# Aug 21, 2025

## 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.

## Weaponizing image scaling against production AI systems

- **Image scaling attacks** exploit vulnerabilities in AI systems by using downscaled images to reveal hidden prompt injections, enabling **data exfiltration** from platforms like Google Gemini CLI and Vertex AI Studio.

## SK hynix dethrones Samsung as world’s top DRAM maker

- **SK hynix** has overtaken **Samsung** as the world's leading **DRAM** manufacturer, marking a significant shift in the semiconductor industry after more than **30 years** of Samsung's dominance.

## "AI first" and the Bus Factor of 0

- The **"Bus Factor"** measures the risk of losing critical project knowledge, and with the rise of **AI-first** approaches, many teams now operate with a **Bus Factor of zero**, relying solely on AI-generated code without human understanding.

## 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.

## Apple Watch wearable foundation model

- **Foundation models** utilizing **behavioral data** from wearables can significantly enhance health predictions, leveraging over **2.5B hours** of data from **162K individuals** to optimize model architectures and tokenization strategies.

## 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.

## DeepSeek-v3.1 Release

- **DeepSeek-V3.1** introduces **hybrid inference** with **two modes**—Think and Non-Think—enhancing both speed and capability in agent tasks.

## In a first, Google has released data on how much energy an AI prompt uses

- Google’s **Gemini AI** prompts consume a median of **0.24 watt-hours** of electricity, equivalent to running a microwave for about one second, marking a significant step in transparency regarding AI energy usage.

## NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model

- **Nemotron-Nano-9B-v2** is a hybrid Mamba-Transformer model that enhances **throughput** for reasoning tasks while maintaining **state-of-the-art accuracy** against similarly-sized models, achieving up to **6x higher inference throughput** in specific settings.

## 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.

## Gearing Up for the Gigawatt Data Center Age

- **AI factories** are transforming data centers into high-performance computing units, utilizing **tens to hundreds of thousands of GPUs** orchestrated as a single entity, which is essential for training advanced AI models.

## Think SMART: How to Optimize AI Factory Inference Performance

- The **Think SMART framework** enables enterprises to optimize AI factory inference by balancing **accuracy, latency, and ROI**, essential for handling complex AI models and diverse workloads effectively.

## 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.

## Applicability vs. job displacement: further notes on our recent research on AI and occupations

- **AI's applicability** in occupations is primarily beneficial for **knowledge work and communication tasks**, such as writing and information gathering, rather than job displacement, which the study explicitly cautions against concluding.