Amazon Science

A new Amazon study reveals that targeting LoRA at one model sublayer — or "module" — delivers 98% of multimodule-LoRA performance while cutting latency 22.6%.amazon.science/blog/optimizin…

@RishavPramanik @amazon Welcome, Rishav!

Late News: I will be joining @amazon @AmazonScience in Seattle for the Summer of 2026, in the months of June to September. I will be with the Trust and Store Integrity Team (TSI-Science) as an Applied Scientist Intern. Hit me up if you are around Seattle around that time!

Amazon's AGI Lab is training AI agents to navigate the outdated legacy systems keeping banks, hospitals, and government agencies running. The goal: help institutions modernize without the risk of tearing everything down first. Learn more. ⬇️ amazon.science/blog/how-agent…

Critical systems are running on software older than the people managing them. Amazon's AGI Lab is training agents to navigate legacy infrastructure. amazon.science/blog/how-agent…

"As scientists studying the theory of machine learning, we’re already seeing a similar transformation in basic scientific methodology, especially for research of a mathematical nature," write Michael Kearns and Aaron Roth in their @AmazonScience op-ed. amazon.science/blog/how-ai-is…

AWS and @cerebras are bringing dramatically faster AI inference to customers through Amazon Bedrock. The solution splits inference into two stages: AWS Trainium3 for prompt processing and Cerebras CS-3 for output generation. AWS will be the first and exclusive cloud provider to deploy Cerebras hardware. Learn more: go.aws/3NBfcFC

🎙️️ Pre-training gives language models broad knowledge—but how do we elicit the structured reasoning needed to handle ambiguous, multi-step tasks? Cognitive scientist Danielle Perszyk sits down with Amazon AI researcher, Meiqi Sun, to explore how agents learn to analyze, explore, backtrack, and self-correct. Episode 6 of "Making a Mind" is live now. Tune in on YouTube: amzn.to/46UsSC9

Trust, control, and transparency in agentic AI all depend on one thing: coordination. Amazon Scholar James Pierce introduces a framework and a shared vocabulary to design for it. amazon.science/blog/designing…

AI is bringing a sea change in scientific research methodology, training, and peer review. Amazon Scholars and Penn professors @mkearnsupenn and @Aaroth on what agentic AI tools mean for the next generation of researchers: amzn.to/47o1ZGR

Excited about our new strategic partnership with @OpenAI. Developers and companies of all kinds are eager to run services powered by OpenAI models on AWS, and our unique collaboration will provide a stateful runtime environment for them that’s powered by OpenAI’s frontier intelligence on Amazon Bedrock. OpenAI is also going big on our custom Trainium chips, which are 30-40% more price performant than comparable GPUs, to power their growth. Both the leading AI labs now have made significant commitments to Trainium, which is gaining a lot of momentum. We’ll be the exclusive 3P cloud distribution provider for OpenAI Frontier (which enables organizations to build, deploy, and manage teams of AI agents). And finally, we’re excited about our investment in OpenAI — it’s an extremely talented team with great products, IP, and vision for how they can continue to serve customers and enterprises. We think they’ll be one of the big winners in AI, we can help them grow, and we believe we’ll earn a strong return for Amazon over the long term. aboutamazon.com/news/aws/amazo…

What makes AI models truly intelligent? AWS VP Stefano Soatto argues it is not the number of parameters, but how quickly they can reason: amazon.science/blog/intellige…

🎧 What if building smarter agents starts with better teachers? In the latest "Making a Mind" podcast, cognitive scientist Dr. Danielle Perszyk and AI researcher Ani Chakravarthy explore how agents learn through experience—drawing parallels between human cognitive development and AI learning curriculums to reveal a unique training method that enables agents to collaborate and overcome challenges on the web. amzn.to/4aOHcgK

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. labs.amazon.science/blog/designing…

📣 Call for posters: Amazon is accepting submissions for two workshops at Amazon Research Day in Palo Alto, CA. Submit extended abstracts by March 13 for AI co-scientist and efficient multimodal AI. #event-item-470005" target="_blank" rel="nofollow noopener">amazon.science/conferences-an…

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

A little over three years ago, @byroncook and I had a chat about automated reasoning. Since then, the landscape has shifted faster than any of us anticipated. AI agents are in production, generating code, making decisions, and we need ways to prove their correctness. I sat down with Byron for a follow-up to understand what’s changed and what we’re doing about it: allthingsdistributed.com/2026/02/a-chat…

Meet Peter DeSantis, a longtime Amazon leader who has helped shape foundational technologies across the company. After 27 years, he now leads a new organization uniting AI models, custom silicon, and quantum computing. Learn more about the visionary behind company's most ambitious long-term technology bets. go.aws/4chQLqz

🎙️ In continued conversation on the "Making a Mind" podcast, cognitive scientist @drperszyk sits down with AI researcher @denizbirlikci from Amazon's AGI Lab to explore how reinforcement learning is transforming AI agents into dependable tools. A system that succeeds once is a demo—a system that succeeds every time is a breakthrough. Episode 4 is live now, listen in to learn more.

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: amzn.to/4qmFB7B