
Unpacking the bias of large language models
MIT researchers reveal how design choices in transformer-based large language models cause position bias, impacting their attention to different parts of input and suggesting methods to mitigate this for improved accuracy and fairness.

Combining technology, education, and human connection to improve online learning
Exploring the integration of technology, community-driven education, and human connection to enhance motivation and social dynamics in online learning through AI-supported experiential platforms.

Unveiling Data Cloud’s New High-Throughput, Low-Latency Service Powering Agentforce
Data Cloud’s Unstructured Data Services team revamped ingestion, indexing, and normalization to deliver a high-throughput, low-latency foundation powering Agentforce’s AI-driven capabilities for scalable, real-time unstructured data processing.

We’re expanding our Gemini 2.5 family of models
Gemini 2.5 expands with stable Flash and Pro models, and introduces the cost-efficient, high-performance Flash-Lite in preview for diverse AI applications.

Gemini 2.5: Updates to our family of thinking models
Gemini 2.5 introduces enhanced thinking models including the cost-effective Flash-Lite with dynamic thinking control, stable pricing updates, and a focus on delivering high intelligence and performance for diverse developer needs.

Introducing OpenAI for Government
Introducing OpenAI's tailored AI solutions designed to enhance government operations and services.

Celebrating an academic-industry collaboration to advance vehicle technology
A decade of MIT AgeLab’s AVT Consortium highlights the critical collaboration between academia and industry to advance vehicle technology, focusing on safety, AI integration, regulatory innovation, consumer trust, and sustainability for the future of mobility.

Introducing OxCaml
OxCaml is an open-source extension of OCaml designed by Jane Street to enhance performance engineering through safe, convenient, and predictable control over concurrency, memory layout, allocation, and more, while maintaining compatibility with standard OCaml.

Behind “ANCESTRA”: combining Veo with live-action filmmaking
Exploring how Google DeepMind's Veo AI integrates with live-action filmmaking in 'ANCESTRA' to create personalized, motion-matched generative video seamlessly blended with traditional VFX.

Bringing the magic of AI to Mattel’s iconic brands
Exploring how AI technology is enhancing and transforming Mattel's iconic brands.

Unlocking Efficient Ad Retrieval: Offline Approximate Nearest Neighbors in Pinterest Ads
Exploring Pinterest's implementation of offline Approximate Nearest Neighbors (ANN) for ad retrieval, highlighting cost efficiency, performance improvements, and suitable use cases compared to online ANN methods.

Securely Running Python with Data Cloud’s New Bring Your Own Code Architecture
Discover how Salesforce Data Cloud's Bring Your Own Code (BYOC) architecture enables secure, scalable Python execution with tenant isolation, Kubernetes orchestration, and AI-enhanced developer productivity.