# Jun 19, 2024

- **Safe Superintelligence Inc. (SSI)** aims to develop **the world's first safe superintelligence**, addressing both safety and capabilities through **revolutionary engineering**.

- **Meta FAIR has released six new research artifacts** focusing on innovation, creativity, efficiency, and responsibility, including models for image-to-text and text-to-music generation, and a technique for detecting AI-generated speech.

- **Refusal behavior in conversational large language models**, which are designed to reject harmful instructions, is controlled by a **single direction** in their operational mechanics, as discovered across 13 popular models.

- **YaFSDP** is a **Sharded Data Parallelism framework** optimized for transformer-like neural networks, offering detailed insights through [Medium](https://medium.com/yandex/yafsdp-a-tool-for-faster-llm-training-and-optimized-gpu-utilization-is-no-632b7539f5b3) and [Habr](https://habr.com/ru/companies/yandex/articles/817509/) blog posts.

- The **3D Gaussian Splatting** method for neural rendering is enhanced by treating the set of 3D Gaussians as **Markov Chain Monte Carlo (MCMC) samples**, leading to higher quality scene reconstructions without the need for precise initial placements.

- **Large language models (LLMs) rely on complex data pipelines** for dataset creation, with **Common Crawl's WARC/WAT/WET formats** serving as a primary data source, highlighting the **importance of data quality and preprocessing** in model training. [Common Crawl](https://commoncrawl.org/)

- Slack's engineering team successfully **automated the conversion of 15,000 unit and integration tests** from Enzyme to React Testing Library, achieving an **80% success rate** by integrating Abstract Syntax Trees (AST) with Large Language Models (LLMs), specifically using Anthropic's Claude 2.1.

- **Early efforts to enhance code completion for Rust in Cody** have shown promising results, particularly in addressing the performance gap observed in languages not well-represented in training datasets.

- The **[paper](https://dl.acm.org/doi/10.1145/3571884.3604316)** by Liesenfeld, Lopez, and Dingemanse (2023) **critically examines the openness of instruction-tuned text generators** like ChatGPT, highlighting the **importance of open research** for scientific progress and informed decision-making in AI deployment.

- **Evolutionary Strategy (ES)** for training neural networks achieves **90% accuracy** without using gradient information, matching the speed of backpropagation on GPUs.

- **Ilya Sutskever**, a prominent figure in AI, has co-founded **Safe Superintelligence Inc.** with a focus on developing **Artificial Superintelligence (ASI)** without the distraction of product cycles.

- **Logit Prisms** extend the logit lens method by mathematically decomposing transformer outputs into contributions from individual components like attention heads and MLP neurons, revealing how each part influences the final decision.

- The **VkFFT library** creator benchmarked **AMD MI300X** and **Nvidia H100** GPUs in FFT tasks, revealing that both GPUs deliver similar performance in single precision, with bandwidths around **3TB/s**, but fall short of their theoretical maximums. [VkFFT](https://github.com/DTolm/VkFFT/)

- **DiTTo-TTS** introduces an **efficient and scalable Diffusion Transformer** for Text-to-Speech, leveraging **pre-trained text and speech encoders** to overcome traditional text-speech alignment challenges.

- **Ilya Sutskever**, co-founder and former chief scientist of OpenAI, is launching **Safe Superintelligence Inc. (SSI)**, an AI startup dedicated to developing a safe and powerful AI system, emphasizing safety alongside capabilities.