# May 16, 2025

## Articles

- **KJ Muldoon**, a 9½-month-old boy, became the first patient to receive a **custom gene-editing treatment** for CPS1 deficiency, a rare genetic disorder affecting 1 in 1.3 million babies, showcasing the potential for personalized medicine in treating rare diseases.

- The **core loop** of an LLM agent with tool use is surprisingly simple, enabling efficient problem-solving with minimal input, as demonstrated in the provided code snippet.

- **Ollama's new engine** enhances support for **multimodal models**, starting with advanced vision models like **Meta Llama 4** and **Google Gemma 3**, enabling complex interactions such as location-based queries and image analysis.

- **LPython** is a **novel Python compiler** that converts type-annotated Python code into optimized machine code, supporting multiple backends like **LLVM, C, C++, WASM**, and **Julia**, while emphasizing **speed** through features like **JIT** and **AoT** compilation.

- **Free-threaded Python** enhances performance by fully utilizing multicore CPUs and GPUs, addressing limitations of the GIL that hinder parallel processing in traditional builds.

- **`llm-min.txt` compresses technical documentation into a machine-optimized format, achieving token reductions of 90-97%, making it ideal for AI assistants that struggle with large context windows.**

- The **RXTX algorithm** computes the product of a matrix and its transpose, **reducing multiplications and additions by 5%** compared to the current State-of-the-Art, demonstrating efficiency even for small matrices.

- **Tinfoil** offers a **verifiable privacy solution** for cloud AI by utilizing secure enclaves, ensuring that no data is accessible to Tinfoil or cloud providers during AI model inference, thus enhancing trust in AI applications.

- **AlphaEvolve** is an **evolutionary coding agent** that enhances LLM capabilities, enabling it to autonomously improve algorithms through iterative feedback, leading to significant scientific and computational advancements.

- **KVSplit** enables **up to 72% memory reduction** while maintaining or improving inference speed, allowing for **2-3x longer context windows** on Apple Silicon devices.

- **Generative AI** techniques now enable the transformation of **2D product images** into **immersive 3D experiences** with as few as three images, enhancing online shopping by replicating the tactile in-store experience.

- **AlphaEvolve** is an **evolutionary coding agent** that leverages **Gemini-powered** large language models to discover and optimize algorithms, enhancing efficiency in areas like data centers and AI training processes.

- Tubi's research reveals that **Tweedie Regression** outperforms the traditional **weighted LogLoss** model, achieving a **+0.4% revenue** and **+0.15% viewing time lift** by directly predicting user watch time.

- **Gemini 2.5** enhances **text-to-SQL** capabilities, enabling users to generate SQL queries from natural language, thus boosting productivity for developers and non-technical users alike.

- **LLM inference optimization** has evolved significantly since 2020, with techniques like **continuous batching** and **FlashAttention** now enhancing model speed and throughput.