# Oct 7, 2025

## Nobel Prize in Physics Awarded to John Clarke, Michel Devoret and John Martinis

- The **2025 Nobel Prize in Physics** was awarded to **John Clarke**, **Michel H. Devoret**, and **John M. Martinis** for their groundbreaking experiments demonstrating **macroscopic quantum tunnelling** and energy quantisation in superconducting circuits, revealing quantum properties on a human scale.

## OpenZL: An open source format-aware compression framework

- **OpenZL** is a novel open source compression framework that achieves **lossless compression** for structured data, outperforming traditional compressors by leveraging data structure awareness to enhance compression ratios and speeds.

## Introducing CodeMender: an AI agent for code security

- **CodeMender** is an AI agent that automatically identifies and patches software vulnerabilities, having already contributed **72 security fixes** to open-source projects, including those with **4.5 million lines of code**.

## How does gradient descent work?

- **Gradient descent** operates effectively in deep learning by navigating regions of weight space where the **sharpness** (S(w)) remains below the critical threshold (2/\eta), despite oscillating dynamics that suggest instability.

## What makes 5% of AI agents work in production?

- **Only 5% of AI agents succeed in production** due to inadequate context engineering and system design, highlighting the need for robust scaffolding around AI models rather than just relying on their intelligence.

## How to tile matrix multiplication (2023)

- **Tiling matrix multiplication** optimizes resource utilization across power, memory, and compute, significantly reducing latency for models like transformers that rely on dense matrix operations.

## Less Is More: Recursive Reasoning with Tiny Networks

- The **Hierarchical Reasoning Model (HRM)** utilizes two small neural networks to outperform **Large Language Models (LLMs)** on complex tasks like Sudoku and ARC-AGI, demonstrating that **smaller models** can achieve significant results with limited data.

## Writing high-performance matrix multiplication kernels for Blackwell

- **High-performance matrix multiplication kernels** for Blackwell can be developed through iterative enhancements, transforming a basic implementation into one that rivals optimized libraries like cuBLAS and CUTLASS, achieving up to **109.6% utilization** of cuBLAS performance with advanced techniques such as **grid tiling** and **collective MMAs**.

## Compressed Convolutional Attention: Efficient Attention in a Compressed Latent Space

- **Compressed Convolutional Attention (CCA)** introduces a method that down-projects queries, keys, and values, enabling efficient attention operations within a shared latent space, significantly reducing parameters and computational costs.

## [R] Predictive control of generative models

- **Generative models**, particularly **flow models**, can be enhanced by integrating **exogenous inputs** through **predictive control algorithms** like MPC and MPPI to guide image generation from Gaussian noise.

## Show HN: Arc – high-throughput time-series warehouse with DuckDB analytics

- **Arc Core** is a **high-performance time-series data warehouse** leveraging DuckDB, Parquet, and MinIO, achieving **1.89M records/sec** with the MessagePack binary protocol, making it significantly faster than traditional methods.

## Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

- **ACE (Agentic Context Engineering)** enhances large language models by treating contexts as evolving playbooks, allowing for the accumulation and refinement of strategies without losing critical details.

## SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs

- **SwiReasoning** introduces a **dynamic switch** between explicit and latent reasoning, enhancing **token efficiency** and accuracy in large language models (LLMs) by addressing challenges in training-free settings.

## Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts

- **dLLMs** exhibit a unique property where they learn a mixture of **semi-autoregressive experts**, allowing for diverse generation behaviors that can be leveraged during inference.

## Codex is now generally available