Top 10 Most-Read Blog Posts of 2025

Amazon Science has compiled its 10 most-read blog posts of 2025, offering a snapshot of key research trends across foundation model, agentic AI, quantum computing, database, and robotics.

• Chronos-2 is a universal forecasting model that goes beyond single time series to jointly handle multiple variables and exogenous factors, emphasizing zero-shot prediction and in-context learning.
• The Ocelot quantum chip uses cat qubits and bosonic error correction to significantly reduce bit-flip errors, implementing a distance-5 error correction code with just 9 qubits.
• Scientific frontiers of agentic AI explores research questions such as inter-agent communication languages, context sharing, and privacy, examining the possibilities of an agentic Web.
• Mitra is a tabular foundation model pretrained solely on synthetic datasets, where the quality of priors—mixing structural causal models with tree-based methods—determines generalization performance.
• Aurora has evolved over a decade with an architecture that separates compute and storage, and as of May 2025, the entire Aurora product family has become serverless.
• SoLM is a lightweight LLM for schema-compliant structured data generation, improving accuracy and suppressing hallucination through denoising training with progressively stronger noise and CABS.
• GENIUS performs multimodal retrieval spanning images and text by directly generating IDs instead of comparing vectors, using semantic quantization and query augmentation.
• Science in the age of foundation models covers the application of foundation models in scientific fields, showing that data scarcity, physical law constraints, and uncertainty quantification are key challenges.
• DeepFleet is a multirobot coordination foundation model trained on billions of hours of real driving data, supporting routing and task allocation that reduce congestion in large-scale robot fleets.
• Demystifying agents explains how agents operate through LLMs, tools, and a planning loop, introducing Bedrock AgentCore, Firecracker microVM, MCP, and Nova Act as infrastructure components.