You’re reading an archived edition. Today’s is here →
You’d think the AI world sleeps, but we got a model duel, a speed hack, and a security scare.
Catch you tomorrow, friend (wink)
Fine‑tuned Gemma‑3‑12B was evaluated against GPT‑4o for multi‑label intracranial hemorrhage extraction from 100 expert‑adjudicated radiology reports.
The study used a 2×2 design crossing discriminative heads and instruction fine‑tuning with two data sources (distilled GPT‑4o labels and synthetic GPT‑4o reports) across five training set sizes. Results were benchmarked against both GPT‑4o and the untuned base model to assess privacy‑friendly alternatives.
EviRover introduces an agentic perception system trained on 5K SFT and 12K RL examples to handle queries requiring information beyond a single image glance.
The authors created two data pipelines to generate training data for perception under insufficient evidence and released the human‑verified EviLens benchmark covering five perception categories. EviRover is positioned as the first system to actively acquire additional visual evidence during inference.
Looped‑DiT adds deep supervision and self‑modulating attention to looped Transformer blocks, enabling iterative refinement without increasing parameter count.
By reusing the same Transformer layers across denoising steps, the method increases computational depth while keeping model size fixed. Naïve looping degraded image quality due to weak intermediate supervision, which the authors address with supervision at each loop and attention regulation.
OpenAI disclosed that it disrupted a coordinated effort to extract protected model reasoning and is enhancing defenses against adversarial distillation.
The blog post outlines the tactics used by the adversarial group and describes new safeguards OpenAI is deploying to prevent future model‑distillation attacks, emphasizing the ongoing security arms race in large‑scale AI deployment.
Magnitude is an inference engine for agents that claims up to 2× speed improvement over llama.cpp on typical hardware.
The engine auto‑optimizes for the host’s CPU/GPU configuration and runs on macOS, Linux, and Windows. It targets single‑session workloads where existing engines trade off batch performance or hardware specificity.
PSSA is a non‑transformer language model written from scratch in Rust, released as an open‑source repository.
The project showcases a Rust‑native implementation of a language model architecture that avoids transformer components, aiming for performance and safety benefits inherent to the language. Source code and documentation are publicly available.
Barclays is scaling Anthropic's Claude model to upgrade its operations and improve client experience.
The deployment aims to enhance internal workflows and customer interactions, leveraging Claude's conversational capabilities across banking services. No specific model size or timeline is disclosed.
DeepMind announced Gemini 4 Argon as the next era of frontier intelligence.
Gemini 4 Argon follows previous Gemini iterations, targeting more advanced reasoning and multimodal abilities. The blog post provides high‑level goals but no quantitative benchmarks.
OpenAI is partnering with the U.S. Small Business Development Center network to deliver hands‑on AI training and local support for small businesses, accompanied by a new usage report.
The collaboration will provide workshops, technical resources, and case studies to help small teams adopt AI tools. The accompanying report surveys current AI adoption patterns among small enterprises.
DeepMind introduced SynthID Bio, a proof‑of‑concept system for watermarking AI‑generated proteins while preserving their biological function.
SynthID Bio embeds invisible identifiers into protein sequences, enabling provenance tracking without altering activity. The approach could help mitigate misuse of synthetic biology outputs.
All videos from the recently concluded Kernel Recipes conference have been posted online, along with caricature drawings of attendees and speakers.
The archive provides full access to the conference talks covering kernel development techniques. The accompanying artwork adds a visual recap of the event.
The Linux Foundation Technical Advisory Board election for 2026 will accept nominations until October 7 to fill five seats, including one vacancy left by Dan Williams.
Voting will occur electronically after the Linux Plumbers Conference. Serving on the TAB offers influence over kernel development policy and community direction.
The 2026 Python Language Summit took place on July 14 in Kraków, Poland, covering topics such as the memory buffer protocol, garbage collection, free‑threaded Python, and the project "Spicycrab".
The summit produced a series of reports that capture the discussions and proposals presented. These reports are now publicly available for developers interested in upcoming Python enhancements. The event highlights ongoing efforts to improve Python's concurrency and memory handling.
Security updates were released for multiple Linux distributions on Wednesday, including AlmaLinux, Debian, and Fedora, with several packages receiving new versions.
The updates address vulnerabilities across a wide range of components, from web servers to desktop environments. Fedora's list spans core libraries, GNOME applications, and system utilities, reflecting a broad security sweep. Users should apply the patches promptly to maintain system integrity.
Sharon Yang presented a comparison of Chromium development practices at Google and Igalia at FOSSY 2026, highlighting operational differences without focusing on complaints.
