“Cooking AGI” turns into a checkpoint that only learns shallow pattern matching

flowersslop · x · 2026-07-26

A joke progress update for an AGI project: the screenshot says the corrected v1 checkpoint only learned superficial pattern matching, with no benefit from extra passes, and that the next step is to train on simple one-hop rule lookup before attempting deeper compositionality.

It reads like a tongue-in-cheek “we’re cooking AGI” status post, with the image making the underlying message clear: the model is still far from genuine reasoning.

Original post →

More from Fun

Fun channel →