Top AI News - April 20, 2026

Google DeepMind

Google DeepMind launched Gemini Robotics-ER 1.6, a reasoning-first AI model that significantly improves robot spatial understanding and introduces instrument reading capabilities, enabling more autonomous physical agents in real-world environments.

Recursive Mode

recursive-mode solves AI agent 'context rot' by storing all development requirements, plans, and decisions in durable repository files instead of chat history, enabling auditable and resumable workflows across sessions. It predates and offers a free, open-source alternative to Factory.ai's Missions feature with stronger recursion mechanisms.

Troutman Privacy

Nebraska, Maryland, and Maine passed AI bills in April 2026 addressing chatbot safety, pricing algorithms, and healthcare AI, while multiple other states advanced legislation covering employment, healthcare, and disclosure requirements, indicating rapid expansion of state-level AI regulation.

Diffusion Language Models collapse during RL training because their intractable likelihood computations require noisy proxy estimates, creating unstable importance ratios that trigger destructive feedback loops. StableDRL proposes unconditional clipping and self-normalization to stabilize training by constraining updates in high-variance environments.

Enterprise software is being rebuilt for AI agents rather than humans, with machine users now outnumbering humans 45-to-1. Companies are shifting from GUIs to CLI tools, skill files, and vertical models to serve agents that bypass interfaces entirely and operate through APIs and structured commands.