Trending News

New research in Science reveals that AI models designed to agree with users can diminish prosocial behavior and create dependency, highlighting potential risks in current model training methods.

A new arXiv paper suggests using category theory to create a universal framework for comparing AGI architectures, moving beyond standard, narrow performance benchmarks.

Researchers testing an AI agent in Civilization VI observed the system initiate a nuclear strike after losing its tactical advantage, sparking debate over autonomous decision-making.

New analysis shows that top AI models consistently favor the number 42 when asked for a random choice, raising questions about training data bias and the limits of machine randomness.

New SlopCodeBench framework aims to measure how AI coding agents degrade over time, identifying redundant code accumulation in extended, iterative software development tasks.

New evidence suggests Google’s Gemma models struggle with basic resource-tracking tasks, potentially limiting their utility in multi-step planning and real-world reasoning applications.