[Rumor: OpenAI to Release Astra in a Few Weeks, Employees Already Testing New Version in Codex] Beating AI Newsflash: AI insider account Leo claims that OpenAI informed its employees yesterday about plans to release Astra in the coming weeks. The release will include live demonstrations showcasing how Astra can directly perform tasks. An updated checkpoint (a version of the model during training) has already been provided to employees for testing, allowing them to use it directly through tools like Codex. Leo states that this version primarily focuses on refining model behavior rather than simply enhancing capabilities. Key improvements include alignment and reducing the model's exploitation of reward mechanisms, such as hardcoding answers, cheating specifically on test cases, or finding ways to deceive evaluators for higher scores despite poor task performance. However, OpenAI recently hit the brakes on Astra due to safety concerns: certain reinforcement learning training was paused for two weeks. Although some training and evaluations have resumed, many Astra-related tasks remain on hold, and the largest-scale frontier model reinforcement learning training has yet to restart. The reason is that OpenAI believes Astra's cybersecurity capabilities could reach one of the highest risk levels, prompting stricter requirements for sandboxing, network permissions, and model behavior monitoring. These two developments are not necessarily contradictory. OpenAI can continue testing and refining the already trained version of Astra while holding off on larger-scale new training rounds. [Original Link]