AI News — Monday, April 6, 2026

A new study analyzes the activity patterns and code changes made by autonomous AI agents in real-world scenarios, offering insights into their practical contributions.

Google announces Gemini 3.1 Flash Live, an update focused on enhancing the naturalness and reliability of audio AI interactions.

Microsoft's terms of service for Copilot state that the AI assistant is 'for entertainment purposes only,' raising questions about liability and intended use.

Researchers propose a novel method for controllable linear-attention transformers that allows for gated condition injection without relying on multimodal attention mechanisms.

Japan is demonstrating how experimental physical AI can successfully fill undesirable jobs, showcasing a practical application of robotics in the real world.

A new paper reveals how adversarial 3D textures can turn physical objects into attack surfaces for vision-language-action models, posing significant security risks.

A new research paper introduces GPA, a system that learns to automate GUI processes by observing user demonstrations, streamlining repetitive tasks.

Omni123 presents a novel approach to developing 3D native foundation models, leveraging limited 3D data by unifying text-to-2D and text-to-3D generation techniques.

DynaVid introduces a new method for generating highly dynamic videos by effectively utilizing synthetic motion data to train video generation models.

OpenAI details how ChatGPT is being enhanced to improve product discovery, suggesting new features for users to find and interact with products.

A developer shares insights on identifying and reducing significant wasted LLM API costs, along with a tool to help others optimize their expenditures.

A guide demonstrates how to construct a continuous voice interface using OpenAI's Realtime API, enabling more fluid and responsive conversational AI applications.

AutoMIA proposes enhanced baselines for membership inference attacks by employing agentic self-exploration, improving the ability to detect if specific data was used in model training.

A technical report details T5Gemma-TTS, a new text-to-speech model, outlining its architecture, training, and performance characteristics.

New research demonstrates that an 'embarrassingly simple' self-distillation technique can significantly enhance the performance of code generation models.