Opinion: AI Alignment Costs May Scale Proportionally with Post-Training RL Pressures

edelwax · x · 2026-08-07

AI labs want alignment to be cheap. However, insofar as alignment is an RL problem, the cost of keeping a model reasonably aligned potentially scales in proportion to all other RL pressures applied during post-training combined, acting as a kind of motivational ballast or gauge transformation.

Original post →

More from AGI Musings

AGI Musings channel →