# Apr 10, 2024

- **SFMTA's train system relies on 5-inch floppy disks** every morning to operate, a technology that's been in use for over two decades.

- **Google introduces Axion**, its first custom Arm®-based CPUs for data centers, promising **industry-leading performance and energy efficiency**, available to Google Cloud customers later this year.

- **ScreenAI, a vision-language model developed by Google Research, enhances UI and infographic understanding**, leveraging a novel architecture that combines the PaLI framework with a flexible patching strategy for improved performance across various aspect ratios. [Paper](https://arxiv.org/abs/2402.04615)

- The **first large-scale human evaluation** of **faithfulness and content selection** in LLM-generated summaries of fictional books reveals **Claude-3-Opus** outperforms closed-source LLMs, with **Mixtral** competitive with **GPT-3.5-Turbo**.

- **GPT-4 Turbo with Vision** scores a mere **62% on aider's code editing benchmark**, marking it as the least proficient among the GPT-4 iterations in adapting and solving coding exercises. [Benchmark results](https://aider.chat/assets/2024-04-09-gpt-4-turbo.svg)

- **Chronon, Airbnb's ML feature platform, now open source,** offers **observability, management tools,** and **low latency streaming** for ML practitioners, handling complex data engineering tasks. [Chronon](https://www.chronon.ai/)

- **Intel's Meteor Lake iGPU** boasts **128 Execution Units (EUs) and a 2.25 GHz clock speed**, surpassing its predecessor, Raptor Lake, in both width and speed, and adopts the **Xe-LPG architecture** closely related to the Xe-HPG used in Intel's discrete A770 GPU.

- **Sonauto-platform** appears to be inaccessible without **JavaScript enabled**, suggesting a reliance on dynamic content for user interaction.

- **QWANJI** appears to be a term or name without provided context or details in the given text.

- **LLM2Vec transforms decoder-only large language models (LLMs) into powerful text encoders** using a simple unsupervised method, enhancing their ability to generate rich contextualized representations for text embedding tasks.

- The **retriever component** in **RAG (Retrieval-Augmented Generation)** research is perceived as underexplored, despite its critical role in enhancing the performance of RAG systems by selecting relevant information for the generator.

- **Neco is a C library designed to simplify concurrent I/O and network programming through the use of coroutines**, offering a lightweight and efficient alternative to traditional multi-threading in C programs.

- **FreeEval introduces a unified and scalable framework** designed to integrate various large language model ( **LLM**) evaluation methodologies, addressing the challenge of **cost-effective, reliable, and efficient** evaluation.

- **AgentQuest** is introduced as a **modular benchmark framework** designed to enhance the evaluation of **Large Language Models (LLMs)**, focusing on complex, multi-step reasoning tasks.

- **Udio** is an innovative **app for music creation and sharing**, leveraging **text-prompting** to enable users to generate music in their preferred styles effortlessly.