I like this figure that plots perf of the final model on non-post-trained tasks against perf on post-trained tasks; this is meant to show base model improvement (or as the blog says non-benchmaxxing + big model smell)
Lucas Beyer praises figure separating base model gains from post-training gains
AISummary
Lucas Beyer says a plot comparing final-model performance on non-post-trained tasks with performance on post-trained tasks shows base model improvement. He calls it evidence of a non-benchmaxxing approach and a big-model effect. The post quotes Muse Wonder, a new frontier model family from finkd for personal and professional agentic work.
Post on XView on X
Lucas Beyer (bl16)Verified on X
@giffmana
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
Source: Lucas Beyer (bl16) · x.comPublished · added here