# Apr 25, 2024

### When do we stop finding new music?
- **Research indicates our music tastes begin to stagnate at age 31**, with a marked decline in the discovery of new music, as highlighted by a [Deezer survey](https://www.businessinsider.com/why-we-stop-discovering-new-music-around-age-30-2018-6#:~:text=A%20survey%20from%20music%20streaming%20service%20Deezer,choice%2C%20and%20busy%20with%20work%20and%20children.) and further supported by Spotify user data analysis.

### CoreNet: A library for training deep neural networks
- **CoreNet is a comprehensive toolkit** designed for training a wide range of deep neural network models, including foundation models like CLIP and LLM, as well as for tasks such as object classification, detection, and semantic segmentation.

### Ex-athletic director arrested for framing principal with AI-generated voice
- **Dazhon Darien**, the former athletic director, was **arrested at an airport with a firearm**, accused of using **AI to fabricate racist remarks** attributed to a principal.

### VideoGigaGAN: Towards detail-rich video super-resolution
- **VideoGigaGAN** revolutionizes video generation with **unprecedented realism and detail**, setting a new benchmark in the field.

### Snowflake Arctic Instruct (128x3B MoE), largest open source model
- **Snowflake AI Research Team** has developed **Arctic**, a **dense-MoE Hybrid transformer architecture**, pre-trained from scratch, available in both base and instruct-tuned versions under an **Apache-2.0 license** for broad use in research and product development. [Snowflake Arctic: The Best LLM for Enterprise AI](https://www.snowflake.com/blog/arctic-open-and-efficient-foundation-language-models-snowflake)

### Tiny GPU: A minimal GPU implementation in Verilog
- **`tiny-gpu` is a minimal GPU implementation in Verilog, designed to demystify the complexities of GPU architecture and operation, focusing on core elements like architecture, parallelization, and memory.**

### Air Force picks Anduril, General Atomics to develop unmanned fighter jets
- The **Air Force selects Anduril and General Atomics** for the next phase of the **Collaborative Combat Aircraft (CCA) program**, focusing on building and testing drone prototypes.

### The Stainless SDK Generator
- **Stainless SDK generator**, now available to all developers, automates the creation of client libraries for APIs, leveraging OpenAPI specifications for **multi-language SDK production** and ensuring **continuous updates**.

### Quaternion Knowledge Graph Embeddings (2019)
- **Quaternion embeddings** introduce **hypercomplex-valued representations** with three imaginary components for entities, and model relations as rotations in quaternion space, offering a more expressive alternative to traditional complex-valued representations.

### NVIDIA to Acquire GPU Orchestration Software Provider Run:ai
- **NVIDIA announces acquisition of Run:ai**, a Kubernetes-based software provider enhancing **AI workload efficiency** across various computing environments, signaling a strategic move to optimize AI infrastructure utilization.

### Australia's spies, cops want accountable encryption- a.k.a. access to backdoors
- Australian intelligence and police chiefs demand **Big Tech cooperation** in decrypting messages for cases involving terrorism and other serious crimes, citing **encryption as both a protector of privacy and a shield for criminals**.

### Show HN: I made a programmable computer from NAND gates
- **NAND is a Turing equivalent 16-bit computer emulated on the web**, featuring its own CPU, machine code, assembly language, and more, based on the Jack-VM-Hack platform from the [Nand to Tetris course](https://www.nand2tetris.org/) and its [associated book](https://www.amazon.com/Elements-Computing-Systems-second-Principles-dp-0262539802/dp/0262539802/ref=dp_ob_title_bk).

### Eric Schmidt-backed Augment, a GitHub Copilot rival, launches out of stealth
- **Augment**, an AI-powered coding platform, has launched with **$252 million** in funding, aiming to enhance software development with AI, backed by notable investors including **Eric Schmidt**.

### Fine tune LLAMA3 on million scale dataset in consumer GPU using QLora, DeepSpeed
- **Fine-tuning LLAMA-8b-instruct on the Openhermes-2.5 dataset with 4 RTX 4090 GPUs was achieved using QLoRA and DeepSpeed**, enabling efficient training on consumer-grade hardware.

### Launch HN: Nango (YC W23) – Source-available unified API
- **Nango** offers a **source-available unified API** that simplifies the creation of custom integrations for SaaS products, handling over **167 external APIs** and eliminating the need for in-depth API knowledge or complex infrastructure development.