Jan 6, 2026

Databases in 2025: A Year in Review

• PostgreSQL's dominance in the database landscape continues, with significant developments such as the release of v18, featuring an asynchronous I/O storage subsystem and support for skip scans, marking a pivotal shift in its architecture and performance capabilities.

• AWS has increased GPU prices by 15%, affecting EC2 Capacity Blocks for machine learning, with specific instance rates rising significantly across various regions, such as the p5e.48xlarge from $34.61 to $39.80 per hour.

• Intel Core Ultra Series 3 is the first platform built on Intel 18A, marking a significant advancement in semiconductor technology developed in the U.S.

• The lottery ticket hypothesis posits that within dense, randomly-initialized neural networks, there exist subnetworks ("winning tickets") that can be trained to achieve comparable accuracy to the original network with significantly fewer parameters and iterations.

• Opus 4.5 revolutionizes coding by enabling developers to create complex applications with minimal input, demonstrating capabilities that surpass previous AI models in efficiency and accuracy.

• jax-js is a pure JavaScript machine learning library that reimplements Google DeepMind’s JAX framework, enabling numerical computing directly in the browser using WebGPU and Wasm for near-native performance.

• Shipping code at inference speed has transformed from a novelty to an expectation, allowing for rapid development primarily constrained by inference time and cognitive effort, as most software tasks are straightforward data manipulations.

• Memchunk is a high-performance chunking library for RAG pipelines, achieving 164 GB/s throughput, significantly outperforming competitors like kiru (36x slower) and langchain (469x slower) by optimizing delimiter-based chunking strategies.

• Tamarind Bio is revolutionizing AI drug discovery by providing a comprehensive platform that integrates leading open-source models, enabling biopharma companies to design new medicines without requiring technical expertise.

• PHOTON introduces a hierarchical autoregressive model that enhances language generation by utilizing vertical, multi-resolution context access, significantly improving throughput and reducing memory constraints during long-context decoding.

• AI agents are poised to revolutionize knowledge work by operating independently of human input, thus enabling businesses to scale operations without the limitations of human attention and availability.

• Falcon-H1R is a 7B-parameter model that achieves competitive reasoning performance, outperforming larger models by 2× to 7× on various benchmarks, showcasing the potential of small language models (SLMs) in reasoning tasks.

• Falcon-H1-Arabic introduces a hybrid architecture combining State Space Models and Transformer attention, enhancing long-context understanding and coherence in Arabic NLP tasks.

• NVIDIA Cosmos Reason 2 enhances physical AI with advanced reasoning capabilities, achieving top rankings in the Physical AI Bench and Physical Reasoning leaderboards, marking it as the leading open model for visual understanding.

• Yuan3.0 Flash is an open-source Mixture-of-Experts (MoE) model with 3.7B activated parameters and 40B total parameters, tailored for enterprise tasks while excelling in general-purpose applications.