Getting a world model to run in real time isn't just a quality problem. It's a latency problem, a memory problem, and a hardware problem, all at once, on every forward pass. Using the Neuron Kernel Interface, Reactor and the Amazon Neuron Science team built a kernel-centric path to real-time autoregressive diffusion video generation on Trainium, directly addressing the dynamic shapes, memory access patterns, and cache management that make these workloads hard for generic compilers. A 3D-RoPE kernel that took 5 seconds now runs in 1.8 milliseconds. Cache copies dropped from 23 milliseconds to 1.9 milliseconds per layer. Crucially, the techniques they built aren't model-specific. They generalize across autoregressive diffusion models with real-time streaming requirements: https://amzn.to/4dWTwO5
Amazon Science
Research Services
Seattle, Washington 395,280 followers
The latest news and research from Amazon’s science community. #AmazonScience
About us
Amazon Science gives you insight into the company’s approach to customer-obsessed scientific innovation. Amazon fundamentally believes that scientific innovation is essential to being the most customer-centric company in the world. It’s the company’s ability to have an impact at scale that allows us to attract some of the brightest minds in artificial intelligence and related fields. Our scientists continue to publish, teach, and engage with the academic community, in addition to utilizing our working backwards method to enrich the way we live and work. Follow us on LinkedIn and visit our website to get a deep dive on innovation at Amazon, and explore the many ways you can engage with our scientific community. #AmazonScience
- Website
-
https://www.amazon.science
External link for Amazon Science
- Industry
- Research Services
- Company size
- 10,001+ employees
- Headquarters
- Seattle, Washington
- Founded
- 2020
- Specialties
- Artificial Intelligence, Machine Learning, Computer Vision, Cloud, Economics, Sustainability, AI, ML, Conversational AI, Natural Language Processing, NLP, Robotics, Security, Privacy, Information, Knowledge Management, Operations, Scientific Research, Search, Amazon, and Alexa
Updates
-
Amazon didn't set out to be a chip designer. That's what Peter DeSantis, Amazon's head of AI, chips, and quantum, shared in a fireside chat with SemiAnalysis founder Dylan Patel at AI Infra Summit. Each generation of custom silicon started with a customer workload that needed something better, and the company found itself innovating at the hardware level because that was the only way to solve the problem. Rather than optimizing for benchmarks, Amazon designs its chips using data from real workloads, so gains translate directly to what customers are running. And because the hardware is delivered through the cloud, customers get the benefits by simply launching an instance. That approach has guided every generation of Amazon's chips from Nitro to Graviton to Trainium: https://amzn.to/4xGi4SD
-
-
Amazon Bio Discovery developed three AI approaches to accelerate antibody drug design: MochiBind for ranking binding candidates by sequence alone, CA-MAP for predicting antibody developability properties while accounting for lab-specific batch effects, and an agent-guided design system validated experimentally on a novel cancer target. The agent-guided design system powered a validated drug development campaign that produced 46 confirmed strong binders against a desmoplastic small round cell tumor target with no prior structural or antibody data available.
-
-
New work on ML research agents offers a concrete answer to a long-standing puzzle: why does benchmark-driven progress keep being real, even after years of iterating against the same held-out data? The answer comes down to compressibility. A model whose strategy fits into a few tokens can't be relying on memorization. Memorized data wouldn't fit. When a compressed strategy enables one model to reproduce another's performance, it's evidence that the original gains came from something structural, not from benchmark-specific tuning. Across eight datasets, 32-token prompts were sufficient for a fresh ML research agent with no access to validation data to reproduce a source agent's performance on the large majority of problems. One language-modeling strategy survived compression to 16 tokens with no loss in held-out performance. The same pipeline identified overfitting when it was deliberately induced: in 38 of 102 runs, validation-specific gains vanished after compression.
-
Amazon and Defense Advanced Research Projects Agency (DARPA) brought together more than 150 researchers from academia, government, and industry for a three-day workshop in Seattle on how AI is transforming mathematical discovery and reasoning. The program featured opening remarks from Amazon Web Services (AWS) CEO Matt Garman, a keynote from Fields Medalist Terence Tao, perspectives from the National Security Agency, lightning talks from research teams, and panels on formal verification and AI-assisted mathematical reasoning. Amazon's investment in this space goes back more than a decade – from formal verification running in production at AWS to the Lean theorem prover led by Leonardo de Moura. The company has also committed $1M in AWS credits to support the research teams doing this work. Workshops like this help connect Amazon scientists with the broader research community and push the boundaries of what's possible.
-
-
-
-
-
+5
-
-
Congratulations to Amazon Redshift researchers for being awarded Best Paper Runner-Up: Industrial Track at VLDB 2026 for "FastCompose: Eliminating Compilation Cold Starts in Query Execution with Composition." The work addresses a fundamental tradeoff in compilation-based query execution: when compiled code isn't cached, queries wait on compilation before any data processing begins. FastCompose resolves this by generating lightweight LLVM IR code that assembles pre-compiled operators into execution plans, cutting compilation time from seconds to milliseconds. On TPC-DS benchmarks, the approach achieves a 7x speedup, with composition and compilation yielding identical results on a single codebase. Read the paper: https://amzn.to/4qRWGIn
-
-
Verus verifies complex projects with thousands of lines of code and proof in the time prior tools took to verify individual functions. That speed enables an interactive development loop and lets AI agents iterate on proofs faster. Amazon has used this open-source, automated program verifier for Rust to prove correctness of key primitives in the Nitro Isolation Engine and critical internal infrastructure.
-
Eight out of ten LLM judges agree. But how independently did they arrive at that answer? When judges share a prompt template, training lineage, or model family, a majority vote can make evidence look far more convincing than it really is. The vote count inflates confidence without adding independent signal. Amazon researchers introduce dependence-aware label aggregation using Ising models to address this. The method models both individual judge reliability and pairwise dependencies, allowing redundant agreement to be discounted without discarding votes entirely. Across three tasks with 10-judge panels, the approach improved accuracy by 9–14% over weighted-majority-vote baselines.
-
-
Amazon Science reposted this
Honored to be included on this year's TIME100 AI list. This one belongs to our customers and to the teams at Amazon Web Services (AWS). The majority of workloads haven't moved to the cloud yet, and when they do, the opportunity to transform with AI is massive. We're just getting started.
AWS CEO Matt Garman has been named to the TIME100 AI list for 2026. 🎊 Matt joined as an MBA intern in 2005, before AWS even had a name, and twenty years later he leads the world's largest cloud provider, now on a $169 billion annualized revenue run rate. Congratulations, Matt. Read the full TIME100 AI profile: https://lnkd.in/eSA6V-rw
-
-
How did a model upgrade make agents worse? SOP-Bench is a new benchmark that pairs genuine enterprise procedures with functioning tools and ground-truth answers across 12 industries and 2000+ tasks. With a reasoning-style (ReAct) agent, the newer Claude 4.5 family scored lower than the older Claude 4 family. A routine upgrade can lower an agent's success rate with no obvious signal that anything changed. The only reliable way to catch it is to test on the procedures the team actually runs. With SOP-Bench, agents earn their scores by completing the work rather than producing text an automatic grader happens to like.
-
