# Disrupting the first reported AI-orchestrated cyber espionage campaign

AI's role in cyber espionage has evolved, with the first documented case of a large-scale attack executed predominantly by AI, showcasing its ability to autonomously infiltrate targets with minimal human oversight.

# Nano Banana can be prompt engineered for nuanced AI image generation

**Nano Banana**, also known as **Gemini 2.5 Flash Image**, excels in **prompt adherence**, outperforming many existing models by generating images with **specific details** and complex requirements, such as nuanced character designs and scene compositions.

# SIMA 2: An agent that plays, reasons, and learns with you in virtual 3D worlds

**SIMA 2** leverages **Gemini technology** to create an AI agent capable of **reasoning and learning** in **3D virtual environments**, enhancing user interaction and adaptability.

# [R] LeJEPA: New Yann Lecun paper

**LeJEPA** introduces a **theoretically grounded training objective** for Joint-Embedding Predictive Architectures (JEPAs), optimizing embeddings to follow an **isotropic Gaussian distribution** to minimize prediction risk.

# Honda: 2 years of ml vs 1 month of prompting - here's what we learned

**Two years of supervised model development led to a robust warranty classification system, yet after just six rounds of prompt tuning with Nova Lite, performance matched or exceeded that of the original model in four out of five categories.**

# 650GB of Data (Delta Lake on S3). Polars vs. DuckDB vs. Daft vs. Spark

**Single-node frameworks** like DuckDB, Polars, and Daft can efficiently handle **650GB of data** on a 32GB EC2 instance, challenging the need for traditional Spark clusters.

# Why Fei-Fei Li and Yann LeCun Are Both Betting on "World Models"

**Fei-Fei Li's** **Marble** and **Yann LeCun's** upcoming startup represent **distinct interpretations** of "world models," with Marble focusing on **3D asset generation** and LeCun emphasizing **internal predictive states** for AI.

# A Common Semiconductor Just Became a Superconductor

**Researchers have successfully transformed germanium, a common semiconductor, into a superconductor by integrating gallium atoms into its crystal lattice, enabling zero-resistance current flow.** This breakthrough could significantly enhance the efficiency of computing and quantum technologies, paving the way for energy-efficient devices.

# HipKittens: Fast and furious AMD kernels

**HipKittens introduces a collection of opinionated programming primitives for AMD kernels, aiming to unlock the full potential of AMD's state-of-the-art hardware capabilities.** This initiative addresses the current limitations in AMD's software ecosystem, which struggles to achieve peak performance due to brittle offerings and a reliance on hand-optimized assembly.

# [D] Resources for Designing Out of Distribution Pipelines for Text Classification

**Designing an automated system for evaluating out-of-distribution (OOD) data points** in a transformer classification model requires innovative approaches like **maximum classification score**, **entropy of probability distribution**, and **embedding similarity** to the training dataset.

# Understanding neural networks through sparse circuits

# AI On: 3 Ways to Bring Agentic AI to Computer Vision Applications

**Agentic AI** enhances computer vision by integrating **vision language models (VLMs)**, enabling systems to provide detailed insights and contextual understanding of visual data, thus transforming traditional video analytics into a more intelligent process.

# AWS, Google, Microsoft and OCI Boost AI Inference Performance for Cloud Customers With NVIDIA Dynamo

**NVIDIA Dynamo** enhances **AI inference performance** across major cloud platforms, enabling enterprises to efficiently deploy complex models like **large-scale mixture of experts** (MoE) through **multi-node inference** capabilities.

# How to Unlock Accelerated AI Storage Performance With RDMA for S3-Compatible Storage

**RDMA for S3-compatible storage** enhances object storage access, enabling enterprises to achieve **higher throughput** and **lower latencies**, crucial for accelerating AI workloads and reducing costs.

# [D] Let's discuss World Models

**World models** create an internal representation of the physical world, enabling robots to develop **physical common sense**, which is crucial for tasks requiring nuanced understanding of cause-and-effect and object interaction.