letters, auto-start on boot, and reconnection after disconnection, while completely avoiding calls to wmic or PowerShell. It depends on WinFsp and SSHFS-Win, is licensed under MIT, and the project is hosted on GitHub.
Mirendil released the coding-specific foundation model Mirendil-Base, with parameters ranging from 100 billion to 500 billion. The pre-training was completed in three and a half months by one researcher collaborating with AI agents, at a total cost of less than $10 million; all infrastructure, architecture, algorithms, and data pipelines were written from scratch by Agents, relying only on general libraries such as PyTorch/JAX; it scored 0.222 in cross-entropy tests, second only to Kimi K3 (0.193), but higher than MoE models such as GLM-5.3 and DeepSeek-V4-Pro-Base; the company stated that its Large and XL versions are currently being trained, targeting 2 trillion to 10 trillion parameters.