OpenAI says it fixed a batch of quality problems that cropped up after Astra launched. The strangest one traces back to old Skills: Skills written for earlier models were misfiring on Astra constantly, and in some cases they even stopped the model from checking its own work.
Another problem came from a context-management experiment that made Astra stop short, or reply to messages from earlier in the conversation. OpenAI has shut the experiment down and estimates 4,000 to 5,000 users were affected.
The team also pulled a set of misconfigured inference engines that were dragging down output quality. With those fixes in place, Astra should stick to the latest instructions more reliably and double-check its own work. OpenAI also announced one more reset for users.
https://twitter.com/thsottiaux/status/2098569976143806918?s=46