# Apr 3, 2024ML Times

## The xz attack shell script
- The **xz attack** involves a sophisticated shell script and object file combination, designed to inject malicious code into the `xz` utility's build process, specifically targeting the `liblzma` library used in Debian and potentially other Linux distributions.

## CityGaussian: Real-time high-quality large-scale scene rendering with Gaussians
- **CityGaussian (CityGS)** introduces a **novel divide-and-conquer training approach** and **Level-of-Detail (LoD) strategy** for efficient large-scale 3D Gaussian Splatting (3DGS) training and rendering.

## Introducing Stable Audio 2.0
- **Stable Audio 2.0** elevates AI-generated audio by producing **high-quality, full tracks** with coherent musical structures up to three minutes long at 44.1kHz stereo, introducing **audio-to-audio generation** capabilities.

## [D] LLMs causing more harm than good for the field?
- **Large Language Models (LLMs)** have significantly increased public interest and awareness in AI, transforming skepticism into enthusiasm, especially in smaller countries.

## 'Lavender': The AI machine directing Israel's bombing in Gaza
- The Israeli army's AI system, **Lavender**, has been instrumental in the **targeting and assassination** of tens of thousands of Gazans, marking a significant shift towards automation in military operations with minimal human oversight.

## Princeton group open sources "SWE-agent", with 12% fix rate for GitHub issues
- **Princeton's SWE-agent** transforms LMs like GPT-4 into software engineering agents capable of resolving **12.29%** of issues in real GitHub repositories, as demonstrated on the [SWE-bench](https://github.com/princeton-nlp/SWE-bench) test set.

## PyTorch Library for Running LLM on Intel CPU and GPU
- **`bigdl-llm` has been renamed to `ipex-llm`**, enhancing its capabilities for running large language models (LLMs) on Intel CPUs and GPUs with low latency, as detailed in the [migration guide](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/bigdl_llm_migration.html).

## [P] SWE-agent: an open source coding agent that achieves 12.29% on SWE-bench
- The **SWE-agent**, now open source, can autonomously convert GitHub issues into pull requests, achieving a **12.29% success rate** on the SWE-bench benchmark.

## ReALM: Reference Resolution as Language Modeling
- **ReALM transforms reference resolution into a language modeling problem**, demonstrating significant advancements in handling context, including non-conversational entities.

## Choose your weapon: Survival strategies for depressed AI academics
- **AI academics face increasing challenges** in keeping pace with advancements due to limited access to computational and human resources, a situation exacerbated by private sector investments.

## Octopus v2: On-device language model for super agent
- **Octopus v2**, an on-device language model with **2 billion parameters**, outperforms GPT-4 in **accuracy and latency**, while significantly reducing context length by **95%**.

## Launch HN: Creo (YC W24) – Build internal tools with NextJS and AI
- **Creo** offers a **code-first approach** for building internal tools using NextJS, targeting engineers who prefer coding in their IDE over low-code, drag-and-drop interfaces.

## Opera becomes the first major browser with built-in access to local AI models
- Opera introduces **experimental support for 150 local Large Language Models (LLMs)** from around 50 families, including Llama from Meta and Gemma from Google, directly accessible within its Opera One browser.

## [D] For those who consider buying AMD GPUs
- **George Hotz** has shifted from recommending **AMD's 7900 XTX GPU** to **NVIDIA**, citing **driver instability** as the primary reason, a significant viewpoint for potential buyers.

## [D] Is RAG just glorified prompt Engineering?
- **RAG (Retrieval-Augmented Generation)** fundamentally enhances prompts by integrating **information-rich documents** into the generation process, making the prompt more informative.