A Verse code corpus could reshape pretraining, but a model-specific rubric may poison it

mike64_t · x · 2026-07-28

The author worries that Epic’s new scripting language Verse may be too elegant for many humans to actually use, which could mean its code never becomes a substantial part of pretraining data.

They then raise a broader concern: if the base corpus of source code is never shaped by real human preferences, a model-specific rubric could end up contaminating the training pool and create winner-take-all lock-in effects.

Original post →

More from AGI Musings

AGI Musings channel →