MIT Associate Professors Jacob Andreas and Brett McGuire win the 2026 Harold E. Edgerton Faculty Achievement Award for groundbreaking work in natural language processing and astrochemistry. Andreas' innovative research bridges foundational theory with real-world impact in language learning and AI.
Amazon Bedrock now offers granular cost attribution, automatically assigning inference costs to IAM principals like IAM users, roles, or federated identities from providers like Okta. Cost allocation tags allow for easy aggregation by team, project, or custom dimension in AWS Cost Explorer and CUR 2.0, simplifying financial planning and optimization.
Alibaba's Qwen team introduces Qwen3.6-35B-A3B, a parameter-efficient AI model outperforming larger models. Its Sparse MoE architecture delivers impressive results across various benchmarks, showcasing significant advancements in agentic coding and frontend code generation.
AWS Marketing's TAA team collaborated with Gradial to create an AI solution on Amazon Bedrock, reducing webpage assembly time by over 95%. The agentic AI solution streamlines content publishing workflows, enabling marketing teams to focus on reaching and serving customers more effectively.
Understanding complex machine learning systems like Large Language Models (LLMs) is crucial for AI. New algorithms like SPEX and ProxySPEX aim to identify critical interactions at scale by measuring influence through ablation, isolating drivers of decisions with the fewest possible perturbations.
PLAID, a model that generates protein sequences and structures, reflects AI's role in biology. The model addresses challenges like all-atom generation and organism specificity, aiming to generate useful proteins efficiently.
Researchers from UC San Diego and Together AI introduce Parcae, a looped transformer architecture that outperforms prior models, using the same parameters and training data. Parcae's design addresses memory constraints and enables more compute per forward pass, solving stability issues seen in past looped models.
Training a modern large language model involves pretraining for general language patterns, followed by supervised fine-tuning for specific tasks. Techniques like LoRA and RLHF refine the model, leading to deployment in real-world systems for optimal performance and value delivery.
An encoder maps objects to noiseless images, quantifying how well measurements distinguish objects. AI can extract useful information even when encoded in ways humans cannot interpret, optimizing imaging systems based on their information content.
Recent advances in Large Language Models (LLMs) enable exciting integrated applications, but prompt injection attacks pose a major threat. StruQ and SecAlign are proposed defenses to mitigate prompt injection threats in LLM systems like Google Docs and ChatGPT.
Researchers have uncovered the learning dynamics of word2vec, revealing its linear structure and sequential steps. The algorithm's minimal neural model provides insights into feature learning in advanced language tasks.
Text-to-SQL challenges are tackled with Amazon Bedrock and Nova Micro models, offering cost-efficient custom solutions. Fine-tuning LoRA adapters for custom SQL dialects ensures performance without persistent hosting costs.
New divide and conquer RL algorithm challenges traditional TD learning, offering scalability to long-horizon tasks. Off-policy RL allows flexibility with old data, crucial for complex domains like robotics and healthcare.
Google DeepMind introduces Gemini Robotics-ER 1.6, an upgrade enhancing robot reasoning capabilities for real-world tasks. The model acts as a high-level strategist, guiding physical actions through advanced spatial reasoning and instrument reading.
Google introduces Skills in Chrome within Gemini, allowing users to save AI prompts as reusable workflows. This feature streamlines tasks across multiple tabs, offering a glimpse into the future of browser-level AI agents.