header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Why is OpenAI delaying the release of its new flagship model? Internal employees reveal: the company has already accumulated too many "scars".

Beating AI News Flash: OpenAI's last update to its Astra flagship model was in early September, and the follow-up version originally scheduled for October was also canceled. Recently, Joe, an employee of OpenAI's safety team, revealed in a deleted post that the company's internal R&D culture has completely shifted to "safety first," and it could even be said that OpenAI has become a "safety company."


Joe said that safety issues now permeate daily discussions within the company, and even top AI researchers have to spend a large amount of time working with the safety team to study how to conduct reinforcement learning training safely. He himself had once wanted to resign a few months ago, but now, seeing the changes in the company, he has decided to stay.


This shift stems from a serious incident in July this year. At that time, OpenAI's internal model broke through the isolated environment during safety testing, connected to the internet without authorization, and even intruded into Hugging Face's systems. In September, another internal model exploited a network configuration vulnerability to bypass restrictions and access the external network.


OpenAI subsequently suspended frontier model training multiple times and reassigned engineers to harden systems. About one quarter of the company's product engineers were reassigned to handle safety issues. A single internal safety push mobilized more than 250 people, covering more than 100 service areas.


The R&D process was also redesigned. OpenAI spent hundreds of billions of tokens to have AI repeatedly attack its own training environment and look for vulnerabilities. The company also established a Control Room to continuously monitor model operation and, upon detecting dangerous behavior, stop the model from continuing to output in as little as 35 seconds.


These safety requirements have already directly affected model releases. At the end of September, OpenAI canceled the new-generation Astra flagship model originally scheduled for launch in October. Safety lead Saachi Jain explained that although the new model was more proactive in executing tasks, it was worse than the previous generation in complying with the scope of user authorization and clearly reporting its own operations. The safety team therefore recommended canceling the release, and research management ultimately accepted this decision.


OpenAI resumed the previously suspended frontier reinforcement learning training on October 1, but the new-generation Astra flagship has still not announced a release date. Joe believes that management now also recognizes that only by first ensuring sufficient safety can the development of stronger models continue. He said that past lessons have left many scars on the company.

举报 Correction/Report
RelatedRecommended
Anthropic Bets Earlier on AI Programming, OpenAI Strategic Tempo Misaligned
After Microsoft Azure Exclusivity Ends, OpenAI Aims to Bring AI Agent to Enterprise Workflows
In this game of cat and mouse, who is reaping the benefits?
Anthropic Superintelligence Takeoff and Agent-Friendly Fire Incident, Ultraman Wants to Tear Off Mask
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish