AI Briefing
Amazon Science news
The latest AI and developer news from Amazon Science, with short summaries.
Visit website →New Benchmark for Evaluating Patient-Facing Healthcare AI Agents
A new benchmark called PatientAgentBench has been released to measure the safety and clinical workflow performance of patient-facing AI agents.
Amazon and University of Michigan Give Robots a Sense of Touch
Amazon and the University of Michigan have improved robots' ability to learn tactile sensing through the physics-based simulator HydroShear.
How Amazon tracks carbon intensity across its operations
To reach net-zero by 2040, Amazon is introducing carbon intensity metrics tailored to each business unit, including carbon emissions per unit shipped.
The fuel of the future is already here: why TRISO matters
Amazon is investing in TRISO, a next-generation nuclear fuel technology, to meet the growing power demands of AI infrastructure.
Graviton5 boosts speed and energy efficiency with improved design — beyond Moore's Law
AWS unveiled Graviton5, its next-generation CPU that maximizes performance and efficiency by applying a 3nm process and a 4-chiplet architecture.
EC2's Formally Verified 'Isolation Engine' Provides Mathematical Guarantees for Virtual Machine Isolation
AWS has introduced the Nitro Isolation Engine, which guarantees isolation between virtual machines through mathematical proof, as the first such deployment in commercial cloud.
Grounding Agentic AI in the Real World
This presents four key grounding techniques that integrate physical laws and external information so agentic AI can operate safely in physical environments.
Bridging the Intent-Execution Gap in Agent Systems
Amazon researchers released the SSA harness to reduce the gap between AI agents' intent and execution.
Ground truth is a process, not a dataset
Amazon raised evaluation accuracy to 90.9% through an 'audit-then-score' protocol in which AI models challenge benchmarks and humans verify the challenges.
AWS Data Center Networks Adopt 'Flat' Structure Instead of 'Fat Tree'
AWS has introduced a quasi-random flat network to replace the existing fat tree structure, improving efficiency and resilience.
Amazon Research Awards (ARA) Winners Announced
Amazon has announced 68 ARA winners who will receive support for research in various fields including AI security and agentic AI.
Improving LLM Decision-Making Through Diverse Reasoning Paths
Amazon researchers unveiled SSFT and GFPO, techniques that train LLMs to generate diverse reasoning paths and select the optimal strategy.
Making LLMs Faster Without Sacrificing Accuracy
Amazon extended the Chinchilla scaling law to present a framework that speeds up LLMs without losing accuracy.
Promptimus: Automatically Optimizing Already-Good LLM Prompts
Promptimus automatically improves already-good LLM prompts without manual work.
Amazon Tackles Uncertainty in Its Middle-Mile Logistics Network
Amazon presented an optimization and ML approach that designs its middle-mile logistics network to account for uncertainty.
How Mechanism Design Theory Optimizes Amazon and Vendor Collaboration
Amazon demonstrated supply chain cost savings in a 9-week pilot with Flo Pro, which combines VCG and CPP.
How to Build Trust in AI
Amazon designs AI trustworthiness through a full-lifecycle RAI pipeline.
Preserving Privacy in AI Training Data
Sensitive AI training data proved vulnerable to membership inference and reconstruction attacks, with DP and MPC serving as defenses.
How Catastrophic Is Your LLM
Amazon and UIUC unveiled C3LLM, which statistically certifies conversational LLM risk.
Isabelle/HOL, the Proof Assistant Behind the Nitro Isolation Engine
AWS formally verified the security and correctness of the Nitro Isolation Engine using Isabelle/HOL.