# Oct 25, 2025

## Diamond Thermal Conductivity: A New Era in Chip Cooling
- **Diamond's exceptional thermal conductivity**—up to **2,400 watts per meter per kelvin**—is now harnessed in chip technology, allowing for significant heat dissipation and improved performance in high-density electronics.

## Why can't transformers learn multiplication?
- **Transformers struggle with multi-digit multiplication** due to their inability to effectively learn and utilize **long-range dependencies**, despite evidence that they can encode these structures through attention mechanisms.

## How was Multi-head Latent Attention not a thing before DeepSeek-V2 came up with it?
- **Multi-head Latent Attention (MLA)**, introduced by DeepSeek-V2 in 2024, **projects keys and values into a latent space**, significantly reducing computational complexity in attention mechanisms.

## ChunkLLM: A Lightweight Pluggable Framework for Accelerating LLMs Inference
- **ChunkLLM** introduces a **lightweight and pluggable framework** that enhances inference speed for large language models (LLMs) by utilizing **QK Adapters** and **Chunk Adapters** to optimize attention mechanisms and chunk detection.

## Startup plans to cool data centers by converting heat to light
- **Photonic cooling** technology by Maxwell Labs aims to revolutionize chip cooling by converting heat into light, allowing for targeted cooling of hot spots with **laser precision**, potentially eliminating the dark silicon problem that limits chip performance.

## Compiler optimizations for 5.8ms GPT-OSS-120B inference (not on GPUs)
- **Performance of gpt-oss-120b** was achieved at **5.8 ms per output token** using two RNGD cards, showcasing significant power efficiency under 180 W per card, enabled by advanced compiler optimizations and hardware integration.

## Torchcomms: A modern PyTorch communications API
- **Torchcomms** is an experimental communication API for **PyTorch Distributed**, designed to support **over 100,000 GPUs** and facilitate rapid prototyping of communication primitives, enhancing scalability and flexibility in distributed training.

## Signal Processing for AI — A New Way to Think About LLMs and ANN Search
- **Signal processing principles** can enhance AI models and embedding spaces, leading to improved efficiency and accuracy in handling noisy data, as demonstrated in collaboration with **Prof. Gunnar Carlsson** from Stanford.

## UFIPC: Physics-based AI Complexity Benchmark - Models with identical MMLU scores differ 29% in complexity
- **Models with identical MMLU scores can differ by 29% in architectural complexity**, highlighting the need for benchmarks like UFIPC that assess AI beyond mere task accuracy.