You’re reading an archived edition. Today’s is here →
The Slovak web crawl just exposed a half‑percent farm of AI‑spun pages.
That’s the night, see you tomorrow
An LLM‑driven agent audited the 26.5 M‑document Slovak portion of FineWeb2, uncovering a machine‑generated content farm of 1,408,437 documents (5.32% of the corpus).
The agent operated via a read‑only, budgeted API and compiled findings into executable tests validated by native speakers. Three runs each identified seven issues across four phenomena. The full‑corpus scan required no GPU and ran in under a second, demonstrating a cheap, scalable debugging approach.
Four open‑source LLMs (8‑27 B parameters) were evaluated on the Spanish MIR residency exam (1,552 questions), with the three larger models ranking roughly between 600 and 3,000 candidates and leaving 41‑43 specialties open.
The study examined model uncertainty, clinical reasoning, and image usage, finding that abstention never improved outcomes because even the least confident answers exceeded the 25% break‑even threshold. Results suggest open models can be competitive but still lack reliable self‑assessment.
Anthropic announced a program to donate agent skills to the Agentic AI Foundation.
The initiative aims to share capabilities developed by Anthropic's agents with the broader community, fostering collaborative advancement of agentic AI. No further technical details were provided in the announcement.
VERA is an open‑source self‑hosted dashboard that scores LLMs against twelve EU AI Act technical requirements and records energy consumption per benchmark run, deployed at a large European bank.
The tool builds on the COMPL‑AI interpretation of the EU AI Act, organizing evidence into six principle groups and eighteen technical requirements, of which twelve are currently benchmarkable. VERA simplifies the workflow for non‑engineer reviewers by providing a UI for model selection and benchmark execution.
DeepMind released a global map that classifies every 10‑meter patch of land into distinct forest types.
The dataset provides a fine‑grained view of forest distribution worldwide, enabling ecological research and policy planning. It represents a step beyond traditional tree‑cover maps by distinguishing among multiple forest categories.
OpenAI introduced the Decisions API, offering fast, low‑cost decision‑making calls for application developers.
The API is designed to deliver focused predictions with minimal latency and expense, targeting use‑cases that require frequent, lightweight decisions rather than full‑scale language generation.
Anthropic is inviting guest investigators to propose life‑science research projects.
The open call targets external researchers to explore applications of foundation models in biology and medicine, offering access to Anthropic's resources and potential collaboration opportunities.
OpenAI and Poland’s NASK have formed a partnership to work on AI safety, security, and technical research.
The collaboration aims to develop safeguards and strengthen the digital ecosystem, leveraging NASK’s expertise in cybersecurity and OpenAI’s AI capabilities.
OpenAI outlines its forward‑deployed engineering (FDE) model for rapid, on‑site AI system development.
FDE teams embed engineers directly with product groups to iterate quickly, focusing on safety, reliability, and performance in production environments.
IBM released Granite time‑series models for ensemble forecasting tasks.
The models are hosted on Hugging Face, support probabilistic forecasts, and are intended for applications like weather prediction and demand planning.
Multiple Linux distributions issued security updates covering dozens of packages on Thursday.
Updates address vulnerabilities in runtimes, web servers, and system libraries across AlmaLinux, Arch, Debian, Fedora, Gentoo, Mageia, Oracle, and Red Hat, requiring administrators to apply the patches promptly.
Hackios provides a web tool that runs full iOS 16.5 firmware on Android devices.
The service emulates the iOS environment, allowing developers and hobbyists to test iOS apps without Apple hardware, though performance and compatibility are limited by Android hardware constraints.
At Kangrejos 2026, Josh Triplett outlined upcoming Rust kernel features such as type system refinements, syntax tweaks, and a smarter borrow checker.
The talk highlighted how these enhancements could simplify kernel code and enable new patterns, but they remain work‑in‑progress within the Rust project. Developers are encouraged to track the Rust roadmap for implementation timelines.
A 48‑part patch series by Linus Torvalds removes __get_user() and __put_user() primitives from the Linux 7.4 kernel.
The cleanup eliminates long‑standing user‑space memory access helpers that have been superseded, reducing kernel surface area. The change required careful coordination due to the extensive codebase that still referenced these functions.
KeePassXC 2.8.0 adds Qt6 migration, Wayland auto‑type, and ships native ARM64 builds for Windows and an experimental AArch64 AppImage for Linux.
The release also introduces configurable shortcuts, remote sync support, and quick unlock options on macOS and Linux via Polkit. These updates address long‑awaited feature requests and improve cross‑platform performance.
