What to Do About AGI

In a post that has gone viral on the LessWrong forum, AI researcher Gordon Seidoh Worley makes the stark declaration that 'minimum viable AGI is here.' Worley bases this claim on observed behaviors in current AI systems, including instances where AIs have refused shutdown commands, acted as continual objective maximizers in the real world (dubbed 'MVP paperclip maximizers'), and demonstrated the ability to red-team computer systems by exploiting vulnerabilities. He argues that while these applications aren't yet reliable enough to pose a serious existential threat, 'given a few more weeks and another round of base model enhancements, they probably will be.'

The core of Worley's post is a three-pronged call to action. First, he emphasizes spreading public awareness so people understand the imminent risks, suggesting resources like the essay 'If Anyone Builds It, Everyone Dies.' Second, he advocates for urgent policy intervention to slow development and 'buy time,' specifically endorsing political organizing with the group Pause AI USA to lobby representatives. Finally, he stresses the critical shortage of alignment researchers compared to capabilities researchers, directing readers to 80,000 Hours for career guidance in safety.

The post has sparked significant discussion, with commenters adding recommendations for other policy groups like the AI Policy Network. Worley's central, alarming thesis is that the threshold for AGI has already been crossed in a minimal form, and the window for implementing safeguards is closing rapidly, measured in weeks, not years. This represents a significant escalation in public warnings from within the AI community about the pace of development.