Amazon Science

How AI is changing the nature of mathematical research What machine learning theorists learned using AI agents to generate proofs — and what comes next.

Michael @mkearnsphilly.bsky.social ) and I wrote a blog post about our experiences using AI for research, and our thoughts on what these developments will mean for research, publication, and education: www.amazon.science/blog/how-ai-...

How AI is changing the nature of mathematical research What machine learning theorists learned using AI agents to generate proofs — and what comes next.

AI is bringing a sea change in scientific research methodology, training, and peer review. Amazon Scholars and Penn professors @mkearnsphilly.bsky.social and @aaroth.bsky.social on what agentic AI tools mean for the next generation of researchers.

Dialogue Boost: How Amazon is using AI to enhance TV and movie dialogue New<b> </b>audio-processing technology is making entertainment more accessible for millions of viewers.

🔊 Audio-processing technology from Amazon separates speech from background sounds using sub-band neural networks compressed to &lt;1% original size.

Results show 86% listener preference and 100% approval from users with hearing loss:

Using LLMs to improve Amazon product listings Large language models are increasing the accuracy, reliability, and consistency of the product catalogue at scale.

Amazon's LLM-based system improves product listing quality by recognizing standard attribute values, collecting synonyms, and detecting errors. The process updates millions of listings within days:

Intelligence isn’t about parameter count. It’s about time. As AI models grow larger, they become less insightful, not more. To ensure that they continue to learn, we need to reduce their inference time.

What makes AI models truly intelligent? AWS VP Stefano Soatto argues it is not the number of parameters, but how quickly they can reason.

Designing AI interfaces that align with how people actually work "If I were to ding the field right now on one thing, it is that there has been a massive lack of creativity on how people interface with these increasingly smart LLMs and agents."

AI models are getting smarter, but we're still interacting with them the same way we did five years ago. Amazon's AGI Lab explains why the interface problem matters as much as the model problem.

Why a 12-year-old forecasting paper has stood the test of time Amazon Scholar Aravind Srinivasan coauthored a 2014 paper about forecasting civil unrest in Latin America, which won a test-of-time award at KDD 2025.

12 years after publication, EMBERS wins the applied-data-science test-of-time award at KDD. The system used open-source indicators like social media posts and satellite imagery to forecast civil unrest across 10 Latin American nations.

A chat with Byron Cook on automated reasoning and trust in AI systems Over the past decade, Byron's team has proven the correctness of our authorization engine, our cryptographic implementations, and our virtualization layer. Now they're taking those same techniques and...

A new Q&A with Amazon VP & CTO @wernervogels.bsky.social and Amazon Distinguished Scientist Byron Cook on why trust, not capability, is the real barrier to deploying agentic AI in production.

Ten years of NFL Next Gen Stats + the latest in reinforcement learning Find the latest news and research from Amazon's science community at Amazon Science. Overview of how Next Gen Stats uses data to make accurate predictions.

🏈 How AWS changed the game with machine learning and what's next in agentic AI. The latest from Amazon Science:

Institute for Assured Autonomy & Computer Science Seminar Series. Talk Trends in Safe and Reliable Reinforcement Learning. February 9, 2026, 1–2 p.m. Zoom. Alec Koppel, Johns Hopkins APL. Pratap Tokekar, UMD & Amazon.

Join us and @johnshopkinsiaa.bsky.social on Monday for a joint talk on trends in safe and reliable reinforcement learning, featuring @jhuapl.bsky.social’s Alec Koppel and @univofmaryland.bsky.social & Amazon Robotics’ Pratap Tokekar. Learn more here: www.cs.jhu.edu/event/iaa-cs...

Amazon and @stanford.edu researchers collaborated to develop cvc5, an open-source software tool that powers Automated Reasoning checks in Amazon Bedrock and other AWS services. The tool now processes ~1B solver calls daily to enhance security for customers: https://amzn.to/3OlTTry

A decade of NFL Next Gen Stats innovation Every NFL game generates millions of tracking data points from 22 RFID-equipped players. Seventy-five machine learning models running on AWS process that data in under a second, transforming football ...

The NFL introduced machine learning to football with Next Gen Stats, transforming how the game is measured. Learn how the league went from basic box scores to producing up to 1,000 stats per play in 10 years:

A decade of NFL Next Gen Stats innovation Every NFL game generates millions of tracking data points from 22 RFID-equipped players. Seventy-five machine learning models running on AWS process that data in under a second, transforming football ...

The NFL introduced machine learning to football with Next Gen Stats, transforming how the game is measured. Learn how the league went from basic box scores to producing up to 1,000 stats per play in 10 years:

Thanks to everyone who contributed to a productive @aaai.org conference in Singapore. Our team enjoyed the conversations with researchers and practitioners advancing AI. See you next year! #AAAI2026

The unseen work of building reliable AI agents "Reinforcement learning gyms" train agents on the many low-level tasks that they must chain together to execute customer requests.

Before an AI agent can book your vacation, it must learn to scroll, click, tab, and navigate other low-level tasks. Amazon's AGI Lab is building "reinforcement learning gyms" where agents practice atomic behaviors, mastering mundane interactions that underpin reliable software operation:

Customizing multiturn AI agents with reinforcement learning Leveraging existing environment simulators and reward functions based on verifiable ground truth boosts task success rate, even with small models and small training datasets.

Reinforcement learning boosts AI agent task success two- to fourfold with small training sets. AWS research shows smaller models can match larger proprietary models at 1% to 2% the cost.

Fine-tuning vision-language models on memory-constrained devices A new hybrid optimization approach allows edge devices to fine-tune vision-language models using only forward passes, achieving up to 7% higher accuracy than existing techniques.

SharpZO enables edge AI fine-tuning using only forward passes. The approach achieves 7% higher accuracy than existing low-memory methods and converges in as little as one-tenth the time: https://amzn.to/4pLQek8

Why AI for good depends on good data New technologies are helping vulnerable communities produce maps that integrate topographical, infrastructural, seasonal, and real-time data — an essential tool for many humanitarian endeavors.

No data, no AI, no progress. My @AmazonScience article explores how multi-layered mapping + petabyte-scale cloud infrastructure helps save lives in time of crisis. Building AI without addressing the fundamental data divide means solving the wrong problems. amazon.science/blog/why-ai-...

Thanks, Danilo! We updated our username without the dash :)

The unseen work of building reliable AI agents "Reinforcement learning gyms" train agents on the many low-level tasks that they must chain together to execute customer requests.

"Normcore agents" are trained by Amazon's AGI Lab to chain together hundreds of micro-interactions to execute customer requests. In reinforcement learning gyms, agents practice atomic behaviors across dozens of application domains, learning to execute complex workflows with near-perfect reliability: