Bird-trained AI model helps decode underwater acoustic mysteries
Google Research repurposes an AI trained on bird calls to analyze ocean sounds, revealing unexpected underwater signals.
AI & ML
Frontier models, research results, agents, and developer tooling.
Google Research tackles scheduling with dynamic capacity
New theoretical work from Google Research addresses how to maximize throughput when scheduling capacity changes over time.
ServiceNow AI Announces SyGra Studio
A new AI/ML tool is introduced, but the announcement lacks specifics.
Google Teaches AI to Read Maps
A Google Research project is applying machine perception to the challenge of interpreting maps.
Why Agent Systems Work: Google Research Seeks a Science of Scaling
Google Research outlines a framework for determining when and why multi-agent systems outperform simpler alternatives.
Google Research Moves AI Past One-on-One Chat
Google Research is exploring how to author, simulate, and test AI systems that take part in dynamic group conversations with multiple humans.
Project Genie: Google DeepMind's Prototype for Interactive Worlds
Google DeepMind has released an experimental research prototype that lets U.S. subscribers to Google AI Ultra create and explore worlds.
Hugging Face Unveils Daggr for Programmatic App Chaining
Daggr lets developers link applications in code while offering a visual way to inspect the resulting chains.
Holo2: H Company's 235B Model for UI Localization
H Company's Holo2, a 235B-parameter model with 22B active parameters, is reported to take the lead in UI localization.
Ablation Study Reveals Key Choices in Text-to-Image Training
A detailed ablation analysis from Photoroom shows which training design decisions matter most for text-to-image models.
Google GIST sampling cuts data prep cost for ML pipelines
Google Research's GIST algorithm reduces the cost of sampling large datasets for machine learning by producing smaller, more representative training sets.
Free model training via Unsloth and Hugging Face Jobs
A new collaboration lets developers train AI models at no cost using Unsloth's optimization and Hugging Face's job infrastructure.
LinkedIn Engineers Share Agentic RL Training Retrospective for GPT-OSS
A practical retrospective from LinkedIn engineers details lessons from applying agentic reinforcement learning to GPT-OSS.
China's Open-Source AI Moves Past DeepSeek With New Architectures
A Hugging Face analysis examines how Chinese open-source models are exploring architectural choices beyond DeepSeek's influence.
ATLAS: Google Research Introduces Practical Scaling Laws for Multilingual Models
A new Google Research blog post outlines ATLAS, a framework aimed at practical scaling laws for multilingual model development.
Gemini Robotics-ER 1.6 Improves Robots’ Spatial and Multi-View Understanding
Google DeepMind’s updated embodied reasoning model aims to strengthen real-world robotics by improving spatial reasoning and multi-view understanding.
Google's ConvApparel measures how realistic AI user simulators are
A new framework from Google Research quantifies the gap between simulated and real user behavior, then uses that signal to improve the simulators.
Sentence Transformers Adds Multimodal Embedding and Reranking
Hugging Face's library now supports models that embed and rerank both text and images, making cross-modal retrieval more practical.