You’re reading an archived edition. Today’s is here →
The AI world woke up with a benchmark that actually talks to Kali Linux, and a Smash bot that beats pros without breaking a sweat.
Catch you tomorrow, and keep those compilers humming.
KaliBench is a fine-grained benchmark for natural‑language‑to‑CLI translation on Kali Linux, containing 8,504 query‑command pairs covering 1,642 tools across 23 capability dimensions and five security phases.
The dataset is built via a manuscript‑grounded pipeline that enforces deterministic canonicalization and accounts for command aliases, enabling precise evaluation of LLMs’ ability to generate correct CLI commands for real‑world cybersecurity tasks.
Faynt introduces 10 M‑parameter and 75 M‑parameter Transformer policies that control all 26 Super Smash Bros. Melee characters from a single checkpoint, achieving a 98.4% win rate against specialist opponents.
The 10 M model defeats 240 of 244 same‑character matches and wins every game against a private zero‑delay Slippi‑AI, while the larger 75 M model is distilled into the smaller one; training leveraged ~840 k human replays and a rank‑ and outcome‑based curriculum.
TACO is a ternary absolute‑max column‑wise one‑sparse optimizer that reduces memory overhead for full‑parameter LLM fine‑tuning without sacrificing accuracy or efficiency.
TACO builds on the Muon optimizer’s operator‑norm steepest‑descent view but computes the exact steepest‑descent direction under a dimension‑normalized 1→1 operator norm, offering a memory‑light alternative to AdamW for large models.
Clef is an open‑weight decision‑model suite and a new platform for reinforcement‑learning fine‑tuning announced on Hacker News.
The announcement highlights Clef’s focus on providing publicly available decision models and a streamlined RL fine‑tuning infrastructure, aiming to lower barriers for research and deployment.
Albertsons Companies is deploying ChatGPT Enterprise and the OpenAI API to accelerate internal workflows and simplify grocery shopping for millions of customers.
The retailer expects the AI integration to help teams work faster, improve operational efficiency, and enhance the shopping experience, though specific performance metrics were not disclosed.
OpenAI’s blog argues that advanced AI will most impact routine execution work that underlies breakthrough ideas, shaping the next economy’s pace.
The piece emphasizes that while headline‑grabbing innovations capture attention, the real economic driver will be AI‑augmented execution of everyday tasks, accelerating progress across industries.
AutoSynthData is a Hugging Face blog post describing a tool that automatically generates synthetic training data for enterprise AI agents.
The article outlines how the system creates diverse, domain‑specific examples without manual labeling, aiming to reduce data acquisition costs and accelerate model deployment in corporate settings. It discusses integration with existing pipelines and potential privacy safeguards.
Olmo‑core 3 is an open, scalable infrastructure released by Hugging Face for training large mixture‑of‑experts (MoE) models.
The platform supports distributed training across heterogeneous hardware and includes tools for expert routing and load balancing. It builds on earlier Olmo releases, emphasizing lower compute cost per parameter and easier experimentation for researchers and enterprises.
The Cloudflare/clef model on Hugging Face is an image‑text‑to‑text transformer with 18 downloads and 466 likes.
Its primary use case is converting visual content into textual descriptions, useful for accessibility and content moderation. The model is publicly available under a permissive license, allowing fine‑tuning for custom domains.
PSRben/VisionHOPE is an image‑classification model on Hugging Face with 305 downloads and 362 likes.
VisionHOPE targets high‑resolution object recognition tasks and is trained on a curated dataset of natural scenes. The repository includes inference scripts and supports conversion to ONNX for edge deployment.
A Hacker News discussion highlights multiple projects that discovered hidden software‑defined radio (SDR) functionality in ESP32 microcontrollers.
Researchers repurposed the ESP32's Wi‑Fi and Bluetooth radios to receive and transmit raw RF signals, enabling low‑cost SDR experiments. The findings suggest firmware updates could expose or mitigate these unintended capabilities.
At the 2026 Kernel Recipes conference, Greg Kroah‑Hartman advised developers not to panic about the surge of security bugs reported by large language models.
He explained that the kernel security team is expanding triage resources and automating reproducibility checks to handle the increased volume. The talk emphasized maintaining rigorous review processes despite the influx of AI‑generated reports.
Rust 1.99.0 was released, adding extern "C" variadic function support and APIs for raw pointer size and alignment.
The update also stabilizes several previously experimental APIs, improving low‑level interoperability and safety checks for pointer arithmetic. It continues Rust's six‑monthly release cadence.
Multiple Linux distributions issued security updates on Thursday covering dozens of packages.
