Model access for third-parties — it's a big deal!

According to a post by Cleo Nardo on LessWrong, frontier AI labs are creating a widening gap between what employees have access to and what outside researchers can use. She calls this the 'model access gap,' and she argues closing it should be a top priority for the AI safety community over the next six to twelve months. The concern: as internal models improve, the gap widens. According to Nardo's analysis, public models lag behind internal ones by three to six months, and as capabilities grow exponentially, that differential could grow from two times to sixty times. She suggests ensuring outside researchers, auditors, and safety organizations have comparable access might be more valuable than scaling those organizations ten times over. The obstacles are mounting. As of late June, the most powerful internal models aren't available to the public at any price, and Nardo points to government restrictions — specifically the Mythos situation — as making labs increasingly reluctant to release their best work. She notes this is a precedent that could stick, making the next six to twelve months critical.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton