# May 7, 2025

## ACE-Step: A step towards music generation foundation model  
- **ACE-Step** is an innovative open-source foundation model for music generation, achieving **15× faster synthesis** than traditional LLM-based models while maintaining superior musical coherence and lyric alignment, capable of generating **up to 4 minutes of music in just 20 seconds** on an A100 GPU.

## Launch HN: Exa (YC S21) – The web as a database  
- **Exa Websets** is an **embeddings-powered search engine** that delivers precise results for complex queries, contrasting with traditional keyword-based search engines that often yield irrelevant content.

## Accents in latent spaces: How AI hears accent strength in English  
- **Accent fingerprints** are generated by a machine learning model to quantify accent strength, revealing how subtle speech patterns can be represented in a latent space.

## Absolute Zero: Reinforced Self-play Reasoning with Zero Data [R]  
- **Absolute Zero** introduces a novel **Reinforcement Learning with Verifiable Rewards (RLVR)** paradigm that enables a model to autonomously generate tasks for its own learning, eliminating the need for external data and human supervision.

## Create and edit images with Gemini 2.0 in preview  
- **Gemini 2.0 Flash** now offers **image generation** capabilities in preview, allowing developers to integrate **conversational image generation and editing** with enhanced features via the Gemini API in [Google AI Studio](https://aistudio.google.com/app/prompts/new_chat?model=gemini-2.0-flash-preview-image-generation) and [Vertex AI](https://console.cloud.google.com/freetrial?redirectPath=/vertex-ai/studio).

## Alignment is not free: How model upgrades can silence your confidence signals  
- **Model upgrades, such as transitioning to GPT-4.1-mini, can obliterate confidence signals**, leading to overconfidence in outputs and complicating the detection of hallucinations in AI systems.

## Perfect Random Floating-Point Numbers  
- The **new algorithm** for generating **perfect random floating-point numbers** overcomes limitations of traditional methods by accessing the entire range of floating-point outputs without significant performance penalties, as detailed in the [preprint on GitHub](https://github.com/specbranch/fp-rand).

## OpenSearch 3.0 Released  
- **OpenSearch 3.0** achieves a **9.5x performance improvement** over version 1.3, significantly enhancing capabilities for AI applications that depend on vector databases to manage billions of data points.

## Jargonic Sets New SOTA for Japanese ASR  
- **Jargonic V2** sets a new benchmark in **Japanese ASR**, achieving a **94.7% recall rate** for domain-specific terms, significantly enhancing transcription accuracy in complex linguistic environments.

## [R] Cracking 40% on SWE-bench with open weights (!): Open-source synth data & model & agent  
- **SWE-smith** enables the generation of **100s to 1000s of task instances** for any GitHub repository, facilitating the creation of training data for software engineering agents, which has historically been a challenge.

## [R] Process Reward Models That Think  
- **ThinkPRM** addresses the challenge of **expensive step-level supervision** in training Process Reward Models (PRMs) by utilizing only **8K process labels** to enhance reasoning verification through long chains-of-thought.

## [R] Hybrid AI for Generating Programs: a Survey  
- **Hybrid AI** combines **symbolic AI** and **neural networks** to enhance program synthesis, aiming to automate software generation from specifications or examples.

## PyTorch Foundation Expands to Umbrella Foundation and Welcomes vLLM and DeepSpeed Projects  
- The **PyTorch Foundation** has evolved into an **umbrella foundation**, now hosting significant projects like **vLLM** and **DeepSpeed**, enhancing its role in the open source AI ecosystem.

## Gemini 2.5 Pro Preview: even better coding performance  
- **Gemini 2.5 Pro Preview** enhances coding performance with improved capabilities for front-end and UI development, enabling developers to create sophisticated workflows and applications more efficiently.

## Your Service Teams Just Got a New Coworker — and It’s a 15B-Parameter Super Genius Built by ServiceNow and NVIDIA  
- **Apriel Nemotron 15B**, a **15 billion-parameter** LLM, was developed by **ServiceNow** and **NVIDIA** using **NVIDIA NeMo** and domain-specific data, optimizing for **real-time reasoning** and enterprise applications.