OpenAI's GPT-Live talks over you now

Cognition

Cognition's SWE-1.7 achieves frontier-level AI capabilities for software engineering at dramatically lower cost through breakthrough RL training techniques including multi-continent distributed training, self-compaction for extended task horizons, and stability improvements that challenge assumptions about post-training limitations. The model demonstrates enhanced trustworthiness, more thorough codebase exploration, and condensed reasoning while being available today through Devin at 1000 tokens per second.

TechCrunch

Meta's new Muse image generator enables AI image creation from public Instagram profiles without user consent or notification, reigniting privacy concerns given the company's troubled data-handling history and pattern of opt-out privacy defaults.

Google Blog

Google Photos introduces Video Remix, a Gemini-powered feature that automatically transforms videos with cinematic effects and artistic styles, available to premium AI subscribers starting today.

OpenAI

OpenAI's GPT-Live introduces full-duplex voice AI that can listen and speak simultaneously while delegating complex reasoning to GPT-5.5 in the background, creating significantly more natural conversations than previous turn-based systems. The technology rolls out globally today across ChatGPT platforms with enhanced safety features and substantial performance improvements.

arXiv

Z.ai introduces SAO, a stable asynchronous reinforcement learning method for training LLMs on long-horizon agentic tasks that outperforms existing approaches on coding and reasoning benchmarks. The technique has been deployed in training the open GLM-5.2 model, addressing key stability challenges in asynchronous RL for agents.