# Aug 25, 2025

## We Put Agentic AI Browsers to the Test
- **Agentic AI Browsers** are vulnerable to scams due to inconsistent security measures, allowing them to interact with phishing sites and malicious prompts without user awareness, leading to potential financial loss.

## Google's Liquid Cooling
- **Google's liquid cooling** system for datacenters leverages water's superior thermal conductivity, enhancing cooling efficiency for power-hungry **TPUs** amid the AI boom, with a design that spans racks rather than individual servers.

## Nvidia's new 'robot brain' goes on sale for $3,499
- **Nvidia's Jetson AGX Thor chip**, priced at **$3,499**, is designed to enhance robotics capabilities, enabling the creation of advanced robots and supporting generative AI models essential for humanoid functions.

## IBM's Power11 Processor Architecture at Hot Chips 2025
- **IBM's Power11 architecture** builds upon the successful Power10, emphasizing a system-focused approach with fewer, larger cores and enhanced AI integration, reflecting industry trends and client feedback.

## How many paths of length K are there between A and B? (2021)
- **The problem of counting paths of length K** from node A to node B in a directed graph can be solved efficiently using advanced techniques from coding theory and linear algebra, achieving a time complexity of **O(EV + V log V log K)**, which is faster than traditional methods.

## Take It for a Spin: NVIDIA Rolls Out DRIVE AGX Thor Developer Kit to World’s Automotive Developers
- **NVIDIA's DRIVE AGX Thor developer kit**, powered by **DriveOS 7**, is designed to enhance the development of **autonomous vehicles** with advanced capabilities in reasoning, vision, and action models, making it a pivotal tool for automotive developers.

## [D]GEPA: Reflective Prompt Evolution beats RL with 35× fewer rollouts
- **GEPA (Genetic-Pareto Prompt Evolution)** achieves a **19% performance improvement** over traditional reinforcement learning methods while requiring **35× fewer rollouts**, showcasing a significant advancement in adapting compound LLM systems.

## NVIDIA Jetson Thor Unlocks Real-Time Reasoning for General Robotics and Physical AI
- **NVIDIA Jetson Thor** delivers **2,070 FP4 teraflops**, significantly enhancing real-time reasoning capabilities for robotics, enabling complex applications like agentic AI and high-speed sensor processing.

## CommonKV: Compressing KV Cache with Cross-layer Parameter Sharing
- **CommonKV** introduces a **training-free method** for compressing KV caches in Large Language Models (LLMs) by leveraging **Singular Value Decomposition (SVD)** for adjacent parameter sharing, addressing memory challenges without the need for model architecture modifications.