# Reflections on OpenAI

- **OpenAI's rapid growth** from 1,000 to over 3,000 employees in just one year has led to significant challenges in communication and organizational structure, impacting team dynamics and project management.

# Voxtral – Frontier open source speech understanding models

- **Voxtral** introduces **state-of-the-art open-source speech understanding models** with a 24B variant for production and a 3B variant for local use, both licensed under Apache 2.0, enhancing accessibility and control in speech technology.

# GPUHammer: Rowhammer attacks on GPU memories are practical

- **GPUHammer** demonstrates the **first practical Rowhammer attack on GDDR6 memory** in NVIDIA GPUs, enabling malicious users to induce bit flips that can degrade DNN model accuracy from **80% to 0.1%** with just a single bit flip.

# Altermagnets: The first new type of magnet in nearly a century

- Researchers have identified a new type of magnetism called **altermagnetism**, which combines properties of both ferromagnets and antiferromagnets, potentially revolutionizing computing technology by enabling faster, more energy-efficient devices.

# OpenAI Vulnerability: 48 Days, No Response

- A **vulnerability** reported to OpenAI allows unauthorized access to chat responses, potentially exposing **personal data** and confidential information, yet remains unaddressed as of mid-July 2025.

# Show HN: Improving RAG with chess Elo scores

- **ELO scores enhance RAG** (Retrieval-Augmented Generation) by providing a dynamic ranking system that improves the relevance of retrieved documents, leading to more accurate and contextually appropriate responses in AI applications.

# Petabit-class transmission over > 1000 km using standard 19-core optical fiber

- **World record** set for optical fiber transmission with **1.02 petabits per second** over **1,808 km** using a **19-core optical fiber**, marking a significant advancement in high-capacity communication technology.

# Hierarchical Modeling (H-Nets)

- **Hierarchical networks (H-Nets)** introduce a dynamic chunking mechanism that enhances AI's ability to model raw data hierarchically, addressing limitations of traditional architectures that treat all inputs equally.

# Show HN: I gave Claude a sundial and it built a calendar

- The **"Passage of Time" Model Context Protocol (MCP)** server enhances language models with **temporal awareness**, enabling them to calculate time differences and understand human contexts, thus improving interaction quality.

# [R][D] Interpretability as a Side Effect? Are Activation Functions Biasing Your Models?

- **Current activation functions induce a representational collapse**, leading to discrete clusters that bias model interpretations, revealing that many interpretability phenomena are artifacts of design choices rather than inherent to deep learning.

# [R] Interesting paper on cost-aware prompt optimization (CAPO)

- The **cost-aware prompt optimization (CAPO)** paper introduces a novel prompt tuning algorithm that leverages **evolutionary algorithms** to enhance the efficiency of prompt optimization by jointly refining both **instructions** and **few-shot examples**.

# [R] Is the Two-Tower Model Hitting Its Limits for RecSys Retrieval?

- **Pinterest's PinRec** introduces a **generative retrieval system** that leverages a transformer to enhance candidate recommendations, incorporating **outcome-conditioning** and **windowed multi-token generation** for improved scalability and efficiency.

# A recent literature review outlines trends, challenges, and taxonomy of Retrieval-Augmented Generation

- **Retrieval-Augmented Generation (RAG)** enhances large language models (LLMs) by integrating external evidence during inference, addressing issues like **factual inconsistency** and **domain inflexibility** while introducing challenges in **retrieval quality** and **robustness**.

# 🤗Seq vs Seq: the Ettin Suite of Paired Encoders and Decoders

- The **Ettin Suite** introduces state-of-the-art paired encoder and decoder models, achieving superior performance by utilizing a unified training recipe across both architectures, with models ranging from **17M to 1B parameters** and trained on **2 trillion tokens** of open data.

# CollabLLM: Teaching LLMs to collaborate with users

- **CollabLLM** enhances LLMs by training them in **simulated multi-turn interactions**, allowing models to learn effective collaboration strategies and improve user engagement through reinforcement learning.