# Aug 20, 2025

## How we exploited CodeRabbit: From simple PR to RCE and write access on 1M repos

- **CodeRabbit's vulnerability allowed for remote code execution (RCE) and unauthorized access to 1 million repositories**, exposing sensitive API tokens and secrets, which could lead to significant data breaches and supply chain attacks.

## How to Scale Your Model: How to Think About GPUs

- **GPUs, particularly NVIDIA's H100 and B200, excel in matrix multiplication with specialized cores and high memory bandwidth, making them versatile for large language models (LLMs) despite their origins in graphics processing.** Each H100 GPU boasts **990 bf16 TFLOPs/s** and **3.35 TB/s** memory bandwidth, while the B200 offers **2250 bf16 TFLOPs/s** and **9 TB/s** bandwidth, enhancing performance for ML tasks.

## Positron, a New Data Science IDE

- **Positron** is a **free, next-generation IDE** for data science that integrates Python and R, allowing seamless transitions from ideation to application, built on insights from over 14 years of RStudio development.

## Learning about GPUs through measuring memory bandwidth

- **Measuring memory bandwidth** in GPUs reveals significant performance differences across architectures, with the Qualcomm Adreno 740 achieving a bandwidth of over **130 GiB/s** when using textures instead of buffers, highlighting the importance of resource selection in software optimization.

## MapLibre Tile: A next generation geospatial format optimized for rendering

- The **MapLibre Tile (MLT)** format is engineered to surpass the **Mapbox Vector Tile (MVT)** standard, achieving **up to three times better compression ratios** and **three times faster decoding speeds**, making it ideal for modern geospatial data demands.

## Show HN: Luminal – Open-source, search-based GPU compiler

- **Luminal** is a **deep learning library** that leverages **search-based compilation** to enhance performance, aiming to be the fastest ML framework across devices, with a focus on simplicity and minimalism in its architecture.

## Monoid-Augmented FIFOs, Deamortised

- **Monoid-augmented FIFOs** enable efficient windowed aggregation in streaming analytics, allowing for constant-time updates and queries while maintaining aggregates like top-K values without requiring inverses, as demonstrated in the provided [Python code](https://pvk.ca/Blog/2025/08/19/monoid-augmented-fifos/monoid-fifo.py).

## [R] azzurra-voice, a new State-of-the-Art Italian Text-to-Speech model

- **`azzurra-voice`** is a cutting-edge **Text-to-Speech model** for Italian, designed to deliver a **natural and expressive** auditory experience by leveraging tens of thousands of hours of diverse speech data.

## Into the Omniverse: How OpenUSD and Digital Twins Are Powering Industrial and Physical AI

- **OpenUSD and digital twins** are revolutionizing industrial and physical AI by enabling the rapid creation of **physically accurate simulations**, which enhance the training of AI agents and autonomous systems.

## MindJourney enables AI to explore simulated 3D worlds to improve spatial interpretation

- **MindJourney** empowers AI to navigate **3D environments** by simulating movement and generating multiple perspectives, enhancing spatial reasoning capabilities beyond static images.

## [R] Virtuous Machines: Towards Artificial General Science

- A **generalizable scientific method** has been integrated into AI, enabling it to navigate the entire scientific process, from ideation to manuscript production, yielding new insights in **cognitive science** as detailed in the [arXiv paper](https://arxiv.org/abs/2508.13421).

## [D] Beyond the cloud: SLMs, local AI, agentic constellations, biology and a high value direction for AI progress

- **Pushing for models under 1 billion parameters** can drive innovation in AI, as it compels researchers to focus on fundamental algorithms rather than relying on sheer scale for performance improvements.

## Chunks as Arms: Multi-Armed Bandit-Guided Sampling for Long-Context LLM Preference Optimization

- **LongMab-PO** introduces a **Multi-Armed Bandit (MAB)** strategy to optimize long-context LLMs by selecting the most informative chunks for generating diverse and high-quality responses.

## Input Time Scaling

- **Input Time Scaling** introduces a novel approach to enhance Large Language Models (LLMs) by focusing on optimizing query inputs during both training and testing, revealing that low-quality datasets can yield high performance contrary to traditional beliefs.

## ZenFlow: Stall-Free Offloading Engine for LLM Training

- **ZenFlow** is a **stall-free offloading engine** for LLM training that achieves over **85% reduction in GPU stalls** and up to **5× speedup** by prioritizing important gradient updates on the GPU while deferring less critical ones to the CPU asynchronously.