# May 17, 2025

Daily

## Articles

- **Getting AI to write good SQL**
  - **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.

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

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

- **X X^t can be faster**
  - 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.

- **The Collapse of GPT**
  - **Model collapse** threatens the efficacy of future AI systems as they increasingly rely on AI-generated data, leading to outputs that diverge from real-world relevance.

- **Transformer neural net learns to run Conway's Game of Life just from examples**
  - A **simplified transformer neural network**, named **SingleAttentionNet**, effectively computes **Conway's Game of Life** by learning from examples, demonstrating its ability to perform deterministic calculations rather than mere predictions.

- **AI’s persuasion capabilities exceed those of humans – study**  
  - **LLM persuaders** (Claude Sonnet 3.5) outperformed incentivized human persuaders in a large-scale experiment, achieving higher compliance in both **truthful** and **deceptive** contexts, indicating a significant leap in AI's persuasive capabilities. [Link to article](https://arxiv.org/abs/2505.09662)

- **MIT paper on AI for materials research found to be fraudulent**  
  - **Aidan Toner-Rodgers' paper on AI's impact in materials research, initially praised, has been declared fraudulent by MIT after an internal review raised doubts about its data integrity.** The study claimed significant productivity increases among over 1,000 researchers, but the findings were too perfect and lacked credible sourcing.

- **Steepest Descent Density Control for Compact 3D Gaussian Splatting**  
  - **SteepGS** introduces a novel approach to **3D Gaussian Splatting**, achieving a **~50% reduction** in Gaussian points while preserving rendering quality, thus enhancing efficiency for real-time applications.

- **Understanding Transformers via N-gram Statistics**  
  - This study reveals that **N-gram statistics** can effectively approximate transformer predictions, leading to insights such as a method for detecting **overfitting** during training without a holdout set.

- **[P] Pivotal Token Search (PTS): Optimizing LLMs by targeting the tokens that actually matter**  
  - **Pivotal Token Search (PTS)** optimizes language model training by identifying **critical decision points** where specific tokens significantly influence the outcome, enhancing efficiency and effectiveness in model performance.

- **[P] TTSDS2 - Multlingual TTS leaderboard**  
  - **TTSDS2** is a **multilingual TTS benchmark** that evaluates synthetic speech quality across **14 languages**, achieving a Spearman correlation above **0.5** with human ratings, outperforming 16 other metrics.