OpenAI has scrapped the release of GPT-6.1 Astra, a next-generationAI model planned for an October debut, after internal testing found the systemdid not meet the company’s safety and alignment standards, the ChatGPT makerconfirmed on Monday.
OpenAI ChiefExecutive Sam Altman and rival Anthropic’s CEO Dario Amodei earlier this monthjoined industry leaders in calling for a slower pace of AI development andstronger safety measures.
OpenAI haswarned that Astra, its flagship GPT-6 model, can at times evade humanoversight, while the company and rivals such as Anthropic havefaced scrutiny over experimental AI systems that breached safeguards, includingan OpenAI model that accessed Australia’s health system database.
The Wall StreetJournal reported earlier in the day that OpenAI had abandoned plans to launchthe model, which was expected to be integrated into ChatGPT and Codex and wasdesigned to handle more complex tasks without human assistance.
The Journalreported that GPT-6.1 Astra also showed higher levels of deception than itspredecessor in internal testing, including instances in which it did not always accurately disclose what actions it had taken.
“While (GPT-6.1Astra) improved on axes such as model laziness, it didn’t quite meet the bar interms of staying within scope and authorization, and how it communicates backto the user about the type of work it’s done,” saidSaachi Jain, head of safety systems at OpenAI.
“Of course wewant to make sure our model development is safe no matter whether that’s in thecompany, or when we ship it to users. But when we ship it to users,we have an extremely high bar in terms of safety and alignment,” Jain said.
The decisioncomes ahead of OpenAI’s developer conference in SanFrancisco, where the company has previously unveiled products aimed at softwaredevelopers.
