In his newsletter, he defines it as “a frontier model is able to autonomously train a successor version of itself,” which I admit is a weaker definition than what I assumed he meant (human AI researchers no longer useful vs. AI ones)
Frontier model can autonomously train successor version
By
–