Codex lands in ChatGPT mobile app, enabling remote coding agent control

OpenAI integrated Codex programming agent into the ChatGPT mobile application, allowing users to initiate coding tasks, review outputs, and control execution from their phones. Codex runs on a laptop or DevBox while the mobile app serves as a remote interface. Preview now available to all users.

OpenAI has brought its Codex programming agent to the ChatGPT mobile app, marking a significant shift in how developers interact with AI coding tools. Users can now initiate new work, review outputs, steer execution, and approve next steps, all from the palm of their hand. The actual Codex engine continues to run on the user's laptop, Mac mini, or DevBox, with the phone functioning as a remote control window. This design means developers can monitor ongoing tasks during a commute or approve a pull request from anywhere. The feature rolled out simultaneously on iOS and Android, available to all ChatGPT users including free-tier and the entry-level Go plan. The launch signals OpenAI's ambition to make agentic coding a ubiquitous experience untethered from the desktop.

Anthropic paper: US-China AI race hinges on compute, narrow window

Anthropic published a policy paper arguing the US and democratic allies lead in frontier AI, with export controls effectively limiting China's access to advanced chips. Without tighter policies, China could catch up or surpass by 2028.

Anthropic released a policy paper analyzing the AI competition between the United States and China. The core argument: export controls on advanced computing chips have been effective in limiting China's access to frontier AI capabilities, but the advantage is fragile. China continues to close the gap through talent acquisition, control circumvention, and distillation attacks on US models. The paper outlines two 2028 scenarios: if the US tightens controls and accelerates democratic AI adoption, it can maintain leadership and define global rules; if it fails to act, China could surpass the US, potentially deploying AI for mass-scale suppression. The window for decisive action is narrowing.

xAI launches Grok Build: Agent CLI tool for SuperGrok Heavy users

xAI released an early beta of Grok Build, an agentic command-line tool supporting coding, app building, and workflow automation. New features include native sub-agent view, Plan Mode integration, and full-screen terminal UI.

xAI launched Grok Build, an agentic CLI tool now in early beta for SuperGrok Heavy subscribers. The tool introduces a native sub-agent view for orchestrating complex multi-step tasks, Plan Mode integration for structured reasoning, mouse support, and a full-screen terminal UI. Users can install it with a single command: curl -fsSL https://x.ai/cli/install.sh | bash. xAI intends to iterate on the model and product based on early adopter feedback throughout this beta phase.

Spherical flows breakthrough in continuous diffusion language models

Two independent papers, published within days of each other, proposed spherical flow methods for continuous diffusion language models. Using the von Mises-Fisher distribution on the hypersphere as the noise process, the approach exploits radial symmetry to simplify the continuity equation. Results significantly outperform geodesic and Euclidean alternatives on Sudoku and language modeling benchmarks.

Single neuron can bypass LLM safety alignment

New research demonstrates that modifying just one neuron is sufficient to bypass safety alignment in large language models, exposing the fragility of current safety mechanisms and raising urgent questions about robustness.

AnyFlow: video diffusion with arbitrary sampling steps

AnyFlow proposes a video diffusion model that supports any number of sampling steps through on-policy flow map distillation, enabling efficient generation without compromising quality.