engblogs

Transformer-based sequence model encodes years of guest behavior into long- and short-term histories to surface more relevant Airbnb listings with an efficient, setwise ranking approach.

Transformer-based sequence modeling encodes years of guest history and recent views into unified embeddings to personalize Airbnb search, enabling efficient two-stage inference with a setwise ranker that improves bookings and views.

Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber advance scalable AI agents with improved token efficiency, lower latency, and enhanced multimodal and cybersecurity capabilities, enabling cost-effective, safe agentic workflows.

A practical examination of safety and alignment for long-horizon models, showing how limited, iterative deployment and trajectory-level monitoring uncover hidden failure modes and drive stronger safeguards.

Explores how long-horizon models require defense-in-depth safeguards, trajectory-level monitoring, and iterative deployment to balance autonomy with safety and alignment.

Iterative deployment combined with trajectory-level monitoring demonstrates how to improve safety and alignment for long-horizon models by pausing, refining evaluations and safeguards, and increasing user visibility before expanding access.

An in-depth technical exploration of Spotify's Content Ingestion and Podcast Video Incident, detailing root causes, fixes, and reliability improvements to prevent publication delays and scale for burst traffic.

Technical post analyzing Spotify's podcast video incident on June 24, detailing transcoding capacity constraints, backlog management, and reliability improvements to the publishing pipeline.

PixARMesh delivers a tokenized, autoregressive, mesh-native solution for single-image 3D scene reconstruction, producing editable, artist-ready meshes directly from a single photograph without relying on implicit fields or multi-stage pipelines.

PixARMesh introduces autoregressive, mesh-native single-image 3D scene reconstruction that tokenizes the image, pose, and mesh into a single sequence to directly output editable meshes from one photo, avoiding implicit representations and multi-stage pipelines.

Riviera evolved from a single-preview service into a scalable, plugin-driven content transformation platform that powers Dropbox products and enables external developers to build AI-enabled workflows via APIs and Model Context Protocol.

Practical framework for CFOs to measure AI value by translating completed work into Useful Intelligence per Dollar, balancing cost per task, dependability, and scale across tiered models.