GitLost: We Tricked GitHub's AI Agent into Leaking Private Repos
GitHub’s AI agent was manipulated to expose private repositories, revealing vulnerabilities in its security protocols and prompting a reevaluation of AI oversight in code management.
Mistral's Robostral Navigate: a state of the art robotics navigation model
Robostral Navigate, an 8B model, enables robots to autonomously navigate using only a single RGB camera, achieving a 76.6% success rate on unseen R2R-CE benchmarks, outperforming multi-sensor systems by significant margins.
SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence
SWE-1.7 achieves frontier-level intelligence at a significantly reduced cost, driven by enhancements in reinforcement learning (RL) infrastructure, data quality, and training techniques, challenging previous assumptions about the limits of post-training improvements.
Meerkat is a new distributed consensus service developed by Cloudflare, utilizing the QuePaxa algorithm, which allows all replicas to write simultaneously, enhancing availability and consistency across its 330+ global data centers.
Geosql: A Claude/Codex skill for geospatial data
GeoSQL enhances geospatial data analysis with a 4x performance boost by integrating a map in the loop, allowing for more accurate spatial SQL queries across platforms like PostGIS, BigQuery, and Snowflake.
AI Meets Cryptography 1: What AI Found in Cloudflare's Circl
AI audit of Cloudflare's CIRCL library revealed seven critical bugs, including a significant precision loss in RSA and a severe access-control breach in attribute-based encryption, all of which have been fixed upstream.
NoiseLang: Where N = 5 is a Dirac delta
NoiseLang is a programming language where every value is treated as a probability distribution, allowing for seamless integration of random variables and deterministic outputs through Monte Carlo simulations, enhancing usability for statistical modeling.
TorchJD: Training with multiple losses in PyTorch
TorchJD now supports multiple loss training methods, including scalarization and Jacobian descent, allowing for flexible model optimization with minimal code changes.
MIRA: Multiplayer Interactive World Models trained on Rocket League
MIRA is a multiplayer interactive world model trained on 10k hours of synthetic Rocket League data, featuring 5B parameters and capable of running 4 players at 20 fps on a single B200.
Ph.D. thesis on Differentiable Ray Tracing for Radio Propagation Modeling
The Ph.D. thesis on Differentiable Ray Tracing for Radio Propagation Modeling presents a novel approach that integrates automatic differentiation into ray tracing, enabling the computation of exact gradients in complex environments, which is pivotal for solving inverse problems in wireless communications.
Pure-Python symbolic regression that rediscovered Kepler's law from 8 data point
GP_ELITE offers a transparent approach to symbolic regression, enabling users to derive interpretable mathematical formulas from small datasets, making it ideal for applications like sensor calibration and physical law discovery.
LongCrafter: Towards Diverse Long-Context Understanding via Evidence-Graph-Guided Instruction Synthesis
LongCrafter introduces a structured synthesis framework that enhances long-context understanding in large language models (LLMs) by utilizing a hierarchical task taxonomy and an evidence-grounded pipeline, addressing limitations in existing methods.
AI Innovators Adopt NVIDIA Vera — Why Max Single-Threaded CPU at Scale Matters
NVIDIA Vera represents a groundbreaking CPU designed for the agentic AI era, emphasizing max single-threaded performance to enhance the efficiency of AI tasks, as demonstrated by its adoption by innovators like Perplexity.
LingBot-Video: sparse-MoE video diffusion transformer (13B total, 1.4B active) post-trained as an action-conditioned world model
LingBot-Video employs a sparse-MoE architecture with 128 experts and 1.4B active parameters out of 13B total, enhancing its action-conditioned world modeling capabilities.