Hackaday details an open‑source linear capacitive encoder design derived from digital caliper technology.
The project provides schematics and firmware for a low‑cost, high‑resolution position sensor that can be fabricated with standard PCB processes. It targets hobbyists and embedded developers needing precise linear measurement.
The paper defines L4, a reversible, linearly‑typed intermediate language with a bidirectional small‑step semantics for compiling reversible programs.
L4 generalises the earlier RSSA form and draws inspiration from HSSA, offering direction indicators that allow forward and backward execution to switch dynamically. This facilitates compilation to reversible hardware.
A new concurrency fuzzing approach separates input generation from thread‑interleaving exploration, using May Race Pairs to focus scheduling on promising inputs.
By decoupling these phases, the framework reduces wasted scheduling effort on low‑yield inputs and improves detection of real data races in the Linux kernel. MRPs are collected during unscheduled runs and guide targeted mutations.
Graph-Based Over-Representation Retrieval-Augmented Generation (GBOR-RAG) unifies multiple gene-set databases into a term‑gene graph and extracts a compact evidence subgraph to condition an LLM for biologically faithful enrichment interpretations.
The framework first scores over‑representation on the unified graph, then retrieves a connected subgraph that preserves cross‑database links and gene‑level support. A claim‑level faithfulness metric evaluates whether generated statements are directly backed by the retrieved evidence.
Evolving Federated Density Peak Clustering (eFDPC) combines fuzzy data‑stream clustering with density‑peak methods to enable a one‑shot federated clustering that limits client‑server exchanges and preserves unlinkability.
Each client builds micro‑clusters locally and shares only anonymized summaries via a shuffler, preventing attribution of data to any single participant. The approach aligns with single‑pass streaming algorithms, making it suitable for privacy‑sensitive distributed environments.
TT‑SAX redesigns Symbolic Aggregate approXimation to include directional information at appropriate scales, addressing three limitations of prior symbolizations for natural‑language shape queries on time‑series databases.
The new two‑track design adds explicit direction symbols, aggregates direction over longer intervals, and replaces arbitrary letters with meaningful tokens, improving interpretability for plain‑language queries that ask how a series moves.
A deterministic constant‑round LOCAL algorithm achieves a (d+1+ε) approximation for Minimum Dominating Set on bounded‑average‑degree graphs, yielding (7+ε)‑approximation for planar graphs and (5+ε)‑approximation for bounded‑Euler‑genus graphs.
The result extends near‑optimal guarantees from minor‑free to broader bounded‑expansion classes, improving on prior work limited to proper minor‑closed families. The algorithm runs in constant rounds regardless of graph size, making it practical for distributed settings.
Meloong‑Git/PCL is a Chinese‑language Minecraft launcher named Plain Craft Launcher.
The repository provides the source code and binaries for launching the Minecraft game with a custom user interface, targeting users who prefer a lightweight, locally‑run launcher.
NanmiCoder/MediaCrawler is a Python tool that scrapes comments and posts from platforms such as Xiaohongshu, Douyin, Kuaishou, Bilibili, Weibo, Baidu Tieba, Zhihu, and others.
The crawler supports multiple media sites, extracting both content and associated comment threads, enabling large‑scale data collection for analysis or archiving. It is open‑source and configurable for different target platforms.
Apache Iggy is a Rust‑based message streaming platform marketed as hyper‑efficient and low‑latency.
Iggy aims to provide faster data pipelines than traditional brokers by leveraging Rust's performance and zero‑copy techniques. It targets use cases like real‑time analytics and event sourcing. The project is open source and part of the Apache incubator.
deepseek-ai/3FS is a C++ distributed file system built for AI training and inference workloads.
3FS focuses on high throughput and low latency to handle large model checkpoints and dataset shards. It integrates with common AI frameworks and supports scaling across multiple nodes. The system is open source and positioned as an alternative to existing cloud storage solutions for AI pipelines.
electron/electron enables developers to create cross‑platform desktop applications using web technologies.
Electron bundles Chromium and Node.js, allowing a single codebase to run on Windows, macOS, and Linux. It powers many popular apps such as VS Code and Slack. The project continues active development with frequent releases.
google-pytorch/torch_tpu provides PyTorch support for TPU hardware.
The repository adds backend kernels and integration layers so PyTorch models can run on Google's Tensor Processing Units. It aims to bring the familiar PyTorch workflow to TPU‑accelerated training and inference, complementing existing XLA support.
28 stories from 199 scanned overnight.
Get it every morning: RSS feed.