# Jul 18, 2025

## Highlights

- **Le Chat** now features **Deep Research mode**, enabling users to generate structured, reference-backed reports on complex topics swiftly, enhancing research efficiency and depth.

- **Fully Homomorphic Encryption (FHE)** enables computations on encrypted data, allowing users to send queries to services like Google without revealing the content of their questions, thus ensuring **complete privacy** during data processing.

- **Matrix multiplication is crucial for AI**, particularly in transformer architectures where it dominates computational load; optimizing these algorithms can significantly enhance model performance and capabilities.

- Apple introduces **two advanced multilingual, multimodal foundation language models**: a compact ∼3B-parameter on-device model optimized for Apple silicon and a scalable server model utilizing a novel **Parallel-Track Mixture-of-Experts (PT-MoE)** architecture, enhancing performance across devices and services.

- **Claude Code** revolutionizes coding efficiency by enabling the simultaneous operation of multiple agents in the cloud, allowing for seamless task management and execution without local constraints.

- **Invertible Bloom Filters (IBFs)** extend the XOR trick to efficiently identify thousands of missing IDs in large datasets by utilizing a space-efficient structure that focuses on the size of the difference between sets.

- **RisingWave** is a **stream processing platform** that efficiently handles millions of events per second, integrating real-time analytics with **Apache Iceberg™** for seamless data management and storage.

- Knostic's study uncovered **1,862 exposed MCP servers** on the internet, with **119 verified** to grant access to internal tools without authentication, highlighting significant security vulnerabilities.

- **Muon** revolutionizes neural network optimization by conceptualizing weight matrices as **geometric objects**, resulting in **35% faster training** and **15% fewer tokens** required compared to traditional methods.

- **Prompt routing** is a strategy that intelligently directs queries to the most suitable and cost-effective **language model**, optimizing for both **quality** and **latency** while minimizing expenses.

- **Liquid neural networks (LNNs)** leverage differentials against time to effectively model **dynamic changes** in time series data, enhancing predictive capabilities in various applications.

- **Minimum Description Length (MDL)** offers a promising alternative to traditional regularization methods, potentially enhancing generalization in challenging domains like transformers and LLMs, as suggested by the paper's findings.

- **Gaussian Splatting** presents a novel approach that, while sharing concepts with Deep Learning, **outperforms** the NERF method in achieving similar objectives.

- **Isambard-AI** is now operational, marking a significant advancement in the UK's AI capabilities, with a focus on high-performance computing for complex simulations and data analysis.

- **Optimizing AI conversation architecture** enhances user experience by addressing **platform limitations** through a user-driven approach, fostering more natural interactions.