The Cardano Foundation has introduced the CIP-0113 token standard, which allows issuers of compliant assets such as stablecoins, funds, and bonds to restrict recipients, freeze or confiscate held assets, and enforce identity and sanctions screening for all transfers without relying on hard forks; rules are selected or customized by issuers and can be updated in line with regulatory changes; tools such as Eternl and GeroWallet already support this standard.
Zit is a Git extension tool designed for simultaneous modifications to the same repository by multiple humans and coding agents (such as Claude Code, Codex, Pi, and Autohand Code).
FluidPD is an SLO-aware large language model serving system with prefill-decode (P/D) disaggregation. It introduces two mechanisms, FluidToken and FluidRole: the former offloads part of the prefill computation to decode workers when there is idle capacity on the decode side.
This paper proposes the GAMEGO framework, which transforms short game seeds into complete product requirement documents compliant with industry practices via task-specific dynamic compression. It constructs the GameGoData dataset containing 55,060 development trajectories and the GameGoBench benchmark with 124 queries. Based on this, the GameGoCoder model is trained, outperforming baseline models and rivaling frontier models in game development evaluations; code, datasets, and models will be open-sourced.
RadOnc-Agent is an agent framework designed for the entire radiotherapy workflow. It divides the clinical pathway into four stages and coordinates 26 callable functions through a large language model controller, supporting cross-stage workflow execution. In tests involving 2,600 single-function requests, 200 synthetic cross-stage scenarios, and 120 real patient workflows, the function selection accuracy reached 98.79%, with cross-stage process completion rates of 96.50% and 96.67%, respectively. Ablation studies indicate that longitudinal state maintenance and schema validation are critical to system reliability.
This paper reports on industrial experience with an event-driven cloud infrastructure for continuous machine learning training that has been running for three years in automotive manufacturing. The system triggers multi-factory GPU-accelerated training via manufacturing events, utilizing Amazon ECS and EC2 GPUs.
The paper proposes the Decision Operators method, training System-1 decision operators via behavior cloning to reduce the cost of injecting skills into frozen language models from millions of parameters to 330K (matching RL-trained 1.33M parameter operators) or even 23K (rank-4 variant).
The paper proposes medical image alignment assessment as a new task for testing the general visual reasoning capabilities of frontier multimodal large language models (MLLMs), examining their ability to judge anatomical consistency without task-specific fine-tuning; GPT-6 achieves over 85% accuracy in most scenarios, while models released several months ago perform near random levels, and fine-tuned local models show significantly worse generalization.
EasyToSTL is an AI tool for 3D printing users that supports generating watertight STL files from JPG/PNG/WebP images and text prompts, directly importable into mainstream slicers such as Bambu Studio, Cura, and PrusaSlicer; it offers a free trial, credits upon registration, and requires no credit card; outputs support STL/GLB/OBJ formats, with the Pro version supporting up to 1.5M triangles and millimeter-level precision; users self-report zero repair warnings for generated models in Bambu Studio and Cura.
ImgKit is a set of image processing tools that run locally in the browser, offering a total of 62 features including compression, resizing, format conversion, background removal, face blurring, and text recognition. All processing does not upload files, requires no account, has no watermarks, and supports offline use; the core model (26 MB) is downloaded once and cached for reuse, supporting JPG, PNG, WebP, AVIF, and iPhone HEIC formats, and is free for files up to 50 MB.
nanoMuse is an open-source personal agent supporting Android, iOS, macOS, Windows, and Linux. The latest version, 0.1.40, includes a fix for the iPhone input box.
cvcat released the open-source command-line tool catbus, supporting operations such as search, details, comments, users, recommendation feeds, downloads, publishing, direct messages, and live stream bullet comments across 10 platforms: Xiaohongshu, Douyin, TikTok, Bilibili, Kuaishou, Weibo, Xianyu, Taobao, JD.com, and X. The command structure is unified (e.g., catbus xhs item search 露营), output is in JSON format, it relies only on Node, supports multi-platform architectures, and ensures local isolation of login states with no telemetry. The capabilities originate from cv-cat's 10 open-source repositories, all rewritten in TypeScript. GitHub address: https://github.com/cv-cat/catbus, MIT license.
Igloo Inc., the parent company of Pudgy Penguins, announced that its Ethereum L2 network, Abstract, will shut down its mainnet on December 15, 2026, after which unmigrated assets will become inaccessible. The chain operated for 18 months, incurring losses of tens of millions of dollars. Despite having 400,000 users, 4 million wallets, 144 applications, and brand partnerships with Red Bull Racing and Disney, it failed to establish a DeFi ecosystem or generate sustainable revenue. Igloo chose not to issue a token to extend its lifespan, instead focusing on Pudgy Penguins and the PENGU token. Concurrently, Blast also shut down, highlighting the industry reality that L2 projects struggle to survive whether they have 'funding but no users' or 'users but no funding.'
