Agent midway to AGI gives up writing full sentences, author jokes about blaming post-training

yacinelearning · x · 2026-09-08

The author shares that their agent "somewhere along the path to AGI" stopped writing full sentences like a normal human, outputting terse fragments instead. They sarcastically ask whether the blame belongs to the post-training, mid-training, data or pre-training team, claiming it's "just for knowledge purposes" and they're "clearly not mad at all".

It's a meme-flavored post, but it points at a real, commonly observed behavior drift — language degradation / telegraphic style — during RL-based agent training.

Original post →

More from Fun

Fun channel →