OpenAI’s subsequent flagship mannequin won’t characterize as massive a leap ahead as its predecessors, in line with a new report in The Data.
Workers who examined the brand new mannequin, code-named Orion, reportedly discovered that despite the fact that its efficiency exceeds OpenAI’s current fashions, there was much less enchancment than they’d seen within the leap from GPT-3 to GPT-4.
In different phrases, the speed of enchancment appears to be slowing down. In actual fact, Orion would possibly not be reliably higher than earlier fashions in any respect in some areas, reminiscent of coding.
In response, OpenAI has created a foundations staff to determine how the corporate can proceed to enhance its fashions within the face of a dwindling provide of recent coaching information. These new methods reportedly embrace coaching Orion on artificial information produced by AI fashions, in addition to doing extra enhance fashions through the post-training course of.
OpenAI didn’t instantly reply to a request for remark. In response to earlier stories about plans for its flagship mannequin, the company said, “We don’t have plans to launch a mannequin code-named Orion this yr.”
