# Oct 6, 2025

## What GPT-OSS leaks about OpenAI's training data

- **OpenAI's GPT-5 model** has been found to include phrases from **adult websites** in its training data, revealing potential vulnerabilities in the model's training stack.

## Managing context on the Claude Developer Platform

- **New capabilities** on the Claude Developer Platform, including **context editing** and the **memory tool**, enhance agent performance by allowing for longer, uninterrupted tasks while preserving critical information across sessions.

## Generative AI's Energy Needs Are Reshaping Our World

- **ChatGPT's energy consumption** is estimated at **850 MWh daily**, enough to charge **14,000 electric vehicles**, highlighting the substantial energy demands of generative AI technologies.

## 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.

## Battering RAM – Low-Cost Interposer Attacks on Confidential Computing

- **Battering RAM** demonstrates that even advanced memory encryption technologies like **Intel SGX** and **AMD SEV-SNP** can be compromised using a low-cost ($50) interposer that manipulates memory signals, allowing attackers to access sensitive data undetected.

## Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR

- The proposed **PACS** framework reformulates **Reinforcement Learning with Verifiable Rewards (RLVR)** into a supervised learning task, enhancing stability and efficiency in training by coupling actor and critic roles implicitly.

## Show HN: Neural emotion matrix for NPCs

- **Neural Affect Matrix** empowers NPCs with **persistent emotions** and **dynamic personalities**, enhancing realism through emotional understanding and memory integration, allowing for more engaging interactions.

## 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**.

## 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**.

## Building Effective Text-to-3D AI Agents: A Hybrid Architecture Approach

- **The Hybrid Architecture** outperformed single-model setups by combining a **"Thinker" LLM** for reasoning with a **"Doer" LLM** for coding, achieving faster convergence and greater reliability in generating complex 3D models.

## Creating a database of motivated proofs

- **A new initiative aims to create a database of "structured motivated proofs,"** addressing the inadequacy of current mathematical data for AI applications, which often lack the underlying thought processes essential for understanding complex proofs.

## The math behind tiled v/s naive matrix multiplication in CUDA

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