The talk covered how each organization structures its codebase contributions, review processes, and release cycles. Yang emphasized the cultural and workflow contrasts that affect developer experience. Attendees gained insight into the varied environments shaping Chromium's evolution.
At Akademy 2026, Marco Martin announced that KDE Plasma will drop X11 support, marking the biggest upcoming change for the desktop environment.
The transition aims to streamline development around Wayland, improving performance and security. Additional plans include enhanced remote‑desktop capabilities and incremental feature refinements. Users are encouraged to test the Wayland session ahead of the X11 deprecation.
The Darpan framework introduces a digital twin that continuously mirrors a running application, enabling real‑time evaluation of placement, scheduling, and recovery decisions.
Physical execution observes runtime state while the digital counterpart simulates alternative actions, addressing the lag between decision making and execution in distributed environments. This approach differs from static simulations by adapting to live conditions, potentially improving resource utilization across edge and cloud layers.
A study of EPR‑pair consumption in distributed quantum computing shows that identical EPR costs can yield widely varying execution times, and lower‑EPR mappings may be slower.
The authors evaluated balanced two‑QPU mappings of QFT, QAOA, and CDKM circuits, finding that runtime depends heavily on workload and communication latency. In regimes with high EPR‑generation latency, the disparity between EPR cost and actual runtime becomes more pronounced, questioning the reliability of EPR count as a sole metric.
FissionReady schedules SMR-powered data centers to cut water use by about one‑third and grid electricity cost by roughly half while avoiding any involuntary shutdowns or batch deadline misses.
The scheduler monitors each reactor module's fuel age, xenon state, and flexibility, then runs a two‑timescale optimization that coordinates load‑following, batch deferral, and grid purchases. Compared to a baseline of equal size, the approach also reduces water consumption and eliminates missed batch deadlines, making SMR integration more reliable.
The paper benchmarks how directed‑acyclic‑graph topology influences task‑level CPU and memory intensity predictions for large‑scale cloud workflows.
While prior work focused on task‑specific features, this study systematically varies graph structures to quantify their effect on prediction accuracy, highlighting cases where topology alone can improve provisioning decisions and reduce over‑ or under‑allocation.
CCV is an LLM‑assisted framework that automatically generates machine‑checked assurance cases linking C code to its requirements.
CCV builds interface protocols describing allowed call sequences and resource assumptions, then produces conditional safety guarantees under verified contracts. It coordinates specification extraction and proof generation to create auditable artifacts for open‑source libraries without a designated client.
A modular quorum‑system abstraction was added to BFT‑SMaRt to enable direct latency comparisons of various Byzantine quorum designs in WAN settings.
The extension isolates quorum composition effects from protocol optimizations, allowing apples‑to‑apples evaluation of threshold, weighted, grid, and other quorum constructions on client‑perceived latency across wide‑area networks.
GhostTrack is a Python utility for tracking a device's location or mobile phone number.
Hosted on GitHub and trending, the tool provides simple APIs to query and monitor geographic coordinates or carrier‑level identifiers, useful for security analysts and investigators.
codegraph offers a locally stored, pre‑indexed code knowledge graph that syncs on changes and supports many LLM code assistants, reducing token usage and tool calls.
The repository integrates with Claude Code, Codex, Gemini, Cursor, OpenCode, AntiGravity, Kiro, CoPilot, and Hermes Agent, enabling faster, token‑efficient code queries while keeping all data on‑premises.
iced-rs/iced is a cross‑platform GUI library for Rust that takes its design cues from the Elm architecture.
The library provides a declarative UI model, supporting Windows, macOS, Linux, and WebAssembly targets. It aims to simplify state management and event handling for Rust developers, offering a consistent API across platforms.
JayWebtech/autoshorts is a local‑first desktop app that converts long video or audio recordings into vertical short‑form clips ranked by AI for viral potential.
The tool extracts candidate moments, formats them to 9:16 portrait orientation, and uses a machine‑learning model to score each clip's likely engagement. All processing runs locally, preserving user privacy while streamlining content repurposing.
Mafifrizi/ARES provides an automated red‑team engagement platform with a dashboard, scoped campaigns, module orchestration, OPSEC controls, encrypted vault, and reporting.
Designed for security teams, ARES coordinates attack simulations, stores credentials securely, and enforces operational security policies. It integrates multiple attack modules and presents results in a unified interface, facilitating repeatable and auditable assessments.
VectifyAI/OpenKB is an open‑source knowledge‑base system built around large language models.
OpenKB stores and retrieves structured information to augment LLM responses, enabling more accurate and context‑aware interactions. The project aims to provide a flexible framework for developers to integrate external knowledge into AI applications.
28 stories from 220 scanned overnight.
Get it every morning: RSS feed.