Affected projects span core system libraries, browsers, development tools, and language runtimes. Administrators should apply the updates promptly to mitigate disclosed vulnerabilities.
A proposal suggests ancient Greeks could have built a purely mechanical radio.
The idea relies on resonant structures and electromagnetic principles known in antiquity, but lacks archaeological evidence. It remains a speculative historical engineering exercise.
FogCam, installed at San Francisco State University, remains the world’s oldest continuously operating webcam, posting a new image roughly every 20 seconds.
The feed has run for decades, providing a unique long‑term visual record of campus weather and lighting. Its persistence demonstrates the durability of early web infrastructure.
A Hacker News post outlines techniques to speed up the Rust compiler in September 2026.
Suggested methods include incremental compilation tweaks, cargo cache pruning, and selective feature gating. The advice targets developers experiencing long build times on large codebases.
DBWorkout is a web platform for formative SQL practice that uses LLM‑assisted task generation and reported 90% positive learner feedback from 170 undergraduates.
The system provides sandboxed execution, automated correctness feedback, and gamified sessions to reduce instructor workload. A pilot with teaching assistants showed increased engagement and perceived learning gains.
GridSMR is a sharded blockchain that executes cross-shard dependent operations within a single block, eliminating latency growth with causal depth.
The system uses an execute‑before‑agree architecture, localizing atomicity to accounts and validating schedules later. Cross‑account work runs asynchronously, allowing causal compression. Compared to prior sharded designs, GridSMR avoids the coordination bottleneck that forces latency to increase with each hop.
JOINGR is a join‑aware table retrieval method that navigates the database join graph to score tables for Text‑to‑SQL tasks.
Columns become graph nodes and foreign‑key relationships are typed edges; a lightweight MLP scorer, built on frozen embeddings, evaluates anchor tables and traversed joins. Training uses a pairwise margin loss on gold tables. This approach captures tables not explicitly mentioned in the query, improving retrieval over dense independent ranking.
JEVDB is a semantic database that routes easy decisions to fast typed models and escalates only uncertain cases to generative LLMs.
It combines exact Yannakakis‑style semijoin reduction with Semantic Bloom Filters to prune candidates across latent semantic edges. Evaluated on SemBench and a TPC‑DS‑derived workload, JEVDB‑Flash outperforms baselines in both latency and monetary cost while preserving result quality.
The paper analyzes GPU‑initiated RDMA communication, detailing the GPU‑side network path and measuring minimal transport implementations against existing libraries.
It breaks down queue placement, work‑request construction, doorbell ordering, and completion semantics, then introduces mini‑gda and mini‑proxy as baseline transports. Measurements span NVSHMEM, NCCL GIN, DeepEP, and others on NVIDIA H100, H200, B200, and GB200 platforms, exposing overheads hidden in library layers.
A work‑in‑progress recompilation of The Simpsons Game (2007) for Xbox 360 using the RexXGlue tool.
The project aims to translate the original binary into a runnable form on modern hardware, leveraging RexXGlue's ability to reconstruct executable code. Progress is ongoing, with early builds capable of launching the game.
fframes is a Rust library for programmatic video rendering that emphasizes high performance.
It provides an API for constructing video frames and pipelines, targeting low‑overhead execution. The crate is positioned as a fast alternative to existing rendering frameworks, suitable for real‑time or batch video generation.
EasyTier is a decentralized mesh VPN that includes native WireGuard support.
The project is written in Rust and aims to simplify secure networking across multiple nodes without relying on central servers. It targets both personal and small‑team use cases, offering easy configuration and peer‑to‑peer connectivity.
UniMate, presented at SIGGRAPH Asia 2026, is a single model capable of animating a wide variety of skeletal rigs.
The system unifies motion capture and procedural animation pipelines, reducing the need for separate models per character type. It demonstrates comparable quality to specialized animators while handling diverse skeleton topologies in a single framework.
Itemzflow is an open‑source PS4 game manager and home‑menu replacement built in C++.
The tool leverages community‑maintained libraries to provide a customizable interface for launching and organizing games. It targets developers and hobbyists who want more control over the console's UI without official SDK access.
prime-agent is a Rust‑based self‑improving reinforcement‑learning‑based agent designed for coding workflows and long‑running autonomous tasks.
The agent iteratively refines its own prompts and code generation strategies, aiming to reduce human oversight in repetitive development cycles. It integrates with existing IDEs and CI pipelines to automate debugging, testing, and deployment steps.
28 stories from 223 scanned overnight.
Get it every morning: RSS feed.