Ten years of NFL Next Gen Stats + the latest in reinforcement learning
Amazon Science
The latest news and research from Amazon’s science community. #AmazonScience
Find the latest news and research from Amazon's science community at Amazon Science.
A decade of NFL Next Gen Stats innovation: 75 machine learning models running on AWS process millions of tracking-data points in under a second, transforming every movement into measurable insight. From helping coaches evaluate completion probability to promoting the new kickoff rule that reduced concussions by over 40%, AWS infrastructure powers the sport's nervous system. Read more >
Customizing multiturn AI agents with reinforcement learning: What if you could match GPT-4's agent performance at 1% the cost? AWS researchers show that reinforcement learning transforms small models into specialized domain experts, achieving 72% task completion with just 72 training examples.
Fine-tuning vision-language models on memory-constrained devices: Edge devices have been locked out of AI fine-tuning — until now. SharpZO breaks through memory barriers with a clever two-stage approach that smooths the optimization landscape, achieving 7% higher accuracy than existing methods and converging up to 10x faster.
The unseen work of building reliable AI agents: Before an AI agent can book your vacation, it must master something far more mundane: scrolling. Amazon's "normcore agents" train in reinforcement learning gyms, practicing hundreds of micro-interactions until they can navigate messy real-world software with near-perfect reliability.
Recommended by LinkedIn
Engaging the AI community through building, research, and shared learning: Amazon launched two new Nova initiatives to advance AI development. Developers can join a 6-week hackathon to build generative-AI apps, while researchers will compete in an 8-month challenge focused on creating trusted AI agents.
How academic collaboration delivers real-world security to Amazon customers: Amazon and Stanford 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.
Introducing the Making a Mind podcast: Hosted by cognitive scientist Danielle Perszyk from Amazon’s AGI Lab, the podcast explores the science of intelligence with leading AI researchers. Tune in bi-weekly on Wednesdays on YouTube, Amazon Music, Apple, or Spotify.
© 1996-2026 Amazon.com, Inc. or its affiliates | Privacy | Conditions of Use
BEWARE. Amazon KDP will steal your money, cancel your account and destroy your work if you ask them any questions. This is the most unprofessional, disgusting and evil company I have every worked with. I had over 1000 Blank Journals, Photo Books, Art Books, and Print books for sale on Amazon through KDP for about 5 years now and regularly made between $20 and $50 in sales while using their advertising platform and spending anywhere between $5-$20 in advertising fees monthly. I recently received this email "We are emailing to let you know that you will see a change in your Kindle Edition Normalized Pages (KENP) read count for the month of January. As part of our monthly process, we removed some pages read from reader accounts that were simulating reading or borrowing activity on your titles. You do not need to take any action, and your account is still in good standing." After sending an email back asking what and why this was happening because I don't know how, or would never somehow use suspicious activities to make my $10 a month after advertising. Immediately after asking a question about this email I WAS TERMINATED AND TOLD THEY WERE KEEPING 3 MONTHS OF MY MONEY AND DELETING ALL MY WORK.