PARCAE: DOING MORE WITH FEWER PARAMETERS USING STABLE LOOPED MODELS

Summary from the TLDR AI newsletter

Parcae is one of the first stable architectures for looped language models. It achieves the quality of a Transformer twice the size with clean, predictable training. Parcae increases the recurrence rather than purely scaling data, creating a new medium to scale quality. The name Parcae is a homage to the three Roman fates: Nona, Decima, and Morta.