engblogs

GIE-Bench introduces a novel benchmark for evaluating text-guided image editing models focusing on functional correctness and image content preservation with a large-scale annotated dataset.

Investigates data processing, synthetic data augmentation, and sequence interleaving strategies to enhance speech-language model pretraining, resulting in a compact yet highly effective SpeechLM named SpeLangy.

Synthetic Bootstrapped Pretraining (SBP) enhances language model performance by synthesizing new training data through learning inter-document relations, achieving significant improvements over traditional pretraining methods.

COSINE introduces a unified open-world segmentation model leveraging multi-modal prompts to enhance flexibility and accuracy in image understanding across diverse applications.

Gemma Scope 2 offers an advanced open-source suite of interpretability tools enabling the AI safety community to analyze and debug complex behaviors in large language models from 270M to 27B parameters.

Celebrating the innovative winners of the first Databricks Free Edition Hackathon, showcasing diverse AI and data engineering projects that leverage Free Edition capabilities for practical and impactful solutions.

Lakebase introduces dynamic compute autoscaling, instant branching, point-in-time recovery, increased storage capacity, and a refreshed UI to enhance serverless Postgres-powered application data management on Databricks.

PayTo enables real-time, cost-effective, and easily integrable payment solutions in Australia, enhancing sales conversions and expanding global payment options through Stripe.

MIT engineers developed a deep-learning dual-graph model that predicts, with 90% accuracy, the minute-by-minute cell folding, division, and rearrangement during early fruit fly development, offering potential applications in understanding tissue formation and early disease detection.

Innovative research on extracting decay heat from high-level nuclear waste offers a sustainable solution to enhance carbon-free energy production while addressing nuclear waste disposal challenges.

Exploring how computational genomic language models can decode microbial diversity to unlock novel biological functions and applications in extreme environments.

Meta leverages AI-driven automation and carefully designed secure-by-default mobile frameworks to enhance security while maintaining developer efficiency across its extensive app ecosystem.