Scholars Debate Scaling Laws: Are Models Less General Despite Growing Stronger?

davidmanheim · x · 2026-07-31

A recent discussion on X has sparked debate over whether AI progress is hitting a wall. @arannayebi shared and commented on a bearish take by @RealAdamHunt.

@RealAdamHunt argued that the thesis of achieving broad new capabilities simply by scaling up models is flawed, observing that recent models have become less general, with degraded language outputs. In contrast, @arannayebi suggested this is a statement about pretraining data rather than the underlying methods (like RL and SSL), which remain sufficiently general.

Original post →

More from AGI Musings

AGI Musings channel →