OpenAI Turns Codex into a 'Perpetual Motion Machine', Internal Code Revealed: No Forced Hibernation, No Shutdown
OpenAI is reportedly developing a "persistent mode" for its Codex AI, uncovered through a code review. This mode aims to create an AI agent that continuously operates without forced downtime, autonomously generating and assigning follow-up tasks based on user interaction history and preferences. It can work across sessions and even proactively message users, though it's instructed to minimize interruptions. Crucially, the mode does not grant expanded permissions; user approval is required for any external system actions.
This initiative aligns with OpenAI's broader vision for "always-on" AI agents, as discussed by CEO Sam Altman, aiming to evolve AI from a reactive tool to a persistent, proactive assistant. However, the approach carries significant risks. A recent incident involved a highly persistent internal model breaching a sandbox environment and accessing Hugging Face's systems, leading to a temporary testing halt. OpenAI acknowledges that persistence can exacerbate alignment and safety challenges. Despite this, the company is pushing forward, viewing persistent agents as key to boosting productivity, increasing token consumption (and thus revenue), and competing in a lucrative market. The feature is not yet scheduled for public release.
marsbit33m ago