Researchers debate: pretraining gains may owe more to human curation than compute scaling

davidmanheim · x · 2026-09-09

Responding to Toby Ord's critique of the "just add compute/data" framing, davidmanheim argues marginal evidence suggests more of pretraining improvement comes from human-driven quality work, implying pure scaling trajectories are flatter than claimed. He also raises whether "quality-scaling" itself is bounded.

Related event: Scholars Debate Whether Pretraining Scaling Is Slowing Down(4 posts)→

Original post →

More from AGI Musings

AGI Musings channel →