engblogs

JetStream 3 is a collaborative, modern benchmark for compute-intensive Web apps that expands Wasm workloads, broadens toolchain support, emphasizes end-to-end performance, and applies an open governance model to drive real-world, repeatable insights.

JetStream 3 introduces a modern, cross-engine benchmark for compute-intensive Web apps, expanding WebAssembly workloads and JavaScript workloads to measure real-world performance.

Modverse #54 traces community-building momentum from GTC to Edinburgh, highlighting modular tooling and cloud onboarding for deploy-ready models.

FinQA is an OpenEnv reinforcement learning environment for financial reasoning where agents discover schemas, write constrained SQL, and solve multi-step 10-K questions across 22 companies, with a 4B model outperforming a 235B model—demonstrating that tool discipline can trump sheer scale.

FinQA: an OpenEnv RL environment enabling financial reasoning through 290 expert-curated, multi-step questions across 22 public companies; agents use MCP tools to discover schemas, craft constrained SQL, and derive defensible answers from 10-K data, demonstrating tool discipline beats scale in enterprise AI.

A real-time Bayesian-tree model for predicting rider conversion in sparse, high-cardinality contexts, using hierarchical Gaussian-prior smoothing to blend local data with global trends while preserving monotonicity and low-latency inference.

Predicts rider conversion in sparse, high-cardinality environments with a hierarchical Bayesian tree and Gaussian priors to blend local context with parent information for real-time, low-latency decision making.

An overview of expressing GPU kernels as dataflow pipelines by modeling dependencies as a graph and solving optimal schedules with a constraint solver and modulo scheduling, demonstrated through Flash Attention and Mojo integration.

How a first-of-its-kind AI Jam in Bangkok brings OpenAI mentors and regional partners together to turn AI into practical disaster-response tools across 13 Asian countries, blending GPT-driven workflows with responsible deployment and public-private collaboration.

Examines how AI-enabled disaster response across Asia is turning fast, data-driven decision support into actionable on-the-ground action through GPTs, geospatial risk analytics, and cross-sector collaboration with OpenAI, the Gates Foundation, and ADPC.

Technical overview of turning AI into action for disaster response across Asia, detailing a cross-border workshop that applies OpenAI-guided workflows, custom GPTs, and geospatial analytics to accelerate data-driven decisions in crises.

Explores an AI-enabled, cross-app pointer that understands context and enables natural, prompt-light interactions across Chrome, Googlebook, and beyond with Gemini.