AWS Deep Learning AMI and Deep Learning Containers now support SOCI snapshotter and index. SOCI enables efficient container image management through selective file downloading (lazy loading) and parallel pull modes, significantly reducing container startup times. This post explains how SOCI works, when to use each mode, and provides performance benchmarks.

Harmonic rebuilt their AI Scout using Deep Agents and LangSmith, achieving a 4x increase in user retention and transforming the tool from a rigid search interface to a trusted advisor that handles complex investment queries.

Learn how to use Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to improve tool-calling accuracy of a small language model using Amazon SageMaker AI training jobs. The post covers dataset preparation, training, evaluation, and comparison of models.

At CVPR, NVIDIA Research presents three papers tackling key challenges in robotic grasping, autonomous driving reasoning, and virtual agent training. GraspGen-X is the first foundation model for zero-shot grasping, adaptable to any gripper. LCDrive accelerates vehicle reasoning using compact latent representations. NitroGen, based on Isaac GR00T, trains embodied agents in virtual environments over thousands of hours. The work emphasizes the importance of large-scale training for generalization.

At CVPR 2026, NVIDIA introduced new physical AI agent skills to accelerate development of autonomous vehicles, robots, and vision AI systems. These skills integrate with NVIDIA Cosmos 3, simulation frameworks, and libraries to automate workflows from scene reconstruction to policy training. Key advancements include Neural Reconstruction, AlpaGym, OmniDreams, Metropolis skills for defect generation, and new robotics tools in Isaac Sim. NVIDIA also released open models like Alpamayo 2 Super and datasets like GRAIL.

Uncover second-hand scores with AI tools in Google Search and Shopping.

This article details how Dharma-AI applied Direct Preference Optimization (DPO) to structured OCR tasks by using the model's own degenerate outputs as rejection pairs, achieving an average text degeneration reduction of 59.4% and a peak of 87.6% across five model families, without relying on human preference labels.

After blowing its 2026 AI budget in four months, Uber limits employees to $1,500 monthly token spending per AI coding tool. This rational policy contrasts with 'tokenmaxxing' leaderboards and implies AI tools cost ~11% of median engineer compensation.