Zizq is a zero-dependency, single-binary persistent background task queue server based on an embedded LSM database, requiring no Redis or relational databases; it provides HTTP/1.1 and HTTP/2 APIs.
Independent developer dongsanxiang1213 released the browser start page "Qingqi er" (qingqier.com), which is now available in the official Microsoft Edge and Google Chrome extension stores. It features an ad-free minimalist interface, one-click access to mainstream AI tools such as DeepSeek, Kimi, Doubao, and Tongyi Qianwen, default search engine switching, and deep customization options including themes, wallpapers, weather, and layout density. A new account system enables bookmark categorization and custom navigation, while keyboard shortcuts allow focusing on the search box.
offstage provides a second logged-in macOS account runtime environment for Claude Code, Codex, and opencode. This allows simulators, Xcode UI tests, and applications built by agents to open on that account's desktop, keeping the main account's screen, keyboard, and mouse undisturbed. The project is free and open-source, with code hosted on GitHub.
YAVCHN is an open-source Hacker News reader supporting keyboard navigation (j/k/Enter/]/p/n/c/f, etc.), multi-window article reading, comment collapsing, Pinned story filtering, note addition, collection management, user reply tracking, domain blocking, and site-wide link search; data is stored locally by default, with cross-device synchronization of pins/collections/notes available after login; the project is open-sourced at github.com/paulmooreparks/yavchn and built on PUDL.
TerrainSR is an open-source model released by an individual developer that upscales terrain heightmaps from 100m resolution to 10m resolution. The training data comes from institutions such as swisstopo, Kartverket, IGN, and USGS, excluding built-up areas. Processing a 50×50km area takes only 0.74 seconds on an RTX 4070. It supports GPU/CPU inference; inputs are 100m heightmaps with corresponding 10m water masks, and outputs are float32 height arrays. Licensed under Apache 2.0.
Qualcomm filed a lawsuit against Arm in Delaware state court, alleging that Arm failed to deliver chip testing tools and leaked a letter terminating the license, thereby damaging Qualcomm's partnership with Meta. Qualcomm is seeking a suspension of royalty payments for the next 5 years (potentially amounting to billions of dollars) and demanding that Arm fulfill its obligation to negotiate in good faith regarding the V10 instruction set architecture. Arm has fully denied the allegations, stating that Qualcomm is using litigation as leverage and that no actual damages have occurred.
According to The Information, Microsoft has cut its internal Claude budget for 2026 by more than one-third, canceling most Claude Code licenses before June 30; the number of Meta employees using Claude Code internally has dropped from approximately 60,000 to about 30,000, with spending still exceeding $105 million within 28 days.
OpenAI has released textGrain, a text watermarking technology that embeds key-controlled statistical signals into generated text using an entropy-calibrated optimal transport algorithm. This achieves high detection rates (over 90% for 400-token psychology responses) with almost no loss in model performance (5 out of 8 benchmarks showed slight improvements on GPT-6 Astra). The detector is not yet available to the public, but textGrain itself will be open-sourced. The technology was jointly developed by Peking University's School of Mathematical Sciences, the University of Pennsylvania, Yale University, and OpenAI.
NavGPT-3 is an embodied agent architecture that connects large language model reasoning with low-latency action policies. Its OS-like runtime supports threaded reasoning, execution, and monitoring; the underlying action policy, NavGPT VLA, achieves 74.51 SR on R2R-CE and 78.19 SR on RxR-CE; the complete system achieves 81.51 SR on R2R-CE and matches human success rates (90.43 vs. 90.4) and path fidelity (78.47 vs. 77.7 nDTW) on RxR-CE at 1 minute 22 seconds per episode; the authors will release all models, code, and evaluation records.
Long-WAM is a world-action model framework for real-time robotic control that improves the utilization efficiency of long historical visual contexts (up to 19.2 seconds) through autoregressive video pre-training. It increases task success rates from 63.3% to 78.7% on RoboCasa GR-1 and outperforms comparative methods on LIBERO-Long, RoboTwin 2.0, and DOMINO. It supports deployment on RTX 5090, DGX Spark, and Jetson AGX Thor, with single-step action inference including future video latent representation prediction taking 107.4 ms. It achieves a 95% success rate on dynamic cup stacking tasks using Unitree G1 and YAM robots.